A conversation text sentiment analysis method for supplier evaluation

Abstract

The present invention discloses a method for sentiment analysis of dialogue texts that can be used for supplier evaluation, and relates to the field of electric power operation and maintenance. Dialogue texts often have diverse syntactic forms, prominent colloquialism, rich semantic emotions, and emotional transitions, which brings difficulties to the sentiment analysis of dialogue texts, thereby affecting the accuracy of supplier evaluation. Based on the ontology dictionary and the general dictionary in the field of electric power, the present invention identifies keywords related to emotions, analyzes the semantic emotions of the basic unit and auxiliary unit of emotional evaluation, not only overcomes the difficulties of syntactic diversification, but also can identify the semantic emotional transitions of dialogue texts. Based on the results of the sentiment analysis, the sentiment evaluation value is calculated according to the established sentiment analysis evaluation model. The technical solution can solve the problems of diverse syntactic forms, prominent colloquialism, rich semantic emotions, and difficult-to-mine emotional transitions in dialogue texts, and realize a reliable supplier evaluation method based on sentiment analysis of dialogue texts.

Description

Technical Field

[0001] The present invention relates to the field of electric power operation and maintenance, and in particular to a conversation text sentiment analysis method that can be used for supplier evaluation. Background Art

[0002] In the process of evaluating power equipment suppliers, my country's power grid companies usually use contract performance, after-sales service and quality supervision as evaluation contents. This year, the State Grid Corporation of China added a satisfaction evaluation to the first supplier evaluation scoring rules in 2020, requiring project units to score suppliers based on quality, service and supply through manual statistics and direct calculation. However, this scoring method is not only inefficient, but its accuracy is greatly affected by the experience of the scorers and the scope of the evaluation materials. It may also cause the evaluation results to be distorted due to the supplier only paying attention to the quality of the evaluation content and neglecting other qualities. With the widespread use of communication software such as Tencent (Real Time eXpert, RTX) and QQ, power grid companies have gradually established work groups on various business platforms. When managers and technicians conduct technical exchanges in work groups through text (voice can be converted to text), they will mention the supplier's more fine-grained equipment quality, service and other information. This type of information comes from front-line personnel, has the characteristics of basic, extensive and high credibility, and can be used as one of the supporting data for evaluating equipment suppliers. At present, these texts have not been mined and utilized. The main difficulties are: dialogue texts often have diverse syntactic forms, prominent colloquialisms, rich semantic emotions, and emotional transitions, which makes it difficult to analyze the sentiment of dialogue texts, thus affecting the accuracy of supplier evaluation. Therefore, it is very important to accurately analyze the sentiment of power dialogue texts.

[0003] Power dialogue text features

[0004] 1) Contains irrelevant and redundant content. Since the text is generated in the form of a chat dialogue, the content of the dialogue includes not only the problems and solutions encountered in the work of the power field, as well as the tendency evaluation of product quality, but also irrelevant content, such as discussions on social news and industry prospects.

[0005] 2) The syntax is diverse and colloquial. The syntax of dialogue texts is often more personalized, free in form, without a fixed format, and mostly short sentences with a strong colloquial style, including a lot of Internet slang and common abbreviations.

[0006] 3) There is an implicit evaluation object. In the semantic context of the upper and lower sentences, the evaluation object may be omitted, such as in the continuous dialogue text "What should I do if the clock of the meter of manufacturer A is abnormal and there is a string of English?" "Their meters often have clock abnormalities." In the second sentence, "they" refers to "manufacturer A", so there is an implicit evaluation object.

[0007] 4) There are cross-interruption phenomena in the conversation. A multi-sentence conversation text is defined as a multi-round conversation, and a conversation text with only two sentences is defined as a single-round conversation. Semantically related content in a multi-round conversation is the same topic, and the conversation content discussing the same supplier is the same supplier topic. The characteristics of a multi-round conversation are that it is long in length and wide in span. Due to the high degree of freedom and flexible topics of chat conversations, there are often cross-conversation or interruption phenomena. For example, when several staff members discuss a meter failure problem, they will cross the chat content of other personnel, or interact through the link @user ID after a long interval of conversation. At this time, it is more difficult to divide the same topic of the multi-round conversation text.

[0008] 5) Semantic sentiment is rich and there are sentiment changes. The sentiment words in the dialogue text are rich and varied, the sentiment expression is more flexible, and there are sentiment changes. For example, "It was fine before the rotation meter, but it is not good after the replacement", which expresses two opposite emotions in one sentence. It is difficult to accurately mine the sentiment of this sentence.

[0009] Dialogue texts often have diverse syntactic forms, prominent colloquialisms, rich semantic emotions, and emotional transitions, which brings difficulties to the sentiment analysis of dialogue texts and thus affects the accuracy of supplier evaluations. Summary of the invention

[0010] The technical problem to be solved and the technical task proposed by the present invention are to improve and perfect the existing technical solutions and provide a dialogue text sentiment analysis method that can be used for supplier evaluation, so as to achieve the purpose of accurately analyzing the style and sentiment of power dialogue. To this end, the present invention adopts the following technical solutions.

[0011] A conversation text sentiment analysis method that can be used for supplier evaluation includes the following steps:

[0012] 1) Dialogue text preprocessing;

[0013] Obtain the dialogue text and perform dialogue text preprocessing based on the power field ontology dictionary and general dictionary, including word segmentation, part-of-speech retrieval extraction and annotation;

[0014] 2) Identify and categorize emotional keywords;

[0015] Find the keywords related to emotions in the sentence and classify them into categories, specifically:

[0016] 201) Based on the ontology dictionary of the power field, the single-sentence dialogue text after word segmentation and part-of-speech tagging is taken as the analysis object, and the event word Inc and the sentiment word Sen are set as the basic units of sentiment evaluation; the conjunction Conj, the degree adverb Ad and the negation word Neg are set as the auxiliary units of sentiment evaluation, which are used for the correction of subsequent text sentiment analysis; the basic unit and the auxiliary unit together constitute the main unit of sentiment evaluation of the dialogue text;

[0017] 202) In a single sentence dialogue text, when at least one of the basic unit event word Inc and the sentiment word Sen appears, the sentence is considered to have sentiment tendency and evaluation opinion, otherwise the sentence is considered to have no value in evaluating the equipment supplier; when only Inc appears in the sentence, it is counted in the event score; when only Sen appears in the sentence, it is counted in the sentiment score; when Inc and Sen coexist in the sentence, the sentiment score and event score are counted at the same time considering the priority; when neither Inc nor Sen exists in the sentence, the sentence is considered to be neutral and has no direct impact on the sentiment evaluation, and is deleted as a redundant sentence;

[0018] 3) Perform semantic sentiment analysis based on the established dialogue sentiment analysis rules;

[0019] The main units of sentiment evaluation in the dialogue text include: event word Inc, sentiment word Sen, conjunction Conj, degree adverb Ad and negation Neg; based on the logical relationship between the evaluation object and the main units, the sentiment keyword score is obtained: Among them, the dialogue sentiment analysis rule is:

[0020] 301) There is an event word Inc in the single sentence text;

[0021] The event word Inc is a highly professional word that is mostly used to describe negative events. The score is set according to the severity. i is the sentiment score of the ith event, which is set according to the severity of the event: general or severe;

[0022] 302) There is a sentiment word Sen in the single sentence text;

[0023] Sentiment word Sen refers to Bosonnlp open source computer semantic analysis and processing library to set scores according to positive and negative sentiments. j Sentiment score for the jth sentiment word;

[0024] 303) There is a degree adverb Ad in a single sentence text; the polarity coefficient is set according to the polarity strength of the degree adverb, Ad k is the sentiment polarity coefficient of the kth degree adverb; if there is no degree adverb, it is represented by Ad0 and the value is fixed;

[0025] 304) There is a negation word Neg in the single sentence text; determine the negation coefficient value according to whether there is a negation word, where Neg l is the negation coefficient of the lth negative word; if there is no negative word, it is represented by Neg0;

[0026] 305) When there are multiple event words Inc or sentiment words Sen in a sentence, it is necessary to judge the sentiment semantics and transition of the sentence; when there is a conjunction Conj, the sentiment semantic evaluation priority is judged according to the position relationship between the conjunction and the basic unit and the transition relationship of the conjunction, and the conjunction relationship semantic evaluation priority coefficient η is determined C ; Among them, the conjunction is the boundary, and the basic unit priority coefficient before the conjunction is η Cb , the priority coefficient of the basic unit after the conjunction is η Ca ;

[0027] 306) When there are both event words Inc and sentiment words Sen in a sentence, the event words, as objective fact-statement words, have higher evaluation credibility than subjective sentiment words. The relationship evaluation priority coefficient η between the basic units of event words Inc and sentiment words Sen is Y are η YI =1.2 and η YS =0.8;

[0028] 307) For each basic unit in the sentence, combining the above methods 305) and 306), its evaluation priority coefficient η is: η = η C ×η Y ; where η C is the semantic evaluation priority coefficient of conjunction relationship, η Y is the priority coefficient for evaluating the basic unit relationship; according to the basic unit type, η I Indicates the priority coefficient of the event word; η S Indicates the priority coefficient of sentiment words;

[0029] 308) The dependency relationship between the basic unit event word Inc and the emotion word Sen and the degree adverb Ad and the negation word Neg. If there is a conjunction Conj, the conjunction is used as the dividing line. If there is no conjunction, each basic unit is used as the dividing line.

[0030] 4) Calculate the sentiment evaluation value according to the established sentiment analysis evaluation model;

[0031] For the conversation set with the same topic, sentiment analysis evaluation is given by combining the relationship characteristics of sentiment evaluation basic unit and auxiliary unit; each basic unit is directly affected by the unit priority coefficient, degree adverb and negation coefficient, and the basic unit score is directly modified by the series product of coefficients; the basic units are independent of each other and connected by parallel addition; for the same evaluation object supplier under the same topic, the sentiment score value in sentence units is calculated. The positive or negative value of the value directly reflects the positive or negative evaluation of the interlocutor, and the size of the value reflects the intensity of the emotion; thus, the sentiment analysis evaluation model of a single sentence is obtained:

[0032]

[0033] Among them, S MANu The sentiment evaluation value of the supplier score for the u-th single sentence; I represents the number of event words in the sentence; η Ii represents the priority coefficient of the ith event word, which is calculated by 307) in step 3; K represents the number of degree adverbs attached to each event word; L represents the number of negative words attached to each event word; J represents the number of sentiment words in the sentence; η Sj represents the priority coefficient of the jth sentiment word, which is also calculated by 307) in step 3; P represents the number of degree adverbs attached to each sentiment word; Q represents the number of negative words attached to each sentiment word; wherein k, l, p, q values ​​start from 0, indicating that when there is no corresponding auxiliary unit, the coefficient takes its initial value 1;

[0034] The total sentiment evaluation value of the supplier is divided into the sum of all single sentence scores: That is, the overall sentiment analysis evaluation model;

[0035] Among them, S MAN is the sentiment evaluation value of the total suppliers; U is the number of evaluation sentences;

[0036] 5) Based on future language development and following natural laws, the dictionary will be expanded regularly and the evaluation model will be verified and revised;

[0037] Obtain multiple power dialogue texts, and use the sentiment analysis evaluation model to perform evaluation calculations and manual analysis to obtain the corresponding theoretical evaluation values ​​and actual evaluation values; compare the theoretical evaluation values ​​and the actual evaluation values. When the difference between the two exceeds the set threshold, return to step 3) to revise the sentiment analysis evaluation model and expand the dictionary; if not, the sentiment analysis evaluation model is considered reasonable.

[0038] As a preferred technical means: In step 301), Inc i is the sentiment score of the ith event, which takes -1 or -2 according to the severity of the general and serious events respectively.

[0039] As a preferred technical means: in step 303), the polarity coefficient is set to 0.5, 1, and 2 according to the polarity strength of the degree adverb. k is the sentiment polarity coefficient of the kth degree adverb; if there is no degree adverb, it is represented by Ad0, and Ad0=1.

[0040] As a preferred technical means: in step 304), when there is a negative word, the negative coefficient is taken as -1, if there is no negative word, it is represented by Neg0, and Neg0=1;

[0041] As a preferred technical means: in step 305), for conjunctions between multiple basic units, the priorities on both sides are set as follows:

[0042] Conjunctions that indicate transitions and causal relationships, including "but" and "so", have a priority coefficient for the evaluation of the basic units on both sides of the conjunction: η Cb =0.8,η Ca =1.2;

[0043] Conjunctions that indicate parallel and continuation relations, including "and" and "then", the evaluation priority coefficient of the basic units on both sides of the conjunction is: η Cb =η Ca =1;

[0044] Conjunctions that indicate concessional relations include "although" and "although". The evaluation priority coefficients of the basic units on both sides of the conjunction are: η Cb =1.2,η Ca =0.8; if there is no conjunction, then for all basic units, η C =1.

[0045] As a preferred technical means: in step 306), the relationship evaluation priority coefficient η between the event word Inc and the sentiment word Sen basic unit Y are η YI and η YS .

[0046] As a preferred technical means: before performing semantic sentiment analysis, a dialogue interruption cross-processing step is also included. When performing dialogue interruption cross-processing, the steps include:

[0047] A1) obtaining a dialogue style set D;

[0048] A2) Sequentially extract the two dialogue texts d in D i and d j ;

[0049] A3) Judge d i and d jWhether the single-round dialogue text turn interval condition is met; if so, proceed to step A4); if not, proceed to step A5);

[0050] A4) Calculate the single-round dialogue relevance matching degree of the two texts;

[0051] A5) Determine whether there is a link. If so, remove the link statement d i Calculate the relevance matching degree of the two conversation texts with the closest interval between the linked users in turn; if not, it is considered that the topic conversation set to which the text belongs has been summarized, and then go to step A7);

[0052] A6) Summarizing the conversations on the same topic based on the matching results;

[0053] A7) Determine whether the dialogue text set D is empty. If not, proceed to step A2); if so, end the process to obtain a set of all dialogues with the same topic.

[0054] As a preferred technical means: when it is necessary to evaluate suppliers, after obtaining the same subject conversation set, the supplier identification is performed, and the supplier information in each subject conversation set is extracted according to the supplier information category in the power business ontology dictionary, and the implicit evaluation object is identified using the upward nearest principle, and then irrelevant redundant subject content is removed;

[0055] Supplier identification uses the following three rules:

[0056] B1) If the supplier information is not identified, the conversation set is considered to be irrelevant and redundant, which is of no value to the evaluation of the equipment supplier and can be screened out;

[0057] B2) If one or more identical supplier information is identified, the evaluation object of the conversation set is considered to be the identified supplier;

[0058] B3) Two or more different supplier information appears, such as manufacturer A, B, ..., X in order of appearance; for each text in the set, the corresponding supplier is determined by the principle of upward proximity, and the evaluation object of the conversation set from the first sentence to the sentence before manufacturer B appears is defined as manufacturer A, the evaluation object from the sentence before manufacturer B appears to the sentence before manufacturer C appears is manufacturer B, and so on; if a manufacturer appears repeatedly, the conversation set of the manufacturer is merged;

[0059] At this point, the topic text summary of each supplier is completed.

[0060] As the preferred technical means: it also includes the step of expanding the power ontology dictionary. The expansion of the power ontology dictionary adopts a semi-supervised word segmentation method. On the basis of the power equipment defect domain ontology dictionary and the general dictionary, the conversation power business domain text is used as the corpus to perform word segmentation based on the HMM model, and new domain ontology vocabulary is selected according to the word frequency sorting. It is manually checked to determine whether it becomes an ontology word and a synonym, and the ontology dictionary is supplemented. As the corpus continues to increase, the above steps are repeated for multiple corrections to achieve the supplementation and enrichment of the power domain ontology dictionary. The newly added ontology dictionary includes supplier names and business domain events. The supplier names and their commonly used abbreviations are expanded to cope with the prominent characteristics of the conversation text in spoken language.

[0061] Beneficial effects: This technical solution targets the content of dialogue texts in the power field, identifies keywords related to emotions, and overcomes the difficulty of syntactic diversity by analyzing the semantic emotions of basic units and auxiliary units of emotional evaluation, and identifies the semantic emotional transitions of dialogue texts; it solves the problems of diverse syntactic forms, prominent colloquialism, rich semantic emotions, and the existence of emotional transitions that make it difficult to conduct emotional mining in dialogue texts, establishes a reliable emotional analysis and evaluation model, and realizes accurate emotional analysis and evaluation of dialogue texts. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a flow chart of the present invention.

[0063] Figure 2 It is a supplier evaluation flow chart of the present invention.

[0064] Figure 3 It is a flow chart of the dialogue interruption cross-processing of the present invention. DETAILED DESCRIPTION

[0065] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings.

[0066] Embodiment 1:

[0067] like Figure 1 The present invention shown comprises the following steps:

[0068] 1) Dialogue text preprocessing;

[0069] Obtain the dialogue text and perform dialogue text preprocessing based on the power field ontology dictionary and general dictionary, including word segmentation, part-of-speech retrieval extraction and annotation;

[0070] 2) Identify and categorize emotional keywords;

[0071] Find the keywords related to emotions in the sentence and classify them into categories, specifically:

[0072] 201) Based on the ontology dictionary of the power field, the single-sentence dialogue text after word segmentation and part-of-speech tagging is taken as the analysis object, and the event word Inc and the sentiment word Sen are set as the basic units of sentiment evaluation; the conjunction Conj, the degree adverb Ad and the negation word Neg are set as the auxiliary units of sentiment evaluation, which are used for the correction of subsequent text sentiment analysis; the basic unit and the auxiliary unit together constitute the main unit of sentiment evaluation of the dialogue text;

[0073] 202) In a single sentence dialogue text, when at least one of the basic unit event word Inc and the sentiment word Sen appears, the sentence is considered to have sentiment tendency and evaluation opinion, otherwise the sentence is considered to have no value in evaluating the equipment supplier; when only Inc appears in the sentence, it is counted in the event score; when only Sen appears in the sentence, it is counted in the sentiment score; when Inc and Sen coexist in the sentence, the sentiment score and event score are counted at the same time considering the priority; when neither Inc nor Sen exists in the sentence, the sentence is considered to be neutral and has no direct impact on the sentiment evaluation, and is deleted as a redundant sentence;

[0074] 3) Perform semantic sentiment analysis based on the established dialogue sentiment analysis rules;

[0075] The main units of sentiment evaluation in the dialogue text include: event word Inc, sentiment word Sen, conjunction Conj, degree adverb Ad and negation Neg; based on the logical relationship between the evaluation object and the main units, the sentiment keyword score is obtained: Among them, the dialogue sentiment analysis rule is:

[0076] 301) There is an event word Inc in the single sentence text;

[0077] The event word Inc is a highly professional word that is mostly used to describe negative events. The score is set according to the severity. i is the sentiment score of the ith event, which is -1 or -2 according to the severity of the general and serious events respectively;

[0078] 302) There is a sentiment word Sen in the single sentence text;

[0079] Sentiment word Sen refers to Bosonnlp open source computer semantic analysis and processing library to set scores according to positive and negative sentiments. j Sentiment score for the jth sentiment word;

[0080] 303) There is a degree adverb Ad in a single sentence text; according to the polarity strength of the degree adverb, the polarity coefficient is set to 0.5, 1, and 2. Ad k is the sentiment polarity coefficient of the kth degree adverb; if there is no degree adverb, it is represented by Ad0, and Ad0=1;

[0081] 304) There is a negative word Neg in the single sentence text; the negation coefficient value is determined according to whether there is a negative word. If there is a negative word, the negation coefficient is -1, where Neg l is the negation coefficient of the lth negative word; if there is no negative word, Neg0 is used to represent it, and Neg0=1;

[0082] 305) When there are multiple event words Inc or sentiment words Sen in a sentence, it is necessary to judge the sentiment semantics and transition of the sentence; when there is a conjunction Conj, the sentiment semantic evaluation priority is judged according to the position relationship between the conjunction and the basic unit and the transition relationship of the conjunction, and the conjunction relationship semantic evaluation priority coefficient η is determined C ; Among them, the conjunction is the boundary, and the basic unit priority coefficient before the conjunction is η Cb , the priority coefficient of the basic unit after the conjunction is η Ca ;

[0083] Conjunctions that indicate transitions and causal relationships, including "but" and "so", have a priority coefficient for the evaluation of the basic units on both sides of the conjunction: η Cb =0.8,η Ca =1.2;

[0084] Conjunctions that indicate parallel and continuation relations, including "and" and "then", the evaluation priority coefficient of the basic units on both sides of the conjunction is: η Cb =η Ca =1;

[0085] Conjunctions that indicate concessional relations include "although" and "although". The evaluation priority coefficients of the basic units on both sides of the conjunction are: η Cb =1.2,η Ca =0.8; if there is no conjunction, then for all basic units, η C =1;

[0086] 306) When there are both event words Inc and sentiment words Sen in a sentence, the event words, as objective fact-statement words, have higher evaluation credibility than subjective sentiment words. The relationship evaluation priority coefficient η between the basic units of event words Inc and sentiment words Sen is Y are η YI =1.2 and η YS =0.8;

[0087] 307) For each basic unit in the sentence, combining the above methods 305) and 306), its evaluation priority coefficient η is: η = η C ×η Y ; where η C is the semantic evaluation priority coefficient of conjunction relationship, η Y It is the priority coefficient for evaluating the relationship between basic units;

[0088] 308) The dependency relationship between the basic unit event word Inc and the emotion word Sen and the degree adverb Ad and the negation word Neg. If there is a conjunction Conj, the conjunction is used as the dividing line. If there is no conjunction, each basic unit is used as the dividing line.

[0089] 4) Calculate the sentiment evaluation value according to the established sentiment analysis evaluation model;

[0090] For the conversation set with the same topic, sentiment analysis evaluation is given by combining the relationship characteristics of sentiment evaluation basic unit and auxiliary unit; each basic unit is directly affected by the unit priority coefficient, degree adverb and negation coefficient, and the basic unit score is directly modified by the series product of coefficients; the basic units are independent of each other and connected by parallel addition; for the same evaluation object supplier under the same topic, the sentiment score value in sentence units is calculated. The positive or negative value of the value directly reflects the positive or negative evaluation of the interlocutor, and the size of the value reflects the intensity of the emotion; thus, the sentiment analysis evaluation model of a single sentence is obtained:

[0091]

[0092] Among them, S MANu The sentiment evaluation value of the supplier score for the u-th single sentence; I represents the number of event words in the sentence; η Ii represents the priority coefficient of the ith event word, which is calculated by 307) in step 3; K represents the number of degree adverbs attached to each event word; L represents the number of negative words attached to each event word; J represents the number of sentiment words in the sentence; η Sj represents the priority coefficient of the jth sentiment word, which is also calculated by 307) in step 3; P represents the number of degree adverbs attached to each sentiment word; Q represents the number of negative words attached to each sentiment word; wherein k, l, p, q values ​​start from 0, indicating that when there is no corresponding auxiliary unit, the coefficient takes its initial value 1;

[0093] The total sentiment evaluation value of the supplier is divided into the sum of all single sentence scores: That is, the overall sentiment analysis evaluation model;

[0094] Among them, S MAN is the sentiment evaluation value of the total suppliers; U is the number of evaluation sentences;

[0095] 5) Regularly verify and correct the sentiment analysis and evaluation model;

[0096] Obtain multiple power dialogue texts, and use the sentiment analysis evaluation model to perform evaluation calculations and manual analysis to obtain the corresponding theoretical evaluation values ​​and actual evaluation values; compare the theoretical evaluation value and the actual evaluation value, and when the difference between the two exceeds the set threshold, return to step 3) to correct the sentiment analysis evaluation model; if not, the sentiment analysis evaluation model is considered reasonable.

[0097] This technical solution targets the content of dialogue texts in the power field, identifies keywords related to emotions, and overcomes the difficulty of syntactic diversity by analyzing the semantic emotions of basic units and auxiliary units of emotional evaluation, and identifies the semantic and emotional transitions of dialogue texts. It solves the problems of diverse syntactic forms, prominent colloquialism, rich semantic emotions, and difficulty in emotion mining in dialogue texts. It establishes a reliable emotion analysis and evaluation model to achieve accurate emotion analysis and evaluation of dialogue texts.

[0098] Embodiment 2:

[0099] The same parts as the first embodiment will not be described in detail, except that:

[0100] like Figure 2 As shown, in order to more accurately perform semantic sentiment analysis and evaluate suppliers, this embodiment also includes power ontology dictionary expansion, single-round dialogue correlation analysis, dialogue interruption cross processing, and supplier identification; wherein single-round dialogue correlation analysis, dialogue interruption cross processing, and supplier identification are dialogue text preprocessing steps. Specifically:

[0101] First, the expansion of the power ontology dictionary is studied. Due to the strong professional characteristics of the power dialogue text, there are certain differences from the common vocabulary. In order to improve the accuracy of text understanding, it is necessary to establish an ontology dictionary in the dialogue business field and add attribute entries to the dictionary, including power-specific vocabulary, supplier name vocabulary and event key vocabulary, to lay the foundation for subsequent topic induction and sentiment analysis evaluation. This technical solution adopts a semi-supervised word segmentation method. On the basis of the initially established power equipment defect field ontology dictionary and general dictionary, the dialogue power business field text is used as the corpus for word segmentation based on the HMM model, and new domain ontology vocabulary is selected according to the word frequency sorting. It is manually checked to determine whether it becomes an ontology word and a synonym, and the ontology dictionary is supplemented. With the continuous increase of the corpus, the above steps are repeated for multiple corrections to achieve the supplementation and enrichment of the power field ontology dictionary. The newly added ontology dictionary mainly includes vocabulary such as supplier name and business field event, and the supplier name and its commonly used abbreviation are expanded to cope with the prominent characteristics of the dialogue text.

[0102] Then, the topic induction study is conducted. Due to the flexibility of multi-round dialogue topics, it is difficult to obtain accurate emotional tendencies towards suppliers. It is necessary to decompose multi-round dialogues into multiple independent dialogue topics. Whether the semantics are related can be judged by the coherence of the previous and next sentences. Therefore, the multi-round dialogues are decomposed into multiple single-round dialogues for correlation judgment, which is an important basis for summarizing the same topic of the dialogue text. However, the irrelevance of the correlation judgment of a single-round dialogue does not necessarily mark the end of the topic. It may be due to the existence of dialogue interruption and intersection, so the dialogue topic induction should be carried out on the basis of meeting the dialogue interruption and intersection processing rules. In the case of multiple suppliers or no suppliers on the same topic, it is necessary to divide the topic for suppliers, identify the implicit evaluation objects in the dialogue set, and remove the redundant content of irrelevant suppliers in the dialogue.

[0103] Finally, we conducted a study on evaluation modeling based on sentiment analysis. After the supplier theme is summarized, sentiment analysis is to identify sentiment-related keywords, analyze the semantic sentiment of the basic and auxiliary units of sentiment evaluation, overcome the difficulties of syntactic diversity, and identify the semantic sentiment transition of the dialogue text. Finally, based on the results of sentiment analysis, we obtained a supplier evaluation model based on the dialogue text.

[0104] The following is a further description of some steps:

[0105] 1. Single-round dialogue correlation analysis based on BERT-NSP and cosine similarity weighting

[0106] In order to facilitate the classification and analysis of multi-round dialogue texts on the same topic, the multi-round dialogue texts are first broken down into single-round dialogue texts with only two sentences. The next sentence is predicted based on the single-round dialogue, and dialogue interruption cross-processing and supplier identification are performed based on this.

[0107] BERT-NSP is a model trained with Wikipedia corpus for judging the coherence of sentences, and has achieved good results in many natural language processing tasks. BERT has better performance on the GLUE benchmark dataset than algorithms such as BiLSTM and OpenAI GPT. At the same time, the performance of BERT-NSP and training with LTR, LTR+BiLSTM and other algorithms without NSP was compared, and the optimal training result of BERT-NSP was obtained. This technical solution adopts a fine-tuning model for sentence coherence prediction based on BERT-NSP. This model is a model that fine-tunes the training parameters during supervised training based on the BERT pre-training model. The structure is as follows: Figure 2 shown.

[0108] The model takes two dialogue texts as input, adds the first token [CLS], and inserts the token [SEP] between the two dialogue texts as a special separator; each Tok in the input layer is the vectorization of each word in the input sentence, which is obtained by adding the three word embedding methods of tag word embedding, segment word embedding, and position word embedding to obtain a one-to-one corresponding E [CLS] , E1,…,E N , E [SEP] , E1',…,E M ', passed into the bidirectional Transformer layer, which transforms it and outputs the hidden vector C, T1, ..., T corresponding to each word N , T [SEP] , T1', ..., T M '. The model output is shown in the following formula:

[0109] p = softmax(CW T ) (1)

[0110] Among them, p is the next sentence prediction matching probability matrix; C is the final hidden state of the first flag [CLS] of the BERT model; W is the fully connected layer weight matrix. This model is actually a binary classification problem, so p is a two-dimensional vector, which represents the probability values ​​of the next sentence prediction as 0 and 1, that is, the probability of irrelevant and relevant, and the next sentence prediction probability P NS Take the value in the vector that represents the correlation between the two sentences.

[0111] from Figure 2 From the process, we can see that the BERT-NSP model predicts deep features based on characters, which invisibly ignores word-level features. Cosine similarity, which is more commonly used in text analysis, has the factor of exploring the correlation between two sentences at the word level. For this reason, we can build a next sentence prediction algorithm based on BERT-NSP and cosine similarity weighting, with the aim of integrating the advantages of BERT-NSP deep distribution analysis and cosine similarity word-level analysis, thereby improving the accuracy of next sentence prediction of dialogue text.

[0112] When the cosine similarity method is used to judge the coherence of a conversation text, the content that appears repeatedly in the text is often used as a coherence judgment criterion. Therefore, the single-round conversation text is segmented and the cosine similarity of adjacent conversations can be calculated using formula (2):

[0113]

[0114] Where S is the cosine similarity of adjacent conversations; A = (a1, a2, ..., a n ) and B=(b1,b2,…,b n) are the n-dimensional word frequency feature vectors obtained after the word frequency vectorization of the two sentences. Based on the deep features and language similarity features, this technical solution defines the semantic relevance matching degree of a single round of dialogue, and uses formula (3) to fuse these two types of features:

[0115] M=(1-α)P NS +αS (3)

[0116] Among them, M is the semantic relevance matching degree of a single round of dialogue; α is the cosine similarity weight coefficient. The M value is a number greater than or equal to 0. The larger the M, the greater the matching relevance of the two sentences. NS The binary classification value standard of α is that if the value is greater than or equal to 0.5, the upper and lower sentences are considered to be related and the two sentences are classified as the same conversation topic. If the value is less than 0.5, they are considered to be unrelated. Its significance lies in that it can integrate deep features and similarity features, comprehensively consider the language connection between the upper and lower sentences, and improve the accuracy of matching judgment. The role of α is to balance the weight ratio of deep features and similarity features. Optimizing this coefficient can obtain the optimal model for single-round conversation text judgment.

[0117] 2. Dialogue Interruption Cross-Processing

[0118] In the electric power dialogue text, there are often topic interruptions and crossovers in multiple rounds of dialogue text. In order to achieve dialogue topic induction based on the next sentence matching of a single round of dialogue, Figure 3 The dialog shown interrupts the cross-processing flow.

[0119] Suppose two texts d are taken sequentially from the dialogue text set D i and d j , judge the interval between two texts. According to the experience of dialogue text research, it is believed that dialogues with intervals beyond 3 rounds generally have no direct correlation. If the two texts are within the allowed interval range, the single-round dialogue correlation matching degree M value of the two texts is judged; if the two texts are not within the allowed interval range, the d i Check the link @ user ID information. If there is link information, the link statement d i The two conversation texts with the closest interval between the linked users are judged for the relevance matching degree M value respectively, and the conversation sets with the same topic are summarized according to the matching degree judgment results; if there is no link information, it is considered that the conversation set with the topic to which the text belongs has been summarized. Continue to repeat the recursive call according to the above steps until the conversation text set D is empty, and all conversation sets with the same topic are obtained.

[0120] 3. Supplier Identification Rules

[0121] After obtaining the same topic dialogue set, the supplier information in each topic dialogue set is extracted according to the supplier information category in the power business ontology dictionary, and the implicit evaluation object is identified using the upward nearest principle, and then irrelevant redundant topic content is removed. The following three rules are used specifically:

[0122] 1) If the supplier information is not identified, the conversation set is considered to be irrelevant and redundant, which is of no value to the evaluation of the equipment supplier and can be screened out;

[0123] 2) If one or more identical supplier information is identified, the evaluation object of the conversation set is considered to be the identified supplier;

[0124] 3) Two or more different supplier information appears, such as manufacturer A, B, ..., X in order of appearance. For each text in the set, the corresponding supplier is determined by the principle of upward proximity. The evaluation object of the conversation set from the first sentence to the sentence where manufacturer B appears (not included) is manufacturer A, the evaluation object from the sentence where manufacturer B appears to the sentence where manufacturer C appears (not included) is manufacturer B, and so on. If a manufacturer appears repeatedly, the conversation set of that manufacturer is merged.

[0125] At this point, the topic text summary of each supplier is completed.

[0126] The following example analysis verifies this technical solution:

[0127] 1. Example Corpus and Ontology Dictionary Expansion

[0128] The verification study was conducted by taking the dialogue text in the field of power collection and operation as an example. First, a corpus was established, and a total of 23.8M text data including the RTX work group chat dialogue of power collection and operation, the "A Handbook of Abnormal Operations for Frontline Employees of Power Grid Enterprises", and abnormal collection text were collected. Then, the corpus was segmented, sorted and adjusted based on the hidden Markov model (HMM). This process is based on the general dictionary and the existing power equipment defect ontology dictionary, identifying new ontology words, and supplementing them after manual verification to form a new domain ontology dictionary. The ontology words are annotated with entry attributes and synonyms and antonyms, and the attributes include proprietary domain nouns, supplier names, and event key words. A total of 752 new collection and operation domain ontology dictionaries were added, including 539 proprietary domain noun entries that appeared in the collection and operation corpus but were not included in the original ontology dictionary, 106 supplier name words, and 107 event key words.

[0129] 2. Verification of the Theme Induction Method

[0130] The dataset selected 347 pairs of single-round dialogue texts from the RTX work group chat dialogue texts of power collection and operation. The BERT-NSP model parameters use the BERT-Chinese pre-trained model structure with 12 layers of Transformer units and 110M parameters, the self-attention mechanism is 12 heads, the hidden layer dimension is 768 dimensions, the maximum sequence length is 128, the learning rate is 3e-5, and the batch size is 32. Based on BERT-NSP and cosine similarity weighted single-round dialogue next sentence prediction analysis, experiments were conducted on the topic dialogue content of 347 pairs of power suppliers, and the cosine similarity weight coefficient α in the single-round dialogue semantic relevance matching degree M was optimized. When α is 0 and 1, it represents the BERT-NSP model and the cosine similarity model respectively. The single-round dialogue judgment accuracy is 80.69% at most when α is 0.04. After exceeding 0.04, the accuracy decreases monotonically. The model accuracy indicators are shown in Table 1. Therefore, the following BERT-NSP and cosine similarity weighted model takes the cosine similarity weight coefficient α as 0.04.

[0131] Table 1. Single-round dialogue text judgment accuracy

[0132] Model Single-round dialogue judgment accuracy / % <![CDATA[Cosine similarity (α＝1) > 57.35 <![CDATA[BERT-NSP (α＝0) ]]> 78.67 <![CDATA[BERT-NSP and Cosine Similarity Weighting (α＝0.04) > 80.69

[0133] Single-round dialogue correlation analysis uses cosine similarity, BERT-NSP, or BERT-NSP and cosine similarity weighted model to perform single-round dialogue correlation analysis. As shown in Table 1, the model of this technical solution is based on BERT-NSP to judge the correlation of two sentences' deep features and weighted with language feature similarity, which can improve the accuracy of single-round dialogue judgment. The establishment of this weighted model also has a certain degree of interpretability. The same text content in the dialogue process is more likely to be a discussion of the same topic.

[0134] Based on the prediction and analysis of the next sentence of a single-round conversation text, the cross-interruption of the conversation is handled. Taking the power conversation text as an example, this conversation contains two conversation topics, namely the discussion topic of the abnormal meter supplier and the discussion topic of the ceramic cup cost performance. The comparison of the multi-round conversation topic division obtained after the cross-interruption of the conversation is compared through no processing, cosine similarity, BERT-NSP, and BERT-NSP and cosine similarity weighted model is shown in Table 2. In the case of the same supplier identification, different conversation topics are divided by slashes. As can be seen from Table 2, in terms of multi-round conversation topic division, the model using only cosine similarity calculation has the lowest topic division accuracy, and the weighted model has the highest accuracy. In comparison, the cosine similarity model mainly focuses on the text repetition rate between two sentences of dialogue, and only relies on the repeated content in the two sentences to judge coherence, ignoring the internal connection. Although the BERT model alone can achieve a high accuracy, it still has incomplete division. According to the analysis of the 4th, 7th, and 13th sentences with incorrect divisions, it can be seen that the two sentences can be judged as the same topic through the same keyword "ceramic cup". Therefore, the weighted model combines the advantages of the above two models, overcomes the deficiency of the BERT model that ignores word-level connections, and more accurately divides the scope of each topic. From the dialogue set matching results of the weighted model, it can be seen that the weighted model can be used to match the word-level connections of the two sentences. Figure 3 The dialogue cross-interruption processing flow divides the content of the same topic. Not only can the same topic be correctly divided if the interval between dialogues is within 3 sentences, but also the long-distance dialogues linked by @user ID, such as sentences 4 and 13, can be accurately divided. In addition, the supplier identification rule can correctly identify the corresponding manufacturer information under the same topic and delete redundant dialogues from the dialogue topic.

[0135] Table 2. Topic division of multi-round dialogues in different models

[0136]

[0137]

[0138] 3. Sentiment Analysis and Supplier Evaluation Verification

[0139] In order to demonstrate the process of sentiment analysis of dialogue text, the sentence "Manufacturer A used to be good, but now there are frequent clock anomalies?" is used as an example for keyword extraction. The text is preprocessed based on the ontology dictionary and general dictionary in the field of power acquisition and operation and maintenance, and the words are segmented and keyword extraction and annotation are performed, as shown in Table 3 below.

[0140] Table 3 Keywords for sentiment analysis of conversation text

[0141]

[0142] This sentence has two basic units, the sentiment word "good" and the event word "clock abnormal", which are connected by the conjunction "but", and the auxiliary units "originally", "quite" and "often" modify them respectively. According to formula (5), for the sentiment word "good" and the event word "clock abnormal", their priority coefficient η is the priority coefficient of the event word η YS <η YI and the conjunction "but" priority η Cb <η Ca Under the priority judgment condition, the two are in a parallel addition relationship; the sentiment word "good" is affected by the degree adverbs "originally" and "quite"; "clock abnormality" is also directly affected by the degree adverb "often", and each has a series product relationship.

[0143] On this basis, a horizontal comparison was conducted, taking semantic transition, semantic extension and ordinary single sentences as examples to test the accuracy of the judgment of semantic transition by the method of this technical solution, and the analysis and scoring results were obtained as shown in Table 4. There are positive evaluations "good" and event words "clock abnormality" in the semantic transition sentences, and negative scores are obtained through sentiment analysis; positive semantic extension sentences have higher scores based on the positive single sentence scores. Therefore, the method of this technical solution can achieve relatively accurate scoring for semantically rich dialogue texts, and can also show obvious differences in scoring for content containing semantic transitions.

[0144] Table 4. Conversation text sentiment analysis semantic comparison scores

[0145]

[0146] Table 5 Dialogue text sentiment analysis vendor ratings

[0147]

[0148] According to the division results, sentiment analysis is performed to verify the supplier evaluation effect, as shown in Table 5. It can be seen that the scoring results obtained by BERT-NSP and BERT-NSP with cosine similarity weighting are similar, which is quite different from the scores under the unprocessed and cosine similarity calculation models. For the scoring of manufacturer A, the scores of BERT-NSP and the weighted model are closer to the degree of negative sentiment expressed in the text. The unprocessed model only selects sentences containing information about manufacturer A as the scoring object, and other related sentences are missing. The cosine similarity model also lacks some information, resulting in poor scoring results; for the scoring of manufacturer B, since only one conversation is involved, the scoring is not affected. It can be seen that the supplier evaluation results based on sentiment analysis are greatly affected by the correctness of topic division. The accuracy of the weighted model in topic division can effectively improve the authenticity and reliability of supplier evaluation.

[0149] According to the characteristics of power dialogue text, single-round dialogue correlation analysis was performed based on BERT-NSP and cosine similarity weighted model, and multi-round dialogue topic division based on multi-round dialogue cross-interruption was constructed, and supplier topics were identified. On this basis, a supplier evaluation model based on sentiment analysis was established. A large number of examples show that this technical solution method is superior in multi-round dialogue topic division, and the supplier evaluation model is reliable, which provides a method for the normalization of supplier evaluation.

[0150] In addition, the evaluation model of power equipment suppliers based on text mining technology has good transplantation capabilities. This technical solution uses the dialogue text of power acquisition and operation and maintenance business as an example to supplement the ontology dictionary and establish the evaluation model. It can be combined with other power business text corpora for corresponding transplantation research. It not only provides new ideas for the in-depth use of business dialogue texts, but also provides new support for the satisfaction evaluation of power grid enterprise equipment suppliers, thereby helping power grid enterprises to more objectively understand the equipment quality and service capabilities of suppliers and make reasonable decisions. At the same time, it is beneficial for equipment suppliers to improve their shortcomings in a targeted manner and enhance their competitiveness.

[0151] above Figure 1-3 The method for sentiment analysis of conversation texts that can be used for supplier evaluation is a specific embodiment of the present invention, which has embodied the substantial characteristics and progress of the present invention. Based on actual use needs and under the guidance of the present invention, equivalent modifications in shape, structure, etc. can be made to the method, which are all within the protection scope of this solution.

Claims

1. A conversation text sentiment analysis method that can be used for supplier evaluation, characterized by The following steps are involved: 1) Dialogue text preprocessing; Obtain the dialogue text and perform dialogue text preprocessing based on the power field ontology dictionary and general dictionary, including word segmentation, part-of-speech retrieval extraction and annotation; 2) Identify and categorize emotional keywords; Find the keywords related to emotions in the sentence and classify them into categories, specifically: 201) Based on the ontology dictionary of the power field, the single-sentence dialogue text after word segmentation and part-of-speech tagging is taken as the analysis object, and the event word Inc and the sentiment word Sen are set as the basic units of sentiment evaluation; the conjunction Conj, the degree adverb Ad and the negation word Neg are set as the auxiliary units of sentiment evaluation, which are used for the correction of subsequent text sentiment analysis; the basic unit and the auxiliary unit together constitute the main unit of sentiment evaluation of the dialogue text; 202) In a single sentence dialogue text, when at least one of the basic unit event word Inc and the sentiment word Sen appears, the sentence is considered to have sentiment tendency and evaluation opinion, otherwise the sentence is considered to have no value in evaluating the equipment supplier; when only Inc appears in the sentence, it is counted in the event score; when only Sen appears in the sentence, it is counted in the sentiment score; When Inc and Sen coexist in a sentence, both the sentiment score and the event score are calculated taking into account the priority; When Inc and Sen do not exist in a sentence, the sentence is considered neutral and has no direct impact on sentiment evaluation, and is deleted as a redundant sentence; 3) Perform semantic sentiment analysis based on the established dialogue sentiment analysis rules; The main units of sentiment evaluation in the dialogue text include: event word Inc, sentiment word Sen, conjunction Conj, degree adverb Ad and negation Neg; based on the logical relationship between the evaluation object and the main units, the sentiment keyword score is obtained: Among them, the dialogue sentiment analysis rule is: 301) There is an event word Inc in the single sentence text; The event word Inc is a highly professional word that is mostly used to describe negative events. The score is set according to the severity. i is the sentiment score of the ith event, which is set according to the severity of the event: general or severe; 302) There is a sentiment word Sen in the single sentence text; Sentiment word Sen refers to Bosonnlp open source computer semantic analysis and processing library to set scores according to positive and negative sentiments. j Sentiment score for the jth sentiment word; 303) There is a degree adverb Ad in a single sentence text; the polarity coefficient is set according to the polarity strength of the degree adverb, Ad k is the sentiment polarity coefficient of the kth degree adverb; if there is no degree adverb, it is represented by Ad0; 304) There is a negation word Neg in the single sentence text; determine the negation coefficient value according to whether there is a negation word, where Neg l is the negation coefficient of the lth negative word; if there is no negative word, it is represented by Neg0; 305) When there are multiple event words Inc or sentiment words Sen in a sentence, it is necessary to judge the sentiment semantics and transition of the sentence; when there is a conjunction Conj, the sentiment semantic evaluation priority is judged according to the position relationship between the conjunction and the basic unit and the transition relationship of the conjunction, and the conjunction relationship semantic evaluation priority coefficient η is determined C ; Among them, the conjunction is the boundary, and the basic unit priority coefficient before the conjunction is η Cb , the priority coefficient of the basic unit after the conjunction is η Ca ; 306) When there are both event words Inc and sentiment words Sen in a sentence, the event words, as objective fact-statement words, have higher evaluation credibility than subjective sentiment words. The relationship evaluation priority coefficient η between the basic units of event words Inc and sentiment words Sen is Y are η YI and η YS ; 307) For each basic unit in the sentence, combining the above methods 305) and 306), its evaluation priority coefficient η is: η = η C ×η Y ; where η C is the semantic evaluation priority coefficient of conjunction relationship, η Y is the priority coefficient for evaluating the basic unit relationship; according to the basic unit type, η I Indicates the priority coefficient of the event word; η S Indicates the priority coefficient of sentiment words; 308) The dependency relationship between the basic unit event word Inc and the emotion word Sen and the degree adverb Ad and the negation word Neg. If there is a conjunction Conj, the conjunction is used as the dividing line. If there is no conjunction, each basic unit is used as the dividing line. 4) Calculate the sentiment evaluation value according to the established sentiment analysis evaluation model; For the conversation set with the same topic, sentiment analysis evaluation is given by combining the relationship characteristics of sentiment evaluation basic unit and auxiliary unit; each basic unit is directly affected by the unit priority coefficient, degree adverb and negation coefficient, and the basic unit score is directly modified by the series product of coefficients; the basic units are independent of each other and connected by parallel addition; for the same evaluation object supplier under the same topic, the sentiment score value in sentence units is calculated. The positive or negative value of the value directly reflects the positive or negative evaluation of the interlocutor, and the size of the value reflects the intensity of the emotion; thus, the sentiment analysis evaluation model of a single sentence is obtained: Among them, S MANu The sentiment evaluation value of the supplier score for the u-th single sentence; I represents the number of event words in the sentence; η Ii represents the priority coefficient of the ith event word, which is calculated by 307) in step 3; K represents the number of degree adverbs attached to each event word; L represents the number of negative words attached to each event word; J represents the number of sentiment words in the sentence; η Sj represents the priority coefficient of the jth sentiment word, which is also calculated by 307) in step 3; P represents the number of degree adverbs attached to each sentiment word; Q represents the number of negative words attached to each sentiment word; wherein k, l, p, q values ​​start from 0, indicating that when there is no corresponding auxiliary unit, the coefficient takes its initial value 1; The total sentiment evaluation value of the supplier is divided into the sum of all single sentence scores: That is, the overall sentiment analysis evaluation model; Among them, S MAN is the sentiment evaluation value of the total suppliers; U is the number of evaluation sentences; 5) Based on future language development and following natural laws, the dictionary will be expanded regularly and the evaluation model will be verified and revised; Obtain multiple power dialogue texts, and use the sentiment analysis evaluation model to perform evaluation calculations and manual analysis to obtain the corresponding theoretical evaluation values ​​and actual evaluation values; compare the theoretical calculation evaluation value and the actual evaluation value, and when the difference between the two exceeds the set threshold, return to step 3) to revise the sentiment analysis evaluation model and expand the dictionary; if not, the sentiment analysis evaluation model is considered reasonable.

2. The method for analyzing the sentiment of a conversation text for supplier evaluation according to claim 1, characterized in that: In step 301), Inc i is the sentiment score of the ith event, which is -1 or -2 according to the severity of the general and serious events respectively.

3. The method for analyzing the sentiment of a conversation text for supplier evaluation according to claim 1, characterized in that: In step 303), the polarity coefficient is set to 0.5, 1, and 2 according to the polarity strength of the degree adverb. k is the sentiment polarity coefficient of the kth degree adverb; if there is no degree adverb, it is represented by Ad0, and Ad0=1.

4. The method for analyzing the sentiment of a conversation text for supplier evaluation according to claim 1, characterized in that: In step 304), when there is a negative word, the negative coefficient is taken as -1. If there is no negative word, it is represented by Neg0, and Neg0=1.

5. The method for analyzing conversational text sentiment for supplier evaluation according to claim 1, characterized in that: In step 305), for conjunctions between multiple basic units, the priorities on both sides are set as follows: Conjunctions that indicate transitions and causal relationships, including "but" and "so", have a priority coefficient for the evaluation of the basic units on both sides of the conjunction: η Cb =0.8,η Ca =1.2; Conjunction indicating parallel and consecutive relationships, including "moreover" and "then". The evaluation priority coefficients of the basic units on both sides of the conjunction are: η Cb = η Ca = 1; Conjunctions that express concessions include "although" and "although". The evaluation priority coefficients of the basic units on both sides of the conjunction are: η Cb =1.2,η Ca =0.8; if there is no conjunction, then for all basic units, η C =1.

6. The method for analyzing conversational text sentiment for supplier evaluation according to claim 1, characterized in that: In step 306), the relationship evaluation priority coefficient η between the event word Inc and the sentiment word Sen basic unit is Y are η YI =1.2 and η YS =0.

8.

7. The method for analyzing conversational text sentiment for supplier evaluation according to claim 6, characterized in that: Before performing semantic sentiment analysis, a dialogue interruption cross-processing step is also included. When performing dialogue interruption cross-processing, the steps include A1) obtaining a dialogue style set D; A2) Sequentially extract the two dialogue texts d in D i and d j ; A3) Judge d i and d j Whether the single-round dialogue text turn interval condition is met; if so, proceed to step A4); if not, proceed to step A5); A4) Calculate the single-round dialogue relevance matching degree of the two texts; A5) Determine whether there is a link. If so, remove the link statement d i Calculate the relevance matching degree of the two conversation texts with the closest interval between the linked users in turn; if not, it is considered that the topic conversation set to which the text belongs has been summarized, and then go to step A7); A6) Summarizing the conversations on the same topic based on the matching results; A7) Determine whether the dialogue style set D is empty. If not, proceed to step A2); if so, end the process to obtain a set of all dialogues with the same topic.

8. The method for analyzing conversational text sentiment for supplier evaluation according to claim 7, characterized in that: When it is necessary to evaluate suppliers, after obtaining the same topic dialogue set, the supplier identification is performed. According to the supplier information category in the power business ontology dictionary, the supplier information in each topic dialogue set is extracted, and the implicit evaluation object is identified using the upward nearest principle, and then irrelevant redundant topic content is removed; Supplier identification uses the following three rules: B1) If the supplier information is not identified, the conversation set is considered to be irrelevant and redundant, which is of no value to the evaluation of the equipment supplier and can be screened out; B2) If one or more identical supplier information is identified, the evaluation object of the conversation set is considered to be the identified supplier; B3) There are two or more different supplier information, such as manufacturers A, B, ..., X in order of appearance; for each text in the set, the corresponding supplier is determined by the principle of upward proximity, and the evaluation object of the dialogue set from the first sentence to the sentence before manufacturer B appears is defined as manufacturer A, the evaluation object from the sentence before manufacturer B appears to the sentence before manufacturer C appears is manufacturer B, and so on; If a manufacturer appears repeatedly, the conversation sets of this manufacturer will be merged; thus, the topic text of each supplier is summarized.

9. The method for analyzing conversational text sentiment for supplier evaluation according to claim 1, characterized in that: The expansion of the power ontology dictionary adopts a semi-supervised word segmentation method. On the basis of the power equipment defect domain ontology dictionary and the general dictionary, the HMM model is used to perform word segmentation based on the power business domain text of the dialogue. New domain ontology words are selected according to the word frequency sorting. Manual verification is performed to determine whether they become ontology words and synonyms, and the ontology dictionary is supplemented. With the continuous increase of the corpus, the above steps are repeated for multiple corrections to achieve the supplement and enrichment of the power domain ontology dictionary. The newly added ontology dictionary includes supplier names and business domain events. The supplier names and their commonly used abbreviations are expanded to cope with the prominent characteristics of the dialogue text in spoken language.

Images