CN110837740B - An opinion-level mining method for review aspects based on dictionary-based improved LDA model - Google Patents

Abstract

The invention discloses an opinion-level mining method based on a dictionary-based improved LDA model for comments, and belongs to the field of network comment text mining. Including: constructing an inverted index list based on the original network comment database; performing stop-word processing on each sentence of the original network comment database to obtain the preprocessed network comment database; inputting the preprocessed network comment database into the improved LDA based on SentiWordNet and WordNet The model uses Gibbs sampling to obtain the sampling results; sort the sampling results, select the top m words belonging to the corresponding evaluation category, and find the specific sentence according to the inverted index of the word. The present invention directly sets the aspect of the network comment database as the seed word without manual annotation. The evaluation object words are separated from the comment opinions, and the LDA model parameters are biased by calculating the similarity between the words and the seed words, so as to improve the effect of the model. Based on the inverted index, the clustering results are linked with the seed words and the original text to improve the readability of the results.

Description

An opinion-level mining method for review aspects based on dictionary-based improved LDA model

technical field

The invention belongs to the field of network comment text mining, and more particularly, relates to an opinion-level mining method for comments based on a dictionary-based improved LDA model.

Background technique

The rapid development of the mobile Internet and the popularization of smart phones have provided favorable conditions for people to post comments and opinions anytime, anywhere. On social platforms such as Twitter and Weibo, and online shopping platforms such as Taobao, Amazon, and JD.com, people can evaluate different products in different fields. Effective analysis of these evaluations can assist manufacturers in making decisions on sales and future development, and can also help consumers select products that meet their expectations. However, simply judging the sentiment polarity of the comment sentence cannot provide effective information, and it is necessary to further determine the object described by the sentiment word. Unlike news reports, blogs, etc., online reviews are generally shorter in content. Due to different service contents, the fields of review objects of online reviews are also different. And the online comment object contains many attributes, only by mining aspect-level viewpoints, can we grasp the effective information in the comment.

Aspect-level opinion mining of reviews can extract aspect-level review objects and review categories from reviews, which has important research significance and value. The aspect-level comment object (Opinion Target Expression Extraction) refers to the entity itself or the attribute modified by the sentiment opinion word. In the mining of comment information, only judging the emotional polarity of the comment sentence is meaningless to the people who read the comment, and people are more concerned about the good or bad of the product at the specific level. Therefore, it is very important to determine the aspect-level comment object of the comment. For example, in a product review, "This phone has a normal appearance, sufficient battery power, and strong signal." If you directly judge the emotional polarity of the sentence, you can only know that there is one comment that indicates that the phone is good without reading the original text. , which is obviously of little value to the user. Therefore, when conducting comment mining, the aspect-level comment object words in the comment sentences should be extracted first. For example, for the above sentences, the words that should be extracted are "appearance", "battery", and "signal". AspectCategory Identification and aspect-level review objects belong to each other. In addition to judging that the word belongs to a certain review category, sentences can also be labeled with a review category.

However, the massive reviews involve a wide variety of commodities, and the process of data labeling required for aspect-level opinion mining is cumbersome. It will consume a lot of resources to build a canonical labeling corpus for reviews in all fields. Supervised methods that rely on annotated datasets will be difficult to apply to the review domain that lacks annotated corpora. How to improve the effect of the model under the condition of less supervision and no supervision, and make the model have domain adaptability (including different domains and different languages), is a very worthwhile research topic. The prior art is the MaxEnt-LDA model, which introduces two distributions to indicate the classification of comment object words and sentiment words and the classification of positive sentiment words and negative sentiment words. However, it has the following drawbacks: the classifier used to indicate the classification of comment object words and sentiment words uses a maximum entropy model, which requires a large number of annotations on the dataset.

SUMMARY OF THE INVENTION

In view of the fact that the prior art aspect-level opinion mining method based on MaxEnt-LDA model reviews requires a large number of data sets to be labeled, the present invention provides a review aspect-level opinion mining method based on dictionary improvement LDA model, the purpose of which is to use as little as possible. Using labeled data to solve the problem of aspect-level opinion mining in online reviews.

In order to achieve the above object, according to the first aspect of the present invention, a method for opinion-level mining of comments based on a dictionary improved LDA model is provided, and the method includes the following steps:

S1. Build an inverted index list based on the original network comment library;

S2. Perform stop-word processing on each sentence of the original online comment database, and obtain the pre-processed network comment database;

S3. Input the preprocessed network comment library into the improved LDA model based on SentiWordNet and WordNet, and use Gibbs sampling to obtain the sampling result;

S4. Sort the sampling results, select the top m words belonging to the corresponding evaluation category in terms of probability, and find the specific sentence according to the inverted index of the word.

Specifically, step S1 includes the following sub-steps:

S11. Number each word in each sentence in the original online comment database in the form of a binary group <a, b>, where a represents the number of the sentence in which it is located, and b represents the number of the word in the sentence;

S12. Remove the repeated words in the original online comment database, and record the number of the remaining words;

S13. Based on the deduplicated word numbers, generate an inverted index list.

Specifically, step S3 includes the following sub-steps:

S31. Directly set the aspect of the online comment base as the seed word;

S32. Divide the comment text in the online comment database into sentence units to form a set of comment text sentences;

S33. Based on the similarity between the word and the seed word, set different parameters α _d for each sentence; based on the semantic similarity between the word and the seed word, set the aspect-level object word, positive comments for each topic separately Set the parameters β _{t, A} , β _{t, P} , β _{t, N} for words and negative comment words respectively;

S34. Using Gibbs sampling comment text sentence set, perform parameter estimation and inference on the improved LDA model based on SentiWordNet and WordNet.

specifically,

β _{t, A} =sim(w, A)*β _base

β _t,P =sim(w,P)*β _base

β _{t, N} =sim(w, N)*β _base

Among them, N _d is the number of all words in the current sentence, T is the number of topics, w _{d, i} is the ith word in the current sentence, t is the seed word, and sim(w, t) represents w and the seed word t Semantic similarity, α _base represents the fixed value parameter α of the topic obeying Dirichlet distribution in the standard LDA model; sim(w, A) represents the probability that w belongs to the object word, sim(w, P) represents that w belongs to the positive word Probability, sim(w, N) represents the probability that w belongs to a negative word, and β _base is the fixed value parameter β of the word obeying Dirichlet distribution in the standard LDA model.

Specifically, step S34 includes the following steps:

(1) When each word in each sentence in the corpus is randomly assigned a topic number, the values of the indicator variable y and the indicator variable v are randomly set for all the words in the sentence, where, when y=A, it means that the current word is The aspect-level object of the comment, when y=O, the current word is a comment point of view, when v=P, it means that the current word is a positive emotion, and when v=N, it means that the current word is a negative emotion;

(2) Rescan the corpus, for each word, resample and update its topic number according to formula (1), update the number of the word in the corpus, according to formulas (2) and (3) for the indicator variables y and indicator The variable v is resampled and updated;

表示主题为t类别为q的词v的个数，

表示主题为t类别为q的词v的狄利克雷分布参数，

表示主题为t类别为u的词v的个数，

表示主题为t类别为u的词v的狄利克雷分布参数，V表示语料库中单词的个数，w_d，n表示第d个评论语句中的第n个词，则

表示主题为t类别为u的词w_d，n的个数，

表示主题为t类别为u的词w_d，n的狄利克雷分布参数，

表示主题为t类别为q的词w_d，n的个数，

表示主题为t类别为q的词w_d，n的狄利克雷分布参数，n_d，t表示第d个句子的主题为t的词个数，α_d，t表示第d个评论语句主题为t的狄利克雷分布参数，

i表示非i；Among them, z _{d, n} indicates the topic of the d-th comment sentence, t indicates the topic number, y _{d, n} indicates that the n-th word of the d-th sentence is an aspect-level object word or sentiment opinion word, and v _{d, n} indicates the d-th sentence. The nth word of each sentence is a positive emotion word or a negative emotion word,

represents the number of words v whose topic is t and whose category is q,

represents the parameters of the Dirichlet distribution of word v with topic t and category q,

Represents the number of words v whose topic is t and whose category is u,

Represents the Dirichlet distribution parameter of the word v with the topic t and the category u, V represents the number of words in the corpus, w _{d, n} represents the nth word in the dth comment sentence, then

Represents the number of words w _{d and n} whose topic is t and whose category is u,

represents the parameters of the Dirichlet distribution of words w _{d, n} with topic t and category u,

Represents the number of words w _{d, n} whose topic is t and whose category is q,

Represents the Dirichlet distribution parameters of the word w _{d, n} with the topic t and the category q, n _{d, t} represents the number of words with the topic t of the d-th sentence, α _{d, t} represents the d-th comment sentence The topic is the parameters of the Dirichlet distribution of t,

i means not i;

(3) Repeat the resampling of the above corpus until the Gibbs sampling converges;

统计语料库中各个主题词的分布，得到主题与单词的概率分布

(4) Count the topics of each word in each sentence in the corpus, and obtain the probability distribution of sentence-topic

Calculate the distribution of each subject word in the corpus to obtain the probability distribution of the subject and the word

Specifically, the formula for calculating the probability distribution of sentence d and topic t is as follows:

Taking t as the subject, and the words w _{d and n} as the evaluation aspect-level object words, the probability distribution calculation formula is as follows:

Taking t as the theme and the words w _{d and n} as the evaluation, the probability distribution of positive opinion words is calculated as follows:

Taking t as the theme and the words w _{d and n} as the evaluation, the probability distribution of negative opinion words is calculated as follows:

Among them, n _d represents the number of words in the d-th sentence.

Specifically, the document generation process based on SentiWordNet and WordNet to improve the LDA model is as follows:

评论的积极观点词分布

评论的消极观点词分布

(1) Sampling from the Dirichlet distribution β _t,A , β _t,P , β _t,N to generate an aspect-level object distribution of reviews

Positive opinion word distribution for reviews

Negative opinion word distribution of reviews

(2) For each sentence, sample from Dirichlet distribution α _d to generate topic distribution θ _d ~Dir(α _d );

(3) Sampling from the multinomial distribution θ _d of the topic, and generating the topic Z _d,n ~Multi(θ _d ) of the words w _d,n in the sentence d;

(4) Calculated by WordNet and SentiWordNet, the Bernoulli distribution on {0,1} with parameter π _{d, n} and the Bernoulli distribution on {0, 1} with parameter Ω _{d, n} are obtained;

(5) Extract from the Bernoulli distribution on {0, 1} with parameters π _{d, n} , and obtain the indicator word w _{d, n} is the aspect-level object word of the review or the review opinion word y _{d, n} , with the parameters as Extract the Bernoulli distribution on {0,1} of Ω _d,n to obtain the indicator word w _d,n is a positive opinion comment word or a negative opinion comment word v _d,n ;

(6) According to the following formula, generate the word w _{d, n}

Specifically, step (4) includes the following steps:

(4.1) Query the semantic interpretation of the current word w _{d, n} in WordNet as s _{d, n, k} , calculate the semantics of w _{d, n and} the semantics s _{d, n, k} of each subword w _{t , k0} The similarity Sim(S _{d, n, k} , s _{t, k0} ), among all the calculation results, take the maximum similarity result, and determine the k' when the result is the largest, which is the current word w _{d, n} in this the semantics s _{d, n, k′} to which the sentence belongs;

(4.2) Query the scores of each emotion of s _{d, n, k'} in SentiWordNet

计算参数π_d，n和Ω_d，n。(4.3) Score based on each emotion

Calculate the parameters π _d,n and Ω _d,n .

specifically,

In order to achieve the above object, according to a second aspect of the present invention, a computer-readable storage medium is provided, wherein a computer program is stored on the computer-readable storage medium, and the computer program is implemented when executed by a processor As described in the first aspect, the opinion-level mining method of the review aspect based on the dictionary-based improved LDA model.

In general, through the above technical solutions conceived by the present invention, the following beneficial effects can be achieved:

(1) For the standard LDA model, in the topic-document probability and topic-word probability finally obtained, the topic word is not determined and needs to be determined manually. Annotated datasets.

(2) On the basis of the LDA model, the present invention adds a classification layer of opinion words and evaluation objects to realize the separation of opinion and aspect. With the help of user-configured seed words and the two tools of WordNet and SentiWordNet to solve the problem of less-supervised aspect-level opinion mining, by setting seed words, using the word similarity calculation tool in WordNet, calculate the similarity between words and seed words in the text, and Reflected on the LDA model parameters. At the same time, the tool for calculating lexical sentiment in SentiWordNet is used to separate the evaluation object words from the comment opinions, and classify the positive and negative polarity of the comment opinions. By calculating the similarity between the words in the corpus and the seed words, the parameters of the LDA model are biased to improve the effect of the model.

(3) For the final result in the standard LDA model, the probability of topic-document and topic-word, lacking the connection between words and documents, the present invention is based on the inverted index, and the sentence to which the word belongs is queried, and the clustering result is the same as the seed word and the original text to establish a link to improve the readability of the results.

Description of drawings

1 is a flowchart of a method for opinion-level mining of comments based on a dictionary-based improved LDA model provided by an embodiment of the present invention;

FIG. 2 is a schematic diagram of a document generation process of a SWLDA model provided by an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

First, the terms and variables involved in the invention are explained as follows:

Comment: A person's opinion about something, and an interpretation of that opinion.

Aspect: A detailed aspect about the properties of the review object that is talked about in the review text. When commenting on the object of the review, the reviewer will first determine the aspect to be discussed, select the corresponding words to represent the aspect, and then select the opinion words with specific emotional tendencies to evaluate it according to their own views on this aspect. For example, when reviewers evaluate a hotel, they will use words such as price and value to express the aspect of "price" they want to talk about, and then use words such as high and cheap to express their opinions.

Aspect mining of comments: Use text analysis and mining technology to find multiple aspects of comments discussed in the comment text, analyze the sentimental tendency of commenters to various aspects of the comments, and display the analysis results in a certain form. In the aspect-level opinion mining of reviews, it is necessary to extract the aspect of the review, the aspect-level review object and the opinion from the textual information. For example, in a restaurant review, "food", "service", and "ambiance" are aspects of the review, and in terms of "food", "steak" is an aspect-level review object , while a "good" rating for "steak" is a review's point of view.

Topic Model: Build the "topic" bridge between comments and words.

Table 1 Symbol Description

The present invention extracts the aspect-level comment object words and sentiment opinion words of various aspects when the aspects of the comments in the network comment database are known, and simultaneously determines the comment aspects to which the sentences in the corpus belong. In order to conduct sentiment analysis for the comment aspect, it is necessary to further study the relationship between sentiment opinion words and aspect-level comment object words, as well as the sentences to which the sentiment opinion words belong.

As shown in FIG. 1 , the present invention proposes a method for opinion-level mining of comments based on a dictionary-based improved LDA model. The method includes the following steps:

Step S1. Construct an inverted index list based on the original network comment database.

Online comment information contains a lot of opinion information, and aspect-level opinion mining needs to be carried out to analyze the comments of commenters on various aspects of the comment object, in order to have a complete understanding of the comment object. Different from textual information such as documents, blogs, news, etc., online comments are often shorter in length and often appear in the form of sentences.

In this example, the English review collection of Restaurant is selected, and the sentence 1{“We,there were four of us,arrived at noon-the place was empty-and the staff acted like we were imposingon them and they were very rude.”“Everything is always cooked to perfection, the service is excellent, the decor cool and understated."} as an example, the aspect of the review is {place,staff,cook,service,decor}; with sentence 2{"This restaurant is quite complete and has a good service attitude, but the location of the restaurant is notconvenient."} as an example, the aspect of the comment is {restaurant, service, location}, and the description of the follow-up processing.

S11. Number each word in each sentence in the original online comment database in the form of a binary group <a, b>, where a represents the number of the sentence in which it is located, and b represents the number of the word in the sentence.

For example, in sentence 1, <1,16> means staff, <1,36> means service, and <1,40> means decor. In sentence 2, <2,10> means service, and <2,14> means location.

S12. Remove the repeated words in the original online comment database, and record the number of the remaining words.

S13. Based on the deduplicated word numbers, generate an inverted index list.

Such as <staff:1,16>, <decor:1,40>,<service:1,36; 2,10>,<location:2,14>. In the inverted index list, the sentence number in which the word is located and the position information in the sentence are retained, which is convenient for subsequent searches using contextual information.

Step S2. Perform stop-word processing on each sentence of the original network comment database to obtain a preprocessed network comment database.

The format of the original data set is xml and csv, and the comment sentences need to be extracted according to the corresponding tags and fields.

The sentences in the online comment base contain many useless stop words, such as there, the, etc. in sentence 1. These stop words are removed before further processing so as not to interfere with the results.

Step S3. Input the preprocessed network comment library into the improved LDA model based on SentiWordNet and WordNet, and use Gibbs sampling to obtain the sampling result.

In the SWLDA (SentiWordNet WordNet-Latent Dirichlet Allocation) model, based on the LDA topic model, an aspect-level comment and sentiment opinion word separation layer is introduced, and the semantic similarity calculation in WordNet and the sentiment factor calculation of SentiWordNet are used to assist the solution. the question. First generate sentence-topic distribution, determine a topic for each sentence through multinomial distribution, and then use WordNet and SentiWordNet to determine the two impact factors y _d and v _d to indicate the category of words, and finally select a topic-word distribution and determine the final words.

As shown in Figure 2, the document generation process of the SWLDA model is as follows:

评论的积极观点词分布

评论的消极观点词分布

(1) Sampling from the Dirichlet distribution β _t,A , β _t,P , β _t,N to generate an aspect-level object distribution of reviews

Positive opinion word distribution for reviews

Negative opinion word distribution of reviews

(2) For each sentence, sample from the Dirichlet distribution α _d to generate topic distributions θ _d to Dir(α _d ).

(3) Sampling from the multinomial distribution θ _d of the topics, and generating the topic z _d,n ∼Multi(θ _d ) of the words w _d,n in the sentence d.

(4) Calculated by WordNet and SentiWordNet, the Bernoulli distribution on {0, 1} with parameters π _{d, n} and the Bernoulli distribution on {0, 1} with parameters Ω _{d, n} are obtained.

(4.1) Query the semantic interpretation of the current word w _{d, n} in WordNet as s _{d, n, k} , calculate the semantics of w _{d, n and} the semantics s _{d, n, k} of each subword w _{t , k0} The similarity Sim(s _{d, n, k} , s _{t, k0} ), among all the calculation results, take the maximum similarity result, and determine the k' when the result is the largest, which is the current word w _{d, n} in this The semantics s _{d, n, k, ,} to which the statement belongs.

(该语义为客观词)、

(该语义为积极情感)、

(该语义为消极情感)。(4.2) Query the scores of each emotion of s _{d, n, k,} in SentiWordNet

(the semantics are objective words),

(the semantic is positive emotion),

(The semantic is negative emotion).

SentiWordNet is a sentiment dictionary based on WordNet. For each explanation of each word in WordNet, a positive sentiment score, a negative sentiment score and a neutral sentiment score are given. The three are in the range of 0 to 1, and the sum is 1. The present invention uses the cosine distance of word vectors to calculate the similarity between words.

(该语义为客观词)、

(该语义为积极情感)、

计算参数π_d，n和Ω_d，n。(4.3) Score based on each emotion

(the semantics are objective words),

(the semantic is positive emotion),

Calculate the parameters π _d,n and Ω _d,n .

Bernoulli distribution with parameters π _{d, n} (for separating aspect-level object words and sentiment opinion words) and Bernoulli distribution with parameters Ω _{d, n} (for separating positive sentiment words and negative sentiment words) and WordNet Related to SentiWordNet. The computation of π _d,n depends on the seed word w _t .

(5) Extract from the Bernoulli distribution on {0, 1} with parameters π _{d, n} , and obtain the indicator word w _{d, n} is the aspect-level object word of the review or the review opinion word y _{d, n} , with the parameters as Ω _d,n is extracted from the Bernoulli distribution on {0,1}, and the indicator word w _d,n is obtained as a positive opinion comment word or a negative opinion comment word v _d,n .

The Bernoulli distribution over {0,1} with parameters π _d,n is used to indicate that the word w _d,n is the aspect-level object word of the review or the review opinion word y _d,n ; the parameter is Ω _d,n The Bernoulli distribution over {0,1} is used to indicate that the words w _d,n are positive opinion comments or negative opinion comments v _d,n .

In order to represent the separation of review aspect-level objects and review viewpoints, a variable y _d ∈ {A, O} is introduced to represent it. When y _d =A, it indicates that the current word is an aspect-level object of the review, and when y _d =O, the current word is a review point of view. When v _d =P, it means that the current word is a positive emotion, and when v _d =N, it means that the current word is a negative emotion. Both y _d and v _d are determined by related algorithms based on WordNet and SentiWordNet.

(6) According to the following formula, the word w _d,n is generated.

Step S3 includes the following sub-steps:

S31. Directly set the aspect of the network comment base as the seed word.

The aspect of the online review database is directly set as the seed word, such as {place, staff, cook, service, decor} in sentence 1, and the seed word is recorded as w _t , t∈{1,...,T}. When setting a seed word, it is necessary to clearly indicate the semantics of the word in WordNet, that is, to determine the semantic interpretation _{st, k0} of the seed word _st in WordNet. When the semantics of various subwords are determined, the seed words can be used as topics.

S32. Divide the comment texts in the online comment database into sentence units to form a set of comment text sentences.

S33. Based on the similarity between the word and the seed word, set different parameters α _d for each sentence; based on the semantic similarity between the word and the seed word, set the aspect-level object word, positive comments for each topic separately The parameters β _t,A , β _t,P , β _t,N are respectively set for words and negative comment words.

β _{t, A} =sim(w, A)*β _base

β _t,P =sim(w,P)*β _base

β _{t, N} =sim(w, N)*β _base

Among them, N _d is the number of all words in the current sentence, T is the number of topics, w _{d, i} is the ith word in the current sentence, t is the seed word, and sim(w, t) represents w and the seed word t Semantic similarity, α _base represents the fixed value parameter α of the topic subject to Dirichlet distribution in the standard LDA model. sim(w, A) indicates that w belongs to the object word, sim(w, P) indicates the probability that w belongs to a positive word, sim(w, N) indicates the probability that w belongs to a negative word, and β _base is the standard LDA model. The fixed value parameter β of the Lickley distribution.

S34. Using Gibbs sampling, parameter estimation and inference for the improved LDA model based on SentiWordNet and WordNet.

The training process is as follows:

(1) For each word in each sentence in the corpus, a topic number is randomly assigned, and the values of the indicator variable y and the indicator variable v are randomly set for all the words in the sentence.

y ∈ {A, O}, v ∈ {P, N}. Here, y and u are digitized, A and P correspond to 0, and O and N correspond to 1, that is, y∈{0,1}, v∈{0,1}.

(2) Rescan the corpus, for each word, resample and update its topic number according to formula 1, update the number of the word in the corpus, and resample the indicator variable y and the indicator variable v according to formulas 2 and 3, and update it.

The Gibbs sampling formula for the SWLDA model is as follows:

The right side is actually p(topic|doc)*p(word|topic), and this probability is actually the path probability of doc→topic→word.

表示非i。in,

means non-i.

(3) Repeat the resampling of the above corpus until the Gibbs sampling converges.

统计语料库中各个主题词的分布，得到主题与单词的概率分布

(4) Count the topics of each word in each sentence in the corpus, and obtain the probability distribution of sentence-topic

Calculate the distribution of each subject word in the corpus to obtain the probability distribution of the subject and the word

The formula for calculating the expectation of the Dirichlet distribution of the topic t of sentence d is as follows:

认为语句d的主题为t，即得评论主题-句子信息。when

Considering that the topic of sentence d is t, the comment topic-sentence information is obtained.

为单词w_d，n的主题为t，类别为y的概率，其中，y＝0表示评论对象词，y＝1表示观点词。具体计算方式如下：

is the probability that the topic of word w _{d, n} is t, and the category is y, where y=0 represents the comment object word, and y=1 represents the opinion word. The specific calculation method is as follows:

Taking t as the subject and the words w _{d and n} as the evaluation aspect-level object words, the Dirichlet distribution expectation calculation formula is as follows:

Taking t as the theme and the words w _{d and n} as the evaluation of positive opinion words, the Dirichlet distribution expectation calculation formula is as follows:

Taking t as the theme and the words w _{d and n} as the evaluation of the negative opinion words, the Dirichlet distribution expectation calculation formula is as follows:

时认为单词的主题为t，类别为y，即得到评论主题-单词信息和评论主题-观点词信息。when

When the topic of the word is considered to be t and the category is y, the comment topic-word information and comment topic-opinion word information are obtained.

Step S4. Sort the sampling results, select the top m words that belong to the corresponding evaluation category in the probability ranking, and find the specific sentence according to the inverted index of the word.

A final result is represented by result. result.tpic is the topic information, that is, the set seed word, which can also be used as a comment category word. result.wor holds the original word. result.wordType is the category of the current word (aspect-level comment object word, positive opinion word, negative opinion word). result.sentices is the sentence to which the word belongs. result.prob is the probability that the word becomes a category word under the review category. Generate the top m results for each category under all topics. Based on the inverted index, the sentence to which it belongs can be queried by word, and the information of <subject, comment object, opinion, original sentence> can be obtained.

Those skilled in the art can easily understand that the above are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, etc., All should be included within the protection scope of the present invention.

Claims (9)

1. a review aspect opinion-level mining method based on dictionary improvement LDA model, it is characterized in that, described aspect is the attribute details of the relevant review object talked about in the review text, and the method comprises the following steps: S1. Build an inverted index list based on the original network comment library; S2. Perform stop-word processing on each sentence of the original online comment database, and obtain the pre-processed network comment database; S3. Input the preprocessed network comment database into the improved LDA model based on the dictionary SentiWordNet and WordNet, and use Gibbs sampling to obtain the sampling result; S4. Sort the sampling results, select the top m words belonging to the probability ranking of the corresponding evaluation category, and find the specific sentence according to the inverted index of the word; Step S3 includes the following sub-steps: S31. Directly set the aspect of the online comment base as the seed word; S32. Divide the comment text in the online comment database into sentence units to form a set of comment text sentences; S33. Based on the similarity between the word and the seed word, set different parameters α _d for each sentence; based on the semantic similarity between the word and the seed word, set the aspect-level object word, positive comments for each topic separately Set parameters β _t,A , β _t,P , β _t,N for words and negative comment words respectively; S34. Using Gibbs sampling comment text sentence set, perform parameter estimation and reasoning on the improved LDA model based on dictionary SentiWordNet and WordNet. 2. The method of claim 1, wherein step S1 comprises the following substeps: S11. Number each word in each sentence in the original online comment database in the form of a binary group <a, b>, where a represents the number of the sentence in which it is located, and b represents the number of the word in the sentence; S12. Remove the repeated words in the original online comment database, and record the number of the remaining words; S13. Based on the deduplicated word numbers, generate an inverted index list. 3. The method of claim 1, wherein

β _{t, A} =sim(w, A)*β _base β _t,P =sim(w,P)*β _base β _{t, N} =sim(w, N)*β _base Among them, N _d is the number of all words in the current sentence, T is the number of topics, w _{d, i} is the ith word in the current sentence, t is the seed word, and sim(w, t) represents w and the seed word t Semantic similarity, α _base represents the fixed value parameter α of the topic obeying Dirichlet distribution in the standard LDA model; sim(w, A) represents the probability that w belongs to the object word, sim(w, P) represents that w belongs to the positive word Probability, sim(w, N) represents the probability that w belongs to a negative word, and β _base is the fixed value parameter β of the word obeying Dirichlet distribution in the standard LDA model. 4. The method of claim 1, wherein step S34 comprises the following steps: (1) When each word in each sentence in the corpus is randomly assigned a topic number, the values of the indicator variable y and the indicator variable v are randomly set for all the words in the sentence. When y=A, it means that the current word is The aspect-level object of the comment, when y=O, it means that the current word is a comment opinion, when v=P, it means that the current word is a positive emotion, and when v=N, it means that the current word is a negative emotion; (2) Rescan the corpus, for each word, resample and update its topic number according to formula (1), update the number of the word in the corpus, and change the indicator variables y and indicator according to formulas (2) and (3). The variable v is resampled and updated;

Among them, z _{d, n} indicates the topic of the d-th comment sentence, t indicates the topic number, y _{d, n} indicates that the n-th word of the d-th sentence is an aspect-level object word or sentiment opinion word, and v _{d, n} indicates the d-th sentence. The nth word of each sentence is a positive emotion word or a negative emotion word,

represents the number of words v whose topic is t and whose category is q,

represents the parameters of the Dirichlet distribution of word v with topic t and category q,

Represents the number of words v whose topic is t and whose category is u,

Represents the Dirichlet distribution parameter of the word v with the topic t and the category u, y represents the number of words in the corpus, w _{d, n} represents the nth word in the dth comment sentence, then

Represents the number of words w _{d and n} whose topic is t and whose category is u,

represents the parameters of the Dirichlet distribution of words w _{d, n} with topic t and category u,

Represents the number of words w _{d, n} whose topic is t and whose category is q,

Represents the Dirichlet distribution parameters of the word w _{d, n} with the topic t and the category q, n _{d, t} represents the number of words with the topic t of the d-th sentence, α _{d, t} represents the d-th comment sentence The topic is the parameters of the Dirichlet distribution of t,

means non-i; (3) Repeat the resampling of the above corpus until the Gibbs sampling converges; (4) Count the topics of each word in each sentence in the corpus, and obtain the probability distribution of sentence-topic

Calculate the distribution of each subject word in the corpus to obtain the probability distribution of the subject and the word

5. The method of claim 4, wherein The formula for calculating the probability distribution of sentence d and topic t is as follows:

Taking t as the subject, and the words w _{d and n} as the evaluation aspect-level object words, the probability distribution calculation formula is as follows:

Taking t as the theme and the words w _{d and n} as the evaluation, the probability distribution of positive opinion words is calculated as follows:

Taking t as the theme and the words w _{d and n} as the evaluation, the probability distribution of negative opinion words is calculated as follows:

Among them, n _d represents the number of words in the d-th sentence. 6. method as claimed in claim 4 is characterized in that, the document generation process that improves LDA model based on dictionary SentiWordNet and WordNet is as follows: (1) Sampling from the Dirichlet distribution β _t,A , β _t,P , β _t,N to generate an aspect-level object distribution of reviews

Positive opinion word distribution for reviews

Negative opinion word distribution of reviews

(2) For each sentence, sample from Dirichlet distribution α _d to generate topic distribution θ _d ~Dir(α _d ); (3) Sampling from the multinomial distribution θ _d of the topic, and generating the topic z _d,n ~Multi(θ _d ) of the words w _{d and n} in the sentence d; (4) Calculated by the dictionary WordNet and SentiWordNet to obtain the Bernoulli distribution on {0, 1} with parameters π _{d, n} and the Bernoulli distribution on {0, 1} with parameters Ω _{d, n} ; (5) Extract from the Bernoulli distribution on {0, 1} with parameters π _{d, n} , and obtain the indication word w _{d, n} is the aspect-level object word of the review or the review opinion word y _{d, n} , with the parameters as Ω _d,n is extracted from the Bernoulli distribution on {0, 1}, and the indicator word w _d,n is obtained as a positive opinion comment word or a negative opinion comment word v _d,n ; (6) According to the following formula, generate the word w _{d, n}

7. The method of claim 6, wherein step (4) comprises the following steps: (4.1) Query the semantic interpretation of the current word w _{d, n} in WordNet as s _{d, n, k} , calculate the semantics of w _{d, n and} the semantics s _{d, n, k} of each subword w _{t , k0} The similarity Sim(s _{d, n, k} , s _{t, k0} ), in all calculation results, take the maximum similarity result, and determine the k' when the result is the largest, which is the current word w _{d, n} in this the semantics s _{d, n, k′} to which the sentence belongs; (4.2) Query the scores of each emotion of s _{d, n, k'} in SentiWordNet

(4.3) Score based on each emotion

Calculate the parameters π _d,n and Ω _d,n . 8. The method of claim 7, wherein

9 . A computer-readable storage medium, characterized in that, a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the computer program according to any one of claims 1 to 8 is implemented. Lexicon-improving opinion-level mining methods for the comment aspect of LDA models.

Images