A text sentiment analysis method and system fusing sentiment object extraction and sentiment polarity inference

Abstract

The present application relates to a kind of fusion sentiment object extraction and sentiment polarity inference text sentiment analysis method and system, belong to natural language processing technical field;The present application is from text semantic structure and the relationship of evaluation word, establishes the correlation mechanism between sentiment object, sentiment evidence and sentiment polarity, realizes the comprehensive analysis to explicit evaluation, implicit expression, negative structure, transition relationship and context semantics.Different from the method of traditional dependence sentiment dictionary, keyword matching or single classification model, the present application is through text structured processing, object level semantic positioning, sentiment evidence identification and polarity inference, realizes the automatic identification of the object to be evaluated in the text to be analyzed, evidence extraction and sentiment judgment.System combines pre-training language model and large language model, carries out linkage analysis to sentiment object, evidence fragment and context information, finally generates the structured sentiment analysis result including sentiment object, evidence fragment, sentiment polarity, confidence and inference explanation.

Description

Technical Field

[0001] This invention relates to a text sentiment analysis method and system that integrates sentiment object extraction and sentiment polarity reasoning, belonging to the field of natural language processing technology. Background Technology

[0002] With the development of the internet, e-commerce platforms, online government services, and intelligent customer service, the amount of unstructured text such as user comments, social media texts, conversation logs, and feedback is constantly increasing. These texts often contain users' evaluations and attitudes towards products, services, events, or specific functions. How to automatically identify the evaluated objects in the text and determine their corresponding sentiment tendencies has become a crucial issue in text sentiment analysis. Especially in fine-grained sentiment analysis scenarios, the same text may simultaneously evaluate multiple objects, and the sentiment tendencies of different objects are not the same. Therefore, traditional methods that only determine whether the entire text is positive or negative are insufficient to meet the needs of object-oriented, precise, and interpretable analysis.

[0003] Currently, text sentiment analysis methods mainly include sentiment lexicon and rule-based methods, traditional machine learning methods, deep learning methods, and pre-trained language models or large language models. Sentiment lexicon and rule-based methods are simple to implement but rely on manual maintenance and struggle to adapt to online slang, domain terminology, and implicit expressions. Traditional machine learning methods require manual feature design and have limited cross-domain adaptability. Deep learning and pre-trained language models can better understand contextual semantics, while large language models further possess strong command understanding, semantic reasoning, and structured generation capabilities, providing a new technological foundation for object-level sentiment analysis and complex text understanding.

[0004] In existing text sentiment analysis technologies, such as BERT, RoBERTa, MacBERT, and sentiment analysis methods based on large language models, progress has been made in tasks such as text classification, attribute word extraction, and sentiment polarity determination. These methods, through pre-training, fine-tuning, or prompting guidance on general or comment corpora, enable the model to identify evaluation words, attribute words, and sentiment tendencies in text, and complete basic tasks such as sentence-level sentiment classification and attribute-level sentiment determination. However, when applied to object-level sentiment analysis in real-world business scenarios, especially the fine-grained scenario of "sentiment object extraction and sentiment polarity inference," existing methods still face several key limitations:

[0005] (1) Lack of an integrated association mechanism between object, evidence and polarity: Existing methods often process object extraction and sentiment classification separately, or rely directly on large language models to judge sentiment polarity. They lack a unified model of the correspondence between “sentiment object - sentiment evidence - sentiment polarity”, which can easily lead to inconsistencies in object positioning, evidence attribution and polarity judgment.

[0006] (2) Insufficient understanding of implicit emotions and complex expressions: Emotions in actual texts are often not only expressed as explicit words such as "good" and "bad", but may also be hidden in factual results, negation, transition or causal descriptions. Existing methods are prone to missing implicit emotions or misunderstanding negation and transition relationships.

[0007] (3) Lack of a unified processing mechanism for multiple objects and conflicting expressions: Real comments often contain multiple evaluation objects at the same time, and different objects may have different emotional tendencies. Existing methods tend to classify the entire text into a single sentiment label, or use negative evaluations in the later text to cover positive evaluations in the earlier text, making it difficult to achieve object-level differentiation and result fusion.

[0008] (4) Insufficient interpretability and traceability of output results: Some methods can only output positive, negative or neutral labels, and cannot clearly give the corresponding emotional object, evidence fragment, source sentence and reasoning basis, making it difficult to verify the results and not convenient for subsequent business analysis. Summary of the Invention

[0009] To address the problems of inaccurate identification of sentiment objects, insufficient understanding of implicit sentiment, weak correlation between objects and evidence, easy confusion of results with multiple objects, and poor interpretability of results in existing text sentiment analysis, this invention provides a text sentiment analysis method and system that integrates sentiment object extraction and sentiment polarity reasoning. Starting from the semantic structure of the text and the relationship between evaluative words, this invention establishes a correlation mechanism between sentiment objects, sentiment evidence, and sentiment polarity, and realizes comprehensive analysis of explicit evaluations, implicit expressions, negation structures, adversative relationships, and contextual semantics.

[0010] Unlike traditional methods that rely on sentiment dictionaries, keyword matching, or single classification models, this invention achieves automatic identification, evidence extraction, and sentiment judgment of the evaluated object in the text to be analyzed through text structuring, object-level semantic localization, sentiment evidence recognition, and polarity inference. The system combines pre-trained language models and large language models to perform linked analysis of sentiment objects, evidence fragments, and contextual information, ultimately generating a structured sentiment analysis result that includes the sentiment object, evidence fragment, sentiment polarity, confidence level, and inference explanation.

[0011] Terminology Explanation:

[0012] 1. Sentiment Object (SO): refers to the target object in the text to be analyzed that is evaluated, described, or carries emotional tendencies, including products, services, events, attributes, functions, behaviors, processes, and processing results.

[0013] 2. Sentence Unit (SU): This refers to an analysis unit with relatively complete semantics obtained after segmenting the original text. It can be a sentence, clause, phrase fragment, dialogue turn fragment, or semantic fragment completed with context. It is used to carry out subsequent tasks such as sentiment object recognition, sentiment evidence extraction, and polarity reasoning.

[0014] 3. Sentiment Evidence (SE): refers to text fragments or language structures that can support the judgment of sentiment polarity, including evaluative words, degree modifiers, negative expressions, contrastive expressions, descriptions of factual results, descriptions of abnormal states, descriptions of repeated processing, etc.

[0015] 4. Sentiment Object Boundary Identification (SOBI): This refers to the process of determining the semantic range of sentiment objects in the text to be analyzed. It is used to determine whether a word, phrase, or fragment truly belongs to the object being evaluated and to determine its start and end range or corresponding semantic fragment in the text, avoiding misidentification of background entities, irrelevant nouns, or adjacent objects as sentiment objects.

[0016] 5. Object-Evidence Semantic Association (OESA): This refers to the technical process of accurately matching emotional evidence to corresponding emotional objects.

[0017] 6. Implicit Sentiment Expression (ISE): This refers to a way of expressing emotions that does not directly use obvious emotional words such as "good," "bad," "satisfied," or "dissatisfied," but conveys emotional inclination through factual descriptions, results, processing procedures, or user experience consequences.

[0018] 7. Object-level Sentiment Polarity Reasoning (OSPR): This refers to the process of judging the emotional tendency of a specific emotional object by taking it as the unit of analysis and combining it with corresponding emotional evidence, contextual information, and domain scene information.

[0019] 8. Result Fusion and Confidence Verification (RFCV): This refers to the processing mechanism for merging, filtering, conflict resolution, and credibility assessment of analysis results when there are multiple sentiment objects, multiple sentiment evidences, or multiple candidate judgment results in the same text.

[0020] 9. Feedback Updating Mechanism (FUM): This refers to a closed-loop optimization mechanism in which the system receives results from manual review, user corrections, or feedback from business systems, and updates the domain object vocabulary, sentiment evidence templates, confidence thresholds, model hint templates, or training sample libraries based on these results. It is used to improve the system's adaptability and analytical accuracy during long-term operation.

[0021] 10. Text sentiment analysis: refers to the process of identifying, classifying, and interpreting subjective attitudes, evaluation tendencies, emotional expressions, and experiential feedback contained in text using technologies such as natural language processing, machine learning, or large language models.

[0022] 11. Dependency parser: A natural language processing tool used to analyze the syntactic dependency relationships between words in a sentence. It can identify grammatical relationships such as subject-verb, verb-object, attributive-head, adverbial-head, and complement, and construct dependency paths between words. This invention utilizes a dependency parser to determine the syntactic association and shortest dependency path between sentiment objects and candidate sentiment evidence, providing syntactic features for object-evidence matching and sentiment polarity judgment.

[0023] The technical solution of the present invention is as follows:

[0024] The first aspect of this invention provides a text sentiment analysis method that integrates sentiment object extraction and sentiment polarity reasoning, comprising:

[0025] Step 1: Perform structured preprocessing on the original text to be analyzed, construct standardized sentence-level structural units, and construct a domain resource set according to the domain type of the original text to be analyzed;

[0026] Step 2: Based on the standardized sentence element-level structural units and the domain resource set, determine the sentiment relevance, screen out the candidate sentence elements for sentiment analysis, and obtain the standardized sentence element-level structural units belonging to the candidate sentence elements for sentiment analysis;

[0027] Semantic parsing is performed on standardized sentence-level structural units that belong to the candidate sentence elements of sentiment analysis to identify the set of candidate sentiment objects and filter the set of sentiment objects;

[0028] Candidate evaluation components are extracted from the selected set of candidate emotional objects to form a set of candidate evidence. After screening the set of candidate evidence, a comprehensive correlation score between the set of candidate emotional objects and the set of candidate evidence is calculated. The final set of evidence is obtained based on the correlation score.

[0029] Step 3: Perform sentiment polarity reasoning on the standardized sentence-level structural units belonging to the sentiment analysis candidate sentence elements, the filtered candidate sentiment object set, the final evidence set, and the corresponding domain resource set to obtain the final sentiment polarity;

[0030] Step 4: Construct object-level structured analysis results, review them, and generate object-level sentiment analysis results;

[0031] Based on the similarity of sentiment objects, the strength of evidence, and the overall confidence level, the results of object-level sentiment analysis are integrated and conflict resolved.

[0032] According to a preferred embodiment of the present invention, the specific implementation process of step 1 includes:

[0033] The original text to be analyzed is segmented, including: automatically recognizing periods, semicolons, question marks, exclamation marks, line breaks, transition words, and semantic boundary markers, splitting the original text into multiple semantically complete sentence elements to form sentence-level structural units; at the same time, the sentence order, contextual adjacency, paragraph number, and text source are preserved.

[0034] Inject the sentence element content, text source, paragraph number, sentence element number, and contextual information of the original text to be analyzed into each sentence element-level structural unit to generate a standardized sentence element-level structural unit, as shown below:

[0035] ;

[0036] in, This represents the i-th standardized sentence-level structural unit. This indicates the text source of the original text to be analyzed corresponding to the i-th standardized sentence-level structural unit. This represents the sentence element content of the i-th standardized sentence element-level structural unit. Indicates context fragment information, This indicates the location identifiers in the original text to be analyzed, including sentence element numbers, paragraph numbers, and arrangement order;

[0037] Based on the text source, business tags, or user settings of the original text to be analyzed, determine the domain type, i.e., the business scenario, to which the original text to be analyzed belongs, and obtain the corresponding domain resource set:

[0038] ;

[0039] in, Indicates the first Domain type refers to the set of domain resources corresponding to a business scenario; It represents the category of sentiment objects in the domain, including the entities, attributes, functions, service links, behaviors, or processing results being evaluated in the business scenario; This represents a domain-specific terminology glossary, including domain object names, synonyms, abbreviations, alternative names, and domain-specific expressions. It represents the rules or templates for identifying sentiment evidence, including the modifying relationship between evaluative words and sentiment objects, negation structure, adversative structure, degree modification relationship, causal relationship and implicit fact expression pattern; This represents structured output fields, including sentiment object, object category, evidence fragment, sentiment polarity, confidence level, reasoning description, and processing status.

[0040] According to a preferred embodiment of the present invention, sentiment relevance is determined based on standardized sentence-level structural units and a set of domain resources to screen candidate sentence elements for sentiment analysis, and standardized sentence-level structural units belonging to the candidate sentence elements for sentiment analysis are obtained; including:

[0041] The standardized sentence-level structural units and the domain resource set are input into the sentiment relevance judgment function to obtain the judgment result, as shown below:

[0042] ;

[0043] in, Indicates the judgment result. Represents standardized sentence-level structural units. Represents a collection of domain resources; The function representing sentiment relevance judgment includes: encoding the sentence content and contextual fragment information of standardized sentence-level structural units using a pre-trained language model encoder to obtain sentence semantic representations. Based on domain resource set Domain Emotional Object Category Domain terminology glossary Emotional evidence recognition rules or templates The system matches the current sentence content to obtain the matching results of the corresponding sentiment object category, domain terminology vocabulary, and sentiment evidence recognition rules or templates, and converts them into numerical features to form a domain feature vector. The semantic representation of sentence elements is concatenated with the domain feature vector, and the probability of a sentence element belonging to a sentiment-related sentence element is calculated using a classification function, as shown below:

[0044] ;

[0045] in, The probability of representing sentiment-related sentence elements. and These represent the weight parameters and bias parameters, respectively. This represents a vector concatenation operation;

[0046] Based on the preset judgment threshold Output the judgment result:

[0047] ;

[0048] when When, it indicates that the current sentence element, i.e., the sentence element content, is a candidate sentence element for sentiment analysis, and the current sentence element contains sentiment objects, evaluation expressions, user experiences, service feedback, or implicit sentiment facts; when When the current sentence element is not a candidate sentence element for sentiment analysis, it indicates that the current sentence element is not a candidate sentence element for sentiment analysis.

[0049] Obtain standardized sentence-level structural units belonging to candidate sentence elements for sentiment analysis from standardized sentence-level structural units. and corresponding domain resource sets ;

[0050] Further optimization involves determining the threshold. It ranges from 0.5 to 0.8.

[0051] According to a preferred embodiment of the present invention, semantic parsing is performed on standardized sentence-level structural units belonging to candidate sentence elements for sentiment analysis to identify a set of candidate sentiment objects, and the set of sentiment objects is then filtered; including:

[0052] Standardized sentence-level structural units belonging to the candidate sentence elements of sentiment analysis and corresponding domain resource sets The input emotion object recognition function generates a candidate emotion object set through preset task prompts and a structured output format, as shown below:

[0053] ;

[0054] in, Indicates the first The set of candidate sentiment objects corresponding to each sentence element. This represents the nth emotional object. This represents an emotion object recognition model;

[0055] The process involves filtering emotional objects from the candidate emotional object set, including boundary identification, duplicate merging, and validity verification.

[0056] Boundary recognition includes: determining the start and end range of the emotion object based on the character position, word combination, head word and modifiers of the candidate object in the sentence unit, and deleting punctuation, function words and redundant modifiers that have no actual meaning;

[0057] Duplicate merging includes: normalizing two sentiment objects with the same text content, overlapping position range, synonymous or abbreviated relationship, or cosine similarity higher than a preset similarity threshold, and merging them into a unified object name;

[0058] Validity verification includes: calculating a validity score based on the correlation between the sentiment object and the evaluation expression, the degree of matching between the domain object category, the degree of contextual relevance, and the confidence score of the model output, to determine whether the candidate object belongs to the actual evaluated object, as shown below:

[0059] ;

[0060] in, Indicates the validity score; Let represent the weight, which lies in the interval [0,1] and satisfies: ;

[0061] The score represents the association between the sentiment object and the evaluation expression, including: obtaining the shortest dependency path between the sentiment object and the evaluation expression through a dependency parser, and obtaining the dependency relationship score based on the shortest dependency path. When the shortest dependency path length is 1, the dependency score is 1; when the shortest dependency path length is 2, the dependency score is 0.8; when the shortest dependency path length is 3, the dependency score is 0.6; when the shortest dependency path length is 4, the dependency score is 0.4; when the shortest dependency path length is greater than or equal to 5, or when there is no identifiable dependency path, the dependency score is 0. Simultaneously, the number of lexical units separating the sentiment object and the evaluation expression within the same sentence content is obtained, and the text distance score is obtained based on the number of lexical units separating them. When the emotional object and the evaluative expression are separated by 0-2 lexical units, the text distance score is 1; when separated by 3-5 lexical units, the text distance score is 0.8; when separated by 6-10 lexical units, the text distance score is 0.6; when separated by 11-20 lexical units, the text distance score is 0.3; when separated by more than 20 lexical units, the text distance score is 0; finally, the association score... ,in For association weights;

[0062] This represents the matching score between the sentiment object and the domain type. If all sentiment objects and domain types match, the score is 1. Otherwise, the cosine similarity between the sentiment object and the domain type is calculated, and the highest similarity value is taken as the matching score.

[0063] The score represents the relevance of the sentiment object to the current sentence element and its context, including: calculating the number of sentence elements separating the sentence element containing the sentiment object from the sentence element containing the related evaluative expression; when the sentiment object and the evaluative expression are located in the same sentence element, When located in an adjacent sentence element, take Take when there is a 1-sentence interval Take when there is a 2-sentence interval When the interval exceeds 2 sentence elements ;

[0064] This represents the confidence level output by the emotion object recognition model;

[0065] When the validity score of an emotional object is higher than the preset validity threshold, the emotional object is retained; when the validity score is lower than the preset threshold, the emotional object is filtered to obtain a set of candidate emotional objects after screening.

[0066] Further preferred, the similarity threshold is 0.80–0.90; association weight =0.5; the validity threshold is 0.6–0.8;

[0067] The optimal validity threshold is 0.7.

[0068] According to a preferred embodiment of the present invention, candidate evaluation components are extracted from the screened candidate emotional object set to form a candidate evidence set. After screening the candidate evidence set, a correlation score is calculated between the candidate emotional object set and the candidate evidence set. The final evidence set is obtained based on the correlation score. This includes:

[0069] For each emotional object in the selected candidate emotional object set, candidate evaluation components are extracted from the current sentence element and contextual information. These components include evaluation words, adjectives, adverbs, degree modifiers, negation words, transition words, causal phrases, factual outcome descriptions, and descriptive components that are dependent on the emotional object, forming a candidate evidence set. ;in, Represents the set of candidate evidence. Indicates the first Each candidate piece of evidence is emotional evidence;

[0070] Meanwhile, candidate evidence is initially screened through stop word filtering, invalid modifier filtering, and domain evidence template filtering.

[0071] The stop word filtering includes: pre-constructing a stop word list, and deleting function words, modal words, conjunctions, and words in the candidate evidence that do not carry evaluative meaning and are in the top 5% of the word frequency and have been manually confirmed not to carry evaluative meaning based on the stop word list;

[0072] Invalid modifier filtering includes: combining part-of-speech information, text distance between candidate evidence and sentiment object, and contextual semantics to remove sentiment objects that are not directly modified in candidate evidence, or to remove text components that are only used to describe time, place, quantity and background information;

[0073] Domain evidence template filtering includes: organizing the original text to be analyzed collected in the current business scenario, extracting evaluation words, negation structures, degree modifiers, transitional relationships, causal expressions, and factual result descriptions with evaluative meanings, and manually reviewing and deduplicating them to form a domain evidence template corresponding to the current business scenario; matching candidate evidence with the domain evidence template, and retaining the candidate evidence if all matches are the same; otherwise, generating vector representations of the candidate evidence and the domain evidence template using a semantic encoding model, and calculating the similarity using cosine similarity; retaining the candidate evidence when the highest similarity reaches a preset threshold, otherwise filtering the candidate evidence.

[0074] The overall relevance score between the sentiment object and the filtered set of candidate evidence is calculated using the following formula:

[0075] ;

[0076] in, This indicates the first element in the filtered set of candidate emotion objects. An emotional object, Indicates the first in the filtered set of candidate evidence. One candidate piece of evidence, express and The overall correlation score between them Indicates semantic similarity. Indicates the strength of the dependency relationship. Indicates the context adjacency relevance. Indicates domain matching degree. This indicates the weight in the overall relevance score, located at... An interval, satisfying: ;

[0077] The candidate evidence scores, i.e., the overall relevance scores, are normalized as follows:

[0078] ;

[0079] in, This represents the normalized candidate evidence score. This represents the minimum overall relevance score among all candidate evidence for the same emotional object. This represents the maximum value among all candidate evidence for the same emotional object, based on their overall relevance score. To prevent tiny constants with a denominator of zero;

[0080] Set evidence retention threshold The normalized candidate evidence scores are then used for screening. And select the candidate evidence with the highest normalized score. The following are the final pieces of evidence:

[0081] ;

[0082] in, Representing the object of affection The corresponding set of sentiment evidence is the final set of evidence. TopK represents the top K candidate evidence pieces with the highest scores after normalization.

[0083] Further preferred, retention threshold It ranges from 0.5 to 0.8;

[0084] Further optimization, semantic similarity This is used to measure the semantic correlation between sentiment objects and candidate evidence; through a pre-trained language model, the sentiment objects and candidate evidence are semantically encoded separately to obtain object vectors. and evidence vector The semantic similarity is calculated using cosine similarity, as shown below:

[0085] ;

[0086] Dependency strength This is used to measure the degree of association between sentiment objects and candidate evidence in syntactic structure; the shortest dependency path length between sentiment objects and candidate evidence is obtained using a dependency parser. And determine whether a direct dependency relationship exists, as shown below:

[0087] ;

[0088] in, This indicates a direct dependency relationship; it is defined when there is a direct dependency relationship between the emotional object and the candidate evidence, i.e., the shortest dependency path length is 1. ;otherwise, ; The weight represents the direct dependency relationship, and its value ranges from 1 to 2. When the sentiment object and the candidate evidence are located in different sentence units and there is no identifiable dependency relationship, the dependency relationship strength is set to 0.

[0089] Contextual adjacency relevance Used to measure the proximity of the sentiment object and candidate evidence in the original text; calculates the number of lexical units separating the sentiment object and candidate evidence according to sentence units and context. The context adjacency relevance is calculated based on the number of lexical units, as shown below:

[0090] ;

[0091] in, This represents the distance attenuation parameter;

[0092] Domain matching degree The formula used to measure the degree of matching between sentiment objects and candidate evidence and the set of domain resources in the current business scenario is as follows:

[0093] ;

[0094] in, Indicates the matching parameter, located in An interval, satisfying: ; Representing the emotional object and the domain emotional object category Match score between them; Glossary of Emotional Objects or Candidate Evidence and Domain Terminology Match score between them; This indicates the rules or templates for identifying candidate evidence and sentiment evidence. Match score between them;

[0095] The matching score includes: a score of 1 when the matched content is completely identical, and a score of 0 when the content is completely different; when the matched content is a synonym, abbreviation, or semantically similar expression, the domain terminology lexicon is queried. If there is a mapping relationship with the domain terminology lexicon, the matching score is 1. If there is no mapping relationship, a vector representation is generated through the semantic encoding model, and the cosine similarity is calculated; the cosine similarity is converted to the [0,1] interval, and the converted value is used as the matching score.

[0096] A further preferred option is a distance attenuation parameter of 10 to 30.

[0097] According to a preferred embodiment of the present invention, step 3 specifically includes the following steps:

[0098] Constructing an object-level sentiment polarity inference function includes: concatenating standardized sentence-level structural units belonging to sentiment analysis candidate sentence elements, the filtered candidate sentiment object set, the final evidence set, and the corresponding domain resource set, as shown below:

[0099] ;

[0100] in, Indicates splicing characteristics, Indicates a splicing operation;

[0101] The concatenated features are encoded using a semantic coding model to obtain an object-level semantic representation. And the probability distribution of candidate sentiment polarity is obtained through a classification function:

[0102] ;

[0103] in, and These represent the weight parameters and bias parameters used in the classification calculation, respectively. Represents the classification function; The probability distribution representing the candidate polarity;

[0104] When the sentiment evidence in the final evidence set is directly an evaluative word, the sentiment evidence is identified as an explicit sentiment expression. The basic polarity of the evaluative word is determined according to the sentiment lexicon, and the basic polarity is adjusted by combining negative expressions and degree modifiers to obtain the explicit sentiment judgment result of the sentiment evidence for the sentiment object. ;

[0105] When the sentiment evidence in the final evidence set does not contain evaluative words, but reflects an evaluative tendency towards the sentiment object through factual outcomes, anomalous states, repeated processing, or user experience consequences, the sentiment evidence is identified as implicit sentiment expression, and the implicit sentiment judgment result is expressed as follows:

[0106] ;

[0107] in, This indicates the implicit sentiment judgment result. Representing implicit sentiment inference functions, including: classifying sentiment objects Emotional evidence and domain resource collection As input, the probability distribution of candidate polarities is obtained using the object-level sentiment polarity inference function, and the sentiment category with the highest probability is selected as the implicit sentiment judgment result.

[0108] When the sentiment evidence in the final evidence set contains negation, degree modification, adversative relationship, multiple object coexistence, or cross-sentence reference, the sentiment evidence is identified as a complex structured sentiment expression. Based on punctuation, negation, degree modification, adversative relationship, and the relationship between the sentiment object and the corresponding sentiment evidence, each piece of sentiment evidence is associated with its corresponding sentiment object. Then, the sentiment object corresponding to the sentiment evidence and the contextual fragment information are input into the object-level sentiment polarity inference function to obtain the probability distribution of the candidate polarity of the complex structured sentiment expression, i.e., the complex structured sentiment judgment result. ;

[0109] When multiple pieces of emotional evidence correspond to the same emotional object, the weight of each piece of emotional evidence is determined based on the comprehensive relevance score:

[0110] ;

[0111] in, Indicates the first Emotional evidence in the emotional object Weights in polarity determination This represents the r-th piece of emotional evidence;

[0112] According to weight Obtain the polarity determination result:

[0113] ;

[0114] in, Indicates based on the emotional object The polarity judgment results of each corresponding piece of emotional evidence; For explicit sentiment judgment results Or implicit sentiment judgment result Or complex structure emotion judgment results ;

[0115] The polarity judgment results of each sentiment evidence are weighted and fused with the probability distribution of the candidate polarities:

[0116] ;

[0117] in, Representing the object of affection The final probability distribution of emotional polarity, This represents the weight of the language model output, with a value range of [value missing]. ;

[0118] Selecting the final emotional polarity probability distribution The sentiment category with the highest probability is selected as the candidate sentiment polarity; when the highest probability is greater than a preset reliability threshold, the candidate sentiment polarity is used as the final sentiment polarity; when the highest probability is lower than the preset threshold, or When the probability difference between the positive and negative categories is less than a preset difference threshold, it is marked as pending review or mixed sentiment.

[0119] Further optimization involves the weighting of the language model output. The default reliability threshold is 0.5; the preset reliability threshold range is 0.7 to 0.9; the preset difference threshold is 0.1 to 0.2.

[0120] The optimal setting is a confidence threshold of 0.8 and a difference threshold of 0.15.

[0121] According to a preferred embodiment of the present invention, step 4 includes the following specific implementation process:

[0122] For each identified sentiment object, generate object-level structured analysis results:

[0123] ;

[0124] in, Indicates the first The structured analysis results corresponding to each emotion object; if the text to be analyzed contains If there are several emotional objects, then the structured result set of the text to be analyzed is represented as follows: , Indicates the first One emotional object; Indicates the overall confidence level; Represents object-level structured results The processing status is such that the selected candidate sentiment object set, the final evidence set, and the final sentiment polarity are the same, and the overall confidence level is... Reaching the preset confidence threshold When the overall confidence level does not reach the preset confidence threshold, the processing status is set to automatic pass; when the overall confidence level does not reach the preset confidence threshold, the processing status is set to automatic pass. Or there may be conflicting evidence or a final probability distribution of emotional polarity. When the difference between the highest probability and the second highest probability is less than the preset difference threshold, and the correspondence between the emotional object and the emotional evidence is unclear, the processing status is set to pending review. Representing the object of affection Corresponding candidate-level sentiment polarity: ;in, This represents a predefined set of emotional polarity categories, including at least positive, negative, neutral, and undetermined. This represents the probability distribution of emotional polarity. Emotional object Belongs to the category of emotional polarity The probability; when the structured analysis results are marked as pending review, Indicates candidate polarity; The inference specification is indicated by confirming the inference specification based on the structured output fields in the domain resource set. If no matching inference specification exists, it is determined based on the sentiment object. Collection of evidence Candidate polarity As well as the scope of negation, degree of modification, adversative relationship and domain evidence template, the reasoning explanation is generated using a preset explanation template or a large language model;

[0125] when When automatic passage is enabled, Directly used as the final object-level sentiment analysis result; when For the purpose of review, The candidate analysis results are submitted for manual review or secondary model verification, and updated based on the review results to form confirmed object-level sentiment analysis results.

[0126] For the set of structured analysis results of the same text to be analyzed There are multiple different emotional objects in it. or a certain emotional object The corresponding final set of evidence In cases where multiple pieces of sentiment evidence exist, the object-level sentiment analysis results are integrated and conflict resolved; including:

[0127] The names of sentiment objects are standardized and converted into standard names, namely abbreviations, aliases, and synonyms in the domain terminology lexicon. When two sentiment objects have the same name, or when there is at least one relationship between them—synonym, abbreviation, alias, reference, overlapping textual scope, or both pointing to the same evaluated entity—they are considered semantically identical sentiment objects. When the sentiment object similarity of two sentiment objects reaches a preset threshold... When both belong to the same category of emotional objects, they are considered to be highly similar emotional objects.

[0128] For sentiment objects identified as semantically identical or highly similar, they are merged. Simultaneously, the standard name from the domain terminology lexicon is selected as the merged object name. If no standard name exists in the domain terminology lexicon, the object name with the highest object recognition confidence, the most complete text expression, or the highest frequency of occurrence is selected as the merged object name. Furthermore, the evidence sets corresponding to the merged objects undergo union processing, deleting evidence with identical text content, overlapping positional ranges, or highly repetitive semantics, while retaining the source sentence elements, positional identifiers, comprehensive relevance scores, and evidence-level judgment confidence for each piece of evidence. The evidence-level judgment confidence is represented by the sentiment polarity probability distribution. In this context, emotional evidence falls under the category of emotional polarity. The maximum probability;

[0129] For the set of structured analysis results of the same text to be analyzed In cases where multiple pieces of emotional evidence relate to the same emotional object, the evidence fusion weight is calculated based on the strength of evidence, the degree of contextual proximity, and the confidence level of each piece of emotional evidence. :

[0130] ;

[0131] in, This represents the strength of evidence, or support strength, of the k-th sentiment evidence. This indicates the degree of contextual proximity between the object of emotion and the evidence of emotion. Indicates the first The confidence level of evidence corresponding to each piece of emotional evidence; This indicates the strength of the emotional evidence presented in the r-th piece. This indicates the degree of contextual proximity between the emotional object and the r-th piece of emotional evidence. Indicates the first The confidence level of evidence corresponding to each piece of emotional evidence;

[0132] We use evidence fusion weights to weight and fuse the probability distributions of sentiment polarity from different sentiment evidence:

[0133] ;

[0134] in, This represents the probability distribution of the emotional polarity after fusion;

[0135] For the set of structured analysis results of the same text to be analyzed For emotional objects with conflicting evidence, calculate conflict resolution scores for each piece of emotional evidence:

[0136] ;

[0137] in, Represents the overall relevance score, parameters lie in An interval, satisfying: ;

[0138] When the probability distribution of emotional polarity after fusion The preset polarity judgment threshold is reached, and When the difference between the highest probability and the second highest probability reaches a preset difference threshold, the sentiment polarity corresponding to the highest probability is taken as the object-level judgment result, and the sentiment polarity corresponding to the highest probability is taken as the object-level sentiment analysis result; when the evidence strength of the sentiment evidence of positive and negative categories reaches the evidence strength threshold, or when the difference between the highest probability and the second highest probability in the fused sentiment polarity probability distribution is lower than the preset difference threshold, the sentiment object is marked as mixed sentiment or pending review;

[0139] Further preferred, a preset threshold The threshold value is 0.85; the preset polarity judgment threshold is 0.7–0.9; the evidence strength threshold is 0.7–0.9.

[0140] Further preferred, the similarity of the emotional object includes: for the emotional object and The formula for calculating the similarity between emotional objects is:

[0141] ;

[0142] in, express and Similarity of emotional objects The vector semantic similarity of two emotion objects is obtained by generating object vectors through a text encoding model and calculating the cosine similarity. The lexical overlap of object names is calculated by Jaccard similarity to the character set, word set, or keyword set of the sentiment object; This indicates the degree of matching of synonyms, abbreviations, alternative names, or referential relationships. It is 1 when two sentiment objects are marked as synonyms in the domain terminology lexicon, or when two sentiment objects point to the same evaluated entity; otherwise, it is 0. This represents the similarity parameter, with a value range of [value range missing]. The interval, and satisfying ;

[0143] The strength of evidence is used to indicate the degree to which a candidate piece of evidence supports the judgment of the target emotional object and emotional polarity; Candidate evidence relative to the emotional object The strength of evidence is expressed as follows:

[0144] ;

[0145] in, This indicates the strength of evidence, or the strength of support. This indicates the overall relevance score; It indicates the clarity of the sentiment orientation of the evidence, and is determined based on the highest polarity probability in the evidence-level sentiment classification results; The semantic similarity indicates the degree of matching between candidate evidence and domain evaluation words, negation structures, factual outcome descriptions, or other sentiment evidence templates. This indicates the degree of polarity consistency between emotional evidence and other emotional evidence concerning the same emotional object; a value of 1 indicates the same emotional polarity as other evidence concerning the same emotional object, and a value of 0 indicates a different polarity. (Parameter) All located in An interval, satisfying: ;

[0146] Contextual proximity between the emotional object and the candidate evidence Represented as:

[0147] ;

[0148] in, The number of lexical units or sentence units representing the interval between the emotional object and the candidate evidence. This represents the distance attenuation parameter;

[0149] Overall confidence level As shown below:

[0150] ;

[0151] in, This represents the confidence level in identifying the emotional object. Indicates the confidence level of the emotional evidence match. Indicates the confidence level of sentiment polarity reasoning. These are weight parameters, located at The interval, and the sum of the weight parameters is 1;

[0152] Among them, the confidence level of emotion object recognition Calculated based on object boundary recognition probability, domain object category matching degree, and the degree of correlation between the object and the evaluation expression:

[0153] ;

[0154] in, This represents the probability of object boundaries or object categories output by the emotion object recognition model; This indicates the degree of matching between the candidate object and the domain object category and the domain terminology vocabulary. When the candidate object matches the domain sentiment object category... or field terminology glossary If the object name, synonym, abbreviation, or alias is the same, the value is 1; if they are different, the semantic similarity between the candidate object and each domain object category or domain term is calculated, and the highest similarity is taken as the result. ; The correlation score represents the degree of association between the candidate object and the evaluation expression; parameters lie in An interval, satisfying: ;

[0155] Confidence of Emotional Evidence Matching The weighted calculation is based on the overall relevance score of each piece of evidence in the final evidence set:

[0156] ;

[0157] in, For evidence fusion weights;

[0158] Confidence of Emotional Polarity Inference Calculated based on the highest probability in the fused polarity probability distribution, and the difference between the highest and second-highest probabilities:

[0159] ;

[0160] in, This represents the highest probability in the polarity probability distribution after fusion. Indicates the second highest probability; parameter All located in An interval, satisfying: ;

[0161] Further optimized, weight parameters They were set to 0.3, 0.3, and 0.4 respectively.

[0162] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the above-described text sentiment analysis method that integrates sentiment object extraction and sentiment polarity reasoning.

[0163] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described text sentiment analysis method that integrates sentiment object extraction and sentiment polarity inference.

[0164] Another aspect of the present invention provides a text sentiment analysis system that integrates sentiment object extraction and sentiment polarity reasoning, comprising:

[0165] The structural unit and domain resource construction module is configured to: perform structured preprocessing on the original text to be analyzed, construct standardized sentence-level structural units, and construct a set of domain resources according to the domain type of the original text to be analyzed;

[0166] The sentiment object and candidate evidence acquisition module is configured to: judge sentiment relevance based on standardized sentence element-level structural units and domain resource set, filter out sentiment analysis candidate sentence elements, and acquire standardized sentence element-level structural units belonging to the sentiment analysis candidate sentence elements;

[0167] Semantic parsing is performed on standardized sentence-level structural units that belong to the candidate sentence elements of sentiment analysis to identify the set of candidate sentiment objects and filter the set of sentiment objects;

[0168] Candidate evaluation components are extracted from the selected set of candidate emotional objects to form a set of candidate evidence. After screening the set of candidate evidence, a comprehensive correlation score between the set of candidate emotional objects and the set of candidate evidence is calculated. The final set of evidence is obtained based on the correlation score.

[0169] The sentiment polarity reasoning module is configured to: perform sentiment polarity reasoning on the standardized sentence-level structural units belonging to the sentiment analysis candidate sentence elements, the filtered candidate sentiment object set, the final evidence set and the corresponding domain resource set, and obtain the final sentiment polarity;

[0170] The fusion and conflict resolution module is configured to: construct object-level structured analysis results, review them, and generate object-level sentiment analysis results;

[0171] Based on the similarity of sentiment objects, the strength of evidence, and the overall confidence level, the results of object-level sentiment analysis are integrated and conflict resolved.

[0172] The beneficial effects of this invention are as follows:

[0173] 1. This invention introduces a text structuring module and a sentiment-related sentence element screening mechanism, which can segment, standardize, mark the source, and preserve the context of unstructured text such as user comments, dialogue records, and feedback, forming traceable sentence element-level analysis units, and filtering out text fragments without clear sentiment analysis value, thereby reducing irrelevant content from entering the subsequent reasoning process and improving the processing efficiency and stability of text sentiment analysis.

[0174] 2. This invention constructs an emotional object recognition module and an emotional evidence recognition module, which can automatically locate the evaluated objects such as product attributes, service links, event behaviors, and processing results in the text, and extract related evaluation words, modifiers, negation structures, transitional relationships and implicit factual descriptions, effectively establishing the correspondence between emotional objects and emotional evidence, and reducing the problems of object recognition errors and evidence mismatch.

[0175] 3. This invention designs an emotion polarity reasoning module and a result fusion and confidence verification module, which can perform object-level emotion judgment for complex scenarios such as explicit emotion, implicit emotion, negative expression, transition structure and multi-object mixed evaluation, and output structured results including emotion object, evidence fragment, emotion polarity, confidence and reasoning explanation, significantly improving the accuracy, interpretability, traceability and verifiability of emotion analysis results. Attached Figure Description

[0176] Figure 1 This is a flowchart illustrating a text sentiment analysis method and system that integrates sentiment object extraction and sentiment polarity reasoning according to the present invention. Detailed Implementation

[0177] The present invention will be further described below with reference to the embodiments and accompanying drawings, but is not limited thereto.

[0178] Example 1

[0179] A text sentiment analysis method that integrates sentiment object extraction and sentiment polarity inference, such as... Figure 1 As shown, it includes:

[0180] Step 1: Perform structured preprocessing on the original text to be analyzed, construct standardized sentence-level structural units, and construct a domain resource set according to the domain type of the original text to be analyzed;

[0181] Step 2: Based on the standardized sentence element-level structural units and the domain resource set, determine the sentiment relevance, screen out the candidate sentence elements for sentiment analysis, and obtain the standardized sentence element-level structural units belonging to the candidate sentence elements for sentiment analysis;

[0182] Semantic parsing is performed on standardized sentence-level structural units that belong to the candidate sentence elements of sentiment analysis to identify the set of candidate sentiment objects and filter the set of sentiment objects;

[0183] Candidate evaluation components are extracted from the selected set of candidate emotional objects to form a set of candidate evidence. After screening the set of candidate evidence, a comprehensive correlation score between the set of candidate emotional objects and the set of candidate evidence is calculated. The final set of evidence is obtained based on the correlation score.

[0184] Step 3: Perform sentiment polarity reasoning on the standardized sentence-level structural units belonging to the sentiment analysis candidate sentence elements, the filtered candidate sentiment object set, the final evidence set, and the corresponding domain resource set to obtain the final sentiment polarity;

[0185] Step 4: Construct object-level structured analysis results, review them, and generate object-level sentiment analysis results;

[0186] Based on the similarity of sentiment objects, the strength of evidence, and the overall confidence level, the results of object-level sentiment analysis are integrated and conflict resolved.

[0187] Example 2

[0188] The text sentiment analysis method that integrates sentiment object extraction and sentiment polarity reasoning as described in Example 1 differs in that:

[0189] The specific implementation process of step 1 includes:

[0190] This step is used to perform structured preprocessing on the original text to be analyzed, and to build standard input units with sentence and segment hierarchy, source information, contextual positional relationship and traceable identifier, so as to provide basic data support for subsequent sentiment object recognition, sentiment evidence recognition and sentiment polarity inference;

[0191] The original text to be analyzed is segmented, including: automatically recognizing periods, semicolons, question marks, exclamation marks, line breaks, transition words, and semantic boundary markers (paragraph start and end markers, topic start and end markers, etc.), splitting the original text to be analyzed into multiple semantically complete sentence elements, forming sentence element-level structural units; at the same time, the sentence order, contextual adjacency relationship, paragraph number, and text source are preserved.

[0192] For example, the original text to be analyzed is:

[0193] "This hotel has a great location, but check-in at the front desk is too slow. The rooms are generally clean, but the soundproofing is just average."

[0194] After structuring, the following sentence structure can be formed:

[0195] [{"Sentence ID": "S1","Text Source": "User Review A","Sentence Content": "This hotel has a great location, but check-in at the front desk is too slow."},{"Sentence ID": "S2","Text Source": "User Review A","Sentence Content": "The room was generally clean, but the soundproofing was just average."}]

[0196] Each sentence-level structural unit is injected with the sentence content, text source, paragraph number, sentence number, and contextual information of the original text to be analyzed, forming a standard input pattern of "[Source: Text Identifier][Location: Sentence Number] + Sentence Content". This method ensures traceability of each analysis unit in subsequent model processing, reducing the context loss problem caused by long text segmentation and generating standardized sentence-level structural units, as shown below:

[0197] ;

[0198] in, This represents the i-th standardized sentence-level structural unit. This indicates the text source of the original text to be analyzed corresponding to the i-th standardized sentence-level structural unit. This represents the sentence element content of the i-th standardized sentence element-level structural unit. Indicates context fragment information, This represents the positional identifiers in the original text to be analyzed, including sentence element numbers, paragraph numbers, and arrangement order; this unified format serves as the standard input interface for subsequent domain adaptation, object recognition, and polarity reasoning.

[0199] Based on the business scenario to which the text to be analyzed belongs, domain constraint information is provided for subsequent model analysis, thereby improving the accuracy of object recognition and the availability of results in different application scenarios.

[0200] Based on the text source, business tags, or user settings of the original text to be analyzed, determine the domain type, i.e., the business scenario, to which the original text to be analyzed belongs, and obtain the corresponding domain resource set:

[0201] ;

[0202] in, Indicates the first Domain types refer to the collection of domain resources corresponding to business scenarios (such as user comment analysis, intelligent customer service quality inspection, public opinion monitoring, product experience analysis, government service feedback, and industry text assistance, etc.); This represents the category of domain sentiment objects, including the entities, attributes, functions, service links, behaviors, or processing results being evaluated in the business scenario. The category of domain sentiment objects can be determined based on the business classification system, historical labeled samples, or manual configuration. A domain terminology glossary, including domain object names, synonyms, abbreviations, alternative names, and domain-specific expressions, can be obtained through domain corpus extraction, business terminology import, model generation, and manual correction. The rules or templates for identifying sentiment evidence include the modifying relationship between evaluation words and sentiment objects, negation structure, transition structure, degree modification relationship, causal relationship and implicit fact expression pattern. The rules or templates can be obtained through preset rules, historical sample induction, model generation and manual review. The structured output fields include sentiment object, object category, evidence fragment, sentiment polarity, confidence level, reasoning explanation, and processing status. The output fields can be pre-configured according to specific business needs.

[0203] For example, in the field of hotel reviews, the domain sentiment object category Includes location, hygiene, soundproofing, facilities and services; glossary of terms in the field. Includes object names such as "front desk," "check-in," and "soundproofing," along with their synonyms; rules or templates for emotional evidence. This includes rules governing the modification relationships of evaluative words, the scope of negation, and the object attribution in contrastive structures; structured output fields. This includes the emotional object, object category, evidence fragment, emotional polarity, confidence level, and reasoning explanation. After determining that the text to be analyzed belongs to the hotel review domain, the system reads the above resources from the domain resource library and inputs them together with structured sentence elements into the emotional object recognition module and the emotional evidence recognition module.

[0204] The aforementioned domain resources are pre-stored in the domain resource library. After determining the domain to which the text belongs, the system calls the corresponding resources and can update the domain resources based on the results of manual review, user corrections, or feedback from the business system.

[0205] The above set allows for dynamic adjustments to be made for specific evaluation objects and expression habits in different fields;

[0206] Sentiment relevance is determined based on standardized sentence-level structural units and domain resource sets. Candidate sentence units for sentiment analysis are selected, and standardized sentence-level structural units belonging to these candidate sentence units are obtained; including:

[0207] This step is used to determine whether structured sentence elements contain subjective evaluations, user experiences, service feedback, attitude expressions, or factual emotional result descriptions, thereby avoiding invalid reasoning on pure factual statements, sentences without evaluation relationships, or sentence elements without analytical value;

[0208] The standardized sentence-level structural units and the domain resource set are input into the sentiment relevance judgment function to obtain the judgment result, as shown below:

[0209] ;

[0210] in, Indicates the judgment result. Represents standardized sentence-level structural units. Represents a collection of domain resources; The function representing sentiment relevance judgment includes: encoding the sentence content and contextual fragment information of standardized sentence-level structural units using a pre-trained language model encoder (such as BERT, RoBERTa, or MacBERT) to obtain the sentence semantic representation. Based on domain resource set Domain Emotional Object Category Domain terminology glossary Emotional evidence recognition rules or templates The system matches the content of the current sentence element (e.g., for domain sentiment object categories, it determines whether a sentiment object belonging to that category appears in the current sentence element, or whether it can be classified into a certain object category based on the semantics of the sentence element; for domain terminology lexicon, it determines whether the corresponding term, synonym, abbreviation, or alternative name directly appears in the sentence element; for sentiment evidence recognition rules or templates, it determines whether the corresponding evaluation words, negative expressions, degree modifiers, factual result descriptions, and other sentiment evidence appear in the sentence element), obtains the matching results of the corresponding sentiment object category, domain terminology lexicon, and sentiment evidence recognition rules or templates, and converts them into numerical features to form a domain feature vector. The semantic representation of sentence elements is concatenated with the domain feature vector, and the probability of a sentence element belonging to a sentiment-related sentence element is calculated using a classification function, as shown below:

[0211] ;

[0212] in, The probability of representing sentiment-related sentence elements. and These represent the weight parameters and bias parameters, respectively. This represents a vector concatenation operation;

[0213] Based on the preset judgment threshold Output the judgment result:

[0214] ;

[0215] when When the current sentence element (i.e., its content) is selected as a candidate sentence element for sentiment analysis, it contains sentiment objects, evaluative expressions (negative expressions, degree modifiers, factual results descriptions, and evaluative content reflected through relationships such as transition and causality; for example: "The screen is very good"—"very good" is an evaluative word and also an evaluative expression; "The battery needs to be charged at 4 pm"—there is no obvious evaluative word, but "needs to be charged at 4 pm" is a factual result description and also an evaluative expression; "The problem has been handled three times but it is still not resolved"—this is an implicit evaluative expression), user experience, service feedback, or implicit sentiment facts. This information is then input into the subsequent sentiment object recognition module. When the current sentence element is not a candidate sentence element for sentiment analysis and does not have clear sentiment analysis value, the system skips further processing.

[0216] Obtain standardized sentence-level structural units belonging to candidate sentence elements for sentiment analysis from standardized sentence-level structural units. and corresponding domain resource sets ;

[0217] This screening mechanism enables rapid filtering of irrelevant sentence elements in large-scale texts, improving overall analysis efficiency and reducing model call costs.

[0218] Determine the threshold The value is 0.5 to 0.8, and can be adjusted based on the classification accuracy, recall, or F1 score in the validation samples.

[0219] Semantic parsing is performed on standardized sentence-level structural units belonging to candidate sentence elements for sentiment analysis to identify a set of candidate sentiment objects, and the set of sentiment objects is then filtered; including:

[0220] This step utilizes a pre-trained language model or a large language model to perform semantic parsing on sentiment-related sentence units, identify the sentiment objects most likely to carry evaluation relationships, and realize the mapping between text content and evaluation targets from surface vocabulary to semantic object level.

[0221] Standardized sentence-level structural units belonging to the candidate sentence elements of sentiment analysis and corresponding domain resource sets The input emotion object recognition function generates a candidate emotion object set through preset task prompts and a structured output format, as shown below:

[0222] ;

[0223] in, Indicates the first The set of candidate sentiment objects corresponding to each sentence element. This represents the nth emotional object. This represents a sentiment object recognition model (pre-trained language models such as BERT, RoBERTa, or MacBERT); the sentiment object recognition function generates candidate sentiment objects that are highly relevant to the content of the sentence by analyzing the semantic intent, noun phrases, modification relations, dependency relations, evaluation structure, and domain object categories of sentence elements.

[0224] The process involves filtering emotional objects from the candidate emotional object set, including boundary identification, duplicate merging, and validity verification.

[0225] Boundary recognition includes: determining the start and end range of the sentiment object based on the character position, word combination, headword, and modifiers of candidate objects within the sentence element, and deleting punctuation, function words, and redundant modifiers without actual meaning; for example, for the sentence element "This phone's battery life is very short," firstly, candidate expressions such as "battery," "battery life," and "battery life time" are identified, and the true sentiment object boundary is determined to be "battery life time" based on character position, word combination, headword, and modifier relationships, while background or non-subjective components such as "this phone's" are deleted. This determines the complete start and end range of the sentiment object within the sentence element.

[0226] The duplicate merging includes: normalizing two emotional objects with the same text content, overlapping position range, synonymous or abbreviated relationship, or cosine similarity (by semantically encoding two emotional objects through language models such as SBERT and SimCSE to obtain two encoding vectors, and calculating the cosine similarity between the two encoding vectors) higher than a preset similarity threshold, and merging them into a unified object name;

[0227] Validity verification includes: calculating a validity score based on the correlation between the sentiment object and the evaluation expression, the degree of matching between the domain object category, the degree of contextual relevance, and the confidence score of the model output, to determine whether the candidate object belongs to the actual evaluated object, as shown below:

[0228] ;

[0229] in, Indicates the validity score; Let represent the weight, which lies in the interval [0,1] and satisfies: ;

[0230] The score represents the association between the sentiment object and the evaluation expression, including: obtaining the shortest dependency path between the sentiment object and the evaluation expression through a dependency parser, and obtaining the dependency relationship score based on the shortest dependency path. When the shortest dependency path length is 1, the dependency score is 1; when the shortest dependency path length is 2, the dependency score is 0.8; when the shortest dependency path length is 3, the dependency score is 0.6; when the shortest dependency path length is 4, the dependency score is 0.4; when the shortest dependency path length is greater than or equal to 5, or when there is no identifiable dependency path, the dependency score is 0. Simultaneously, the number of lexical units separating the sentiment object and the evaluation expression within the same sentence content is obtained, and the text distance score is obtained based on the number of lexical units separating them. When the emotional object and the evaluative expression are separated by 0-2 lexical units, the text distance score is 1; when separated by 3-5 lexical units, the text distance score is 0.8; when separated by 6-10 lexical units, the text distance score is 0.6; when separated by 11-20 lexical units, the text distance score is 0.3; when separated by more than 20 lexical units, the text distance score is 0; finally, the association score... ,in The association score is determined by the number of terms in the dependency path; the shorter the dependency path and the fewer the number of terms in between, the higher the association score.

[0231] This represents the matching score between the sentiment object and the domain type. If all sentiment objects and domain types match, the score is 1. Otherwise, the cosine similarity between the sentiment object and the domain type is calculated, and the highest similarity value is taken as the matching score.

[0232] The score represents the relevance of the sentiment object to the current sentence element and its context, including: calculating the number of sentence elements separating the sentence element containing the sentiment object from the sentence element containing the related evaluative expression; when the sentiment object and the evaluative expression are located in the same sentence element, When located in an adjacent sentence element, take Take when there is a 1-sentence interval Take when there is a 2-sentence interval When the interval exceeds 2 sentence elements ;

[0233] This represents the confidence level output by the emotion object recognition model (the confidence level directly output by pre-trained language models such as BERT, RoBERTa, or MacBERT).

[0234] When the validity score of an emotional object is higher than the preset validity threshold, the emotional object is retained; when the validity score is lower than the preset threshold, the emotional object is filtered to obtain a set of candidate emotional objects after screening.

[0235] Through the above processing, the system can filter out common nouns, background entities, and irrelevant objects to obtain a set of sentiment objects with clear boundaries, consistent names, and evaluation relationships.

[0236] The similarity threshold is 0.80–0.90, and can be adjusted based on the object identification and merging results in the validation samples; association weight. =0.5; the validity threshold is 0.6 to 0.8, and can be adjusted according to the object recognition results in the verification samples;

[0237] The validity threshold is 0.7.

[0238] Candidate evaluation components are extracted from the screened set of candidate sentiment objects to form a set of candidate evidence. After further screening of the candidate evidence set, a correlation score is calculated between the set of candidate sentiment objects and the set of candidate evidence. The final evidence set is obtained based on the correlation score; including:

[0239] This module is used to achieve semantic alignment between the emotional object and its evaluation basis, and automatically filter out the emotional evidence fragments most relevant to the target object, providing a basis for subsequent emotional polarity judgment.

[0240] For each emotional object in the selected candidate emotional object set, candidate evaluation components are extracted from the current sentence element and contextual information. These components include evaluation words, adjectives, adverbs, degree modifiers, negation words, transition words, causal phrases, factual outcome descriptions, and descriptive components that are dependent on the emotional object, forming a candidate evidence set. ;in, Represents the set of candidate evidence. Indicates the first Each candidate piece of evidence is emotional evidence;

[0241] Meanwhile, candidate evidence is initially screened through stop word filtering, invalid modifier filtering, and domain evidence template filtering.

[0242] The stop word filtering process includes: pre-constructing a stop word list (such as the publicly available Harbin Institute of Technology stop word list, Baidu stop word list, etc., combined with domain corpus statistics and manual configuration), and deleting function words, modal words, conjunction words, and words in the candidate evidence that do not carry evaluation meaning and are in the top 5% of the word frequency and have been manually confirmed not to carry evaluation meaning based on the stop word list;

[0243] Invalid modifier filtering includes: combining part-of-speech information, text distance between candidate evidence and sentiment object, and contextual semantics to remove sentiment objects that are not directly modified in candidate evidence, or to remove text components that are only used to describe time, place, quantity and background information;

[0244] Domain evidence template filtering includes: organizing the original text to be analyzed collected in the current business scenario, extracting evaluation words, negation structures, degree modifiers, transitional relationships, causal expressions, and factual result descriptions with evaluative meanings, and manually reviewing and deduplicating them to form a domain evidence template corresponding to the current business scenario; matching candidate evidence with the domain evidence template, and retaining the candidate evidence if all matches are the same; otherwise, generating vector representations of the candidate evidence and the domain evidence template using a semantic encoding model (pre-trained language models such as BERT, RoBERTa, and MacBERT), and calculating the similarity using cosine similarity; retaining the candidate evidence when the highest similarity reaches a preset threshold (which can be set to 0.70-0.90, for example, 0.80, and can be adjusted according to the evidence recognition effect of the verification samples); otherwise, filtering the candidate evidence.

[0245] The overall relevance score between the sentiment object and the filtered set of candidate evidence is calculated using the following formula:

[0246] ;

[0247] in, This indicates the first element in the filtered set of candidate emotion objects. An emotional object, Indicates the first in the filtered set of candidate evidence. One candidate piece of evidence, express and The overall correlation score between them Indicates semantic similarity. Indicates the strength of the dependency relationship. Indicates the context adjacency relevance. Indicates domain matching degree. This indicates the weight in the overall relevance score, located at... An interval, satisfying: ;

[0248] To enhance the comparability of scores across different subjects, the candidate evidence scores, i.e., the overall relevance scores, are normalized, as shown below:

[0249] ;

[0250] in, This represents the normalized candidate evidence score. This represents the minimum overall relevance score among all candidate evidence for the same emotional object. This represents the maximum value among all candidate evidence for the same emotional object, based on their overall relevance score. To prevent the use of tiny constants with a denominator of zero (which can be taken as...) Its value range can be set to to When there is only one candidate piece of evidence for the same emotional object, or when all candidate pieces of evidence have the same score, the corresponding normalized score can be directly set to 1).

[0251] Set evidence retention threshold The normalized candidate evidence scores are then used for screening. And select the candidate evidence with the highest normalized score. The following are the final pieces of evidence:

[0252] ;

[0253] in, Representing the object of affection The corresponding sentiment evidence set is the final evidence set. TopK represents the top K candidate evidence scores with the highest normalized scores (its value can be set from 1 to 5 and can be adjusted according to the text length and business scenario; when the number of candidate evidences that meet the threshold condition is less than 5, all candidate evidences that meet the condition are retained).

[0254] Retention threshold It ranges from 0.5 to 0.8;

[0255] semantic similarity This is used to measure the semantic relevance between sentiment objects and candidate evidence; through a pre-trained language model (SBERT, SimCSE, or a language model with equivalent text vector representation capabilities), the sentiment objects and candidate evidence are semantically encoded to obtain object vectors. and evidence vector The semantic similarity is calculated using cosine similarity, as shown below:

[0256] ;

[0257] Through the above transformation, semantic similarity is limited to The higher the semantic similarity, the stronger the semantic association between the candidate evidence and the emotional object.

[0258] Dependency strength This is used to measure the degree of association between sentiment objects and candidate evidence in syntactic structure; the shortest dependency path length between sentiment objects and candidate evidence is obtained using a dependency parser. And determine whether there is a direct dependency relationship (that is, there is a direct parent-child node relationship between the emotional object and the candidate evidence. Direct dependency relationships include attributive-head relationship, adverbial-head relationship, subject-verb relationship, verb-object relationship, and supplementary relationship, etc.), as shown below:

[0259] ;

[0260] in, This indicates a direct dependency relationship; it is defined when there is a direct dependency relationship between the emotional object and the candidate evidence, i.e., the shortest dependency path length is 1. ;otherwise, ; The weight represents the direct dependency relationship, and its value ranges from 1 to 2. The shorter the dependency path, the stronger the dependency relationship between the two; when the emotional object and the candidate evidence are located in different sentence units and there is no identifiable dependency relationship, the dependency relationship strength is set to 0.

[0261] Contextual adjacency relevance Used to measure the proximity of the sentiment object and candidate evidence in the original text; calculates the number of lexical units separating the sentiment object and candidate evidence according to sentence units and context. The context adjacency relevance is calculated based on the number of lexical units, as shown below:

[0262] ;

[0263] in, The distance decay parameter can be determined based on the validation samples. The closer the emotional object is to the candidate evidence, the higher the context adjacency relevance. For candidate evidence that crosses sentence elements, a decay coefficient can be added based on the number of intervals between the two sentence elements to reduce the score of irrelevant evidence that is far away.

[0264] Domain matching degree The formula used to measure the degree of matching between sentiment objects and candidate evidence and the set of domain resources in the current business scenario is as follows:

[0265] ;

[0266] in, Indicates the matching parameter, located in An interval, satisfying: ; Representing the emotional object and the domain emotional object category Match score between them; Glossary of Emotional Objects or Candidate Evidence and Domain Terminology Match score between them; This indicates the rules or templates for identifying candidate evidence and sentiment evidence. Match score between them;

[0267] The matching score includes: a score of 1 when the matched content is completely identical, and a score of 0 when all content is different; when the matched content is a synonym (an expression with different words but the same meaning), an abbreviation (a shortened expression corresponding to the full name), or a semantically similar expression (content with different words but similar meaning), the domain terminology lexicon is queried. If there is a mapping relationship with the domain terminology lexicon, the matching score is 1. If there is no mapping relationship, a vector representation is generated by a semantic encoding model (pre-trained language models such as BERT, RoBERTa, and MacBERT), and the cosine similarity is calculated. The cosine similarity is converted to the [0,1] interval, and the converted value is used as the matching score.

[0268] The distance attenuation parameter is 10–30;

[0269] The specific implementation process of step 3 includes:

[0270] This module is used to perform unified sentiment analysis and polarity judgment on structured sentence elements, sentiment objects and their corresponding evidence. It uses pre-trained language models, large language models or multi-task reasoning models for semantic reasoning and polarity induction, and outputs object-level sentiment analysis results.

[0271] Constructing an object-level sentiment polarity inference function includes: concatenating standardized sentence-level structural units belonging to sentiment analysis candidate sentence elements, the filtered candidate sentiment object set, the final evidence set, and the corresponding domain resource set, as shown below:

[0272] ;

[0273] in, Indicates splicing characteristics, Indicates a splicing operation;

[0274] Encode the concatenated features using a semantic encoding model (pre-trained language models such as BERT, RoBERTa, and MacBERT) to obtain an object-level semantic representation. And the probability distribution of candidate sentiment polarity is obtained through a classification function:

[0275] ;

[0276] in, and These represent the weight parameters and bias parameters used in the classification calculation, respectively. Represents the classification function; The probability distribution representing the candidate polarity (positive, negative, neutral, or undetermined candidate polarity);

[0277] The system further identifies and processes different types of emotional expressions in the following ways;

[0278] When the sentiment evidence in the final evidence set is directly an evaluation word (such as "satisfied," "very good," "poor," "terrible," etc.), the sentiment evidence is identified as an explicit sentiment expression. The basic polarity of the evaluation word is determined according to the sentiment lexicon (DUTIR Emotion Ontology Lexicon). This basic polarity is then adjusted using negative expressions and degree modifiers. (When the evaluation word is within the influence of an odd number of negative words, the basic sentiment polarity is reversed; for example, "satisfied" is positive, and "dissatisfied" is adjusted to negative. When there are no negative words or an even number of negative words, the basic sentiment polarity remains unchanged; for example, "not dissatisfied" is still judged as positive.) Subsequently, the sentiment intensity is adjusted according to degree modifiers. Weakening degree modifiers such as "slightly" and "somewhat" reduce sentiment intensity, while strengthening degree modifiers such as "very" and "extremely" increase sentiment intensity, but do not change the positive or negative polarity direction.) This yields the explicit sentiment judgment result of the sentiment evidence for the sentiment object. ;

[0279] When the emotional evidence in the final evidence set does not contain evaluative words (such as "satisfied," "very good," "poor," "terrible," etc.), but reflects the evaluation tendency of the emotional object through factual results (descriptions of usage effect, service result, or handling result, such as "it needs to be charged at 4 pm"), abnormal status (descriptions of abnormal situations such as failure, delay, or malfunction, such as "it crashed after two days of use"), repeated handling (descriptions of the same problem that has not been resolved after multiple handlings, such as "it was handled three times but still not resolved"), or user experience consequences (results such as inconvenience in use, increased time cost, or decreased service experience caused by facts or status), the emotional evidence is identified as implicit emotional expression, and the implicit emotional judgment result is expressed as:

[0280] ;

[0281] in, This indicates the implicit sentiment judgment result. Representing implicit sentiment inference functions, including: classifying sentiment objects Emotional evidence and domain resource collection As input, the probability distribution of candidate polarities is obtained using the object-level sentiment polarity inference function, and the sentiment category with the highest probability is selected as the implicit sentiment judgment result.

[0282] When the emotional evidence in the final evidence set contains negation, degree modification, adversative relationship, multiple object coexistence, or cross-sentence reference, the emotional evidence is identified as a complex structured emotional expression. Based on punctuation, negation, degree modification, adversative relationship, and the relationship between the emotional object and the corresponding emotional evidence (the correspondence of "which emotional evidence belongs to which emotional object"), each piece of emotional evidence is associated with its corresponding emotional object. Then, the emotional object corresponding to the emotional evidence and the contextual fragment information are input into the object-level emotional polarity inference function to obtain the probability distribution of the candidate polarity of the complex structured emotional expression, i.e., the complex structured emotional judgment result. ;

[0283] Based on punctuation, conjunctions, and syntactic structure, the system divides sentiment evidence into multiple evaluation segments. Then, it determines the scope of negation words, degree adverbs, and conjunctions, and assigns each evaluation segment to the corresponding sentiment object based on the object-evidence association results. For texts like "The location is good, but the service is not so good," the system associates "good" with "location" and "not so good" with "service," avoiding the latter negative evaluation from overriding the former positive evaluation.

[0284] When multiple pieces of emotional evidence correspond to the same emotional object, the weight of each piece of emotional evidence is determined based on the comprehensive relevance score:

[0285] ;

[0286] in, Indicates the first Emotional evidence in the emotional object Weights in polarity determination This represents the r-th piece of emotional evidence;

[0287] According to weight Obtain the polarity determination result:

[0288] ;

[0289] in, Indicates based on the emotional object The polarity judgment results of each corresponding piece of emotional evidence; For explicit sentiment judgment results Or implicit sentiment judgment result Or complex structure emotion judgment results ;

[0290] The polarity judgment results of each sentiment evidence are weighted and fused with the probability distribution of the candidate polarities:

[0291] ;

[0292] in, Representing the object of affection The final probability distribution of emotional polarity, This represents the weight of the language model output, with a value range of [value missing]. ;

[0293] Selecting the final emotional polarity probability distribution The sentiment category with the highest probability is selected as the candidate sentiment polarity; when the highest probability is greater than a preset reliability threshold, the candidate sentiment polarity is used as the final sentiment polarity (positive, negative, neutral, etc.); when the highest probability is lower than the preset threshold, or When the probability difference between the positive and negative categories is less than a preset difference threshold, it is marked as pending review or mixed sentiment.

[0294] The final object-level sentiment analysis results It includes at least the emotional object, emotional polarity, emotional evidence, polarity probability, confidence level, reasoning explanation, and processing status. Through the above processing, the system can distinguish and reason about explicit emotions, implicit emotions, negative expressions, degree modifications, adversative relationships, and mixed evaluations of multiple objects, avoiding errors in object attribution or misjudgment of emotional polarity caused by relying solely on the single output of the language model.

[0295] Weights of language model output The default reliability threshold is 0.5; the preset reliability threshold range is 0.7 to 0.9; the preset difference threshold is 0.1 to 0.2.

[0296] The preset reliability threshold is 0.8; the preset difference threshold is 0.15.

[0297] The specific implementation process of step 4 includes:

[0298] This step is used to unify, resolve conflicts, and verify consistency among multiple emotional objects, multiple pieces of emotional evidence, and potentially conflicting judgments in the same text.

[0299] For each identified sentiment object, generate object-level structured analysis results:

[0300] ;

[0301] in, Indicates the first The structured analysis results corresponding to each sentiment object are a component of the final sentiment analysis results of the text to be analyzed; if the text to be analyzed contains If there are several emotional objects, then the structured result set of the text to be analyzed is represented as follows: , Indicates the first One emotional object; Indicates the overall confidence level; Represents object-level structured results The processing status includes at least two initial states: automatic approval and pending review. The status is determined when the selected candidate sentiment object set, the final evidence set, and the final sentiment polarity are the same, and the overall confidence level is... Reaching the preset confidence threshold When the confidence level is between 0.7 and 0.9, for example, set to 0.8, the processing status is set to automatic pass; when the overall confidence level does not reach the preset confidence threshold... Or there may be conflicting evidence (conflict of evidence refers to different pieces of evidence supporting different emotional polarities for the same normalized emotional object, and the support strength of at least two different polarities reaches a preset threshold; for example, there are strong positive and negative evidence for the same object at the same time, or the output of the language model is significantly inconsistent with the judgment result of the evidence rules), and the final emotional polarity probability distribution. When the difference between the highest probability and the second highest probability is less than the preset difference threshold (0.1 to 0.2, for example, 0.15), and the correspondence between the emotional object and the emotional evidence is unclear (the same candidate evidence has a high object-evidence comprehensive relevance score with two or more emotional objects at the same time, and the difference between each score is small), the processing status is set to pending review. Representing the object of affection Corresponding candidate-level sentiment polarity: ;in, This represents a predefined set of emotional polarity categories, which includes at least positive, negative, neutral, and undetermined, and may also include weakly positive, weakly negative, mixed, and undetermined. This represents the probability distribution of emotional polarity. Emotional object Belongs to the category of emotional polarity The probability; when the structured analysis results are marked as pending review, This indicates the candidate polarity, rather than the final polarity confirmed manually. This section describes the reasoning specification, used to explain the judgment relationship between the emotional object, emotional evidence, and emotional polarity. The reasoning specification is confirmed based on the structured output fields in the domain resource set; if no matching reasoning specification exists, it is determined based on the emotional object. Collection of evidence Candidate polarity In addition to the scope of negation, degree of modification, adversative relationship and domain evidence template, the reasoning explanation is generated using a preset explanation template or a large language model (such as Qwen). For example, for the emotional object "battery life", the evidence fragment "it needs to be charged at four o'clock in the afternoon" and negative polarity, the reasoning explanation can be generated as "This fact shows that the battery life cannot meet the normal use needs, so it is judged as negative".

[0302] when When automatic passage is enabled, Directly used as the final object-level sentiment analysis result; when For the purpose of review, As a candidate analysis result, it should be submitted for manual review or secondary model validation (manual review involves reviewers evaluating the emotional object). Collection of evidence Candidate polarity and reasoning explanation Confirmation or modification is required; secondary model verification refers to calling the object-level sentiment polarity reasoning model again, re-inputting the sentiment object, evidence set, and context information into the model for judgment, comparing the new result with the original judgment result, and updating the model based on the verification result to form a confirmed object-level sentiment analysis result.

[0303] For the set of structured analysis results of the same text to be analyzed There are multiple different emotional objects in it. or a certain emotional object The corresponding final set of evidence In cases where multiple pieces of sentiment evidence exist, the object-level sentiment analysis results are integrated and conflict resolved; including:

[0304] The names of sentiment objects are standardized (consistent capitalization, removal of punctuation and redundant modifiers without actual object meaning), and converted into standard names, i.e., abbreviations, aliases, and synonyms in the domain terminology lexicon. When two sentiment objects have the same name, or when there is at least one relationship between them, such as synonymy, abbreviation, alias, reference, overlapping text scope, or both pointing to the same evaluated entity, they are considered to be semantically identical sentiment objects. When the sentiment object similarity of two sentiment objects reaches a preset threshold... When two objects belong to the same category of emotional objects in the same domain, they are considered to be highly similar emotional objects. For example, "customer service response speed" and "problem-solving effect" both belong to the customer service scenario, but they correspond to service response and service result respectively. They should not be merged into the same emotional object because they are semantically related.

[0305] For sentiment objects identified as semantically identical or highly similar, merging is performed. This merging process does not involve directly concatenating object names; instead, object normalization is applied. Simultaneously, standard names from the domain terminology lexicon are selected as the merged object names. If no standard names exist in the domain terminology lexicon, the object name with the highest object recognition confidence, the most complete textual expression, or the highest frequency of occurrence is selected as the merged object name. Furthermore, the evidence sets corresponding to the merged objects undergo union processing, removing evidence with identical text content, overlapping location ranges, or highly repetitive semantics, while retaining the source sentence elements, location identifiers, comprehensive relevance scores, and evidence-level judgment confidence for each piece of evidence. The evidence-level judgment confidence is represented by a sentiment polarity probability distribution. In this context, emotional evidence falls under the category of emotional polarity. The maximum probability;

[0306] For the set of structured analysis results of the same text to be analyzed In cases where multiple pieces of emotional evidence relate to the same emotional object, the evidence fusion weight is calculated based on the strength of evidence, the degree of contextual proximity, and the confidence level of each piece of emotional evidence. :

[0307] ;

[0308] in, This represents the strength of evidence, or support strength, of the k-th sentiment evidence. This indicates the degree of contextual proximity between the object of emotion and the evidence of emotion. Indicates the first The confidence level of evidence corresponding to each piece of emotional evidence; This indicates the strength of the emotional evidence presented in the r-th piece. This indicates the degree of contextual proximity between the emotional object and the r-th piece of emotional evidence. Indicates the first The confidence level of evidence corresponding to each piece of emotional evidence;

[0309] We use evidence fusion weights to weight and fuse the probability distributions of sentiment polarity from different sentiment evidence:

[0310] ;

[0311] in, This represents the probability distribution of the emotional polarity after fusion;

[0312] For the set of structured analysis results of the same text to be analyzed For emotional objects with conflicting evidence, instead of directly deleting other evidence with similar evidence, or simply selecting the single piece of evidence with the highest score, a conflict resolution score is calculated for each piece of emotional evidence:

[0313] ;

[0314] in, This represents the overall relevance score. It reflects the degree of relevance between candidate evidence and the emotional object, that is, the strength of the relationship between the evidence and the target object; This reflects the strength of the evidence supporting the polarity judgment; It reflects the contextual proximity between the emotional object and the candidate evidence; Reflects the confidence level of the evidentiary sentiment judgment; parameters lie in An interval, satisfying: ;

[0315] When the probability distribution of emotional polarity after fusion The preset polarity judgment threshold is reached, and When the difference between the highest probability and the second highest probability reaches a preset difference threshold, the sentiment polarity corresponding to the highest probability is taken as the object-level judgment result, and the sentiment polarity corresponding to the highest probability is taken as the object-level sentiment analysis result; when the evidence strength of the sentiment evidence of positive and negative categories reaches the evidence strength threshold, or when the difference between the highest probability and the second highest probability in the fused sentiment polarity probability distribution is lower than the preset difference threshold, a single polarity result is not forced to be generated, and the sentiment object is marked as mixed sentiment or pending review;

[0316] Preset threshold The threshold value is 0.85; the preset polarity judgment threshold is 0.7–0.9; the evidence strength threshold is 0.7–0.9.

[0317] Emotional object similarity differs from semantic similarity between an emotional object and candidate evidence; it is used to determine whether two candidate emotional objects refer to the same evaluated object. This includes: for emotional objects... and The formula for calculating the similarity between emotional objects is:

[0318] ;

[0319] in, express and Similarity of emotional objects The vector semantic similarity of two emotion objects is obtained by generating object vectors and calculating cosine similarity through a text encoding model (pre-trained language models such as BERT, RoBERTa, and MacBERT). The lexical overlap of object names is calculated using Jaccard similarity on the character set (character-by-character segmentation of the emotional object name), the word set (segmented by a word segmenter on the emotional object name), or the keyword set (segmented by the emotional object name and its context, removing stop words and modifiers without actual object meaning, and retaining words that can represent the meaning of the object name or attribute). This indicates the degree of matching of synonyms, abbreviations, alternative names, or referential relationships. It is 1 when two sentiment objects are marked as synonyms in the domain terminology lexicon, or when two sentiment objects point to the same evaluated entity; otherwise, it is 0. This represents the similarity parameter, with a value range of [value range missing]. The interval, and satisfying ;

[0320] The strength of evidence indicates the degree to which a candidate piece of evidence supports the judgment of the target emotional object and emotional polarity, and is not equivalent to the aforementioned strength of dependency; Candidate evidence relative to the emotional object The strength of evidence is expressed as follows:

[0321] ;

[0322] in, This indicates the strength of evidence, or the strength of support. This indicates the overall relevance score; It indicates the clarity of the sentiment orientation of the evidence, and is determined based on the highest polarity probability in the evidence-level sentiment classification results; Semantic similarity refers to the degree of match between candidate evidence and domain evaluation terms, negation structures, factual outcome descriptions, or other sentiment evidence templates. (The cosine similarity method is used to calculate semantic similarity, and the highest similarity is taken as the mean.) ); This indicates the degree of polarity consistency between emotional evidence and other emotional evidence concerning the same emotional object; a value of 1 indicates the same emotional polarity as other evidence concerning the same emotional object, and a value of 0 indicates a different polarity. (Parameter) All located in An interval, satisfying: ;

[0323] The context-related scope refers to the text window used to determine the relationship between the emotional object and the candidate evidence, including the current sentence element, several sentences adjacent to the current sentence element, and text fragments related to the current sentence element through referential relationships, adversative relationships, causal relationships, or dialogue succession relationships; in one implementation, the system selects 1 to 3 sentences before and after the current sentence element as the basic context window, and supplements relevant text according to paragraph number, sentence element number, dialogue turn, and referential relationship;

[0324] Contextual proximity between the emotional object and the candidate evidence Represented as:

[0325] ;

[0326] in, The number of lexical units or sentence units representing the interval between the emotional object and the candidate evidence. This represents the distance decay parameter; the closer the emotional object is to the candidate evidence, the higher the degree of contextual proximity.

[0327] Overall confidence level As shown below:

[0328] ;

[0329] in, This represents the confidence level in identifying the emotional object. Indicates the confidence level of the emotional evidence match. Indicates the confidence level of sentiment polarity reasoning. These are weight parameters, located at The interval, and the sum of the weight parameters is 1;

[0330] Among them, the confidence level of emotion object recognition Calculated based on object boundary recognition probability, domain object category matching degree, and the degree of correlation between the object and the evaluation expression:

[0331] ;

[0332] in, This represents the probability of object boundaries or object categories output by the sentiment object recognition model (e.g., language models like BERT, RoBERTa, or MacBERT perform multiple sentiment object recognitions on the same sentence, and the proportion of the number of times the same sentiment object is identified out of the total number of recognitions). ); This indicates the degree of matching between the candidate object and the domain object category and the domain terminology vocabulary. When the candidate object matches the domain sentiment object category... or field terminology glossary If the object name, synonym, abbreviation, or alias is the same, the value is 1; if they are different, the semantic similarity between the candidate object and each domain object category or domain term is calculated, and the highest similarity is taken as the result. ; The relevance score represents the degree of association between the candidate object and the evaluation expression (determined by calculating dependency path and text distance); parameters lie in An interval, satisfying: ;

[0333] Confidence of Emotional Evidence Matching The weighted calculation is based on the overall relevance score of each piece of evidence in the final evidence set:

[0334] ;

[0335] in, The evidence fusion weight has comprehensively considered evidence strength, contextual proximity, and evidence-level judgment confidence. Therefore, evidence with high relevance, high support strength, close contextual distance, and high judgment confidence has a larger weight in the sentiment evidence matching confidence calculation; when there is no candidate evidence that meets the evidence retention threshold, the sentiment evidence matching confidence is set to 0.

[0336] Confidence of Emotional Polarity Inference Calculated based on the highest probability in the fused polarity probability distribution, and the difference between the highest and second-highest probabilities:

[0337] ;

[0338] in, This represents the highest probability in the polarity probability distribution after fusion. The second highest probability is indicated; the higher the highest polarity probability, and the greater the difference between it and the second highest probability, the more stable the polarity judgment; parameters All located in An interval, satisfying: ;

[0339] Weight parameters The values ​​were set to 0.3, 0.3, and 0.4 respectively; the values ​​could be determined based on the accuracy, recall, or F1 score in the validation samples, through grid search, parameter optimization, or manual configuration.

[0340] This step is used to improve the system's adaptability and business usability during long-term operation;

[0341] Receive manual review results, user correction results, or results returned by the business system, and write the corrected sentiment object, candidate evidence, object-level sentiment polarity, reasoning explanation, and processing status into the feedback sample library.

[0342] Among them, the manual review result refers to the result formed after the manual reviewer confirms or modifies the sentiment object, sentiment evidence, candidate sentiment polarity, reasoning explanation or processing status marked as "pending review"; the user correction result refers to the result formed after the system user supplements, deletes or corrects the automatically generated object-level sentiment analysis result; the business system feedback result refers to the information such as object category, problem type, sentiment tendency or processing status returned by the customer service ticket system, public opinion management system, product evaluation system or other downstream business systems based on the actual processing result.

[0343] The term "correction" refers to the modification of the generated object-level structured results. At least one field needs to be modified, said field including the sentiment object. Collection of evidence Object-level emotional polarity Reasoning Explanation and processing status The corrected object-level structuring result can be represented as:

[0344] ;

[0345] in, This represents the object-level structured result after manual review, user correction, or business system verification. The system will write the result before correction, the result after correction, the source of correction, the correction time, and the corresponding original sentence elements into the feedback sample library to retain complete correction records and data sources.

[0346] The feedback sample library can be used to update the domain sentiment object category. and field terminology glossary Emotional evidence recognition rules or templates and comprehensive confidence threshold Among them, the domain emotion object category Domain terminology glossary used to supplement or adjust object types in different business scenarios. Used to supplement the names, synonyms, abbreviations, aliases, and domain-specific expressions of the emotion object; emotion evidence recognition rules or templates. Used to supplement evaluative words, negative structures, adversative relationships, factual outcome descriptions, and implicit sentiment expression patterns. The confidence threshold, excluding the comprehensive confidence threshold... In addition, it may include evidence retention threshold, object validity threshold, and polarity judgment threshold.

[0347] The model prompt templates refer to the task prompts used when calling the large language model, including prompt templates for sentiment-related sentence elements screening, sentiment object recognition, sentiment evidence extraction, and sentiment polarity inference. The training sample library refers to a set of labeled samples composed of standardized sentence elements, sentiment objects, sentiment evidence, sentiment polarity, inference descriptions, and processing states, used to train or fine-tune the sentiment relevance judgment function, sentiment object recognition function, sentiment evidence recognition model, and object-level sentiment polarity inference function, thereby forming a closed-loop mechanism for continuous optimization.

[0348] The system summarizes the object-level sentiment analysis results of batch texts according to text source, business scenario, statistical time interval, or user-defined statistical conditions. First, it normalizes the names of sentiment objects by synonyms, abbreviations, and alternative names, and counts the frequency of each sentiment object within the statistical range, generating a list of high-frequency sentiment objects in descending order of frequency.

[0349] Emotional object The frequency of occurrence can be expressed as:

[0350] ;

[0351] in, This represents the total number of texts within the statistical range. Indicates the first The set of object-level sentiment analysis results corresponding to each text. Indicates the indicator function. When the first... The text contains emotional objects When the condition is met, the indicator function takes a value of 1; otherwise, it takes a value of 0. The system then... Generate a list of high-frequency emotion objects from highest to lowest.

[0352] Secondly, the system statistically analyzes the number of negative results, average negative probability, and average overall confidence level for each emotional object, and calculates the ranking score for negative questions:

[0353]

[0354] in, Representing the object of affection The number of times it was judged as negative within the statistical range; This represents the average negative probability of the outcome corresponding to this emotional object; This represents the average overall confidence level of the results corresponding to this sentiment object. The system follows... A ranking of negative issues is generated from highest to lowest. By considering the number of negative results, their degree of negativity, and the reliability of the results simultaneously, ranking bias caused by relying solely on frequency of occurrence can be avoided.

[0355] Secondly, the system counts the number and proportion of positive, negative, and neutral results for each sentiment object according to daily, weekly, monthly, or other preset time windows, and compares the changes in sentiment proportion between adjacent time windows to generate sentiment trend change results. Emotional object within a time window Corresponding polarity The proportion can be expressed as:

[0356]

[0357] in, Indicates the first Emotional object within a time window Determined to be polar Number of times, Indicates the emotional object within this time window. The total number of results This indicates a small positive number that prevents the denominator from being zero. The system compares adjacent time windows... This determines whether the proportion of positive, negative, or neutral emotions is rising, falling, or remaining stable.

[0358] Simultaneously, the system selects evidence fragments from the evidence sets corresponding to each sentiment object that have high object-evidence comprehensive relevance scores, clear polarity orientations, high comprehensive confidence levels, and non-repetitive text content. It then generates typical evidence summaries using preset summary templates or a large language model. When generating summaries using a large language model, the sentiment object, sentiment polarity, candidate evidence fragments, and corresponding source sentence elements are input into the model. A preset prompt template requires the model to retain the original facts and not add information not included in the original text.

[0359] The scenario-based analysis report should include at least the statistical time range, total text volume, data source, high-frequency sentiment objects, sentiment polarity distribution of each object, ranking of negative issues, sentiment trend changes, summary of typical evidence, number of low-confidence results, and number of results pending verification. The report can be output in the form of structured data, tables, trend charts, or text summaries, and can be used for downstream applications such as public opinion monitoring, product improvement, customer service optimization, and industry-specific text-based analysis.

[0360] When a candidate sentiment object has a clear semantic boundary, can match the domain sentiment object category or domain terminology, and has a clear modification relationship, dependency relationship, referential relationship or semantic association with the evaluation words, factual result descriptions, negative expressions or other evaluation components in the current sentence element or context, and the object recognition confidence reaches the preset object validity threshold, the system determines it as a valid sentiment object.

[0361] If the set of candidate sentiment objects is empty, or if all candidate sentiment objects are filtered out due to unclear boundaries, appearing only as background entities, lack of evaluation relationship support, mismatch with domain object categories, or object recognition confidence scores below the preset object validity threshold, then the system determines that there are no valid sentiment objects in the current sentence element and outputs the following structured result: [

[0363] {

[0364] "Emotional object": null,

[0365] "Emotional polarity": "To be determined"

[0366] "Excerpt of evidence": [],

[0367] Reasoning Explanation: "No emotional object satisfying the validity criteria was identified".

[0368] Processing Status: "Archived"

[0369] } ]

[0371] For a valid emotional object that has been identified, when its final evidence set is not empty, the object-evidence comprehensive relevance score of at least one piece of evidence reaches the evidence retention threshold, the emotional evidence matching confidence reaches the preset evidence confidence threshold, and the highest probability in the emotional polarity probability distribution reaches the preset polarity judgment threshold, and the difference between the highest probability and the second highest probability reaches the preset difference threshold, the system determines that the existing evidence can support a reliable emotional polarity judgment.

[0372] If the final evidence set is empty, the object-evidence comprehensive relevance score is lower than the evidence retention threshold, the sentiment evidence matching confidence is lower than the preset threshold, the highest polarity probability is lower than the preset polarity judgment threshold, the difference between the highest probability and the second highest probability is too small, or there is evidence with similar support strength but opposite polarity, then the system determines that the existing evidence is insufficient to support a reliable sentiment polarity judgment and outputs the following structured result: [

[0374] {

[0375] "Emotional Object": "Identified emotional objects",

[0376] "Emotional polarity": "To be determined"

[0377] "Excerpt of evidence": [],

[0378] Reasoning Explanation: "The existing emotional evidence is insufficient to support a reliable judgment of emotional polarity."

[0379] Processing Status: "Pending Review"

[0380] } ]

[0382] The corresponding sentence elements, candidate sentiment objects, candidate evidence, polarity probabilities, and judgment criteria are written into the list to be reviewed, for manual review, secondary model verification, or subsequent feedback updates.

[0383] Object-level sentiment analysis under multiple objects, multiple evidence, and transitional structures:

[0384] To verify the system's analytical capabilities in scenarios involving multiple objects, multiple pieces of evidence, and transitional structures, the following product and service review texts were selected as input samples:

[0385] "The screen display of this phone is indeed very delicate, and the system runs quite smoothly, but the battery life is a bit short. I have to charge it at 4 pm. Customer service responds quickly, but the problem is still not resolved after three attempts."

[0386] This text contains product attribute evaluations, service response evaluations, service result evaluations, and transitional expressions, making it suitable for verifying the object-level sentiment analysis capabilities of this invention in complex text. The text is processed sequentially through the TextScope, DomainAdapter, SentiLocator, SentEvidence, SentReasoner, and ResultVerifier modules.

[0387] (1) The TextScope module performs sentence segmentation, source tagging, and context preservation on the original text to form standardized analysis units. For example, the sentence “But the battery life is a bit short, so I have to charge it at 4 pm” in the text can be identified as an independent sentence unit for subsequent object extraction and evidence matching.

[0388] (2) The DomainAdapter module determines that the text content belongs to the "product review and after-sales service evaluation" scenario and loads the corresponding product attribute vocabulary, service evaluation vocabulary, and sentiment evidence template. The key objects of focus in this scenario include screen display, system operation, battery life, customer service response, and problem handling effectiveness.

[0389] (3) The SentimentLocator module performs sentiment relevance judgment on each sentence element and identifies the sentiment object within it. For example, the system can identify the product attribute object "battery life" and service objects such as "customer service response" and "problem handling effect" from the text, thereby avoiding mixing multiple evaluation objects into a single overall result.

[0390] (4) The SentEvidence module extracts corresponding evidence around the identified sentiment object and establishes the association between the object and the evidence. For example, for the sentiment object "battery life", the system extracts "a little weak" and "needs to charge at 4 pm" as sentiment evidence, where the latter is an implicit negative fact description.

[0391] (5) The SentReasoner module combines the sentiment object, evidence fragments and contextual information to perform object-level sentiment polarity reasoning. For example, "screen display" corresponds to the evidence "very delicate", which is judged as positive; "battery life" corresponds to the evidence "must be charged at 4 pm", which is judged as negative; "customer service reply" corresponds to the evidence "quite fast", which is judged as positive; "problem handling effect" corresponds to the evidence "it was handled three times but still not resolved", which is judged as negative.

[0392] (6) The ResultVerifier module merges and verifies the confidence level of the analysis results of multiple objects. For semantically different objects such as "customer service response" and "problem handling effect", the system retains their sentiment results separately and does not simply merge them. Finally, the system outputs a structured result containing sentiment objects, evidence fragments, sentiment polarity, confidence level and reasoning explanation.

[0393] As can be seen from this embodiment, the present invention can identify multiple emotional objects in the same text and accurately associate different pieces of evidence with the corresponding objects, avoiding the simple classification of the entire text into a single emotional tag, and improving the analysis accuracy and interpretability of complex comment texts.

[0394] The above embodiments are only used to illustrate the technical solutions of the present invention and do not constitute a limitation on the scope of protection of the present invention. Equivalent substitutions, modifications, or combinations made by those skilled in the art to the above embodiments without departing from the technical concept of the present invention should all fall within the scope of protection of the present invention.

[0395] Example 3

[0396] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the text sentiment analysis method that integrates sentiment object extraction and sentiment polarity reasoning as described in Embodiment 1 or 2.

[0397] Example 4

[0398] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the text sentiment analysis method that integrates sentiment object extraction and sentiment polarity inference as described in Embodiment 1 or 2.

[0399] Example 5

[0400] A text sentiment analysis system that integrates sentiment object extraction and sentiment polarity reasoning includes:

[0401] The structural unit and domain resource construction module is configured to: perform structured preprocessing on the original text to be analyzed, construct standardized sentence-level structural units, and construct a set of domain resources according to the domain type of the original text to be analyzed;

[0402] The sentiment object and candidate evidence acquisition module is configured to: judge sentiment relevance based on standardized sentence element-level structural units and domain resource set, filter out sentiment analysis candidate sentence elements, and acquire standardized sentence element-level structural units belonging to the sentiment analysis candidate sentence elements;

[0403] Semantic parsing is performed on standardized sentence-level structural units that belong to the candidate sentence elements of sentiment analysis to identify the set of candidate sentiment objects and filter the set of sentiment objects;

[0404] Candidate evaluation components are extracted from the selected set of candidate emotional objects to form a set of candidate evidence. After screening the set of candidate evidence, a comprehensive correlation score between the set of candidate emotional objects and the set of candidate evidence is calculated. The final set of evidence is obtained based on the correlation score.

[0405] The sentiment polarity reasoning module is configured to: perform sentiment polarity reasoning on the standardized sentence-level structural units belonging to the sentiment analysis candidate sentence elements, the filtered candidate sentiment object set, the final evidence set and the corresponding domain resource set, and obtain the final sentiment polarity;

[0406] The fusion and conflict resolution module is configured to: construct object-level structured analysis results, review them, and generate object-level sentiment analysis results;

[0407] Based on the similarity of sentiment objects, the strength of evidence, and the overall confidence level, the results of object-level sentiment analysis are integrated and conflict resolved.

Claims

1. A text sentiment analysis method that integrates sentiment object extraction and sentiment polarity reasoning, characterized in that, include: Step 1: Perform structured preprocessing on the original text to be analyzed, construct standardized sentence-level structural units, and construct a domain resource set according to the domain type of the original text to be analyzed; Step 2: Based on the standardized sentence element-level structural units and the domain resource set, determine the sentiment relevance, screen out the candidate sentence elements for sentiment analysis, and obtain the standardized sentence element-level structural units belonging to the candidate sentence elements for sentiment analysis; Semantic parsing is performed on standardized sentence-level structural units that belong to the candidate sentence elements of sentiment analysis to identify the set of candidate sentiment objects and filter the set of sentiment objects; Candidate evaluation components are extracted from the selected set of candidate emotional objects to form a set of candidate evidence. After screening the set of candidate evidence, a comprehensive correlation score between the set of candidate emotional objects and the set of candidate evidence is calculated. The final set of evidence is obtained based on the correlation score. Step 3: Perform sentiment polarity reasoning on the standardized sentence-level structural units belonging to the sentiment analysis candidate sentence elements, the filtered candidate sentiment object set, the final evidence set, and the corresponding domain resource set to obtain the final sentiment polarity; Step 4: Construct object-level structured analysis results, review them, and generate object-level sentiment analysis results; Based on the similarity of sentiment objects, the strength of evidence, and the overall confidence level, the results of object-level sentiment analysis are integrated and conflict resolved.

2. The text sentiment analysis method integrating sentiment object extraction and sentiment polarity reasoning according to claim 1, characterized in that, The specific implementation process of step 1 includes: The original text to be analyzed is segmented, including: automatically recognizing periods, semicolons, question marks, exclamation marks, line breaks, transition words, and semantic boundary markers, splitting the original text into multiple semantically complete sentence elements to form sentence-level structural units; at the same time, the sentence order, contextual adjacency, paragraph number, and text source are preserved. Inject the sentence element content, text source, paragraph number, sentence element number, and contextual information of the original text to be analyzed into each sentence element-level structural unit to generate a standardized sentence element-level structural unit, as shown below: ； in, This represents the i-th standardized sentence-level structural unit. This indicates the text source of the original text to be analyzed corresponding to the i-th standardized sentence-level structural unit. This represents the sentence element content of the i-th standardized sentence element-level structural unit. Indicates context fragment information, This indicates the location identifiers in the original text to be analyzed, including sentence element numbers, paragraph numbers, and arrangement order; Based on the text source, business tags, or user settings of the original text to be analyzed, determine the domain type, i.e., the business scenario, to which the original text to be analyzed belongs, and obtain the corresponding domain resource set: ； in, Indicates the first Domain type refers to the set of domain resources corresponding to a business scenario; It represents the category of sentiment objects in the domain, including the entities, attributes, functions, service links, behaviors, or processing results being evaluated in the business scenario; This represents a domain-specific terminology glossary, including domain object names, synonyms, abbreviations, alternative names, and domain-specific expressions. It represents the rules or templates for identifying sentiment evidence, including the modifying relationship between evaluative words and sentiment objects, negation structure, adversative structure, degree modification relationship, causal relationship and implicit fact expression pattern; This represents structured output fields, including sentiment object, object category, evidence fragment, sentiment polarity, confidence level, reasoning description, and processing status.

3. The text sentiment analysis method integrating sentiment object extraction and sentiment polarity reasoning according to claim 2, characterized in that, Sentiment relevance is determined based on standardized sentence-level structural units and domain resource sets. Candidate sentence units for sentiment analysis are selected, and standardized sentence-level structural units belonging to these candidate sentence units are obtained; including: The standardized sentence-level structural units and the domain resource set are input into the sentiment relevance judgment function to obtain the judgment result, as shown below: ； in, Indicates the judgment result. Represents standardized sentence-level structural units. Represents a collection of domain resources; The function representing sentiment relevance judgment includes: encoding the sentence content and contextual fragment information of standardized sentence-level structural units using a pre-trained language model encoder to obtain sentence semantic representations. Based on domain resource set Domain Emotional Object Category Domain terminology glossary Emotional evidence recognition rules or templates The system matches the current sentence content to obtain the matching results of the corresponding sentiment object category, domain terminology vocabulary, and sentiment evidence recognition rules or templates, and converts them into numerical features to form a domain feature vector. The semantic representation of sentence elements is concatenated with the domain feature vector, and the probability of a sentence element belonging to a sentiment-related sentence element is calculated using a classification function, as shown below: ； in, The probability of representing sentiment-related sentence elements. and These represent the weight parameters and bias parameters, respectively. This represents a vector concatenation operation; Based on the preset judgment threshold Output the judgment result: ； when When, it indicates that the current sentence element, i.e., the sentence element content, is a candidate sentence element for sentiment analysis, and the current sentence element contains sentiment objects, evaluation expressions, user experiences, service feedback, or implicit sentiment facts; when When the current sentence element is not a candidate sentence element for sentiment analysis, it indicates that the current sentence element is not a candidate sentence element for sentiment analysis. Obtain standardized sentence-level structural units belonging to candidate sentence elements for sentiment analysis from standardized sentence-level structural units. and corresponding domain resource sets ; Further optimization involves determining the threshold. It ranges from 0.5 to 0.

8.

4. The text sentiment analysis method integrating sentiment object extraction and sentiment polarity reasoning according to claim 3, characterized in that, Semantic parsing is performed on standardized sentence-level structural units that belong to the candidate sentence elements of sentiment analysis to identify the set of candidate sentiment objects and filter the set of sentiment objects; include: Standardized sentence-level structural units belonging to the candidate sentence elements of sentiment analysis and corresponding domain resource sets The input emotion object recognition function generates a candidate emotion object set through preset task prompts and a structured output format, as shown below: ； in, Indicates the first The set of candidate sentiment objects corresponding to each sentence element. This represents the nth emotional object. This represents an emotion object recognition model; The process involves filtering emotional objects from the candidate emotional object set, including boundary identification, duplicate merging, and validity verification. Boundary recognition includes: determining the start and end range of the emotion object based on the character position, word combination, head word and modifiers of the candidate object in the sentence unit, and deleting punctuation, function words and redundant modifiers that have no actual meaning; Duplicate merging includes: normalizing two sentiment objects with the same text content, overlapping position range, synonymous or abbreviated relationship, or cosine similarity higher than a preset similarity threshold, and merging them into a unified object name; Validity verification includes: calculating a validity score based on the correlation between the sentiment object and the evaluation expression, the degree of matching between the domain object category, the degree of contextual relevance, and the confidence score of the model output, to determine whether the candidate object belongs to the actual evaluated object, as shown below: ； in, Indicates the validity score; Let represent the weight, which lies in the interval [0,1] and satisfies: ; The score represents the association between the sentiment object and the evaluation expression, including: obtaining the shortest dependency path between the sentiment object and the evaluation expression through a dependency parser, and obtaining the dependency relationship score based on the shortest dependency path. When the shortest dependency path length is 1, the dependency score is 1; when the shortest dependency path length is 2, the dependency score is 0.8; when the shortest dependency path length is 3, the dependency score is 0.6; when the shortest dependency path length is 4, the dependency score is 0.4; when the shortest dependency path length is greater than or equal to 5, or when there is no identifiable dependency path, the dependency score is 0. Simultaneously, the number of lexical units separating the sentiment object and the evaluation expression within the same sentence content is obtained, and the text distance score is obtained based on the number of lexical units separating them. When the emotional object and the evaluative expression are separated by 0-2 lexical units, the text distance score is 1; when separated by 3-5 lexical units, the text distance score is 0.8; when separated by 6-10 lexical units, the text distance score is 0.6; when separated by 11-20 lexical units, the text distance score is 0.3; when separated by more than 20 lexical units, the text distance score is 0; finally, the association score... ,in For association weights; This represents the matching score between the sentiment object and the domain type. If all sentiment objects and domain types match, the score is 1. Otherwise, the cosine similarity between the sentiment object and the domain type is calculated, and the highest similarity value is taken as the matching score. The score represents the relevance of the sentiment object to the current sentence element and its context, including: calculating the number of sentence elements separating the sentence element containing the sentiment object from the sentence element containing the related evaluative expression; when the sentiment object and the evaluative expression are located in the same sentence element, When located in an adjacent sentence element, take Take when there is a 1-sentence interval Take when there is a 2-sentence interval When the interval exceeds 2 sentence elements ; This represents the confidence level output by the emotion object recognition model; When the validity score of an emotional object is higher than the preset validity threshold, the emotional object is retained; when the validity score is lower than the preset threshold, the emotional object is filtered to obtain a set of candidate emotional objects after screening. Further preferred, the similarity threshold is 0.80–0.90; association weight =0.5; the validity threshold is 0.6–0.8; The optimal validity threshold is 0.

7.

5. The text sentiment analysis method integrating sentiment object extraction and sentiment polarity reasoning according to claim 4, characterized in that, Candidate evaluation components are extracted from the screened set of candidate sentiment objects to form a set of candidate evidence. After further screening of the candidate evidence set, a correlation score is calculated between the set of candidate sentiment objects and the set of candidate evidence. The final evidence set is obtained based on the correlation score; including: For each emotional object in the selected candidate emotional object set, candidate evaluation components are extracted from the current sentence element and contextual information. These components include evaluation words, adjectives, adverbs, degree modifiers, negation words, transition words, causal phrases, factual outcome descriptions, and descriptive components that are dependent on the emotional object, forming a candidate evidence set. ;in, Represents the set of candidate evidence. Indicates the first Each candidate piece of evidence is emotional evidence; Meanwhile, candidate evidence is initially screened through stop word filtering, invalid modifier filtering, and domain evidence template filtering. The stop word filtering includes: pre-constructing a stop word list, and deleting function words, modal words, conjunctions, and words in the candidate evidence that do not carry evaluative meaning and are in the top 5% of the word frequency and have been manually confirmed not to carry evaluative meaning based on the stop word list; Invalid modifier filtering includes: combining part-of-speech information, text distance between candidate evidence and sentiment object, and contextual semantics to remove sentiment objects that are not directly modified in candidate evidence, or to remove text components that are only used to describe time, place, quantity and background information; Domain evidence template filtering includes: organizing the original text to be analyzed collected in the current business scenario, extracting evaluation words, negation structures, degree modifiers, transitional relationships, causal expressions, and factual result descriptions with evaluative meanings, and manually reviewing and deduplicating them to form a domain evidence template corresponding to the current business scenario; matching candidate evidence with the domain evidence template, and retaining the candidate evidence if all matches are the same; otherwise, generating vector representations of the candidate evidence and the domain evidence template using a semantic encoding model, and calculating the similarity using cosine similarity; retaining the candidate evidence when the highest similarity reaches a preset threshold, otherwise filtering the candidate evidence. The overall relevance score between the sentiment object and the filtered set of candidate evidence is calculated using the following formula: ； in, This indicates the first element in the filtered set of candidate emotion objects. An emotional object, Indicates the first in the filtered set of candidate evidence. One candidate piece of evidence, express and The overall correlation score between them Indicates semantic similarity. Indicates the strength of the dependency relationship. Indicates the context adjacency relevance. Indicates domain matching degree. This indicates the weight in the overall relevance score, located at... An interval, satisfying: ; The candidate evidence scores, i.e., the overall relevance scores, are normalized as follows: ； in, This represents the normalized candidate evidence score. This represents the minimum overall relevance score among all candidate evidence for the same emotional object. This represents the maximum value among all candidate evidence for the same emotional object, based on their overall relevance score. To prevent tiny constants with a denominator of zero; Set evidence retention threshold The normalized candidate evidence scores are then used for screening. And select the candidate evidence with the highest normalized score. The following are the final pieces of evidence: ； in, Representing the object of affection The corresponding set of sentiment evidence is the final set of evidence. TopK represents the top K normalized candidate evidence scores. Further preferred, retention threshold The value ranges from 0.5 to 0.

8. Further optimization, semantic similarity This is used to measure the semantic correlation between sentiment objects and candidate evidence; through a pre-trained language model, the sentiment objects and candidate evidence are semantically encoded separately to obtain object vectors. and evidence vector The semantic similarity is calculated using cosine similarity, as shown below: ； Dependency strength This is used to measure the degree of association between sentiment objects and candidate evidence in syntactic structure; the shortest dependency path length between sentiment objects and candidate evidence is obtained using a dependency parser. And determine whether a direct dependency relationship exists, as shown below: ； in, This indicates a direct dependency relationship; it is defined when there is a direct dependency relationship between the emotional object and the candidate evidence, i.e., the shortest dependency path length is 1. ;otherwise, ; The weight represents the direct dependency relationship, and its value ranges from 1 to 2. When the sentiment object and the candidate evidence are located in different sentence units and there is no identifiable dependency relationship, the dependency relationship strength is set to 0. Contextual adjacency relevance Used to measure the proximity of the sentiment object and candidate evidence in the original text; calculates the number of lexical units separating the sentiment object and candidate evidence according to sentence units and context. The context adjacency relevance is calculated by the number of lexical units, as shown below: ； in, This represents the distance attenuation parameter; Domain matching degree The formula used to measure the degree of matching between sentiment objects and candidate evidence and the set of domain resources in the current business scenario is as follows: ； in, Indicates the matching parameter, located in An interval, satisfying: ; Representing the emotional object and the domain emotional object category Match score between them; Glossary of Emotional Objects or Candidate Evidence and Domain Terminology Match score between them; This indicates the rules or templates for identifying candidate evidence and sentiment evidence. Match score between them; The matching score includes: a score of 1 when the matched content is completely identical, and a score of 0 when all content is different; when the matched content is a synonym, abbreviation, or semantically similar expression, the domain terminology lexicon is queried. If a mapping relationship exists between the lexicon and the domain terminology lexicon, the matching score is 1; otherwise, a vector representation is generated using a semantic encoding model, and cosine similarity is calculated; the cosine similarity is then converted to... The range is determined, and the converted value is used as the matching score. A further preferred option is a distance attenuation parameter of 10 to 30.

6. The text sentiment analysis method integrating sentiment object extraction and sentiment polarity reasoning according to claim 5, characterized in that, The specific implementation process of step 3 includes: Constructing an object-level sentiment polarity inference function includes: concatenating standardized sentence-level structural units belonging to sentiment analysis candidate sentence elements, the filtered candidate sentiment object set, the final evidence set, and the corresponding domain resource set, as shown below: ； in, Indicates splicing characteristics, Indicates a splicing operation; The concatenated features are encoded using a semantic coding model to obtain an object-level semantic representation. And the probability distribution of candidate sentiment polarity is obtained through a classification function: ； in, and These represent the weight parameters and bias parameters used in the classification calculation, respectively. Represents the classification function; The probability distribution representing the candidate polarity; When the sentiment evidence in the final evidence set is directly an evaluative word, the sentiment evidence is identified as an explicit sentiment expression. The basic polarity of the evaluative word is determined according to the sentiment lexicon, and the basic polarity is adjusted by combining negative expressions and degree modifiers to obtain the explicit sentiment judgment result of the sentiment evidence for the sentiment object. ; When the sentiment evidence in the final evidence set does not contain evaluative words, but reflects an evaluative tendency towards the sentiment object through factual outcomes, anomalous states, repeated processing, or user experience consequences, the sentiment evidence is identified as implicit sentiment expression, and the implicit sentiment judgment result is expressed as follows: ； in, This indicates the implicit sentiment judgment result. Representing implicit sentiment inference functions, including: classifying sentiment objects Emotional evidence and domain resource collection As input, the probability distribution of candidate polarities is obtained using the object-level sentiment polarity inference function, and the sentiment category with the highest probability is selected as the implicit sentiment judgment result. When the sentiment evidence in the final evidence set contains negation, degree modification, adversative relationship, multiple object coexistence, or cross-sentence reference, the sentiment evidence is identified as a complex structured sentiment expression. Based on punctuation, negation, degree modification, adversative relationship, and the relationship between the sentiment object and the corresponding sentiment evidence, each piece of sentiment evidence is associated with its corresponding sentiment object. Then, the sentiment object corresponding to the sentiment evidence and the contextual fragment information are input into the object-level sentiment polarity inference function to obtain the probability distribution of the candidate polarity of the complex structured sentiment expression, i.e., the complex structured sentiment judgment result. ; When multiple pieces of emotional evidence correspond to the same emotional object, the weight of each piece of emotional evidence is determined based on the comprehensive relevance score: ； in, Indicates the first Emotional evidence in the emotional object Weights in polarity determination This represents the r-th piece of emotional evidence; According to weight Obtain the polarity determination result: ； in, Indicates based on the emotional object The polarity judgment results of each corresponding piece of emotional evidence; For explicit sentiment judgment results or implicit sentiment judgment result Or complex structure emotion judgment results ; The polarity judgment results of each sentiment evidence are weighted and fused with the probability distribution of the candidate polarities: ； in, Representing the object of affection The final probability distribution of emotional polarity, This represents the weight of the language model output, with a value range of [value range missing]. ; Selecting the final emotional polarity probability distribution The sentiment category with the highest probability is selected as the candidate sentiment polarity; when the highest probability is greater than a preset reliability threshold, the candidate sentiment polarity is used as the final sentiment polarity; when the highest probability is lower than the preset threshold, or When the probability difference between the positive and negative categories is less than a preset difference threshold, it is marked as pending review or mixed sentiment. Further optimization involves the weighting of the language model output. The default reliability threshold is 0.5; the preset reliability threshold range is 0.7 to 0.9; the preset difference threshold is 0.1 to 0.

2. The optimal setting is a confidence threshold of 0.8 and a difference threshold of 0.

15.

7. The text sentiment analysis method integrating sentiment object extraction and sentiment polarity reasoning according to claim 6, characterized in that, The specific implementation process of step 4 includes: For each identified sentiment object, generate object-level structured analysis results: ； in, Indicates the first The structured analysis results corresponding to each emotion object; if the text to be analyzed contains If there are several emotional objects, then the structured result set of the text to be analyzed is represented as follows: , Indicates the first One emotional object; Indicates the overall confidence level; Represents object-level structured results The processing status is such that the selected candidate sentiment object set, the final evidence set, and the final sentiment polarity are the same, and the overall confidence level is... Reaching the preset confidence threshold When the overall confidence level does not reach the preset confidence threshold, the processing status is set to automatic pass; when the overall confidence level does not reach the preset confidence threshold, the processing status is set to automatic pass. Or there may be conflicting evidence or a final probability distribution of emotional polarity. When the difference between the highest probability and the second highest probability is less than the preset difference threshold, and the correspondence between the emotional object and the emotional evidence is unclear, the processing status is set to pending review. Representing the object of affection Corresponding candidate-level sentiment polarity: ;in, This represents a predefined set of emotional polarity categories, including at least positive, negative, neutral, and undetermined. This represents the probability distribution of emotional polarity. Emotional object Belongs to the category of emotional polarity The probability; when the structured analysis results are marked as pending review, Indicates candidate polarity; The inference specification is indicated by confirming the inference specification based on the structured output fields in the domain resource set. If no matching inference specification exists, it is determined based on the sentiment object. Collection of evidence Candidate polarity As well as the scope of negation, degree of modification, adversative relationship and domain evidence template, the reasoning explanation is generated using a preset explanation template or a large language model; when When automatic passage is enabled, Directly used as the final object-level sentiment analysis result; when For the purpose of review, The candidate analysis results are submitted for manual review or secondary model verification, and updated based on the review results to form confirmed object-level sentiment analysis results. For the set of structured analysis results of the same text to be analyzed There are multiple different emotional objects in it. or a certain emotional object The corresponding final set of evidence In cases where multiple pieces of sentiment evidence exist, the object-level sentiment analysis results are integrated and conflict resolved; including: The names of sentiment objects are standardized and converted into standard names, namely abbreviations, aliases, and synonyms in the domain terminology lexicon. When two sentiment objects have the same name, or when there is at least one relationship between them—synonym, abbreviation, alias, reference, overlapping textual scope, or both pointing to the same evaluated entity—they are considered semantically identical sentiment objects. When the sentiment object similarity of two sentiment objects reaches a preset threshold... When both belong to the same category of emotional objects, they are considered to be highly similar emotional objects. For sentiment objects identified as semantically identical or highly similar, they are merged. Simultaneously, the standard name from the domain terminology lexicon is selected as the merged object name. If no standard name exists in the domain terminology lexicon, the object name with the highest object recognition confidence, the most complete text expression, or the highest frequency of occurrence is selected as the merged object name. Furthermore, the evidence sets corresponding to the merged objects undergo union processing, deleting evidence with identical text content, overlapping positional ranges, or highly repetitive semantics, while retaining the source sentence elements, positional identifiers, comprehensive relevance scores, and evidence-level judgment confidence for each piece of evidence. The evidence-level judgment confidence is represented by the sentiment polarity probability distribution. In this context, emotional evidence falls under the category of emotional polarity. The maximum probability; For the set of structured analysis results of the same text to be analyzed In cases where multiple pieces of emotional evidence relate to the same emotional object, the evidence fusion weight is calculated based on the strength of evidence, the degree of contextual proximity, and the confidence level of each piece of emotional evidence. : ； in, This represents the strength of evidence, or support strength, of the k-th sentiment evidence. This indicates the degree of contextual proximity between the object of emotion and the evidence of emotion. Indicates the first The confidence level of evidence corresponding to each piece of emotional evidence; This indicates the strength of the emotional evidence presented in the r-th piece. This indicates the degree of contextual proximity between the emotional object and the r-th piece of emotional evidence. Indicates the first The confidence level of evidence corresponding to each piece of emotional evidence; We use evidence fusion weights to weight and fuse the probability distributions of sentiment polarity from different sentiment evidence: ； in, This represents the probability distribution of the emotional polarity after fusion; For the set of structured analysis results of the same text to be analyzed For emotional objects where there are conflicting pieces of evidence, calculate conflict resolution scores for each piece of emotional evidence: ； in, Represents the overall relevance score, parameters lie in An interval, satisfying: ; When the probability distribution of emotional polarity after fusion The preset polarity judgment threshold is reached, and When the difference between the highest probability and the second highest probability reaches a preset difference threshold, the sentiment polarity corresponding to the highest probability is taken as the object-level judgment result, and the sentiment polarity corresponding to the highest probability is taken as the object-level sentiment analysis result; when the evidence strength of the sentiment evidence of positive and negative categories reaches the evidence strength threshold, or when the difference between the highest probability and the second highest probability in the fused sentiment polarity probability distribution is lower than the preset difference threshold, the sentiment object is marked as mixed sentiment or pending review; Further preferred, a preset threshold The threshold value is 0.85; the preset polarity judgment threshold is 0.7–0.9; the evidence strength threshold is 0.7–0.

9. Further preferred, the similarity of the emotional object includes: for the emotional object and The formula for calculating the similarity between emotional objects is: ； in, express and Similarity of emotional objects The vector semantic similarity of two emotion objects is obtained by generating object vectors through a text encoding model and calculating cosine similarity. The lexical overlap of object names is calculated by Jaccard similarity to the character set, word set, or keyword set of the sentiment object; This indicates the degree of matching of synonyms, abbreviations, alternative names, or referential relationships. It is 1 when two sentiment objects are marked as synonyms in the domain terminology lexicon, or when two sentiment objects point to the same evaluated entity; otherwise, it is 0. This represents the similarity parameter, with a value range of [value range missing]. The interval, and satisfying ; The strength of evidence is used to indicate the degree to which a candidate piece of evidence supports the judgment of the target emotional object and emotional polarity; Candidate evidence relative to the emotional object The strength of evidence is expressed as follows: ； in, This indicates the strength of evidence, or the strength of support. This indicates the overall relevance score; It indicates the clarity of the sentiment orientation of the evidence, and is determined based on the highest polarity probability in the evidence-level sentiment classification results; The semantic similarity indicates the degree of matching between candidate evidence and domain evaluation words, negation structures, factual outcome descriptions, or other sentiment evidence templates. This indicates the degree of polarity consistency between emotional evidence and other emotional evidence concerning the same emotional object; a value of 1 indicates the same emotional polarity as other evidence concerning the same emotional object, and a value of 0 indicates a different polarity. (Parameter) All located in An interval, satisfying: ; Contextual proximity between the emotional object and the candidate evidence Represented as: ； in, The number of lexical units or sentence units representing the interval between the emotional object and the candidate evidence. This represents the distance attenuation parameter; Overall confidence level As shown below: ； in, This represents the confidence level in identifying the emotional object. Indicates the confidence level of the emotional evidence match. Indicates the confidence level of sentiment polarity reasoning. These are weight parameters, located at The interval, and the sum of the weight parameters is 1; Among them, the confidence level of emotion object recognition Calculated based on object boundary recognition probability, domain object category matching degree, and the degree of correlation between the object and the evaluation expression: ； in, This represents the probability of object boundaries or object categories output by the emotion object recognition model; This indicates the degree of matching between the candidate object and the domain object category and the domain terminology vocabulary. When the candidate object matches the domain sentiment object category... or field terminology glossary If the object name, synonym, abbreviation, or alias is the same, the value is 1; if they are different, the semantic similarity between the candidate object and each domain object category or domain term is calculated, and the highest similarity is taken as the result. ; The correlation score represents the degree of association between the candidate object and the evaluation expression; parameters lie in An interval, satisfying: ; Confidence of Emotional Evidence Matching The weighted calculation is based on the overall relevance score of each piece of evidence in the final evidence set: ； in, For evidence fusion weights; Confidence of Emotional Polarity Inference Calculated based on the highest probability in the fused polarity probability distribution, and the difference between the highest and second-highest probabilities: ； in, This represents the highest probability in the polarity probability distribution after fusion. Indicates the second highest probability; parameter All located in An interval, satisfying: ; Further optimized, weight parameters They were set to 0.3, 0.3, and 0.4 respectively.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the text sentiment analysis method that integrates sentiment object extraction and sentiment polarity reasoning as described in any one of claims 1-7.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the text sentiment analysis method that integrates sentiment object extraction and sentiment polarity reasoning as described in any one of claims 1-7.

10. A text sentiment analysis system integrating sentiment object extraction and sentiment polarity reasoning, characterized in that, include: The structural unit and domain resource construction module is configured to: perform structured preprocessing on the original text to be analyzed, construct standardized sentence-level structural units, and construct a set of domain resources according to the domain type of the original text to be analyzed; The sentiment object and candidate evidence acquisition module is configured to: judge sentiment relevance based on standardized sentence element-level structural units and domain resource set, filter out sentiment analysis candidate sentence elements, and acquire standardized sentence element-level structural units belonging to the sentiment analysis candidate sentence elements; Semantic parsing is performed on standardized sentence-level structural units that belong to the candidate sentence elements of sentiment analysis to identify the set of candidate sentiment objects and filter the set of sentiment objects; Candidate evaluation components are extracted from the selected set of candidate emotional objects to form a set of candidate evidence. After screening the set of candidate evidence, a comprehensive correlation score between the set of candidate emotional objects and the set of candidate evidence is calculated. The final set of evidence is obtained based on the correlation score. The sentiment polarity reasoning module is configured to: perform sentiment polarity reasoning on the standardized sentence-level structural units belonging to the sentiment analysis candidate sentence elements, the filtered candidate sentiment object set, the final evidence set and the corresponding domain resource set, and obtain the final sentiment polarity; The fusion and conflict resolution module is configured to: construct object-level structured analysis results, review them, and generate object-level sentiment analysis results; Based on the similarity of sentiment objects, the strength of evidence, and the overall confidence level, the results of object-level sentiment analysis are integrated and conflict resolved.

Images