CN110175325A - The comment and analysis method and Visual Intelligent Interface Model of word-based vector sum syntactic feature - Google Patents

Abstract

The present invention proposes a comment analysis method based on word vectors and syntactic features in the field of data analysis, including: obtaining the comment data on the product page of the e-commerce website; preprocessing the acquired target data set; extracting the praise and criticism provided by Hownet and NTU The word set forms the basic emotional dictionary; the preprocessed data set is used for word vector training through the Word2Vec tool; the semantic similarity matrix is used to establish the probability transition matrix; the obtained product review text is processed based on the core sentence rules; Preprocess the obtained redundant text; evaluate the matching pair through part-of-speech extraction <commodity attribute, negative word, degree word, emotional word> on the obtained dependency relationship pair; combine the obtained evaluation matching pair with the sentiment dictionary, and evaluate the evaluation object Computational value calculation, pros and cons ranking, and finally realized through a visual interactive interface to realize accurate, real-time, automatic, and convenient processing and analysis of product review data, which can be used in e-commerce platforms.

Description

Comment analysis method and visual interactive interface based on word vector and syntactic features

technical field

The invention belongs to the technical field of data analysis, and in particular relates to an emotional dictionary suitable for commodity reviews constructed using word vectors trained by a neural network model, an attribute recognition algorithm, and a comment analysis system based on word vectors and syntactic features.

Background technique

With the popularization of the Internet and the development of e-commerce, Internet e-commerce websites such as JD.com and Taobao have developed rapidly, and more and more consumers have begun to choose online shopping; these e-commerce websites have a large number of commodities and a large user base , resulting in a huge amount of comment data. The comments given by consumers often carry the user's subjective feelings about the consumption, including the degree of preference for the purchased goods, the degree of satisfaction with the merchant's service, and so on. For consumers, these review texts can help them understand the information of related products or services more objectively, so as to give more suitable choices; Help it to further improve the quality of services or goods in a targeted manner, so as to obtain more customers and profits. However, with the explosive growth of data volume, the cost for users to obtain useful information from massive comment data is also increasing. Therefore, how to process and analyze user comment text quickly and effectively, and extract It has important application value and research significance.

At present, a large amount of review data cannot be fully utilized, and it is difficult for consumers to obtain valuable information from massive review data. Therefore, we researched a comment analysis system based on word vectors and syntactic features, obtained the user's satisfaction with each attribute of the product according to the analysis results, and then summarized the advantages and disadvantages of the product, and then visualized the analysis results.

Contents of the invention

The technical problem to be solved by the present invention is how to realize accurate, real-time, automatic and convenient processing and analysis of commodity review data, and provide a comment analysis method based on word vector and syntactic features to overcome the deficiencies of the prior art.

The present invention provides a comment analysis method based on word vectors and syntactic features, comprising the following steps:

1) Obtain the comment data on the product page of the e-commerce website;

2) Preprocess the acquired target data set and construct a candidate emotional word set;

3) Extract the praise and derogation words provided by Hownet and NTU to form the basic sentiment dictionary;

4) Perform word vector training on the obtained preprocessed data set through the Word2Vec tool, obtain word vectors and generate a semantic similarity matrix;

5) Use the semantic similarity matrix to establish a probability transition matrix, and combine the seed word set to generate the final sentiment dictionary through the LPA label propagation algorithm and the basic sentiment dictionary test;

6) Process the obtained product review text based on the core sentence rules to obtain a redundant review text;

7) Preprocess the obtained redundant text, form a dependency tree based on the dependency and syntactic features of the obtained word segmentation data set, and generate SBV, VOB, ATT, CMP, COO dependency pairs;

8) Extracting <commodity attributes, negative words, degree words, emotional words> through part-of-speech evaluation on the resulting dependency pairs;

9) Combining the obtained evaluations with the emotional dictionary, calculate the value of the evaluation object, rank the pros and cons, and finally realize it through the visual interactive interface.

As a further definition of the present invention, step 2) specifically includes:

2-1) Use character matching algorithm to remove illegal characters;

2-2) Use LTP for word segmentation and part-of-speech tagging on the original data set;

2-3) Extract the words that match the part of speech, and form the candidate emotional word set 1 after deduplication;

2-4) Use NLPIR for word segmentation and part-of-speech tagging on the original data set;

2-5) Extract the words that match the part of speech, and form the candidate emotional word set 2 after deduplication;

2-6) Combine the candidate emotional word set 1 and the candidate emotional word set 2, and obtain the candidate emotional word set after deduplication.

As a further limitation of the present invention, step 3) specifically includes: using the hownet emotion dictionary and the ntu evaluation word dictionary to extract the praise and derogatory words in them respectively, and deduplicate them after merging to form a basic emotion dictionary.

As a further definition of the present invention, step 4) specifically includes:

4-1) Use the Word2Vec training data set to get the word vector of the word;

4-2) Combined with the candidate emotional word set, the following formula is used to calculate the semantic similarity between words:

4-3) For example, for two n-dimensional word vectors a (x ₁₁ , x ₁₂ , … , x _1n ) and b (x ₂₁ , x ₂₂ , … , x _2n ), the formula for calculating the semantic similarity is as follows:

in, Indicates the semantic similarity value; Indicates the value of the kth dimension of the word vector a; Indicates the value of the kth dimension of the word vector b;

4-4) Construct a semantic similarity matrix based on the calculated semantic similarity.

As a further definition of the present invention, step 5) specifically includes:

5-1) Each word is regarded as a node of a graph, and the weight of an edge between two nodes is represented by the semantic similarity between the words it represents;

5-2) Establish the probability transition matrix P according to the following formula:

Among them, P[i][j] represents the similarity transfer probability between words i and j, SIM(w _i , w _j ) represents the similarity between words i and j, and m represents the word with the highest semantic similarity with word i the number of

5-3) Count the word frequency of all emotional words in the original comment data in the candidate emotional word set, and filter out the N words with the highest word frequency to form the seed word set 1; use the emotional vocabulary ontology database to filter out the emotional vocabulary ontology strength>m and The words in the candidate emotional word set form the seed word set 2; the seed word set 1 and the seed word set 2 are merged and removed to form the seed word set for artificial emotion labeling;

5-4) Using a small amount of artificially labeled seed words to establish a label matrix Y _L of LxC, wherein: L represents the number of seed words; C represents the number of classes, which are divided into 3 categories, respectively commendatory, derogatory, and neutral;

5-5) At the same time, use the unlabeled sample words to establish the label matrix Y _U of UxC, where: U indicates the number of unlabeled sample words; C indicates the number of classes, which are divided into 3 categories, respectively commendatory, derogatory, and neutral ;

5-6) Finally, the part-of-speech tagging of the sample words is carried out using the LPA label propagation algorithm, and the final sentiment dictionary is formed after passing the basic sentiment dictionary test.

As a further definition of the present invention, step 6) specifically includes:

The core sentence mainly refers to deleting redundancy and retaining the main components related to the evaluation collocation; if the original sentence does not meet any rules, it will remain unchanged. This method uses the core sentence to improve the accuracy of the syntactic dependency analysis of the evaluation text. The rules include the following :

Rule 1: Delete the initial adverbial components in the sentence, such as the sequence of "advantages of...", "disadvantages of...", "deficiencies of...", "advantages of...", "benefits of...";

Rule 2: Delete sentences with hypothetical tendencies, such as "if...", "hope...", "if...", "hope...", "suggest...";

Rule 3: Delete the sequence beginning with "that is", "it is", "especially", "there is", "especially";

Rule 4: Delete the words "feeling" and "thinking";

Rule 5: Delete consecutive punctuation marks except the first punctuation mark and abnormal characters such as emoticons, emoticons, and brackets.

As a further definition of the present invention, step 7) specifically includes:

The five axioms of dependency syntax:

(1) A sentence can only have one and only one independent component;

(2) Any element in the sentence must be dependent on a certain element at the same time;

(3) Any element in a sentence cannot depend on two or more elements at the same time;

(4) In a sentence, if component a directly depends on component b, and component c is located between components a and b, then component c depends on a or b or other components between a and b;

(5) There is no interdependent relationship between the components on the left and right sides of the central component;

The characteristics of the dependency tree are:

(1) The nodes in the tree are acted by the components in the sentence;

(2) The root node of the tree is the central component of the entire sentence;

(3) The edges formed between the nodes in the tree are directional, which reflects the asymmetric dependency between components;

(4) Satisfy the five axioms of dependency syntax;

Most of the sentences in the comments depend on five types of relationship: main-predicate relationship (SBV), verb-object relationship (VOB/FOB), fixed-center relationship (ATT), verb-complement relationship (CMP), and parallel relationship (COO). The LTP dependency syntax analyzer performs dependency syntax analysis, and extracts the dependency relationship in conjunction with the COO algorithm that identifies parallel evaluation objects and parallel evaluation words; the COO algorithm for identifying parallel evaluation objects and parallel evaluation words specifically includes:

Traversing all words between two nodes in the SBV, VOB, ATT, and CMP dependency pairs obtained based on the dependency relationship and syntactic features, and all words related to it in the dependency syntax tree;

Determine whether there is a COO relationship in all the words traversed;

Expand the parallel evaluation objects and evaluation words of the COO relationship.

As a further definition of the present invention, step 8) specifically includes:

8-1) According to the characteristics of the Chinese language, evaluation objects are mostly nouns or verbs, and evaluation words are mostly adjectives or verbs;

8-2) Extract evaluation objects and evaluation words according to part of speech, that is, commodity attributes and emotional words;

8-3) According to the dependency syntax tree, whether there is a negative word between the evaluation object obtained by traversal and the evaluation word, if there is, the number of negative words is +1, and if multiple negative words are traversed, they are accumulated and added until the end of the traversal. The number of negative words is used for parity judgment. If it is an odd number, the corresponding negative word privative is assigned a value of -1, and if it is an even number, the corresponding negative word privative is assigned a value of +1;

8-4) According to the dependency syntax tree, whether there are degree words between the evaluation object obtained by traversing and the evaluation word, if more than one is traversed, the number is accumulated to obtain the number of degree words of this matching pair;

8-5) Finally form <commodity attributes, negative words, degree words, emotional words> evaluation collocation pairs.

As a further limitation of the present invention, step 9) specifically includes:

According to the product attribute a that appears n times, the calculation formula of its praise and depreciation value is as follows:

Where a.score is the emotional value of product attribute a, is the i-th occurrence of the commodity attribute, privative is the obtained value (-1 or +1) of the negative word corresponding to the i-th commodity attribute, and degree is the number of degree adverbs corresponding to the i-th commodity attribute; thus calculated The emotional value of commodity attributes is calculated cumulatively for the same evaluation object;

All the evaluation objects extracted are divided into two categories, and the final results are arranged by bubble sorting.

A visual interactive interface that can perform all the above-mentioned steps in the claims, and can display the emotional value in the form of a histogram well, and also adds many friendly interactive functions, including: loading, logging in, logging out, and changing passwords And user login usage status, etc.

Compared with the prior art, the present invention adopts the above technical scheme and has the following technical effects:

The present invention constructs a basic sentiment dictionary by obtaining and preprocessing the commodity page comment data of an e-commerce website; and then performs word vector training on the obtained preprocessed data set through the Word2Vec tool to generate a semantic similarity matrix and then establish a probability transfer matrix , and combined with the seed word set to generate the final sentiment dictionary through the LPA label propagation algorithm; the obtained product review text is processed based on the core sentence rules to obtain the redundant comment text; and then the redundant redundant text is obtained Perform preprocessing, form a dependency tree based on the dependency relationship and syntactic features of the obtained word segmentation data set, generate SBV, VOB, ATT, CMP, COO dependency relationship pairs and extract <commodity attributes, negative words, degree words, emotional words> evaluation Matching pairs, combined with the emotional dictionary, calculates the value of product attributes, ranks the pros and cons, and finally realizes it through a visual interactive interface; it can simultaneously realize accurate, real-time, automatic and convenient analysis of comment data.

Description of drawings

Fig. 1 is the flow chart of the present invention.

Detailed ways

Below in conjunction with accompanying drawing, technical scheme of the present invention is described in further detail:

The technical solution of the present invention uses a word vector trained by a neural network model, combined with the LTP label propagation algorithm to construct an emotional dictionary suitable for commodity reviews; and designs a commodity attribute recognition based on core sentence rules, dependency relationships and syntactic features Extraction algorithm; combined with the above technical solutions, a comment analysis system based on word vectors and syntactic features was constructed, and the user's satisfaction with each attribute of the product was obtained according to the analysis results, and then the advantages and disadvantages of the product were summarized, and then the analysis results were analyzed. visualization.

Referring to Fig. 1, the present invention implements a comment analysis method based on word vectors and syntactic features, and the specific implementation steps are as follows:

Step S101: Acquiring review data on product pages of an e-commerce website.

In the specific implementation, a comment data crawling algorithm is designed to obtain the comment data of various commodities on the e-commerce website, and generate the original comment data set.

Step S102: preprocessing the acquired target data set, and building a basic sentiment dictionary.

In the specific implementation, the character matching algorithm is used to remove illegal characters on the original data set; firstly, LTP is used for word segmentation and part-of-speech tagging, and words whose part-of-speech identifier is "a" (adj) are extracted, and the candidate emotional word set 1 is formed after deduplication; Then use NLPIR for word segmentation and part-of-speech tagging, extract the words with the part-of-speech tag "a" (adj), and form candidate emotional word set 2 after deduplication; merge candidate emotional word set 1 and candidate emotional word set 2, after deduplication, Compose the final candidate sentiment word set.

Step S103: extract the praise and derogation words provided by Hownet and NTU to form the basic sentiment dictionary.

In the specific implementation, the hownet emotion dictionary and the NTU evaluation word dictionary are used to extract the praise and derogatory words in them respectively, and then merge them to remove the duplicates to form the basic emotion dictionary.

Step S104: Perform word vector training on the obtained preprocessed data set through the Word2Vec tool to obtain word vectors and generate a semantic similarity matrix.

In the specific implementation, use the Word2Vec training data set, set the training parameters size=100, window=5, sg=0, min_count=0 respectively, and get the word vector of the word after training.

Combined with the candidate emotional word set, the following formula is used to calculate the semantic similarity between words.

For example, for two n-dimensional word vectors a (x ₁₁ , x ₁₂ , … , x _1n ) and b (x ₂₁ , x ₂₂ , … , x _2n ), the formula for calculating the semantic similarity is as follows:

in, Indicates the semantic similarity value; Indicates the value of the kth dimension of the word vector a; Indicates the value of the kth dimension of the word vector b;

Traverse all the emotional words in the candidate emotional word set in order, fix one, and calculate its similarity with all other emotional words; suppose there are m candidate emotional words, after m*m calculations, get a m*m semantic similarity matrix.

For the convenience of the following operations, it is stipulated that the similarity between the same emotional words is 0.

Construct a semantic similarity matrix based on the calculated semantic similarity.

Step S105: Use the semantic similarity matrix to establish a probability transition matrix, and combine the seed word set to generate the final sentiment dictionary through the LPA label propagation algorithm and the basic sentiment dictionary test.

In the specific implementation, each word is regarded as a node of the graph, and the weight of an edge between two nodes is represented by the semantic similarity between the words it represents.

The probability transition matrix P is established according to the following formula:

Among them, P[i][j] represents the similarity transfer probability between words i and j, SIM(w _i , w _j ) represents the similarity between words i and j, and m represents the word with the highest semantic similarity with word i The number of (manually set); establish the probability transition matrix P according to the above formula.

The word frequency of all emotional words in the original comment data in the candidate emotional word set was counted, and the 100 words with the highest word frequency were selected to form the seed word set 1; using the emotional vocabulary ontology database of Dalian University of Technology, the emotional vocabulary ontology strength > 7 and in The words in the candidate emotional word set form the seed word set 2; the seed word set 1 and the seed word set 2 are combined and removed to form the seed word set for artificial emotion labeling.

Then use a small number of artificially labeled seed words to establish the LxC label matrix Y _L , where: L represents the number of seed words; C represents the number of classes, generally 3 types (commendative, derogatory, neutral); The sample word establishes the label matrix Y _U of UxC, where: U represents the number of unlabeled sample words; C represents the number of classes, generally 3 classes (commendative, derogatory, neutral); combine the two label matrices to obtain a NxC's soft label matrix F=[Y _L ; Y _U ].

Execute the label propagation algorithm, the specific operation is: 1) Execute propagation: F=PF; 2) Reset the label of the labeled sample in F: F _L =Y _L ; 3) Repeat steps 1) and 2) until F converges.

Among them, the purpose of step 1 is to propagate the label (emotional attribute) of each node (emotional word) to other nodes with the probability determined by the probability transfer matrix. If the similarity between the two nodes is greater, the probability of propagation is greater; step The purpose of 2 is to reset the label of the tagged seed word to the tagged value to avoid changes due to the operation process of step 1; the method of determining the convergence of F in step 3 is to calculate the latest F and the F ₀ after the last operation Matrix similarity, until the similarity does not change, it is considered that F has converged.

The three values in a single row in the final matrix F represent the attribute propagation value of the corresponding emotional word, select the largest value among them, judge its corresponding attribute, and determine the attribute of the emotional word.

Deriving the sentiment word of confirming property, form sentiment dictionary 1, traverse all sentiment words in sentiment dictionary 1, if the basic sentiment dictionary described in step S103 contains this word and contradicts with the property in the basic sentiment dictionary, change its property, with basic sentiment The attribute in the dictionary prevails; otherwise, the attribute remains unchanged.

After the above steps are completed, the modified sentiment dictionary 1 is the final sentiment dictionary.

Step S106: Process the acquired commodity review text based on core sentence rules to obtain a redundant review text.

In the specific implementation, on the interactive interface of the webpage of this system, input the website address of the product, and crawl the comment data of the input product on the e-commerce platform through the web crawler mechanism designed in the background, and the system is set to crawl the first 1000 high-quality articles of the product. comment data.

The commodity review data obtained by crawling will be processed based on the core sentence rules to remove redundancy, and the main components related to the evaluation collocation will be retained; for example: "The mobile phone has been received, it is very good, the pixel and sound quality are very good, especially the express delivery Very good (next day delivery), the only downside is that the packaging is not very good, I hope the store can improve it..." The processing is as follows:

(1) Matching rule 1. In the example sentence, "…'s lack of" is matched, and after processing, it becomes "The mobile phone has been received, it is very good, the pixel and sound quality are very good, especially the express delivery is very good (next day delivery), it is the packaging Not very good, I hope the store can improve it...";

(2) Matching rule 2, "hope" is matched in the example sentence, and after processing, it becomes "The mobile phone has been received, it is very good, the pixel and sound quality are very good, especially the express delivery is very good (next day delivery), but the packaging is not very good Well, the store can improve it...";

(3) Matching rule 3. In the example sentence, "is" and "especially" are matched, and after processing, it becomes "The mobile phone has been received, it is very good, the pixel and sound quality are very good, the express delivery is very good (the next day), the packaging is not Very good, the store can improve it...";

(4) Matching rule 5, delete consecutive punctuation marks in the example sentence, and the core sentence obtained in the final processing is "The mobile phone has been received, it is very good, the pixel and sound quality are very good, the express delivery is very good, the packaging is not very good, the store It can be improved." This embodiment is recorded as the embodiment sentence Sentences.

Step S107: Preprocess the obtained redundant text, form a dependency tree based on the dependency and syntactic features of the obtained word segmentation data set, and generate SBV, VOB, ATT, CMP, COO dependency pairs.

In a specific implementation, the redundant text obtained in the above step S106 is preprocessed to divide sentences with punctuation marks to obtain 6 clauses. Use the LTP tool to segment each sentence, mark part of speech, and form a dependency tree based on dependency and syntactic features. Get the dependency on SBV<mobile phone, received>, SBV<pixel, good>, COO<sound quality, pixel>, SBV<express delivery, awesome>, SBV<packaging, good>, SBV<store, improvement>.

For example, for the small sentence "The pixels and sound quality are both very good", after the above steps, the COO algorithm for identifying parallel evaluation objects and parallel evaluation words is used to extract the dependency relationship pair again, and the obtained dependency relationship pair is <pixel, good> , <sound quality, good>.

Step S108: Evaluate the matching pair by extracting <commodity attribute, negative word, degree word, emotion word> from the obtained dependency relationship pair.

In the specific implementation, for each extracted relationship pair, whether there is a negative word between the evaluation object and the evaluation word is traversed, and the number is calculated. value, that is, if the number of negative words is judged to be an odd number, the corresponding privative value is -1; if the number of negative words is judged to be an even number, the corresponding privative value is +1. Then traverse whether there are degree words between the evaluation object and the evaluation words, and calculate the number of degree words. Finally, an evaluation pair of <commodity attribute, privative, degree, emotional word> is formed. In the example sentence Sentences in step S106, a negative word "no" is identified between the relationship pair <package, good>, then the corresponding privative value is -1; then traverse the degree adverbs between "package" and "good" , if "very" is recognized, the corresponding degree value is 1. Then the evaluation collocation pair extracted from this clause is <package, -1, 1, good>.

Step S109: Combining the obtained evaluation matching pairs with the sentiment dictionary, calculate the value of the evaluation object, rank the superior and inferior, and finally realize it through the visual interactive interface.

In the specific implementation, for the extracted evaluation collocation pairs, the praise and derogation attributes of the sentiment words are obtained through the sentiment dictionary. Then calculate the appreciation and depreciation value of commodity attributes according to the following formula:

For the evaluation collocation obtained in step S107, the value of the product attribute "package" in <package, -1, 1, good> is calculated to obtain its emotional value .

Traverse all the comment data of the acquired product, perform the above steps, accumulate the same evaluation objects, and finally extract all the product attributes of the product, and then divide them into two categories, using bubble sorting to get the final result. Finally, through the front and back ends, it is realized on the web page with a visual interactive interface.

The above is only a specific implementation mode in the present invention, but the scope of protection of the present invention is not limited thereto. Anyone familiar with the technology can understand the conceivable transformation or replacement within the technical scope disclosed in the present invention. All should be covered within the scope of the present invention, therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims (10)

1. A comment analysis method based on word vector and syntactic feature, is characterized in that, comprises the following steps: 1) Obtain the comment data on the product page of the e-commerce website; 2) Preprocess the acquired target data set and construct a candidate emotional word set; 3) Extract the praise and derogation words provided by Hownet and NTU to form the basic sentiment dictionary; 4) Perform word vector training on the obtained preprocessed data set through the Word2Vec tool, obtain word vectors and generate a semantic similarity matrix; 5) Use the semantic similarity matrix to establish a probability transition matrix, and combine the seed word set to generate the final sentiment dictionary through the LPA label propagation algorithm and the basic sentiment dictionary test; 6) Process the obtained product review text based on the core sentence rules to obtain a redundant review text; 7) Preprocess the obtained redundant text, form a dependency tree based on the dependency and syntactic features of the obtained word segmentation data set, and generate SBV, VOB, ATT, CMP, COO dependency pairs; 8) Extracting <commodity attributes, negative words, degree words, emotional words> through part-of-speech evaluation on the resulting dependency pairs; 9) Combining the obtained evaluations with the emotional dictionary, calculate the value of the evaluation object, rank the pros and cons, and finally realize it through the visual interactive interface. 2. The comment analysis method based on word vectors and syntactic features according to claim 1, wherein step 2) specifically includes: 2-1) Use character matching algorithm to remove illegal characters; 2-2) Use LTP for word segmentation and part-of-speech tagging on the original data set; 2-3) Extract the words that match the part of speech, and form the candidate emotional word set 1 after deduplication; 2-4) Use NLPIR for word segmentation and part-of-speech tagging on the original data set; 2-5) Extract the words that match the part of speech, and form the candidate emotional word set 2 after deduplication; 2-6) Combine the candidate emotional word set 1 and the candidate emotional word set 2, and obtain the candidate emotional word set after deduplication. 3. The comment analysis method based on word vectors and syntactic features according to claim 1, characterized in that, step 3) specifically includes: using hownet sentiment dictionary and ntu evaluation word dictionary, extracting the praise and derogation words therein respectively, and removing them after merging Heavy, forming a basic sentiment dictionary. 4. The comment analysis method based on word vectors and syntactic features according to claim 1, wherein step 4) specifically includes: 4-1) Use the Word2Vec training data set to get the word vector of the word; 4-2) Combined with the candidate emotional word set, the following formula is used to calculate the semantic similarity between words: 4-3) For example, for two n-dimensional word vectors a (x ₁₁ , x ₁₂ , … , x _1n ) and b (x ₂₁ , x ₂₂ , … , x _2n ), the formula for calculating the semantic similarity is as follows: in, Indicates the semantic similarity value; Indicates the value of the kth dimension of the word vector a; Indicates the value of the kth dimension of the word vector b; 4-4) Construct a semantic similarity matrix based on the calculated semantic similarity. 5. The comment analysis method based on word vectors and syntactic features according to claim 1, wherein step 5) specifically includes: 5-1) Each word is regarded as a node of a graph, and the weight of an edge between two nodes is represented by the semantic similarity between the words it represents; 5-2) Establish the probability transition matrix P according to the following formula: Among them, P[i][j] represents the similarity transfer probability between words i and j, SIM(w _i , w _j ) represents the similarity between words i and j, and m represents the word with the highest semantic similarity with word i the number of 5-3) Count the word frequency of all emotional words in the original comment data in the candidate emotional word set, and filter out the N words with the highest word frequency to form the seed word set 1; use the emotional vocabulary ontology database to filter out the emotional vocabulary ontology strength>m and The words in the candidate emotional word set form the seed word set 2; the seed word set 1 and the seed word set 2 are merged and removed to form the seed word set for artificial emotion labeling; 5-4) Using a small amount of artificially labeled seed words to establish a label matrix Y _L of LxC, wherein: L represents the number of seed words; C represents the number of classes, which are divided into 3 categories, respectively commendatory, derogatory, and neutral; 5-5) At the same time, use the unlabeled sample words to establish the label matrix Y _U of UxC, where: U indicates the number of unlabeled sample words; C indicates the number of classes, which are divided into 3 categories, respectively commendatory, derogatory, and neutral ; 5-6) Finally, the part-of-speech tagging of the sample words is carried out using the LPA label propagation algorithm, and the final sentiment dictionary is formed after passing the basic sentiment dictionary test. 6. The comment analysis method based on word vectors and syntactic features according to claim 1, characterized in that, step 6) specifically includes: The core sentence mainly refers to deleting redundancy and retaining the main components related to the evaluation collocation; if the original sentence does not meet any rules, it will remain unchanged. This method uses the core sentence to improve the accuracy of the syntactic dependency analysis of the evaluation text. The rules include the following : Rule 1: Delete the initial adverbial components in the sentence, such as the sequence of "advantages of...", "disadvantages of...", "deficiencies of...", "advantages of...", "benefits of..."; Rule 2: Delete sentences with hypothetical tendencies, such as "if...", "hope...", "if...", "hope...", "suggest..."; Rule 3: Delete the sequence of "is", "is actually", "especially", "there is", "especially" at the beginning of the sentence; Rule 4: Delete the words "feeling" and "thinking"; Rule 5: Delete consecutive punctuation marks except the first punctuation mark and abnormal characters such as emoticons, emoticons, and brackets. 7. The comment analysis method based on word vectors and syntactic features according to claim 1, characterized in that, step 7) specifically includes: The five axioms of dependency syntax: (1) A sentence can only have one and only one independent component; (2) Any element in a sentence must be dependent on a certain element at the same time; (3) Any element in a sentence cannot depend on two or more elements at the same time; (4) In a sentence, if component a directly depends on component b, and component c is located between components a and b, then component c depends on a or b or other components between a and b; (5) There is no interdependent relationship between the components on the left and right sides of the central component; The characteristics of the dependency tree are: (1) The nodes in the tree are acted by the components in the sentence; (2) The root node of the tree is the central component of the entire sentence; (3) The edges formed between the nodes in the tree are directional, which reflects the asymmetric dependency between components; (4) Satisfy the five axioms of dependency syntax; Most of the sentences in the comments depend on five types of relationship: main-predicate relationship (SBV), verb-object relationship (VOB/FOB), fixed-center relationship (ATT), verb-complement relationship (CMP), and parallel relationship (COO). The LTP dependency syntax analyzer performs dependency syntax analysis, and extracts the dependency relationship in conjunction with the COO algorithm that identifies parallel evaluation objects and parallel evaluation words; the COO algorithm for identifying parallel evaluation objects and parallel evaluation words specifically includes: Traversing all words between two nodes in the SBV, VOB, ATT, and CMP dependency pairs obtained based on the dependency relationship and syntactic features, and all words related to it in the dependency syntax tree; Determine whether there is a COO relationship in all the words traversed; Expand the parallel evaluation objects and evaluation words of the COO relationship. 8. The comment analysis method based on word vectors and syntactic features according to claim 1, wherein step 8) specifically includes: 8-1) According to the characteristics of the Chinese language, evaluation objects are mostly nouns or verbs, and evaluation words are mostly adjectives or verbs; 8-2) Extract evaluation objects and evaluation words according to part of speech, that is, commodity attributes and emotional words; 8-3) According to the dependency syntax tree, whether there is a negative word between the evaluation object obtained by traversal and the evaluation word, if there is, the number of negative words is +1, and if multiple negative words are traversed, they are accumulated and added until the end of the traversal. The number of negative words is used for parity judgment; If it is an odd number, the corresponding negative word privative is assigned a value of -1, and if it is an even number, the corresponding negative word privative is assigned a value of +1; 8-4) According to the dependency syntax tree, whether there are degree words between the evaluation object obtained by traversing and the evaluation word, if more than one is traversed, the number is accumulated to obtain the number of degree words of this matching pair; 8-5) Finally form <commodity attributes, negative words, degree words, emotional words> evaluation collocation pairs. 9. The comment analysis method based on word vectors and syntactic features according to claim 1, wherein step 9) specifically includes: According to the product attribute a that appears n times, the calculation formula of its praise and depreciation value is as follows: Among them, a.score is the emotional value of product attribute a, X _i is the i-th occurrence of the product attribute, privative is the obtained value (-1 or +1) of the negative word corresponding to the i-th product attribute, and degree is the ith time The number of degree adverbs corresponding to the product attributes; from this, the emotional value of the product attributes is calculated, and the same evaluation objects are cumulatively calculated; All the evaluation objects extracted are divided into two categories, and the final results are arranged by bubble sorting. 10. A visual interactive interface, characterized in that it can perform all the steps of claims 1 to 9, can display the emotional value in the form of a histogram well, and adds many friendly interactive functions, including : Loading, login, logout, password change, user login status, etc.