WO2023103914A1

WO2023103914A1 - Text sentiment analysis method and device, and computer-readable storage medium

Info

Publication number: WO2023103914A1
Application number: PCT/CN2022/136328
Authority: WO
Inventors: 夏睿; 李成路; 周祥生; 董修岗; 孙文卿
Original assignee: 中兴通讯股份有限公司
Priority date: 2021-12-07
Filing date: 2022-12-02
Publication date: 2023-06-15
Also published as: CN114201957A

Abstract

The present application relates to the field of big data, and discloses a text sentiment analysis method and device, and a computer-readable storage medium. The text sentiment analysis method comprises: obtaining each word in target text (S101); obtaining word vector representation of each word (S102); obtaining implicit dependency syntactic structure information representation of each word (S103); splicing the word vector representation and the implicit dependency syntactic structure information representation of each word to obtain an input matrix (S104); and inputting the input matrix into an attribute sentiment analysis model to obtain attribute sentiment classification of the target text (S105).

Description

Text sentiment analysis method, device and computer-readable storage medium

Cross References to Related Applications

This application is based on a Chinese patent application with application number 202111486407.6 and a filing date of December 07, 2021, and claims the priority of this Chinese patent application. The entire content of this Chinese patent application is hereby incorporated by reference into this application.

technical field

The present application relates to the field of big data, in particular to a text sentiment analysis method, device and computer-readable storage medium.

Background technique

The rapid development of the Internet era has led to a massive increase in various types of data, and text data, as an important carrier of people's communication and expression, contains a large amount of valuable information, among which rich user emotional information is reflected in the text. However, the Internet environment is complex, and the amount of data can be imagined. How to intelligently and efficiently analyze the value behind these data has become extremely important. Therefore, for Internet comment resources, related research on text sentiment analysis has also followed. Having attention. Text sentiment analysis, also known as opinion mining, is a classic research task in the field of natural language processing.

However, the inventors of the present application found that the accuracy of text sentiment analysis results in some cases is low.

Contents of the invention

Embodiments of the present application provide a text sentiment analysis method, device, and computer-readable storage medium.

The embodiment of the present application provides a text sentiment analysis method, including: obtaining each word in the target text; obtaining the word vector representation of each word; obtaining the implicit dependency syntax structure information representation of each word; The word vector representation of the word and the implicit dependency syntactic structure information representation are concatenated to obtain an input matrix; the input matrix is input into an attribute sentiment analysis model to obtain the attribute emotion classification of the target text.

Embodiments of the present application also provide a text sentiment analysis device, including: at least one processor; and a memory connected in communication with the at least one processor; wherein, the memory stores information that can be processed by the at least one processor. Instructions executed by a processor, the instructions are executed by the at least one processor, so that the at least one processor can execute the text sentiment analysis method as described above.

Embodiments of the present application also provide a computer-readable storage medium storing a computer program, and implementing the aforementioned text sentiment analysis method when the computer program is executed by a processor.

Description of drawings

Fig. 1 is a program flow chart of the text sentiment analysis method provided by an embodiment of the present application;

Fig. 2 is a schematic diagram of the operational flow of the Biaffine parser model in the text sentiment analysis method provided by an embodiment of the present application;

Fig. 3 is a program flow chart of the steps of obtaining the attribute sentiment classification of the target text in the text sentiment analysis method provided by an embodiment of the present application;

Fig. 4 is a schematic diagram of the operation flow of the attribute-level sentiment analysis model in the text sentiment analysis method provided by an embodiment of the present application;

5 is a program flow chart of a text sentiment analysis method provided in another embodiment of the present application;

Fig. 6 is a schematic structural diagram of a text sentiment analysis device provided by another embodiment of the present application.

Implementation

In order to make the purpose, technical solution and advantages of the present application clearer, various embodiments of the present application will be described in detail below in conjunction with the accompanying drawings. However, those of ordinary skill in the art can understand that, in each implementation manner of the present application, many technical details are provided for readers to better understand the present application. However, even without these technical details and various changes and modifications based on the following implementation modes, the technical solution claimed in this application can also be realized.

Attribute-level sentiment analysis aims to identify the attribute words that appear in the text first, and then judge the emotional tendency of the entire text for the specific attribute words based on the identified attribute words. In actual application scenarios, attribute-level sentiment analysis has a wide range of application scenarios and room for research and development. In e-commerce platforms represented by Taobao, Amazon, and Dangdang, attribute-level sentiment analysis technology can be used to understand the user's perception of a product. Valuable business information is mined from a large amount of review data. In recent years, many works have begun to mine and utilize the dependency syntactic structure information of text in attribute-level sentiment analysis. In these research works that integrate text-dependent syntactic structure information into attribute-level sentiment analysis tasks, the storage method of dependency syntax tree is Stored in the form of an adjacency matrix.

However, the inventors of the present application have found that in some cases, there are errors in the process of text-dependent syntactic analysis, and the existence of this error leads to a lower accuracy of the final text sentiment analysis result.

In order to solve this technical problem, an embodiment of the present application relates to a text sentiment analysis method, as shown in Figure 1, at least including the following steps:

Step S101: Obtain each word in the target text.

In this embodiment, firstly, the target text is segmented into individual words according to the word segmentation model. The target text can be the result text of Optical Character Recognition (OCR) (that is, OCR text), or it can be ordinary text, that is, as long as It only needs to be text and has a wide range of applications. The length of the target text is not limited, and at least one word can be obtained after segmentation by the word segmentation model. The word segmentation model can be an N-gram model (n-gram). The N-gram model is a relatively mature model for word segmentation. The n-th item can be inferred based on the first n-1 items, and the word segmentation of the text is more accurate. It can be understood that the foregoing is only an example of a method for word segmentation of the target text in this embodiment, and does not constitute a limitation. In other embodiments of the application, other methods can also be used, which can be implemented according to actual needs. Flexible settings, for example, when the target text is English text, NLTK (Natural Language Toolkit, Natural Language Toolkit) can also be used to perform precise word segmentation on the target text and remove stop words in the target text.

Step S102: Obtain the word vector representation of each word.

In this embodiment, a pre-trained GloVe (Global Vectors, global vector) word embedding model is used to convert each word into its corresponding 300-dimensional vector. That is, there is a one-to-one correspondence between each word in the target text and its word vector representation. It can be understood that the foregoing is only an example of obtaining the word vector representation of each word in this embodiment, and does not constitute a limitation. In other embodiments of the application, the Skip-gram model or continuous Word bag model (ContinuousBag-of-Words) and other algorithm models to obtain the word vector representation of each word, or use the Skip-gram model and the continuous word bag model to obtain the word vector representation of each word, etc., according to A flexible setting is actually required.

Step S103: Obtain the implicit dependency syntax structure information representation of each word.

In this embodiment, each word in the target text is input into the deep neural dependency syntactic analysis model, the hidden state representation generated by the deep neural dependency syntactic analysis model in the encoding stage is obtained, and the implicit dependency syntactic structure of each word is obtained according to the hidden state representation information representation. The implicitly dependent syntactic structure information of each word refers to the hidden state generated by the deep neural dependency syntactic analysis model in the encoding layer stage. In this embodiment, the deep neural dependent syntactic analysis model is, for example, the Biaffine parser model, and the Biaffine parser model is Including three layers of Bi-LSTM (Long Short-Term Memory, two-way long-short-term memory) network, as shown in Figure 2, in this embodiment, in the coding phase, the three layers of Bi-LSTM network of Biaffine parser model extracts the content of target text feature information and output the hidden state of the target text, and the hidden state representation of each word in the target text can be regarded as a representation with implicitly dependent syntactic structure information. In the decoding stage, a matrix with decimals on the right side of Figure 2 can be obtained. The value in this matrix represents the biaffine score between words, that is, the probability value of a dependency edge between two words. The formed matrix is called It is a probability matrix (Probability Matrix), which can also be regarded as a weighted directed graph. In some cases, after decoding the probability matrix using the MST algorithm, the dependency syntax tree at the top of Figure 2 is obtained, also known as the best dependency syntax tree (1-best dependency tree), and the right element values are only discrete with 0 and 1 The matrix is the adjacency matrix representation of the dependency syntax tree. For example, there are 5 words in the input text in Figure 2, where the "$" symbol represents the root node of the dependency syntax tree, which is used to point to the predicate in the text. The size of the adjacency matrix is 5×5. Each row indicates whether there is a dependency edge between the word and other words. If there is a dependency edge between words, the value of the element at the corresponding position is 1, otherwise it is 0. For example, "like" has A dependency edge pointing to "eating", and the positions of the two words in the original text are 3 and 4 respectively, then the value of the element in the third row and fourth column in the adjacency matrix is 1. According to this rule, the Element values elsewhere in the adjacency matrix.

It can be understood that, in this embodiment, the hidden state generated by the deep neural dependency syntax analysis model at the encoding layer stage is directly used as the implicit dependency syntax structure information of each word, without constructing the dependency syntax tree.

Step S104: Concatenate the word vector representation of each word and the implicit dependency syntax structure information representation to obtain an input matrix.

In this embodiment, the word vector representation of each word is directly concatenated with its corresponding implicit dependency syntax structure information representation to obtain an input matrix. It can be understood that the foregoing is only an example of a splicing method in this embodiment, and does not constitute a limitation. Formula-dependent syntactic structure information is added to obtain an input matrix, etc., which can be flexibly set according to actual needs, and will not be listed here.

Step S105: Input the input matrix into the attribute sentiment analysis model to obtain the attribute sentiment classification of the target text.

In this embodiment, as shown in Figure 3, at least the following steps are included:

Step S201: The attribute sentiment analysis model obtains the attribute words in each word according to the input matrix.

In this embodiment, as shown in Figure 4, after the attribute sentiment analysis model obtains the input matrix, the input matrix passes through an attribute mask layer, and the output matrix of the attribute word in the target text can represent the hidden information of the attribute word, Thus, the attribute words are obtained according to the output matrix.

For example, with

To represent the target text, the target text contains n words, from the beginning of the τ+1 word to the end of the τ+k word is the range of attribute words, the number of attribute words is k, and the aforementioned step S101 is performed on the target text The input matrix obtained after processing to step S104 is

The output matrix obtained after passing the input matrix through the attribute mask layer is

The words corresponding to the non-zero values in the output matrix are the attribute words.

Step S202: The attribute sentiment analysis model uses the attention mechanism to extract the context information related to the attribute words.

In this embodiment, the input matrix

and the output matrix

Performing an attention operation can extract the semantic information most related to the attribute word from the context information, that is, get the weight score of each word in the final emotional representation. The calculation formula of the weight score α is as follows:

Step S203: The attribute sentiment analysis model obtains the attribute sentiment classification of the target text according to the context information.

With the weight scores of each attribute word in each context, the hidden representation used as the final attribute sentiment classification is to combine the implicit dependency syntactic structure information of each word in the context and the concatenation value of the word vector according to the weight score. and obtained, and finally use the fully connected Softmax classifier to predict the probability of the emotional category of r, assuming that the emotional prediction probability distribution is represented by p, the calculation formula is as follows:

p＝softmax(Wp _r +b _p )

Wherein W _p and b _p are fixed value parameters, which can be set according to actual needs, and can also be set according to the method of model training.

In the text sentiment analysis method provided by an embodiment of the present application, by obtaining the word vector representation of each word and the representation of implicitly dependent syntactic structure information, the word vector representation of each word and the representation of implicitly dependent syntactic structure information are directly spliced to obtain The input matrix is input into the attribute sentiment analysis model to obtain the attribute sentiment classification of the target text. In this process, the optimal dependency syntax tree is not directly modeled, but the implicit dependency syntax structure information of the target text is used to represent the input Performing sentiment analysis in the attribute sentiment analysis model not only improves the effect of the attribute sentiment analysis model on the attribute-level sentiment analysis data set, but also reduces the error propagation problem caused by the dependency syntax tree, and improves the attribute-level sentiment of the target text Analyze the effect.

An implementation manner of the present application relates to a text sentiment analysis method, as shown in Figure 5, including:

Step S301: Obtain each word in the target text.

Step S302: Obtain the word vector representation of each word.

It can be understood that steps S301 to S302 in the text sentiment analysis method proposed in this embodiment are substantially the same as steps S101 to S102 in the foregoing embodiments, and will not be repeated here, and can refer to the descriptions of the foregoing embodiments.

Step S303: Input each word into the deep neural-dependent syntactic analysis model, and obtain the hidden state representation generated by the deep neural-dependent syntactic analysis model in the encoding stage.

In this embodiment, each word in the target text is input into the deep neural-dependent syntactic analysis model, and the hidden state representation generated by the deep neural-dependent syntactic analysis model in the encoding stage is obtained.

Step S304: Map the hidden state representation through the linear mapping layer to obtain the implicit dependency syntax structure information representation of each word.

In this step, the three-layer hidden state generated by the deep neural dependency parsing model in the encoding layer is denoted as

in

Represents the hidden state of the nth word in the target text in the first layer of the Bi-LSTM network, and passes the hidden state of the output L layer through a linear mapping layer to obtain the final syntax of each word with implicitly dependent syntactic structure information Perceptual word representation, denoted as s={s ₁ ,...,s _n }, the calculation formula of the linear mapping layer is,

Among them, W ^l and b ^l are fixed parameters, which can be set flexibly according to actual needs, for example, through model training, l is any layer in the Bi-LSTM network, and L is the total number of Bi-LSTM networks. layers. That is, in this step, the three-layer hidden state of the target text is linearly mapped and the mapping results are accumulated to obtain a syntax-aware word representation with implicitly dependent syntactic structure information for each word in the text.

Step S305: Concatenate the word vector representation of each word and the implicit dependency syntax structure information representation to obtain the input matrix.

Step S306: Input the input matrix into the attribute sentiment analysis model to obtain the attribute sentiment classification of the target text.

It can be understood that steps S305 to S306 in the text sentiment analysis method proposed in this embodiment are substantially the same as steps S104 to S105 in the foregoing embodiments, and will not be repeated here, and can refer to the descriptions of the foregoing embodiments.

In this embodiment, while retaining the technical effects of the foregoing embodiments, the three-layer hidden state of the target text is linearly mapped and then the mapping results are accumulated, so that each word in the target text finally has implicitly dependent syntactic structure information The result of the syntax-aware word representation is more accurate, which further improves the accuracy of the final text sentiment analysis result.

The division of the steps of the above methods is only for the sake of clarity of description. During implementation, they can be combined into one step or some steps can be split and decomposed into multiple steps. As long as they contain the same logical relationship, they are all within the scope of protection of this patent. ; Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but not changing the core design of the algorithm and process are all within the scope of protection of this patent.

One embodiment of the present application relates to a text sentiment analysis device, as shown in FIG. 6 , including: at least one processor 401; and a memory 402 communicatively connected to at least one processor 401; Instructions executed by at least one processor 401 , the instructions are executed by at least one processor 401 , so that at least one processor 401 can execute the text sentiment analysis method provided in the foregoing embodiments.

Wherein, the memory 402 and the processor 401 are connected by a bus, and the bus may include any number of interconnected buses and bridges, and the bus connects one or more processors 401 and various circuits of the memory 402 together. The bus may also connect together various other circuits such as peripherals, voltage regulators, and power management circuits, all of which are well known in the art and therefore will not be further described herein. The bus interface provides an interface between the bus and the transceivers. A transceiver may be a single element or multiple elements, such as multiple receivers and transmitters, providing means for communicating with various other devices over a transmission medium. The data processed by the processor 401 is transmitted on the wireless medium through the antenna, and the antenna also receives the data and transmits the data to the processor 401 .

Processor 401 is responsible for managing the bus and general processing, and may also provide various functions including timing, peripheral interface, voltage regulation, power management, and other control functions. And the memory 402 may be used to store data used by the processor 401 when performing operations.

An embodiment of the present application relates to a computer-readable storage medium storing a computer program. When the computer program is executed by the processor, the text sentiment analysis method provided in the foregoing embodiments is realized.

That is, those skilled in the art can understand that all or part of the steps in the method of the above-mentioned embodiments can be completed by instructing related hardware through a program, the program is stored in a storage medium, and includes several instructions to make a device ( It may be a single-chip microcomputer, a chip, etc.) or a processor (processor) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disc, etc., which can store program codes. .

Embodiments of the present application provide a text sentiment analysis method, device, and computer-readable storage medium, so that the accuracy of text sentiment analysis results can be improved.

According to the embodiment of the present application, for each word in the target text, by obtaining the word vector representation and implicit dependency syntax structure information representation of each word, the word vector representation of each word and the implicit dependency syntax Structural information indicates that the input matrix is obtained after splicing, and the input matrix is input into the attribute sentiment analysis model to obtain the attribute sentiment classification of the target text. In this process, the optimal dependency syntax tree is not directly modeled, but the implicit structure of the target text is used. Dependent syntactic structure information is input into the attribute sentiment analysis model for sentiment analysis, which not only improves the effect of the attribute sentiment analysis model on the attribute-level sentiment analysis data set, but also reduces the error propagation problem caused by the dependency syntax tree, and improves the performance of the attribute sentiment analysis model. Attribute-Level Sentiment Analysis Effects on Target Text.

Those of ordinary skill in the art can understand that the above-mentioned embodiments are several embodiments of the present application, and in practical applications, various changes can be made in form and details without departing from the spirit and spirit of the present application. scope.

Claims

A text sentiment analysis method, comprising:

Obtain each word in the target text;

Obtain the word vector representation of each word;

Obtaining the implicit dependency syntax structure information representation of each word;

splicing the word vector representation of each of the words and the implicit dependency syntactic structure information representation to obtain an input matrix;

Inputting the input matrix into the attribute sentiment analysis model to obtain the attribute sentiment classification of the target text.
The text sentiment analysis method according to claim 1, wherein said acquisition of the implicit dependency syntax structure information representation of each word comprises:

Input the target text into the deep neural dependency syntactic analysis model, obtain the hidden state representation generated by the deep neural dependency syntactic analysis model in the encoding stage, and obtain the implicit dependency syntactic structure information representation of each word according to the hidden state representation .
The text sentiment analysis method according to claim 2, wherein said acquisition of the hidden state representation produced by the deep neural dependency syntax analysis model in the encoding stage comprises:

The target text is encoded through a three-layer bidirectional long-short-term memory network to obtain the hidden state representation.
The text sentiment analysis method according to claim 3, wherein said obtaining the implicit dependency syntax structure information representation of each word according to said hidden state representation comprises:

The hidden state representation is mapped through a linear mapping layer to obtain the implicit dependency syntax structure information representation of each word.
The text sentiment analysis method according to claim 4, wherein the implicitly dependent syntactic structure information representation of each word is obtained after the hidden state representation is mapped through a linear mapping layer, including:

Map different levels of the hidden state representation through the linear mapping layer to obtain multiple mapping results, and accumulate the multiple mapping results to obtain the implicit dependency syntax structure information representation of each word.
The text sentiment analysis method according to claim 1, wherein, the described input matrix input attribute sentiment analysis model, obtains the attribute sentiment classification of the target text, comprising:

Obtaining attribute words in the target text according to the input matrix;

Utilize the attention mechanism to extract the contextual information relevant to the described attribute words;

The attribute sentiment classification of the target text is obtained according to the context information.
The text sentiment analysis method according to claim 6, wherein said obtaining attribute words in said respective words according to said input matrix comprises:

passing the input matrix through an attribute mask layer, and obtaining the attribute words according to an output matrix of the attribute mask layer.
The text sentiment analysis method according to claim 7, wherein said utilizing attention mechanism to extract contextual information relevant to said attribute word comprises:

performing an attention operation on the input matrix and the output matrix to obtain context information related to the attribute word.
A text sentiment analysis device, comprising:

at least one processor; and,

a memory communicatively coupled to the at least one processor; wherein,

The memory is stored with instructions executable by the at least one processor, the instructions are executed by the at least one processor, so that the at least one processor can perform any one of claims 1 to 8 text sentiment analysis method.
A computer-readable storage medium storing a computer program, which implements the text sentiment analysis method according to any one of claims 1 to 8 when the computer program is executed by a processor.