CN110569355B - Viewpoint target extraction and target emotion classification combined method and system based on word blocks - Google Patents

Abstract

The present invention proposes a combined method and system for word-block-based opinion target extraction and target sentiment classification. Specifically, for each continuous word block, a word-block-level feature is designed to make full use of the overall information among multiple words. ; Calculate the emotional information of each word block instead of calculating the emotional information of each word separately, so as to ensure the consistency of the emotional tendencies of multiple words in the word block. The present invention improves the accuracy of extraction and classification by effectively utilizing the overall information of multiple words, and by calculating an emotional information representation for word blocks composed of multiple words to avoid the problem of emotional inconsistency, thereby improving the accuracy of extraction and classification, and has good practicability. .

Description

A joint method and system for opinion target extraction and target sentiment classification based on word chunks

Technical field:

The invention relates to deep learning and natural language processing technologies, in particular to a combined method and system for word-block-based opinion target extraction and target emotion classification.

Background technique:

In recent years, with the rapid development of Internet information technology, a large amount of new data is generated every day on news, social networking and other websites, and these data contain various information expressing opinions or emotions. Analysis of opinions, positions, attitudes, etc. on these data can help people make better judgments and decisions. For example, by analyzing product reviews, you can understand users' satisfaction with products, so as to formulate more reasonable marketing strategies. . However, since the amount of data on the Internet has grown exponentially, how to find useful data from these massive amounts of information and conduct correct analysis has become a very useful research topic.

Sentiment analysis technology is a technology that analyzes and studies the sentiment tendency of information generated by users. According to the granularity of sentiment analysis, it is mainly divided into: text-level sentiment analysis, sentence-level sentiment analysis and target-level sentiment analysis. The sentiment analysis for the target mainly includes two sub-tasks, one is to find the opinion target in the sentence, and the other is to judge the emotional tendency of the target. The traditional target-based sentiment analysis studies one of the sub-tasks separately, but in practical applications, it is often not only necessary to complete one of the sub-tasks, but also needs to extract the opinion target and classify the target's emotional tendency. A more intuitive approach is to execute the two subtasks in series in a pipelined manner, but this cannot take advantage of the mutual connection between the two subtasks. To fully exploit the connection between these two subtasks, some joint methods based on word-level sequence tagging are proposed. Many opinion targets are composed of multiple words rather than a single word. For example, "left mouse button" is composed of two words "mouse" and "left button", so the joint method based on word-level sequence tagging deals with such opinion targets. However, there are still some limitations when it comes to the analysis. First, it is difficult to use the features of the opinion target level. Second, there may be inconsistencies in the emotional tendencies between the predicted words of the same opinion target. For example, for the "left mouse button" Sentiment target, maybe the label sentiment classification given to the word "mouse" is positive, but the sentiment classification given to the word "left button" is negative.

Invention content:

In view of the above technical problems, the present invention proposes a combined method and system for opinion target extraction and target sentiment classification based on word blocks, to utilize the features of sentiment target levels while avoiding the problem of inconsistency in sentiment classification.

In order to solve the above-mentioned technical problems, the technical scheme of the present invention is as follows:

A joint method of opinion target extraction and target sentiment classification based on word chunks, comprising the following steps:

Process the sentences that need to extract opinions and target sentiment classification, and get all the continuous word blocks in each sentence;

Input the obtained sentence information and word block information into the joint model of opinion target extraction and target sentiment classification, and classify and predict the word block;

According to the category of the word block, the opinion target in the sentence and its corresponding sentiment category are obtained; wherein, the construction method of the joint model of the opinion target extraction and target sentiment classification includes:

(1) Convert the sentences in the training data and the viewpoint targets and target sentiments marked therein into corresponding word blocks and the categories corresponding to the word blocks; at the same time, through large-scale unlabeled corpus, word vectors with semantic information are obtained by training;

(2) Map each word in the sentence of the training data into a corresponding word vector, input a joint neural network model based on word block target extraction and target sentiment classification, and train through a back-propagation algorithm;

(3) Input the sentences that need to extract the opinion target and classify the target sentiment into the trained joint prediction model, predict the category corresponding to each word block, and obtain the opinion target in the sentence and the sentiment of the target according to the category of the word block. category.

Further, the specific method of the construction method step (1) of the joint model includes:

(1-1) Set the maximum length value of the word block, and enumerate all continuous text word blocks that do not exceed the set maximum length in the input sentence of the training data;

(1-2) Mark the categories for all word chunks according to the viewpoint target and target sentiment classification of sentences in the marked corpus;

(1-3) Use traditional word vectors and context-based word vectors to represent each word in the input sentence: obtain traditional word vectors through word2vec training on a large-scale unlabeled corpus, and input the sentence into the pre-trained Context-based word vectors are obtained from the ELMo (Embeddings from Language Models) model.

Further, the setting range of the maximum length value N in step (1-1) is 1≤N≤L, where N is an integer, L is the maximum length of the input sentence, and the preferred length is 4.

Further, in step (1-2), the category set of the word block is defined as 4 categories {TPOS, TNEG, TNEU, O}, and the meanings of these 4 categories are: TPOS indicates that the word block is an opinion target and its Sentiment orientation is positive, TNEG means that the chunk is an opinion target and its sentiment orientation is negative, TNEU means that the chunk is an opinion target and its sentiment orientation is neutral, and O means that the chunk is not an sentiment target.

Further, the model input described in the construction method step (2) of the joint model is:

A sentence X={w ₁ ,w ₂ ,...,w _T }, which contains T words, where w _t represents the t-th word in the input sentence, and the goal of the model is to predict the category Y of each word block in the word block set ={(i,j,l)|1≤i≤j≤T; j-i+1≤L; l∈C}, where i, j represent the starting position and ending position of the word block in the sentence, l Represents the label corresponding to the word chunk, and C represents the category set.

Further, the model training process described in step (2) of the construction method of the joint model includes:

2-1) splicing the traditional word vector of the above-mentioned input sentence and the word vector based on the context, as the input of the next layer;

2-2) In the context representation layer, the word vector corresponding to each word in the sentence is used as input, and a multi-layer bidirectional long short-term memory neural network (stacked Bi-LSTM) is used to learn the context representation vector of each word in the sentence;

2-3) Each word chunk is represented by two chunk-level information: one is the boundary information of the word chunk, and the other is the overall information of the word chunk;

2-4) A word chunk-based attention mechanism is used to calculate the sentiment information in the context associated with each word chunk in the input sentence;

2-5) At the output layer, the vector representation of each word block and the contextual emotion information representation based on the word block attention mechanism are spliced together to predict the category of the word block;

2-6) Select cross entropy as the loss function of model training;

2-7) Train the model through the back-propagation algorithm, update all the parameters in the model, and finally obtain the word block classification model.

Further, the context described in step 2-2) is represented as:

表示第m层t个隐藏单元的状态，第M层的隐藏层状态

作为每个词的上下文表示。in,

Represents the state of the t hidden unit of the mth layer, the hidden layer state of the Mth layer

as a contextual representation of each word.

Further, the boundary information described in step 2-3) is represented by the output of the stack BiLSTM layer corresponding to the boundary word; the overall information is represented by the context information and of all words in the word block; any word block ( i,j) chunks are represented as:

是多层双向长短记忆神经网络的输出，这两类信息的表示向量拼接起来来作为词块的表示。in,

It is the output of the multi-layer bidirectional long-short-term memory neural network, and the representation vectors of the two types of information are spliced together as the representation of the word block.

Further, in step 2-4), since the task not only needs to identify whether the word block is a viewpoint target, but also needs to determine the emotional information corresponding to the word block, and these emotional information are often in the context, so the use of continuous word based block attention mechanism to calculate the sentiment information related to the target in the text. Intuitively, words that are closer to a continuous word block may have a greater impact on the continuous word block. The context information based on distance weights is used to represent this impact, and the weights set for words that are closer to the word block are set. The larger the value, the less weight is set for words farther away from it.

Step 2-4) The specific method includes:

2-4-1) The distance between each word and the word block (i, j) to define the weight w' _t of the word:

where lt represents the distance from the _t -th word to the word block (i, _j ); for the words in the word block, set the value of the distance lt to 0; for the words on the left side of the word block, the distance _lt is the distance to the most word block. The distance of the left word; for the word on the right side of the word block, the distance _lt is the distance to the rightmost word of the word block;

2-4-2) According to the weight value obtained above, the model calculates the context representation of each word block (i, j) based on the position weight:

表示第t个词的上下文；where e _t is the distance weight-based representation corresponding to the t-th word, w′ _t represents the weight value of the t-th word,

Represents the context of the t-th word;

2-4-3) Use the chunk-based attention mechanism to calculate the context information related to the chunk, including:

a) For word block (i, _j ), calculate each context word et and its relevance weight, the calculation formula is:

为第t个词和词块(i，j)的相关相关程度向量；where s _{(i, j)} is the vector representation of the word block (i, j), e _t is the distance weight-based representation corresponding to the t-th word, W _a is the weight matrix, b _a is the offset vector,

is the correlation degree vector of the t-th word and the word block (i, j);

b) Multiply these weights with these word context representation vectors and weight them to obtain the context information representation c _{(i, j) related to the word block (i, j)} :

Further, in step 2-5), the vector representation of the word block (i, j) obtained above and its related context representation are spliced together to predict the category of the word block (i, j):

r _(i,j) =[s _(i,j) ;c _(i,j) ]

_y =softmax(W _y r _(i,j) +by )

Among them, y is the probability distribution of the word block corresponding to each category, W _y is the weight matrix, and b _y is the offset vector.

The classification of word blocks is processed. Those in category O are considered not to be opinion targets, those in category TPOS are those with positive sentiment in the sentence, and those in category TNEG are those with negative sentiment in sentences. The opinion target of , in the category TNEU is the opinion target whose sentiment tendency is neutral in the sentence.

Further, for all training samples in step 2-6), the model is trained by maximizing the maximum likelihood function of the sample, and the parameters in the model are updated, and the training objective function loss is defined as follows:

where _gi is the vector representation of the true classification corresponding to the word block, _yi is the predicted probability distribution, _H represents all the training sentences, and Sh represents all the word blocks in the hth sentence.

A joint system of opinion target extraction and target sentiment classification based on word chunks, including:

Data preprocessing module to be tested: process the sentences that need to be extracted and classified to target sentiment, and get all the continuous word blocks in each sentence;

Word block classification prediction module: input the obtained sentence information and word block information into the joint model of opinion target extraction and target sentiment classification, and classify and predict the word block; the joint model maps the input sentence to a word vector as a network Input, build a joint neural network model based on the word block sequence, and use the back propagation algorithm to train the result;

Result acquisition module: According to the category of the word block, the opinion target in the sentence and its corresponding sentiment category are obtained.

The beneficial effects of the present invention are: for the two problems in the method of opinion target extraction and target sentiment tendency classification based on word-level sequence labeling: first, it is difficult to use the overall information of opinion targets composed of multiple words; There may be emotional inconsistencies among opinion targets composed of words; a joint model based on continuous word blocks is proposed to simultaneously extract opinion targets and classify target sentiment tendency. Specifically, for each continuous word block, the block-level features are designed to In this way, the overall information between multiple words can be fully utilized; the sentiment information of each word block is calculated instead of separately calculating the sentiment information of each word, so as to ensure the consistency of the emotional tendencies of multiple words in the word block. In this way, the present invention improves the accuracy of extraction and classification by effectively utilizing the overall information of multiple words, and by calculating an emotional information representation for word blocks composed of multiple words to avoid the problem of emotional inconsistency, thereby improving the accuracy of extraction and classification. practicality.

Description of drawings:

FIG. 1 is a flowchart of opinion target extraction and target sentiment classification based on word block level provided by an embodiment of the present invention;

FIG. 2 is a structural diagram of a neural network model according to an embodiment of the present invention.

Detailed ways:

In order to make the above objects, features and advantages of the present invention more obvious and easy to understand, the present invention will be further described in detail below through specific implementation cases and in conjunction with the accompanying drawings.

Fig. 1 is a flowchart of the method for the flow chart of opinion target extraction and target sentiment classification based on the word block level in this embodiment. As shown in the figure, the method mainly includes three stages, which are: a data preprocessing stage, based on the word block In the training stage of the joint model training stage of high-level opinion target extraction and target sentiment classification, the predicted classification is matched to obtain the opinion target in the sentence and its sentiment tendency category stage.

(1) Data preprocessing stage

Step 1: For each sentence, exhaust all consecutive word blocks in it (the length of the word block sets an upper limit). According to the opinion target and target sentiment classification data given in the annotated corpus, the classification of all word blocks is obtained. There are 4 categories of classification: TPOS indicates that the word chunk is an opinion target and the sentiment category of the target is positive, TNEG indicates that the word chunk is an opinion target and the sentiment category of the target is negative, TNEU indicates that the word chunk is an opinion target and the sentiment category for that target is neutral, O indicates that the chunk is not an opinion target. For example, for the sentence "hard disk is broken", the upper limit of the length is set to 2, then all consecutive word blocks and their corresponding labels are "(hard, O), (disk, O), (bad, O), (now, O ), (hard disk, TNEG), (disk bad, O), (bad, O)".

Step 2: Using unlabeled corpus, the word vector representation with semantic information is obtained through word2vec training and provided to the model for use.

(2) Model training stage

Combined with Figure 2, the joint model of opinion target extraction and target sentiment classification based on word block level includes the following specific steps:

Step 1, formalize the output and output, the input is a sentence containing T words X={w ₁ ,w ₂ ,...,w _T }, where w _t represents the t-th word in the input sentence, and the goal is to predict the word block The category of each chunk in the set Y = {(i,j,l)|1≤i≤j≤T; j-i+1≤L; l∈C}, where i, j represent the chunk in the sentence The starting position and ending position of , l represents the label corresponding to the word block, and C represents the category set;

Step 2, using the word vector representation obtained by word2vec training and the pre-trained language model, map each word in the input sentence into the corresponding word vector two types of word vectors, and the word vectors of these two categories are processed. splicing;

Step 3, the context representation layer, takes the word vector corresponding to each word in the sentence as the input, and uses the M-layer bidirectional long short-term memory neural network (stackedBi-LSTM) to learn the context information of each word in the input sentence, where the mth (m The calculation formula of the hidden layer state of the ∈{1,…,M}) layer is as follows:

表示第m层t个隐藏单元的状态。把第M层的隐藏层状态

作为每个词的上下文表示；in,

represents the state of the t hidden unit in the mth layer. Put the hidden layer state of the Mth layer

as a contextual representation of each word;

Step 3: The representation layer of the word block, for any word block (i, j), it is represented by two kinds of word block level information: one is the boundary information of the word block, and the other is the overall information of the word block . The boundary information is directly obtained by the vector representation of the boundary words obtained by the previous layer, and the overall information is the sum of the representations of all words in the word block obtained by the previous layer:

是多层双向长短记忆神经网络的输出。把这两类信息的表示向量拼接起来来作为词块的表示；in,

is the output of a multilayer bidirectional long short-term memory neural network. The representation vectors of the two types of information are spliced together as the representation of the word block;

Step 4 is based on the attention mechanism of word blocks, because the model not only needs to predict whether a word block is an emotional target, but also needs to predict the emotional tendency of the target in the sentence, and these emotional information often exist in the context of the word block, in order to obtain and The emotional information related to word chunks, and a word chunk-based attention mechanism is proposed to learn these relevant information;

Step 4-1, intuitively speaking, the closer the context word is to a word block, the greater the influence on the word block. The model uses a context based on location weights to simulate this effect. First, according to the distance between each word and the word block (i, j), the weight w' _t of the word is defined:

where l _t represents the distance from the t-th word to the word block (i, j). For the words in the word block, set the value of the distance _lt to 0; for the words on the left side of the word block, the distance _lt is the distance to the leftmost word of the word block; for the words on the right side of the word block, the distance _lt is the distance to the word block The distance of the rightmost word.

Step 4-2 According to the weight value obtained above, the model calculates the context representation of each word block (i, j) based on the position weight:

表示第t个词的上下文。where e _t is the distance weight-based representation corresponding to the t-th word, w′ _t represents the weight value of the t-th word,

Represents the context of the t-th word.

In step 4-3, after obtaining the context representation based on the position weight, the word block-based attention mechanism is used to calculate the context information related to the word block. First, for the word block (i, _j ), calculate each context word et and its relevance weight, the calculation formula is:

为第t个词和词块(i，j)的相关相关程度向量。把这些权重和这些词上下文表示向量相乘并进行加权，即可得到和词块(i，j)相关的上下文信息表示c_(i,j)：where s _{(i, j)} is the vector representation of the word block (i, j), e _t is the distance weight-based representation corresponding to the t-th word, W _a is the weight matrix, b _a is the offset vector,

is the correlation degree vector of the t-th word and the word block (i, j). Multiply these weights with these word context representation vectors and weight them, and then the context information representation c _(i, j) related to the word block (i, j) can be obtained:

Step 5, the output layer, concatenates the vector representation of the word block (i, j) obtained above and its related context representation to predict the category of the word block (i, j):

r _(i,j) =[s _(i,j) ;c _(i,j) ]

_y =softmax(W _y r _(i,j) +by )

Among them, y is the probability distribution of the word block corresponding to each category, W _y is the weight matrix, and b _y is the offset vector.

Step 6: For all training samples, the model is trained by maximizing the maximum likelihood function of the samples, and the parameters in the model are updated. The training objective function loss is defined as follows:

where _gi is the vector representation of the true classification corresponding to the word block, _yi is the predicted probability distribution, _H represents all the training sentences, and Sh represents all the word blocks in the hth sentence.

(3) Result processing stage

Step 1: Process the sentences that need to extract the viewpoint target and classify the target sentiment, and obtain all the continuous word blocks in each sentence;

Step 2, input the obtained sentence information and word block information into the joint model of opinion target extraction and target sentiment classification obtained above, and classify and predict the word block;

Step 3, according to the category of the word block, the viewpoint target in the sentence and its corresponding sentiment category are obtained

It can be seen from the above scheme that this scheme addresses two problems in the joint target extraction and target sentiment classification model based on single-word-level sequence labeling: firstly, the overall information of multiple words that make up the target cannot be used; There may be inconsistencies in sentiment classification. A joint extraction and classification model based on word blocks is proposed, which can improve the performance of model prediction and has good practicability.

The above are only the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the concept of the present invention, several improvements and modifications can also be made, and these improvements and modifications should also be regarded as are within the protection scope of the present invention.

Claims (10)

1. A joint method of opinion target extraction and target sentiment classification based on word chunks, comprising the following steps: Process the sentences that need to extract opinions and target sentiment classification, and get all the continuous word blocks in each sentence; Input the obtained sentence information and word block information into the joint model of opinion target extraction and target sentiment classification, and obtain the probability distribution of each category corresponding to each word block in the sentence; According to the probability distribution of each category corresponding to each word block, the opinion target in the sentence and its corresponding emotion category are obtained; Wherein, the construction method of the joint model of opinion target extraction and target sentiment classification includes: (1) Obtain the word category label of each word in the training data input sentence and the word block category label of each word block, and input the quantity data into the sentence and the corresponding word block information into the joint model; (2) Using the word vector matrix obtained by word2vec training and the pre-trained language model, map each word into a traditional word vector and a context-based word vector, and perform vector splicing to generate the word corresponding to each word vector; (3) Based on the word vector, word category label and word block information corresponding to each word, the joint model outputs the probability distribution of each category corresponding to each word block; (4) Calculate the value of the cross-entropy loss function based on the probability distribution of each category and the corresponding word block category label to perform reverse gradient propagation to obtain model parameters to construct a joint model of opinion target extraction and target sentiment classification. 2. a kind of viewpoint target extraction and target sentiment classification joint method based on word block as claimed in claim 1, is characterized in that, the construction method step (1) concrete method of joint model comprises: (1-1) Set the maximum length value of the word block, and enumerate all continuous text word blocks that do not exceed the set maximum length in the input sentence of the training data; the setting range of the maximum length value N is 1≤N ≤L, where N is an integer and L is the maximum length of the input sentence; (1-2) Define the category set of word chunks as 4 categories {TPOS, TNEG, TNEU, O}, where TPOS indicates that the word chunk is an opinion target and its sentiment tendency is positive, TNEG indicates that the word chunk is an opinion target and Its sentiment tendency is negative, TNEU indicates that the word block is an opinion target and its sentiment tendency is neutral, and O indicates that the word block is not an emotional target; Tag category. 3. A joint method of opinion target extraction and target sentiment classification based on word chunks according to claim 2, characterized in that, the training data input sentence X={w ₁ , w ₂ ,...,w _t ... , w _T }, where w _t represents the t-th word in the input sentence; the goal of the model is to predict the category of each chunk in the set of chunks Y={(i, J, l)|1≤i≤j≤ T; j-i+1≤L; l∈C}, where i, j represent the start and end positions of the word block in the sentence, l represents the label corresponding to the word block, and C represents the category set of the word block. 4. a kind of viewpoint target extraction and target sentiment classification joint method based on word block as claimed in claim 3, it is characterized in that, obtain each class probability distribution corresponding to each word block by the following steps: 3-1) In the context representation layer, the word vector corresponding to each word in the sentence is used as input, and a multi-layer bidirectional long-short memory neural network is used to learn the context representation vector of each word in the sentence; 3-2) Each word chunk is represented by two chunk-level information: one is the boundary information of the word chunk, and the other is the overall information of the word chunk; 3-3) A word chunk-based attention mechanism is used to calculate the sentiment information in the context associated with each word chunk in the input sentence; 3-4) In the output layer, the vector representation of each word block and the contextual emotion information representation based on the word block attention mechanism are spliced together, and the probability distribution of each category corresponding to each word block is obtained. 5. a kind of viewpoint target extraction and target sentiment classification joint method based on word block as claimed in claim 4, is characterized in that, the context described in step 3-1) is expressed as:

in,

Represents the state of the t-th hidden unit in the m-th layer, the hidden layer state of the M-th layer

As the contextual representation of each word, where m∈{1,...,M}, M is the number of layers of the multi-layer bidirectional long short-term memory neural network. 6. a kind of viewpoint target extraction and target sentiment classification joint method based on word block as claimed in claim 4, is characterized in that, described in step 3-2), boundary information is carried out with the output of the stack BiLSTM layer corresponding to boundary word Representation; the overall information is represented by the context information sum of all words in the word block; any word block (i, j) word block is represented as:

in,

It is the output of the multi-layer bidirectional long-short-term memory neural network, and the representation vectors of the two types of information are spliced together as the representation of the word block. 7. a kind of viewpoint target extraction and target emotion classification joint method based on word block as claimed in claim 4, is characterized in that, step 3-3) concrete method comprises: 3-3-1) The distance between each word and the word block (i, j) to define the weight w't of the word:

where lt represents the distance from the _t -th word to the word block (i, _j ); for the words in the word block, set the value of the distance lt to 0; for the words on the left side of the word block, the distance _lt is the distance to the most word block. The distance of the left word; for the word on the right side of the word block, the distance _lt is the distance to the rightmost word of the word block; 3-3-2) According to the weight value obtained above, the model calculates the context representation of each word block (i, j) based on the position weight:

where e _t is the distance weight-based representation corresponding to the t-th word, w′ _t represents the weight value of the t-th word,

Represents the context of the t-th word; 3-3-3) Use the chunk-based attention mechanism to calculate the context information related to the chunk, including: a) For word block (i, _j ), calculate each context word et and its relevance weight, the calculation formula is:

where s _{(i, j)} is the vector representation of the word block (i, j), e _t is the representation based on the distance weight corresponding to the t-th word, W _α is the weight matrix, b _α is the offset vector,

is the correlation degree vector of the t-th word and the word block (i, j); b) Multiply these weights with these word context representation vectors and perform weighting to obtain the context information representation c _{(i, j) related to the word block (i, j)} :

8. a kind of viewpoint target extraction and target sentiment classification joint method based on word block as claimed in claim 4 is characterized in that, in step 3-4), the vector representation of above-mentioned word block (i, j) obtained and with Its related contextual representations are concatenated together to predict the class of the chunk (i, j): r _(i,j) =[s _(i,j) ;c _(i,j) ] _y =softmax(W _y r _{(i, j)} +by ) Among them, y is the probability distribution of the word block corresponding to each category, W _y is the weight matrix, and b _y is the offset vector. 9. a kind of opinion target extraction and target sentiment classification joint method based on word block as claimed in claim 1, is characterized in that, in step (4), to all training samples, train the model by maximizing the maximum likelihood function of the sample , update the parameters in the model, the training objective function loss is defined as follows:

Among them, _gi is the vector representation of the true classification corresponding to the word block, _yi is the predicted probability distribution, _H represents all the training sentences, and Sh represents all the word blocks in the hth sentence. 10. A joint system of opinion target extraction and target sentiment classification based on word chunks, comprising: Data preprocessing module to be tested: process the sentences that need to be extracted from viewpoints and classified by target sentiment, and get all the consecutive word blocks in each sentence; Word block classification prediction module: input the obtained sentence information and word block information into the joint model of opinion target extraction and target sentiment classification, and obtain the probability distribution of each category corresponding to each word block in the sentence; Result acquisition module: According to the probability distribution of each category corresponding to each word block, the opinion target in the sentence and its corresponding emotion category are acquired; Wherein, the construction method of the joint model of opinion target extraction and target sentiment classification includes: (1) Obtain the word category label of each word in the input sentence of the training data and the word block category label of each word block, and input the quantity data into the sentence and the corresponding word block information into the joint model (2) Using the word vector matrix obtained by word2vec training and the pre-trained language model, map each word into a traditional word vector and a context-based word vector, and perform vector splicing to generate the word corresponding to each word vector; (3) Based on the word vector, word category label and word block information corresponding to each word, the joint model outputs the probability distribution of each category corresponding to each word block; (4) Calculate the value of the cross-entropy loss function based on the probability distribution of each category and the corresponding word block category label to perform reverse gradient propagation to obtain model parameters to construct a joint model of opinion target extraction and target sentiment classification.

Images