A method for extracting aspect-sentiment triples from large language model annotation datasets

Abstract

The present invention discloses a method for extracting aspect sentiment triples from a large language model annotated dataset, relating to the field of natural language processing technology. The method comprises the following steps: performing initial triple annotation on a standardized text dataset, calculating a cognitive bias index based on an annotation reference set, and generating a bias-labeled dataset; employing a dynamic attenuation suppression method to generate a bias mask matrix, performing exponential mask enhancement on samples exceeding a preset bias threshold, and obtaining a bias mask enhanced representation; inputting the bias mask enhanced representation into a differentiable grammar parser, calculating a grammatical dependency matrix using an attention mechanism, and optimizing the topology of the grammatical dependency matrix using structural entropy loss to generate a boundary decision matrix; extracting candidate triple embeddings based on the boundary decision matrix, calculating the contradiction coefficient using quantum emotional entangled states, and generating a corrected triple set. The present invention enhances the accuracy and usability of sentiment recognition results in fields such as medical evaluation.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a method for extracting aspect sentiment triples from a large language model annotation dataset. Background Art

[0002] With the rapid development of the internet and social media, user-generated text data has surged, and sentiment analysis technology has become a key research direction in natural language processing. In recent years, sentiment triple extraction, a key technology for fine-grained sentiment analysis, aims to automatically identify "aspect-sentiment-emotion polarity" triplets from text, providing strong data support for accurate user profiling, public opinion monitoring, and intelligent recommendations. Traditional sentiment triple extraction methods rely on rule matching, statistical machine learning, and shallow neural network models. They are typically based on lexical and syntactic features, and triples are extracted through hand-designed rules or shallow feature engineering.

[0003] While existing methods have achieved some success on standard datasets, they still have limitations in practical applications. The triplet annotation process is susceptible to inherent cognitive biases in the training data, leading to biases in sentiment attribution and polarity judgment when large language models handle polysemous words, ambiguous modifiers, or complex sentence structures. This is particularly true in social media comments or e-commerce user feedback, where comments are often short and unstructured, and the language expressions exhibit ambiguity, irony, and double negation. Current large language model-based automatic annotation methods, lacking external cognitive control or bias recognition mechanisms, struggle to identify errors in annotation (e.g., "positive words modifying negative aspects" are misclassified as positive polarity), which directly impacts the accuracy of triplet extraction. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method for extracting aspect sentiment triples from a large language model annotation dataset, which solves the problem of sentiment judgment bias caused by cognitive bias in the prior art.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, the present invention provides a method for extracting aspect sentiment triples from a large language model annotated dataset, which includes: collecting social media comment text and preprocessing it to obtain a standardized text dataset; performing initial triple annotation on the standardized text dataset, and calculating a cognitive bias index based on a labeled reference set to generate a biased labeled dataset; using a dynamic attenuation suppression method to generate a bias mask matrix, performing exponential mask enhancement on samples exceeding a preset bias threshold, and obtaining a bias mask enhanced representation; inputting the bias mask enhanced representation into a differentiable grammar parser, using an attention mechanism to calculate a grammatical dependency matrix, and optimizing the topological structure of the grammatical dependency matrix in combination with structural entropy loss to generate a boundary decision matrix; extracting candidate triple embeddings based on the boundary decision matrix, calculating the contradiction coefficient using quantum emotional entangled states, and generating a corrected triple set; inputting the corrected triple set into a multi-view graph neural network to fuse multi-view features, generating an optimized representation vector, and mapping it to an axial sparse attention table structure, and using row and column independent decoding to generate structured sentiment triples.

[0008] As a preferred solution of the method for extracting aspect sentiment triples from the large language model annotation dataset of the present invention, wherein: the standardized text dataset is obtained, the specific steps are as follows:

[0009] Collecting social media user comment content, timestamps, and user IDs, extracting comment fields to form social media comment text;

[0010] The rule cleaning and regular expression matching method is used to denoise the social media comment text to obtain the cleaned comment text;

[0011] The part-of-speech tagging method and unified coding standards are used to convert the format and regularize the language of the cleaned comment text to obtain a standardized text dataset.

[0012] As a preferred solution of the method for extracting aspect sentiment triples from a large language model annotation dataset of the present invention, the specific steps of generating a biased labeled dataset are as follows:

[0013] Based on the standardized text dataset, some samples are selected by random sampling, and triple annotation data is generated by combining the rule extraction method to construct the annotation reference set;

[0014] Use the labeled reference set to perform supervised fine-tuning on the pre-trained large language model to obtain the fine-tuned large language model;

[0015] The cleaned comment text of the standardized text dataset is input into the fine-tuned large language model, and the initial triples are annotated on the cleaned comment text using the sequence labeling method to obtain the initial triple set;

[0016] Referring to the annotated reference set, the cognitive bias index of each triple sample in the initial triple set is calculated, and the high-bias samples are marked to generate a biased labeled dataset.

[0017] As a preferred solution for the method of extracting aspect sentiment triples from the large language model annotated dataset described in the present invention, a bias threshold is set according to the bias-tagged dataset, and a dynamic attenuation suppression method is used to perform exponential mask enhancement processing on the cognitive bias index exceeding the bias threshold, and a bias mask matrix is ​​generated and fused with the representation of the cleaned comment text to obtain a bias mask enhanced representation.

[0018] As a preferred solution of the method for extracting aspect emotion triples from a large language model annotation dataset of the present invention, the specific steps of generating a boundary decision matrix are as follows:

[0019] Taking the bias mask enhanced representation as input, an optimization algorithm with a structured entropy loss function is used to train the differentiable grammar parser to obtain the trained differentiable grammar parser.

[0020] Input the bias mask enhanced representation of the text to be processed into the trained differentiable grammar parser for calculation to obtain the grammatical dependency matrix;

[0021] The structural entropy loss of the grammatical dependency matrix is ​​optimized, and the matrix topology is adjusted. According to the optimized grammatical dependency matrix and the preset boundary judgment rules, the boundary relationship between the candidate entity nodes is extracted to generate a boundary judgment matrix.

[0022] As a preferred solution of the method for extracting aspect emotion triples from a large language model annotation dataset of the present invention, the specific steps of generating a correction triple set are as follows:

[0023] Extract the context embedding of candidate triples from the boundary decision matrix and construct the quantum sentiment state representation using word vector encoding method;

[0024] The contradiction coefficient of quantum emotional state representation is calculated using the entanglement metric function and compared with the entanglement dissociation threshold;

[0025] For the contradictory coefficients represented by quantum emotional states exceeding the entanglement dissociation threshold, the quantum state decomposition method is used to split them into multiple atomic emotional paths;

[0026] The emotional contradiction detection and adjustment method is used to correct the emotional contradictions in the atomic emotional path and generate a set of correction triples.

[0027] As a preferred solution of the method for extracting aspect sentiment triples from a large language model annotation dataset of the present invention, the specific steps of generating the optimized representation vector are as follows:

[0028] Convert the corrected triples into graph-structured data, and construct a grammatical dependency view, a semantic enhancement view, and an emotional bridge view based on the graph-structured data. Extract the feature representations of nodes and edges in multiple views to obtain the initial feature set of multiple views.

[0029] A supervised learning method is used to train the multi-view graph neural network, and the parameter weights of the node representation are optimized to obtain the trained multi-view graph neural network.

[0030] The initial multi-view feature set is input into the trained multi-view graph neural network, and the feature representations of nodes and edges in the multi-views are weightedly fused using the attention mechanism to generate a comprehensive multi-view feature representation.

[0031] Through iterative calculations of a multi-layer graph neural network, the comprehensive multi-view feature representations corresponding to neighbor nodes are aggregated layer by layer to generate an optimized representation vector.

[0032] As a preferred solution of the method for extracting aspect emotion triples from a large language model annotation dataset of the present invention, the specific steps of generating structured emotion triples are as follows:

[0033] Using the axial sparse attention method, the optimized representation vector is mapped into a two-dimensional representation tensor;

[0034] Apply sparse attention to the two-dimensional representation tensor along the row axis to extract local context features between entities and obtain the entity combination tensor;

[0035] Apply sparse attention to the entity combination tensor along the column axis to extract the weighted features of the relationship dimension and obtain the relationship representation tensor;

[0036] The entity combination tensor and the relationship representation tensor are decoded independently in rows and columns to generate structured sentiment triples.

[0037] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the method for extracting aspect emotion triples from a large language model annotated dataset as described in the first aspect of the present invention.

[0038] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the method for extracting aspect sentiment triples from a large language model annotated dataset as described in the first aspect of the present invention.

[0039] The present invention has the following beneficial effects: through quantum emotional entanglement state calculation and splitting mechanism, it can effectively model emotional conflicts in comment texts. By precisely triggering the splitting operation of emotional states, it improves the accuracy of triple extraction in complex emotional scenarios and significantly reduces the extraction error rate caused by emotional polarity ambiguity. The atomic emotional paths generated by the splitting mechanism can correspond to independent structured emotional triplets, effectively resolving the confusion problem in multi-polar emotional expressions, improving the refinement of emotional expression, and enhancing the accuracy and usability of emotion recognition results in fields such as medical evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0041] Figure 1 Flowchart of the method for extracting aspect sentiment triples from a large language model annotation dataset in the present invention.

[0042] Figure 2 This is a flow chart of data preprocessing and initial labeling in the present invention.

[0043] Figure 3 This is a flow chart of the quantum emotion correction process in the present invention.

[0044] Figure 4 This is a flowchart of multi-view fusion and structured generation in the present invention. DETAILED DESCRIPTION

[0045] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0046] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0047] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0048] Reference Figures 1 to 4, is an embodiment of the present invention, which provides a method for extracting aspect-emotion triples from a large language model annotation dataset, comprising the following steps:

[0049] S1. Collect and preprocess social media comment texts to obtain a standardized text dataset.

[0050] S1.1. Collect social media user comment content, timestamps, and user IDs, and extract comment fields to form social media comment text.

[0051] Specifically, the original data stream containing user comment content is obtained according to preset keywords through the social media data interface; the comment content field, timestamp field and user identification field in the original data stream are extracted according to the field structure; text preprocessing operations are performed on the comment content field, including removing HTML tags, emoticons, special symbols and redundant spaces; the timestamp field is standardized, and the original time format is uniformly converted into the year-month-day hour-minute-second format; the user identification field is uniquely verified, and the valid user identification string is retained; the extracted comment content field, standardized timestamp field and user identification field are combined into record entries in a unified format to form social media comment text.

[0052] S1.2. Use rule cleaning and regular expression matching methods to denoise the social media comment text to obtain the cleaned comment text.

[0053] Specifically, a rule-based cleaning method is used to delete hyperlinks, phone numbers, email addresses, and obvious advertising slogans from social media comment texts; a regular expression matching method is used to identify and remove consecutive repeated characters, meaningless character stacking (such as "hahahahaha" or "!!!!!!"), non-natural language characters (such as "*&^%$#@!"), and incomplete sentence fragments; emoticons and emojis in social media comment texts are uniformly replaced or removed; redundant spaces and line breaks in social media comment texts are deleted, retaining the semantically complete text content to form the cleaned comment text.

[0054] S1.3. Use part-of-speech tagging methods and unified coding standards to convert the format and regularize the language of the cleaned comment text to obtain a standardized text dataset.

[0055] Specifically, based on the existing natural language processing tools, the part-of-speech tagging method is performed on the cleaned comment text to identify the grammatical role of words in the sentence and mark the corresponding part-of-speech tags; the tagged words are encoded according to the unified coding specification, for example, "noun", "verb", "adjective" and so on are encoded as "NN", "VV", "JJ" respectively using a unified part-of-speech tag set; the synonymous and heterogeneous expressions in the tagging results are unified, for example, "bought", "purchased", and "ordered" are unified into "purchase" and the unified coding is retained; the structured text with the output format of "word / code" is formed to form a standardized text dataset.

[0056] S2. Perform initial triple annotation on the standardized text dataset and calculate the cognitive bias index based on the annotation reference set to generate a biased labeled dataset.

[0057] S2.1. Based on the standardized text dataset, some samples are selected by random sampling, and triple annotation data is generated by combining the rule extraction method to construct the annotation reference set.

[0058] Specifically, some samples are selected from the standardized text dataset by random sampling. For example, the sampling ratio can be set to 10% to obtain a representative text subset; the three elements of subject, opinion word and emotion polarity in emotional expression are identified one by one using a rule extraction method, and triple annotation data is generated according to the structure of "subject-opinion word-emotion polarity". For example, the triples {mobile phone-appearance-positive} and {mobile phone-battery-negative} are extracted from "the mobile phone has a good appearance but a poor battery"; the generated triple annotation data are sorted according to the original text index and summarized to form an annotation reference set.

[0059] S2.2. Use the labeled reference set to perform supervised fine-tuning on the pre-trained large language model to obtain the fine-tuned large language model.

[0060] Specifically, each set of original text in the annotated reference set and the corresponding triplet annotated data are constructed into an instruction-type input-output pair. For example, "the phone has a good appearance but a poor battery" is mapped to the input "extract triplet: the phone has a good appearance but a poor battery", and the output is "{phone-appearance-positive}, {phone-battery-negative}"; the constructed input-output pair is processed by word segmentation encoding to generate tensor format data that can be recognized by the pre-trained large language model; then the supervised fine-tuning parameters are set, such as setting the learning rate to 1e-5, the batch size to 8, adopting the cross-entropy loss function and using the Adam optimizer for training; finally, the pre-trained large language model is iteratively updated according to the specified supervised fine-tuning parameters to obtain the fine-tuned large language model.

[0061] The specific steps for setting the supervised fine-tuning parameters should also be explained: select the initial learning rate based on the sample size of the labeled reference set and the capacity of the pre-trained large language model, for example, set the learning rate to 1e-5; determine the number of samples for each gradient update, that is, set the batch size, for example, set the batch size to 8; then set the loss function to the cross-entropy loss function to measure the difference between the output and the triplet labeled data; finally, select the optimizer as Adam, and set the weight decay parameter and gradient clipping threshold, for example, set them to 0.01 and 1.0 respectively, to stabilize the training process and prevent gradient explosion.

[0062] S2.3. Input the cleaned comment text of the standardized text dataset into the fine-tuned large language model, and use the sequence labeling method to perform initial triple labeling on the cleaned comment text to obtain the initial triple set.

[0063] Specifically, the cleaned comment text in the standardized text dataset is input into the fine-tuned large language model, and each word or character in the cleaned comment text is annotated according to the BIO encoding rules. The sequence annotation method is called to generate the mark label sequence corresponding to the subject, predicate, and object word by word. Then, based on the continuous and subject-predicate-object grammatical structure of the mark label sequence, the triple candidate items are extracted, and all the triple candidate items are organized into an initial triple set.

[0064] S2.4. Referring to the annotated reference set, calculate the cognitive bias index of each triple sample in the initial triple set, mark the high-bias samples, and generate a biased labeled dataset.

[0065] Specifically, by comparing the attribute matching, semantic consistency and context relevance of the initial triple set with the corresponding triples in the annotated reference set, a weighted scoring method is used to obtain the cognitive bias index; a bias threshold is set, for example, the bias threshold is 0.7, and it is judged whether the cognitive bias index exceeds the bias threshold. If the cognitive bias index is higher than the bias threshold, the corresponding triple sample is marked as a high-bias sample, and a biased labeled dataset containing normal samples and high-bias samples is generated.

[0066] The specific steps for setting the bias threshold should also be explained: based on the annotated reference set and the initial triple set, the distribution of the cognitive bias index is statistically analyzed, and the mean, variance, and quantile of the cognitive bias index are calculated; based on the statistical results and actual needs, an appropriate bias threshold is selected, for example, the bias threshold is set to the 75th percentile of the cognitive bias index or a fixed value such as 0.7 to distinguish high-bias samples from low-bias samples; finally, the rationality and distinguishing effect of the set bias threshold in sample labeling are verified to ensure that the bias threshold can effectively mark abnormal bias.

[0067] S3. Use the dynamic attenuation suppression method to generate a bias mask matrix, perform exponential mask enhancement on samples exceeding the preset bias threshold, and obtain a bias mask enhanced representation.

[0068] S3.1. Set a bias threshold based on the bias-labeled dataset. For cognitive bias indices exceeding the bias threshold, perform exponential mask enhancement processing using a dynamic attenuation suppression method. Generate a bias mask matrix and fuse it with the representation of the cleaned comment text to obtain a bias mask enhanced representation.

[0069] Specifically, according to the bias-labeled dataset, the cognitive bias index of each sample with a bias label is calculated, and based on the set bias threshold, for example, the bias threshold is set to (The example value is 0.05); for cognitive bias index exceeding the bias threshold For samples of the cleaned review text, a dynamic attenuation suppression method is used to adjust the cognitive bias index according to the exponential attenuation formula to generate exponential attenuation weights; a bias mask matrix is ​​constructed based on the exponential attenuation weights, where the values ​​at the corresponding positions in the bias mask matrix are the attenuation weights, and the other positions are zero; finally, the bias mask matrix is ​​fused element by element with the representation of the cleaned review text by multiplying or adding the corresponding elements to obtain the bias mask enhanced representation.

[0070] It should also be explained that the representation of the cleaned comment text comes from the text feature representation obtained by preprocessing the original comment text. The specific steps include: word segmentation of the original comment text, removal of stop words, removal of special symbols and noise data, and using a pre-trained word vector model (such as Word2Vec, GloVe or a pre-trained large language model such as BERT) to convert the processed text into a vector representation, and finally obtain the representation of the cleaned comment text.

[0071] S4. Input the bias mask enhanced representation into the differentiable grammar parser, use the attention mechanism to calculate the grammatical dependency matrix, combine the structural entropy loss to optimize the topological structure of the grammatical dependency matrix, and generate the boundary decision matrix.

[0072] S4.1. Taking the bias mask enhanced representation as input, an optimization algorithm with a structured entropy loss function is used to train the differentiable grammar parser to obtain the trained differentiable grammar parser.

[0073] Specifically, the parameters of the differentiable grammatical parser are initialized, the objective function is defined using a structured entropy loss function, and the weighted sum of the entropy term of the predicted grammatical structure probability distribution and the cross-entropy loss of the true labeled grammatical structure is calculated; the loss value is calculated based on the objective function, and the gradient information of the differentiable grammatical parser parameters is calculated using the backpropagation algorithm; then the optimization algorithm is used to iteratively update the differentiable grammatical parser parameters according to the set learning rate. The update rule uses, for example, gradient descent with momentum or the Adam optimization algorithm. The training process is repeated for multiple iterations until the loss function converges or reaches a preset number of iterations, and finally a trained differentiable grammatical parser is obtained.

[0074] It should also be explained that until the loss function converges or reaches a preset number of iterations, the specific steps include: setting the maximum number of iterations and the convergence threshold, for example, the maximum number of iterations is 1000 times, and the convergence threshold is 0.001 in the example; calculating the current loss function value after each iteration and comparing it with the loss function value of the previous iteration; if the difference between the two loss values ​​is less than the convergence threshold, the loss function is determined to have converged and the iteration is stopped; if the current number of iterations reaches the maximum number of iterations, the iteration is stopped; otherwise, continue to the next update iteration.

[0075] S4.2. Input the bias mask enhanced representation of the text to be processed into the trained differentiable grammar parser for calculation to obtain a grammar dependency matrix.

[0076] Specifically, the biased mask enhanced representation of the processed text is encoded word by word to generate a word vector sequence. Then, the word vector sequence is calculated layer by layer through the structure of the trained differentiable grammar parser, including the self-attention mechanism and the feedforward network, to obtain the feature representation of the grammatical relationship between words. Finally, the grammatical dependency matrix is ​​calculated based on the output layer of the trained differentiable grammar parser. Each element of the grammatical dependency matrix represents the strength of the dependency relationship between the corresponding word pairs. In the example, the grammatical dependency matrix is ​​a grammatical dependency matrix with a dimension of N×N, where N is the number of words in the cleaned comment text.

[0077] It should also be noted that the biased masked enhanced representation of the processed text is derived from word vector encoding of the original review text, combined with preset bias mask information for feature enhancement. The bias mask forms enhanced features by marking the semantic emphasis or contextual relevance of specific locations or words in the original review text, which is used to improve the expressive power of subsequent grammatical parsing. The biased masked enhanced representation is generated based on publicly available pre-trained word vector models and manually defined masking rules.

[0078] S4.3. Optimize the structural entropy loss of the grammatical dependency matrix, adjust the matrix topology, and extract the boundary relationship between the candidate entity nodes to generate a boundary decision matrix based on the optimized grammatical dependency matrix and the preset boundary decision rules.

[0079] Specifically, based on the initial dependency type adjacency matrix, the structural entropy loss is calculated, and the expression is:

[0080] ;

[0081] in, represents the structural entropy loss of the grammatical dependency matrix, represents the trace operation of the grammatical dependency matrix, represents the weight adjustment parameter matrix, represents the dependency weight between nodes, a one-hot encoding matrix representing the dependency type partitions, Represents the grammatical dependency matrix Sum all elements, represents the size of the common dependency label set, Represents the grammatical dependency matrix The number of nodes, Represents a transpose operation;

[0082] The gradient descent algorithm is used to iteratively optimize the weight distribution of the initial dependency type adjacency matrix, minimize the structural entropy loss, and obtain the optimized grammatical dependency matrix; based on the optimized grammatical dependency matrix and combined with the part-of-speech tagging results, nodes that meet the definition of noun phrases are located as candidate entity nodes, such as word nodes with part-of-speech tags as nouns (NN) or proper nouns (NNP); according to the preset boundary judgment rules, it is judged whether there is a direct dependency path between any two candidate entity nodes and the path length does not exceed the preset boundary judgment threshold. The boundary judgment threshold is an integer value parameter set based on corpus statistics and entity dependency characteristics; define the boundary judgment matrix ,in Represents the boundary decision matrix, if and only if the candidate entity node and When the boundary judgment rule is met, let ,otherwise , an example is the noun phrase node "screen" (position ) and "mobile phone" (location ) The length of the dependency path is 2, then set .

[0083] The specific steps of the preset boundary judgment rule are as follows: compare the dependency weights of any two candidate entity nodes in the grammatical dependency matrix. If the dependency weight is greater than the dependency strength judgment reference value, it is considered that there is a boundary relationship between the two nodes. The dependency strength judgment reference value is the 75% quantile of the dependency weight distribution of entity node pairs with dependency relationships in the training corpus; further combine the word order positions of the two candidate entity nodes in the cleaned comment text, if the word order spacing is less than or equal to the word order spacing judgment reference value, it is considered to meet the word order continuity condition. The word order spacing judgment reference value is taken from the average word order spacing of entity pairs with boundary relationships in the training corpus, for example, it is set to 3; finally, when the dependency strength judgment reference value and the word order spacing judgment reference value conditions are met at the same time, the corresponding element in the boundary judgment matrix is ​​assigned a value of 1, otherwise it is assigned a value of 0.

[0084] S5. Extract candidate triple embeddings based on the boundary judgment matrix, calculate the contradiction coefficient using the quantum emotional entangled state, and generate a set of corrected triples.

[0085] S5.1. Extract the context embedding of candidate triples from the boundary decision matrix and construct the quantum sentiment state representation using word vector encoding method.

[0086] Specifically, the step of extracting the context embedding is: according to the position of all elements with a value of 1 in the boundary judgment matrix, determine the starting position index of the entity node in the candidate triple. Based on the starting position index, extract the word vector of the word at the corresponding position from the pre-trained word vector library. The dimension of the word vector is, for example, 768 dimensions, and BERT word vector representation is often used. With each pair of position indexes as the center, intercept the word vector sequence of several words before and after. In this example, take the first three and last three words. Concatenate the word vector sequence to form the context embedding vector corresponding to the candidate triple.

[0087] The steps for constructing the quantum emotion state representation are as follows: the obtained context embedding vector is input into the fully connected layer, and a three-dimensional coefficient vector is obtained after mapping. Using the three-dimensional coefficient vector, according to the principle of quantum state superposition, the three coefficients are mapped to the standard orthogonal basis vectors of the three emotional ground states of positive, neutral, and negative, respectively, thereby constructing the quantum emotion state representation of the candidate triples, expressed as:

[0088] ;

[0089] in, represents the quantum emotional state representation, Represents the corresponding positive emotional ground state in the quantum emotional state The amplitude coefficient of is a real number. Represents the corresponding neutral emotional ground state in the quantum emotional state The amplitude coefficient of is a real number. Represents the corresponding negative emotional ground state in the quantum emotional state The amplitude coefficient of is a real number. Indicates positive emotional ground state, Indicates a neutral emotional ground state, Indicates negative emotional base state.

[0090] S5.2. Use the entanglement metric function to calculate the contradiction coefficient represented by the quantum emotional state and compare it with the entanglement dissociation threshold.

[0091] Specifically, according to the calculation formula of the entanglement metric function, the contradiction coefficient represented by the quantum emotional state is calculated. , the expression is:

[0092] ;

[0093] in, represents the contradiction coefficient of the quantum emotional state representation, represents the entanglement strength between the positive emotional ground state and the neutral emotional ground state, represents the entanglement strength between the neutral emotional ground state and the negative emotional ground state, Represents the entanglement strength between the negative emotional ground state and the positive emotional ground state; For example, when the quantum emotional state represents When , the contradiction coefficient M is calculated to be 0.17;

[0094] Setting the Disentanglement Threshold , based on the statistical data of the annotated reference set, a preset entanglement dissociation threshold is determined . Entanglement dissociation threshold The statistical quantiles derived from the annotated reference set where the proportion of contradictory triples exceeds 90%; the entanglement dissociation threshold comparison operation is performed, and the calculated contradiction coefficient is and the preset disentanglement threshold To judge the value. Greater than When the quantum state is above the threshold, the "exceeding the entanglement dissociation threshold" flag is output; otherwise, the "not exceeding the entanglement dissociation threshold" flag is output. The comparison result is stored for subsequent quantum state decomposition decision processing.

[0095] S5.3. For the contradictory coefficients represented by quantum emotional states that exceed the entanglement dissociation threshold, the quantum state decomposition method is used to split them into multiple atomic emotional paths.

[0096] Specifically, a numerical comparison is performed between the corresponding contradiction coefficient value in the quantum emotional state representation and a preset entanglement dissociation threshold. If the contradiction coefficient value is greater than the entanglement dissociation threshold, the quantum emotional state representation is determined to be in an entangled unstable state. The quantum state decomposition method is applied to the quantum emotional state representation, and the quantum emotional state representation is mapped into a number of orthogonal or approximately orthogonal atomic emotional paths in the quantum state representation space. The quantum state decomposition method adopts a standard linear algebra solution method based on eigenstate expansion, specifically including constructing an emotional Hamiltonian matrix corresponding to the quantum emotional state representation, performing a decomposition operation, and extracting the eigenvectors obtained from the decomposition of the corresponding emotional Hamiltonian matrix as the atomic emotional path expression. Finally, all atomic emotional paths are sorted and numbered according to the projection intensity in the original quantum emotional state representation to form multiple atomic emotional paths that can be tracked individually.

[0097] It should also be explained that the specific steps for constructing the emotional Hamiltonian matrix corresponding to the quantum emotional state representation are as follows: extract the eigenvectors of all atomic emotional paths contained in the quantum emotional state representation to form a set of eigenvectors; calculate the coupling strength between each pair of paths based on the emotional coupling relationship between the atomic emotional paths, and the coupling strength is calculated as the dot product value or cosine similarity of the eigenvectors of the atomic emotional paths; construct the emotional Hamiltonian matrix, and assign the diagonal elements of the emotional Hamiltonian matrix to the self-energy values ​​of the corresponding atomic emotional paths, for example, according to the projection intensity, and assign the non-diagonal elements to the coupling strength between the corresponding paths; form a symmetric real matrix as the emotional Hamiltonian matrix, which is used to characterize the energy distribution and interaction structure between different paths of the quantum emotional state representation.

[0098] S5.4. Use the emotional contradiction detection and adjustment method to correct the emotional contradictions in the atomic emotional path and generate a set of correction triples.

[0099] Specifically, based on the atomic emotion path, the emotion similarity index between the atomic emotion paths is obtained by calculating the cosine value of the angle between the atomic emotion paths in the emotion ground state vector space; when any two atomic emotion paths correspond to the positive emotion ground state and the negative emotion ground state respectively, and the projection coefficients between the atomic emotion paths are greater than the example set value of 0.8, it is judged whether there is an emotion polarity conflict; for the atomic emotion path pairs with emotion polarity conflict, the atomic emotion path merging operation is performed, and the atomic emotion path merging method is to merge according to the emotion ground state direction corresponding to the atomic emotion path with the larger projection coefficient modulus value; after the atomic emotion paths are merged, the atomic emotion path merging is recalculated. The amplitude coefficient after merging is calculated by performing a weighted average operation on the amplitude coefficients of the two atomic emotion paths; a new quantum emotion state representation is constructed using the merged atomic emotion path; based on the merging results of the atomic emotion paths, a set of correction triples is generated, and the correction triplet set consists of the emotion subject entity, the emotion object entity, the merged emotion ground state direction label and the merged amplitude coefficient. When a positive emotion ground state atomic emotion path and a negative emotion ground state atomic emotion path appear at the same time and their projection coefficients are both greater than the example value 0.8, the atomic emotion paths are merged into a neutral emotion ground state path, and a correction triplet containing the neutral emotion ground state path is output.

[0100] It should also be explained how to determine whether there is an emotional polarity conflict: for each atomic emotional path, the dominant emotional polarity is calculated based on the corresponding emotional base state projection coefficient, and the dominant emotional polarity is the emotional polarity corresponding to the emotional base state with the largest projection coefficient; for all atomic emotional paths, combine them in pairs to determine whether the dominant emotional polarity has a combination of "positive emotional base state" and "negative emotional base state"; if there is a group of paths, one of which has a positive emotional base state as the dominant emotional polarity and the other has a negative emotional base state as the dominant emotional polarity, and the corresponding projection coefficients are both greater than 0.8 (for example), then it is determined that there is an emotional polarity conflict.

[0101] S6. Input the corrected triple set into the multi-view graph neural network to fuse the multi-view features, generate an optimized representation vector, and map it to the axial sparse attention table structure, and use row and column independent decoding to generate structured emotion triples.

[0102] S6.1. Convert the corrected triple set into graph-structured data, and construct a grammatical dependency view, a semantic enhancement view, and an emotional bridging view based on the graph-structured data. Extract the feature representations of nodes and edges in multiple views to obtain the initial feature set of multiple views.

[0103] Specifically, each emotional element in the corrected triple set is taken as a node in the graph structure data, and edges are constructed to connect the corresponding nodes according to the association relationship between the emotional elements in the triples; based on the graph structure data, the grammatical dependency edges and types between nodes are determined by analyzing the syntactic dependency relationship, and a grammatical dependency view is constructed; the semantic edges and features between nodes are determined by semantic role labeling and contextual semantic relationships, and a semantic enhancement view is constructed; the emotional transmission rules and emotional similarity calculation are combined to construct emotional bridging edges, and an emotional bridging view is constructed; finally, the feature vectors of the nodes and the weight features of the edges are extracted from the grammatical dependency view, the semantic enhancement view and the emotional bridging view, and the multi-view initial feature set is obtained by integration.

[0104] S6.2. Use supervised learning methods to train the multi-view graph neural network, optimize the parameter weights of the node representation, and obtain the trained multi-view graph neural network.

[0105] Specifically, labeled training sample data is prepared. The labeled training sample data comes from a publicly available annotated comment corpus. The corpus content includes user comment text and corresponding annotation entity boundary labels and grammatical dependency labels, which constitute data input containing a multi-view initial feature set and corresponding real node category labels; a multi-view graph neural network constructed based on existing graph neural network technology is adopted. The multi-view graph neural network consists of multiple shared layers and interactive fusion layers, which are used to capture structural dependency information and semantic coupling relationships under different feature perspectives; the multi-view initial feature set is input into the multi-view graph neural network in a node position alignment manner, and forward propagation is performed to calculate the node representation; based on the current node representation and the real node category label, the loss function value is calculated, and the cross entropy loss function is used in the example; the gradient descent algorithm is used to perform the back propagation process to update the parameter weights of each layer in the multi-view graph neural network; the forward propagation, loss function calculation and parameter weight update steps are repeated until the loss function converges; the trained multi-view graph neural network is obtained, and the weight parameters of the node representation have been optimized.

[0106] The specific steps until the loss function converges should also be explained: set numerical stability conditions to determine whether the training process is terminated, including that the change in the loss function value does not exceed the stability judgment error limit (for example, a value of 0.001), or that the change in the loss function value in several consecutive rounds (for example, 5 rounds) of iterations is not sufficient to form a downward trend; in each training iteration, calculate the current loss function value and compare it with the loss function value of the previous round of iteration; if the change in the loss function value is not lower than the stability judgment error limit and the iteration stability number requirement is not met, continue to the next round of iteration; if the change is lower than the error limit, or the iteration stability number condition has been met continuously, the training process is determined to be stable, the iteration is terminated, and the current parameters are deemed to have reached the optimal or approximately optimal state.

[0107] S6.3. Input the multi-view initial feature set into the trained multi-view graph neural network, and use the attention mechanism to perform weighted fusion on the feature representations of nodes and edges in the multi-views to generate a comprehensive multi-view feature representation.

[0108] Specifically, the initial feature set of multiple views is used as input and passed into the syntax dependency view subnetwork, semantic enhancement view subnetwork and emotion bridging view subnetwork in the trained multi-view graph neural network respectively to calculate the feature representation of nodes and edges in each view; the attention mechanism is used to assign different weights to the feature representation of nodes and edges in each view, and the weights are determined by the trained attention weight parameters, specifically including calculating the attention score of node features and the attention score of edge features; based on the attention score, the multi-view node features and edge features are weightedly summed to form comprehensive node features and comprehensive edge features; finally, the weighted fused node features and edge features are merged to generate a comprehensive multi-view feature representation.

[0109] S6.4. Perform iterative calculations through a multi-layer graph neural network, aggregate the comprehensive multi-view feature representations corresponding to neighbor nodes layer by layer, and generate an optimized representation vector.

[0110] Specifically, the comprehensive multi-view feature representation is input into the first layer of the multi-layer graph neural network, the feature representation of each node and the neighbor nodes is calculated, and the neighbor aggregation vector is obtained by the weighted sum of the features of the neighbor nodes; the neighbor aggregation vector is spliced ​​or added with the node's own features, and processed by a nonlinear activation function to generate the node representation vector of the first layer; the first layer node representation vector is used as the input of the next layer, and the neighbor aggregation, feature fusion and activation operations are repeated, and the calculation is iterated layer by layer until the optimized representation vector of the node is outputted at the last layer. The number of iteration layers and activation function are determined based on the example that the number of iteration layers is three, and the activation function adopts the ReLU function.

[0111] S6.5. Use the axial sparse attention method to map the optimized representation vector into a two-dimensional representation tensor.

[0112] Specifically, the optimized representation vector is first reshaped according to the preset two-dimensional mapping dimension parameters Dimension_Row (row dimension parameter) and Dimension_Column (column dimension parameter) to form an initial two-dimensional tensor; sparse attention weight matrices are constructed on the row and column axes of the two-dimensional tensor respectively. When calculating the row axis sparse attention weight, the elements in the row are weighted summed, and when calculating the column axis sparse attention weight, the elements in the column are weighted summed; the sparse attention weight matrices of the row and column axes are standardized respectively by the normalization function; the row axis sparse attention weighted results and the column axis sparse attention weighted results are multiplied and fused element by element to obtain the axial sparse attention fusion tensor; finally, the axial sparse attention fusion tensor is output as a two-dimensional representation tensor to complete the mapping process from the optimized representation vector to the two-dimensional representation tensor.

[0113] It should also be noted that Dimension_Row and Dimension_Column are preset two-dimensional mapping dimension parameters that determine the number of rows and columns when converting the optimized representation vector into a two-dimensional tensor. For example, Dimension_Row can be set to 16 and Dimension_Column can be set to 32, indicating that the one-dimensional optimized representation vector is converted into a two-dimensional tensor with 16 rows and 32 columns.

[0114] S6.6. Apply sparse attention to the two-dimensional representation tensor along the row axis to extract local context features between entities and obtain the entity combination tensor.

[0115] Specifically, obtain the representation vector group corresponding to each row in the two-dimensional representation tensor, set the attention window size of sparse attention to a fixed value; calculate the attention weight between the representation vector in each row and the representation vector in the adjacent window based on the sparse attention method, use the dot product attention mechanism to obtain the attention score, and perform Softmax normalization; use the normalized attention weight to perform weighted summation on the representation vector of each row in the window to form the context representation corresponding to the current row; splice all row-wise context representations in the original order to form an entity combination tensor.

[0116] It should also be explained that the specific steps for setting the attention window size of sparse attention to a fixed value include: pre-defining the attention window size parameter row-direction attention window size as a hyperparameter and fixing it unchanged, for example, the row-direction attention window size is 3; when applying sparse attention to the two-dimensional representation tensor along the row axis, for the representation vector in each row, the attention calculation is limited to the representation vectors that are no more than the row-direction attention window size positions before and after, thereby constructing a sparse attention connection range to ensure that the calculation is only performed within the set fixed window.

[0117] S6.7. Apply sparse attention to the entity combination tensor along the column axis to extract the weighted features of the relationship dimension and obtain the relationship representation tensor.

[0118] Specifically, set the attention window size parameter column-wise attention window size, for example, the column-wise attention window size is 3; for each column in the entity combination tensor, select the current position representation vector and the representation vectors of the positions no more than the column-wise attention window size above and below as the participants in the attention calculation; calculate the attention weight for each representation vector participating in the attention calculation in the selected window; perform weighted summation on all representation vectors in the window according to the attention weight to generate a weighted feature vector of the current column position; concatenate the weighted feature vectors corresponding to all column positions to form a relational representation tensor.

[0119] S6.8. Perform row-column independent decoding on the entity combination tensor and the relationship representation tensor to generate structured sentiment triples.

[0120] Specifically, the entity combination tensor is expanded in the row direction and the column direction respectively to obtain the corresponding row vector sequence and column vector sequence; then the row vector sequence and the column vector sequence are element-by-element multiplied with the relationship representation tensor respectively to obtain the row direction matching score tensor and the column direction matching score tensor; the row direction matching score tensor and the column direction matching score tensor are normalized respectively, for example, the score value is made to fall into the range of 0 to 1 through the normalization function; finally, according to the maximum score position in the row direction matching score tensor and the column direction matching score tensor, the corresponding entity combination tensor and the relationship representation tensor index are combined to determine the subject, object and emotion relationship in the emotion triplet to form a structured emotion triplet.

[0121] This embodiment also provides a computer device, which is suitable for the method of extracting aspect emotion triples from a large language model annotated dataset, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the method of extracting aspect emotion triples from a large language model annotated dataset proposed in the above embodiment.

[0122] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0123] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for extracting aspect sentiment triples from a large language model annotation dataset as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, disk or optical disk.

[0124] In summary, the present invention utilizes quantum emotional entanglement state calculation and a splitting mechanism to effectively model emotional conflicts in review texts. By precisely triggering the splitting operation of emotional states, it improves the accuracy of triple extraction in complex emotional scenarios and significantly reduces the extraction error rate caused by emotional polarity ambiguity. The atomic emotional paths generated by the splitting mechanism correspond to independent structured emotional triplets, effectively resolving the confusion problem in multi-polar emotional expressions, improving the refinement of emotional expressions, and enhancing the accuracy and usability of emotion recognition results in fields such as medical evaluation.

[0125] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for extracting aspect-sentiment triples from a large language model annotation dataset, characterized by: include, Collect and preprocess social media comment text to obtain a standardized text dataset; Perform initial triple annotation on the standardized text dataset and calculate the cognitive bias index based on the annotation reference set to generate a biased labeled dataset; A dynamic attenuation suppression method is used to generate a bias mask matrix, and exponential mask enhancement is performed on samples exceeding a preset bias threshold to obtain a bias mask enhanced representation. The bias mask enhanced representation is input into the differentiable grammar parser, the grammar dependency matrix is ​​calculated using the attention mechanism, and the topological structure of the grammar dependency matrix is ​​optimized by combining the structural entropy loss to generate the boundary decision matrix. Extract candidate triple embeddings based on the boundary decision matrix, calculate the contradiction coefficient using quantum emotional entanglement state, and generate a set of corrected triples; The corrected triple set is input into the multi-view graph neural network to fuse multi-view features to generate an optimized representation vector, which is mapped to the axial sparse attention table structure, and row and column independent decoding is used to generate structured emotion triples.

2. The method for extracting aspect-emotion triples from a large language model annotation dataset according to claim 1, wherein: The specific steps for obtaining the standardized text dataset are as follows: Collecting social media user comment content, timestamps, and user IDs, extracting comment fields to form social media comment text; The rule cleaning and regular expression matching method is used to denoise the social media comment text to obtain the cleaned comment text; The part-of-speech tagging method and unified coding standards are used to convert the format and regularize the language of the cleaned comment text to obtain a standardized text dataset.

3. The method for extracting aspect-emotion triples from a large language model annotation dataset according to claim 1, wherein: The specific steps for generating a biased labeled dataset are as follows: Based on the standardized text dataset, some samples are selected by random sampling, and triple annotation data is generated by combining the rule extraction method to construct the annotation reference set; Use the labeled reference set to perform supervised fine-tuning on the pre-trained large language model to obtain the fine-tuned large language model; The cleaned comment text of the standardized text dataset is input into the fine-tuned large language model, and the initial triples are annotated on the cleaned comment text using the sequence labeling method to obtain the initial triple set; Referring to the annotated reference set, the cognitive bias index of each triple sample in the initial triple set is calculated, and the high-bias samples are marked to generate a biased labeled dataset.

4. The method for extracting aspect-emotion triples from a large language model annotation dataset according to claim 1, wherein: A bias threshold is set according to the bias-labeled dataset. For the cognitive bias index exceeding the bias threshold, a dynamic attenuation suppression method is used to perform exponential mask enhancement processing. A bias mask matrix is ​​generated and fused with the representation of the cleaned comment text to obtain a bias mask enhanced representation.

5. The method for extracting aspect-emotion triples from a large language model annotation dataset according to claim 1, wherein: The specific steps of generating the boundary judgment matrix are as follows: Taking the bias mask enhanced representation as input, an optimization algorithm with a structured entropy loss function is used to train the differentiable grammar parser to obtain the trained differentiable grammar parser. Input the bias mask enhanced representation of the text to be processed into the trained differentiable grammar parser for calculation to obtain the grammatical dependency matrix; The structural entropy loss of the grammatical dependency matrix is ​​optimized, and the matrix topology is adjusted. According to the optimized grammatical dependency matrix and the preset boundary judgment rules, the boundary relationship between the candidate entity nodes is extracted to generate a boundary judgment matrix.

6. The method for extracting aspect-emotion triples from a large language model annotation dataset according to claim 1, wherein: The specific steps of generating the correction triple set are as follows: Extract the context embedding of candidate triples from the boundary decision matrix and construct the quantum sentiment state representation using word vector encoding method; The contradiction coefficient of quantum emotional state representation is calculated using the entanglement metric function and compared with the entanglement dissociation threshold; For the contradictory coefficients represented by quantum emotional states exceeding the entanglement dissociation threshold, the quantum state decomposition method is used to split them into multiple atomic emotional paths; The emotional contradiction detection and adjustment method is used to correct the emotional contradictions in the atomic emotional path and generate a set of correction triples.

7. The method for extracting aspect-emotion triples from a large language model annotation dataset according to claim 1, wherein: The specific steps of generating the optimized representation vector are as follows: Convert the corrected triples into graph-structured data, and construct a grammatical dependency view, a semantic enhancement view, and an emotional bridge view based on the graph-structured data. Extract the feature representations of nodes and edges in multiple views to obtain the initial feature set of multiple views. A supervised learning method is used to train the multi-view graph neural network, and the parameter weights of the node representation are optimized to obtain the trained multi-view graph neural network. The initial multi-view feature set is input into the trained multi-view graph neural network, and the feature representations of nodes and edges in the multi-views are weightedly fused using the attention mechanism to generate a comprehensive multi-view feature representation. Through iterative calculations of a multi-layer graph neural network, the comprehensive multi-view feature representations corresponding to neighbor nodes are aggregated layer by layer to generate an optimized representation vector.

8. The method for extracting aspect-emotion triples from a large language model annotation dataset according to claim 1, wherein: The specific steps of generating structured sentiment triples are as follows: Using the axial sparse attention method, the optimized representation vector is mapped into a two-dimensional representation tensor; Apply sparse attention to the two-dimensional representation tensor along the row axis to extract local context features between entities and obtain the entity combination tensor; Apply sparse attention to the entity combination tensor along the column axis to extract the weighted features of the relationship dimension and obtain the relationship representation tensor; The entity combination tensor and the relationship representation tensor are decoded independently in rows and columns to generate structured sentiment triples.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for extracting aspect emotion triples from a large language model annotated dataset according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for extracting aspect emotion triples from a large language model annotated dataset according to any one of claims 1 to 8 are implemented.

Images