WO2023173541A1

WO2023173541A1 - Text-based emotion recognition method and apparatus, device, and storage medium

Info

Publication number: WO2023173541A1
Application number: PCT/CN2022/089998
Authority: WO
Inventors: 王彦; 马骏; 王少军
Original assignee: 平安科技（深圳）有限公司
Priority date: 2022-03-17
Filing date: 2022-04-28
Publication date: 2023-09-21
Also published as: CN114662499A

Abstract

A text-based emotion recognition method and apparatus, a device, and a storage medium, applicable to the technical field of artificial intelligence. The method comprises: extracting, from a text to be recognized, a feature keyword for representing the text to be recognized, comparing the feature keyword with a preset pre-recognition rule, performing first-level emotion recognition processing on the text to be recognized, and determining whether the feature keyword satisfies the pre-recognition rule; if not, inputting the text to be recognized into a first emotion recognition model for second-level emotion recognition processing, and determining whether the text to be recognized is a negative emotion text; and if so, inputting the text to be recognized into a second emotion recognition model for third-level emotion recognition processing, recognizing an emotion classification category corresponding to the text to be recognized, and generating, according to the emotion classification category, an emotion recognition result of the text to be recognized. The method employs multi-level emotion recognition to perform hierarchical detection on the text to be recognized, so that the accuracy and efficiency of emotion recognition can be improved simultaneously.

Description

Text-based emotion recognition methods, devices, equipment and storage media

This application requests the priority of the Chinese patent application submitted to the China Patent Office on March 17, 2022, with the application number 202210262203.2, and the invention name is "Text-based emotion recognition method, device, equipment and storage medium", and its entire content is approved by This reference is incorporated into this application.

Technical field

This application relates to the field of artificial intelligence decision-making technology, and in particular to a text-based emotion recognition method, device, equipment and storage medium.

Background technique

With the development of artificial intelligence technology, intelligent customer service robots have been widely used. In order to make the customer service services of intelligent customer service robots more humane, emotion recognition, as a key technology in human-computer interaction, has also become a hot research direction. In customer service scenarios, compared with non-negative emotions, the proportion of negative emotions is very low. However, the inventor realized that customers’ negative emotions tend to attract more attention. Most of the current emotion recognition technologies do not consider the special characteristics of customer service scenarios. sex. Moreover, in customer service scenarios, existing text-based emotion recognition technology generally uses a single model to identify negative and non-negative emotions. On the one hand, because the sample size of non-negative emotions is much larger than that of negative emotions, the model is used to detect people most of the time. Non-negative emotions that we don’t care about consume a lot of computing resources, seriously affect service throughput, and have low emotion recognition efficiency. On the other hand, because the negative emotions detected by a single model are often not accurate enough, the detected data need to be further reviewed. Work The amount is large, resulting in the effect of emotion recognition being unsatisfactory.

Contents of the invention

In view of this, embodiments of the present application provide a text-based emotion recognition method, device, equipment and storage medium, which can simultaneously improve the accuracy and efficiency of emotion recognition in text and take into account the accuracy and efficiency of emotion recognition.

The first aspect of the embodiment of the present application provides a text-based emotion recognition method, including:

Extract the characteristic keywords used to characterize the text to be recognized from the text to be recognized, and compare the characteristic keywords with the preset pre-recognition rules to perform first-level emotion recognition on the text to be recognized. Processing to determine whether the feature keyword hits the pre-identification rule;

If not, input the text to be recognized into the pre-trained first emotion recognition model for secondary emotion recognition processing, and determine whether the text to be recognized is a non-negative emotion text;

If not, the text to be recognized is input into the pre-trained second emotion recognition model to perform three-level emotion recognition processing, the emotion classification category corresponding to the text to be recognized is identified, and the emotion classification category is generated according to the emotion classification category The emotion recognition result of the text to be recognized.

A second aspect of the embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the electronic device. When the processor executes the computer program, the first Each step of the text-based emotion recognition method provided on the one hand:

A fourth aspect of the embodiments of the present application provides a computer-readable storage medium that stores a computer program. When the computer program is executed by a processor, the text-based emotion recognition provided by the first aspect is implemented. Each step of the method:

The fourth aspect of the embodiments of the present application provides a text-based emotion recognition device. The text-based emotion recognition device includes:

The first-level emotion recognition module is used to extract characteristic keywords used to characterize the text to be recognized from the text to be recognized, and compare the characteristic keywords with the preset pre-recognition rules to compare the characteristic keywords. The text to be recognized undergoes first-level emotion recognition processing to determine whether the characteristic keyword hits the pre-identification rule;

A secondary emotion recognition module, configured to use a first emotion recognition model based on the FastText algorithm to perform secondary emotion recognition processing on the text to be recognized when the text to be recognized does not hit the pre-recognition rule, Determine whether the text to be identified is non-negative emotional text;

A three-level emotion recognition module, used to perform three-level emotion recognition processing on the text to be recognized using a second emotion recognition model based on the Bert algorithm when the text to be recognized is not judged to be a non-negative emotion text, Identify the emotion classification category corresponding to the text to be recognized, and generate an emotion recognition result of the text to be recognized according to the emotion classification category.

The text-based emotion recognition method, device, electronic device and storage medium provided by the embodiments of this application have the following beneficial effects:

This application extracts feature keywords used to characterize the text to be recognized from the text to be recognized, compares the feature keywords with the preset pre-recognition rules, performs first-level emotion recognition processing on the text to be recognized, and determines the key features Whether the word hits the pre-recognition rule; if not, the first emotion recognition model is used to perform secondary emotion recognition processing on the text to be recognized, and it is judged whether the text to be recognized is a non-negative emotion text; if not, the second emotion recognition model is used to process it. The recognized text undergoes three-level emotion recognition processing to identify the emotion classification category corresponding to the text to be recognized, and generates the emotion recognition result of the text to be recognized based on the emotion classification category. Multi-level emotion recognition processing is used to conduct hierarchical detection of the text to be recognized, which combines the dual advantages of the rule engine and the machine learning algorithm model, while improving the accuracy and efficiency of emotion recognition of text, taking into account the accuracy and efficiency of emotion recognition. sex.

Description of the drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the purpose of the present application. For some embodiments, for those of ordinary skill in the art, other drawings can be obtained based on these drawings without exerting creative efforts.

Figure 1 is an implementation flow chart of a text-based emotion recognition method provided by an embodiment of the present application;

Figure 2 is a schematic flowchart of a method for performing first-level emotion recognition processing in the text-based emotion recognition method provided by the embodiment of the present application;

Figure 3 is a schematic flowchart of a method for performing secondary emotion recognition processing in the text-based emotion recognition method provided by the embodiment of the present application;

Figure 4 is a schematic flowchart of a method for performing three-level emotion recognition processing in the text-based emotion recognition method provided by the embodiment of the present application;

Figure 5 is a schematic flowchart of a method for bidirectional encoding and characterization processing in a text-based emotion recognition method provided by an embodiment of the present application;

Figure 6 is a basic structural block diagram of a text-based emotion recognition device provided by an embodiment of the present application;

Figure 7 is a basic structural block diagram of an electronic device provided by an embodiment of the present application.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present application more clear, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present application and are not used to limit the present application.

The text-based emotion recognition method provided by the embodiments of this application is applied in an intelligent customer service robot system. During operation, the intelligent customer service robot system will collect the customer's voice data and perform speech and semantic understanding of the voice data, identify and respond to the customer's needs, realize human-computer interaction, and thus provide customers with anthropomorphic services.

Please refer to Figure 1, which is an implementation flow chart of a text-based emotion recognition method provided by an embodiment of the present application. Details are as follows:

S11: Extract the characteristic keywords used to characterize the text to be recognized from the text to be recognized, and compare the characteristic keywords with the preset pre-recognition rules to conduct a first-level analysis of the text to be recognized. Emotion recognition processing determines whether the characteristic keyword hits the pre-identification rule.

In this embodiment, the text to be recognized is obtained by text conversion of the customer's voice data collected by the intelligent robot. In this embodiment, natural language processing technology is used to perform text feature extraction processing on the text to be recognized, and feature keywords used to characterize the features of the text to be recognized are obtained. For example, in the intelligent customer service robot system, a rule list for performing first-level emotion recognition is preset. This rule list stores multiple sets of pre-recognition rules for identifying the emotions corresponding to the text to be recognized. All pre-recognition rules The rules are written and checked manually. Each set of pre-recognition rules contains an emotion label and the applicable conditions set for the emotion label. In this embodiment, emotion labels include but are not limited to "non-negative" emotion labels, "slightly negative" emotion labels, "severely negative" emotion labels, etc. The applicable conditions set for emotion labels can be expressed as A characteristic or set of characteristics for this emotion label. When the intelligent customer service robot system performs first-level emotion recognition processing, it extracts the characteristic keywords used to characterize the text to be recognized from the text to be recognized, and then compares the characteristic keywords with all pre-set pre-recognition rules one by one. Yes, it is determined whether the feature keyword meets the usage conditions of a certain pre-identification rule record among all pre-identification rules. If the feature keyword meets the applicable conditions of a certain pre-identification rule, it is judged that the feature keyword hits the Pre-recognition rule, at this time, the emotion label recorded in the pre-recognition rule hit by the feature keyword is output as the final emotion recognition result. If the feature keyword does not meet the applicable conditions of any of the pre-recognition rules among all pre-recognition rules, it is judged that the feature keyword does not hit the pre-recognition rule, and at this time, the text to be recognized is subjected to further secondary emotion recognition processing.

In some embodiments of the present application, for example, the prefix identification rule may be set as a text matching rule. Specifically, a short text and an emotion label are recorded in a set of text matching rules. In the text matching rule, the short text is used as an applicable condition to verify whether the emotion label is applicable to the text to be recognized. In a set of text matching rules, short texts are mapped and associated with emotion labels, such that a short text corresponds to a unique emotion label. In this embodiment, the preset rule list for performing first-level emotion recognition can be represented as a list of correspondences between short texts and emotion tags. For example, the pre-identification rule can also be set as a regular matching rule. Specifically, a regular expression and an emotion label are recorded in a set of regular matching rules. In the regular matching rule, the regular expression is used as a verification emotion label. Whether the applicable conditions apply to the text to be recognized. In a set of regular matching rules, regular expressions are mapped to emotion labels, such that a regular expression corresponds to a unique emotion label. In this embodiment, the preset rule list for performing first-level emotion recognition can be represented as a correspondence table between regular expressions and emotion tags.

In some embodiments of the present application, please refer to FIG. 2 . FIG. 2 is a schematic flowchart of a method for performing first-level emotion recognition processing in the text-based emotion recognition method provided by the embodiment of the present application. Details are as follows:

S21: Calculate the text correlation between the characteristic keywords and the short text recorded in the short text matching rule, and obtain the text correlation value between the characteristic keywords and the short text;

S22: Compare the text relevance value with a preset relevance threshold. If the text relevance value is less than the relevance threshold, it is determined that the feature keyword does not hit the short text matching rule.

In this embodiment, the preset pre-identification rules can be set as short text matching rules. A set of short text matching rules is represented as a correspondence between short texts and emotion tags, in which a short text and a short text are recorded. This has emotion tags that map association relationships. In this embodiment, after the feature keywords used to characterize the text to be recognized are extracted from the text to be recognized, the obtained feature keywords and the short text recorded in the short text matching rules can be vectorized in advance to obtain A first vector used to represent feature keywords and a second vector used to represent short text. The text correlation between feature keywords and short text is calculated through the representation vector. Specifically, a cosine similarity algorithm is used to calculate the cosine value of the angle between the first vector and the second vector, and the calculated cosine value of the angle is used as the text correlation value between the feature keyword and the short text. After obtaining the text correlation value between the feature keyword and the short text, the calculated text correlation value can be compared with the preset correlation threshold. When the text correlation value is less than the correlation threshold, it is judged Feature keyword misses short text matching rules. It should be noted that when the preset pre-identification rules include multiple sets of short text matching rules, the text correlation between the feature keywords and the short texts recorded in each short text matching rule will be calculated one by one, and then one by one. Compare and determine whether the feature keyword does not satisfy any short text matching rule among all short text matching rules. If so, determine whether the feature keyword hits the short text matching rule. In this embodiment, the preset correlation threshold is obtained by the user customizing the setting in the intelligent customer service robot system according to the actual demand for emotion recognition accuracy.

For example, in this embodiment, when vectorizing the feature keywords and the short text respectively, vectorization can be performed in combination with the semantic dimension and the literal dimension to obtain the first vector used to represent the feature keywords and The second vector used to represent the short text, both the first vector and the second vector contain features of two dimensions: semantic dimension and literal dimension.

In some embodiments of the present application, the preset pre-identification rules can be set as regular matching rules. A set of regular matching rules is expressed as a correspondence between regular expressions and emotion tags, in which a regular expression and a Sentiment tags that are mapped to this regular expression. In this embodiment, the regular expression is specifically represented by ASCII (full name: American Standard Code for Information Interchange, American Standard Code for Information Interchange) characters. After the feature keywords used to characterize the text to be recognized are extracted from the text to be recognized, the feature keywords can be converted into corresponding ASCII character representations in advance based on the ASCII character encoding table, and then based on the ASCII used to represent regular expressions Characters are compared one-to-one with the ASCII characters corresponding to the feature keywords from left to right to determine whether the ASCII characters used to represent the regular expression are consistent with the ASCII characters corresponding to the feature keywords. If they are inconsistent, the feature is The keyword does not meet the requirements of the regular expression. At this time, it can be judged that the feature keyword does not hit the regular matching rule. It can be understood that in this embodiment, if the preset pre-identification rules include multiple sets of regular matching rules, the characteristic keywords will be compared one by one with the regular expressions recorded in each regular matching rule. Yes, it is judged whether the characteristic keyword does not satisfy any of the regular matching rules among all the regular matching rules. If so, it is judged that the characteristic keyword hits the regular matching rule.

S12: If not, use the first emotion recognition model obtained by training based on the FastText model to perform secondary emotion recognition processing on the text to be recognized, and determine whether the text to be recognized is a non-negative emotion text.

In this embodiment, the intelligent customer service robot system can quickly and accurately identify the emotions corresponding to certain texts to be recognized by using pre-recognition rules to perform first-level emotion recognition processing on the text to be recognized. In order to solve the problem of poor generalization of emotion recognition by pre-recognition rules, in this embodiment, the intelligent customer service robot system fails to recognize the text to be recognized corresponding to the emotion using pre-recognition rules, and further identifies it through two-category processing. Check whether the text to be identified is a text with negative emotions. For example, two-classification training is performed in advance based on the FastText (Fast Text Classification) model to generate a first emotion recognition model for performing secondary emotion recognition processing on the text to be recognized. Furthermore, the text to be recognized whose corresponding emotion cannot be recognized by the pre-recognition rule is input into the first emotion recognition model for secondary emotion recognition processing, and the first emotion recognition model outputs whether the text to be recognized is a non-negative emotion. Text results. In robot customer service scenarios, the proportion of non-negative emotions is much greater than the proportion of negative emotions, and negative emotions require more attention. Therefore, by performing binary classification processing on the text to be recognized through the first emotion recognition model, most of the non-negative emotion texts can be filtered out, and for most of the filtered non-negative emotion texts, the "non-negative" emotion label can be directly output as the final emotion recognition result. The binary classification process in the first emotion recognition model can identify non-negative emotion text simply and with high accuracy.

In some embodiments of the present application, please refer to FIG. 3 . FIG. 3 is a schematic flowchart of a method for performing secondary emotion recognition processing in the text-based emotion recognition method provided by an embodiment of the present application. Details are as follows:

S31: Use the first emotion recognition model to perform first word vector representation processing on each word in the text to be recognized, and obtain a first word vector set corresponding to the text to be recognized;

S32: Input the first word vector set into the sentence representation layer of the first emotion recognition model for sentence vector representation, and obtain the first sentence vector used to characterize the text to be recognized;

S33: Input the first sentence vector to the linear layer of the first emotion recognition model for linear transformation processing, and obtain the text to be recognized and the non-negative emotion categories and negative emotions preset in the first emotion recognition model respectively. Match probability value between categories;

S34: If the matching probability value between the text to be recognized and the non-negative emotion category is greater than the matching probability value between the text to be recognized and the negative emotion category, determine that the text to be recognized is a non-negative emotion text.

In this embodiment, when building the first emotion recognition model, a network architecture for two classifications is built based on the FastText model, one of which is configured as a non-negative emotion category, and the other is configured as a negative emotion category, and then uses A large number of texts marked with non-negative emotion labels or negative emotion labels are expected to be used as training samples for network training of the network architecture based on the FastText model, and the network architecture is trained to a convergence state, so that the network architecture has judgment based on text content. Whether the text is a non-negative emotional text, thus obtaining the first emotion recognition model. Specifically, the network architecture trained in the first emotion recognition model includes a sentence representation layer and a linear layer. In this embodiment, when using the first emotion recognition model to perform secondary emotion recognition on the text to be recognized, the text to be recognized can be input into the first emotion recognition model, and the first emotion recognition model is first used to identify the text to be recognized. All the words in the text are disassembled, and then the first word vector representation processing is performed on each word in the text to be recognized. Each word is represented by a vector to obtain the word vector corresponding to each word, and then the word vector to be recognized is obtained. The word vectors corresponding to all words in the recognized text are gathered together, thereby obtaining the first set of word vectors corresponding to the text to be recognized. Then, the first word vector set is input into the sentence representation layer of the first emotion recognition model for sentence vector representation, and a first sentence vector used to characterize the text to be recognized is obtained. Specifically, by calculating the mean value of word vectors corresponding to all words in the first word vector set, the mean value is used as the first sentence vector used to characterize the text to be recognized. After obtaining the first sentence vector, input the first sentence vector into the linear layer of the first emotion recognition model for linear transformation processing, and obtain the two emotion classification categories of the non-negative emotion category and the negative emotion category in the first emotion recognition model. Probability distribution. Based on the probability distribution, the matching probability values between the text to be recognized and the preset non-negative emotion categories and negative emotion categories in the first emotion recognition model can be obtained. At this time, if the matching probability value between the text to be recognized and the non-negative emotion category is greater than the matching probability value between the text to be recognized and the negative emotion category, it can be determined that the text to be recognized is a non-negative emotion text.

S13: If yes, use the second emotion recognition model obtained based on Bert model training to perform three-level emotion recognition processing on the text to be recognized, identify the emotion classification category corresponding to the text to be recognized, and generate a The emotion recognition result of the text to be recognized.

In this embodiment, the intelligent customer service robot system uses the first emotion recognition model to perform secondary emotion recognition on the text to be recognized that fails to recognize the corresponding emotion according to the recognition rules, and can obtain a small portion of the text that has not been judged by the first emotion recognition model. is the text to be identified that is a non-negative emotion text. In this embodiment, for the text to be recognized that is not judged to be non-negative emotional text by the first emotion recognition model, the emotion classification category corresponding to the text to be recognized can be identified through bidirectional coding representation. In this embodiment, emotion classification category recognition training is performed in advance based on the Bert (full name Bidirectional Encoder Representations from Transformers) model to generate a second emotion recognition model for identifying the emotion classification category corresponding to the text to be recognized. Then, the text to be recognized that is not judged as a non-negative emotion text by the first emotion recognition model is input into the second emotion recognition model for three-level emotion recognition processing, so that the second emotion recognition model outputs the emotion corresponding to the text to be recognized. Classification category results. After obtaining the emotion classification category corresponding to the text to be recognized, the emotion classification category is output as the final emotion recognition result, achieving hierarchical and progressive detection of the text using multiple models, while improving the accuracy of emotion recognition of the text. accuracy and efficiency, taking into account the accuracy and efficiency of emotion recognition. Among them, when training the second emotion recognition model, the training samples used can be labeled with more detailed emotion classification category information, such as non-negative, slightly negative, severe negative emotions, etc. Through the second emotion recognition model, the judgment results of the first emotion recognition model can be reviewed and more detailed and accurate emotion classification can be made.

In some embodiments of the present application, please refer to FIG. 4 , which is a schematic flowchart of a method for performing three-level emotion recognition processing in the text-based emotion recognition method provided by an embodiment of the present application. Details are as follows:

S41: Use the second emotion recognition model to perform second sub-vector representation processing on each word in the text to be recognized, and obtain a second set of word vectors corresponding to the text to be recognized;

S42: Input the second word vector set into the Transformer layer of the second emotion recognition model for bidirectional encoding and characterization processing, and obtain a second sentence vector used to characterize the text to be recognized;

S43: Input the first sentence vector to the linear layer of the second emotion recognition model for linear transformation processing, and obtain the probability distribution data of the text to be recognized in each emotion classification category preset by the second emotion recognition model. ;

S44: According to the probability distribution data, select the emotion classification category corresponding to the maximum probability value in the probability distribution data as the emotion classification category of the text to be recognized.

In this embodiment, when building the second emotion recognition model, a network architecture that can perform bidirectional encoding and representation of text is built based on the Bert model. In this network architecture, a Transformer layer and a Transformer layer are provided for bidirectional encoding and representation of text. Linear layer. It can be understood that the Transformer layer is a network layer containing multiple sets of encoding-decoding layers that utilizes the Self-Attention mechanism. Then a large number of text predictions marked with various emotion classification categories are used as training texts to perform model training on the network architecture based on the Bert model, and the network architecture is trained to a convergence state, so that the network architecture can recognize based on text content. The ability to classify text corresponding to emotion categories, thereby obtaining a second emotion recognition model. In this embodiment, when using the second emotion recognition model to perform three-level emotion recognition processing on the text to be recognized, the text to be recognized can be input into the second emotion recognition model. First, the second emotion recognition model is used to process the text to be recognized. All words in the recognition text are disassembled, and then each word in the text to be recognized is separately processed as a second word vector representation. Specifically, for each word, it is obtained from the three dimensions of word embedding, segment embedding and position embedding. The three sub-vectors of the word are then added to obtain the word vector corresponding to the word. After obtaining the word vector corresponding to each word, the word vectors corresponding to all the words in the text to be recognized are gathered together. , thereby obtaining the first word vector set corresponding to the text to be recognized. Then, the second set of word vectors is input into the Transformer layer of the second emotion recognition model for bidirectional encoding and characterization processing, and a second sentence vector used to characterize the text to be recognized is obtained. By inputting the first sentence vector into the linear layer of the second emotion recognition model for linear transformation processing, the probability distribution data of the text to be recognized in each emotion classification category preset by the first emotion recognition model can be obtained. Finally, according to the probability distribution data, the emotion classification category corresponding to the maximum probability value in the probability distribution data is selected as the emotion classification category of the text to be recognized.

In some embodiments of the present application, please refer to FIG. 5 . FIG. 5 is a schematic flowchart of a method for performing bidirectional coding representation processing in a text-based emotion recognition method provided by an embodiment of the present application. Details are as follows:

S51: Perform self-attention calculation on each word vector in the second word vector set, and obtain the self-attention data value corresponding to each word vector in the second word vector set;

S52: Normalize the self-attention data value corresponding to each word vector to obtain a second sentence vector used to characterize the text to be recognized.

In this embodiment, when the second set of word vectors is input into the Transformer layer of the second emotion recognition model for bidirectional encoding and representation processing, specifically, self-attention is performed on each word vector in the second set of word vectors in the Transformer layer. Force calculation is performed to obtain the self-attention data value corresponding to each word vector in the second word vector set. Furthermore, the self-attention data values corresponding to each word vector are normalized to obtain a second sentence vector used to characterize the text to be recognized.

It can be seen from the above that the text-based emotion recognition method provided by this embodiment first performs first-level emotion recognition on the text through pre-recognition rules. As long as the rule is hit, the emotion recognition result of the text is generated based on the hit rule, which is fast. When the first-level emotion recognition does not detect negative emotions, the first emotion recognition model built with the FastText algorithm is used to perform secondary emotion recognition on the text, which can quickly identify the non-negative emotion texts that account for the vast majority of customer service scenarios. Improved system throughput and ability to support concurrency. Furthermore, when the secondary emotion recognition detects that the text is a negative emotional text, the second emotion recognition model based on the Bert algorithm is used to perform third-level emotion recognition on the text to obtain the emotion classification category corresponding to the text and classify it according to the emotion. Emotion recognition results for category generated text. In this way, by using multiple models to perform hierarchical detection of text, integrating the dual advantages of rule engines and machine learning algorithm models to perform emotion recognition on text, while improving the accuracy and efficiency of emotion recognition on text, taking into account the accuracy of emotion recognition. sex and efficiency.

It can be understood that the sequence number of each step in the above embodiment does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any influence on the implementation process of the embodiment of the present application. limited.

In some embodiments of the present application, please refer to FIG. 6 , which is a basic structural block diagram of a text-based emotion recognition device provided by an embodiment of the present application. Each unit included in the device in this embodiment is used to perform each step in the above method embodiment. For details, please refer to the relevant descriptions in the above method embodiments. For convenience of explanation, only parts related to this embodiment are shown. As shown in FIG. 6 , the text-based emotion recognition device includes: a first-level emotion recognition module 61 , a second-level emotion recognition module 62 , and a third-level emotion recognition module 63 . Wherein: the first-level emotion recognition module 61 is used to extract characteristic keywords used to characterize the text to be recognized from the text to be recognized, and compare the characteristic keywords with preset pre-recognition rules, A first-level emotion recognition process is performed on the text to be recognized to determine whether the characteristic keyword hits the pre-identification rule. The secondary emotion recognition module 62 is used to perform secondary emotion recognition on the text to be recognized by using a first emotion recognition model based on the FastText algorithm when the text to be recognized does not hit the pre-recognition rule. Processing to determine whether the text to be recognized is non-negative emotional text. The three-level emotion recognition module 63 is used to perform three-level emotion recognition on the text to be recognized by using a second emotion recognition model based on the Bert algorithm when the text to be recognized is not judged to be a non-negative emotion text. Processing: identifying the emotion classification category corresponding to the text to be recognized, and generating an emotion recognition result of the text to be recognized according to the emotion classification category.

It should be understood that the above-mentioned text-based emotion recognition device corresponds to the above-mentioned text-based emotion recognition method, and will not be described again here.

In some embodiments of the present application, please refer to FIG. 7 , which is a basic structural block diagram of an electronic device provided by an embodiment of the present application. As shown in Figure 7, the electronic device 7 of this embodiment includes: a processor 71, a memory 72, and a computer program 73 stored in the memory 72 and executable on the processor 71, such as text-based emotion recognition. method procedure. When the processor 71 executes the computer program 73, the steps in each embodiment of the above text-based emotion recognition method are implemented. Alternatively, when the processor 71 executes the computer program 73, it implements the functions of each module in the corresponding embodiment of the above text-based emotion recognition device. For details, please refer to the relevant descriptions in the embodiments and will not be repeated here.

Exemplarily, the computer program 73 can be divided into one or more modules (units), and the one or more modules are stored in the memory 72 and executed by the processor 71 to complete the present invention. Apply. The one or more modules may be a series of computer program instruction segments capable of completing specific functions. The instruction segments are used to describe the execution process of the computer program 73 in the electronic device 7 . For example, the computer program 73 can be divided into a first-level emotion recognition module, a second-level emotion recognition module, and a third-level emotion recognition module. The specific functions of each module are as described above.

The electronic device may include, but is not limited to, a processor 71 and a memory 72 . Those skilled in the art can understand that FIG. 7 is only an example of the electronic device 7 and does not constitute a limitation of the electronic device 7. It may include more or fewer components than shown in the figure, or some components may be combined, or different components may be used. , for example, the electronic device may also include input and output devices, network access devices, buses, etc.

The processor 71 can be a central processing unit (Central Processing Unit, CPU), or other general-purpose processor, digital signal processor (Digital Signal Processor, DSP), application specific integrated circuit (Application Specific Integrated Circuit, ASIC), Ready-made field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.

The memory 72 may be an internal storage unit of the electronic device 7 , such as a hard disk or memory of the electronic device 7 . The memory 72 may also be an external storage device of the electronic device 7, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), or a secure digital (SD) equipped on the electronic device 7. Card, Flash Card, etc. Further, the memory 72 may also include both an internal storage unit of the electronic device 7 and an external storage device. The memory 72 is used to store the computer program and other programs and data required by the electronic device. The memory 72 can also be used to temporarily store data that has been output or is to be output.

It should be noted that the information interaction, execution process, etc. between the above-mentioned devices/units are based on the same concept as the method embodiments of the present application. For details of their specific functions and technical effects, please refer to the method embodiments section. No further details will be given.

Embodiments of the present application also provide a computer-readable storage medium. The computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the steps in each of the above method embodiments can be implemented. In this embodiment, the computer-readable storage medium may be a non-volatile storage medium or a volatile storage medium.

Embodiments of the present application provide a computer program product. When the computer program product is run on a mobile terminal, the steps in each of the above method embodiments can be implemented when the mobile terminal is executed.

Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above functional units and modules is used as an example. In actual applications, the above functions can be allocated to different functional units and modules according to needs. Module completion means dividing the internal structure of the device into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be hardware-based. It can also be implemented in the form of software functional units. In addition, the specific names of each functional unit and module are only for the convenience of distinguishing each other and are not used to limit the scope of protection of the present application. For the specific working processes of the units and modules in the above system, please refer to the corresponding processes in the foregoing method embodiments, and will not be described again here.

If the integrated module/unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application can implement all or part of the processes in the methods of the above embodiments, which can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer can When the program is executed by the processor, the steps of each of the above method embodiments can be implemented. Wherein, the computer program includes computer program code, which may be in the form of source code, object code, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory) , Random Access Memory (RAM, Random Access Memory), electrical carrier signals, telecommunications signals, and software distribution media, etc. It should be noted that the content contained in the computer-readable medium can be appropriately added or deleted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium Excluded are electrical carrier signals and telecommunications signals.

In the above embodiments, each embodiment is described with its own emphasis. For parts that are not detailed or documented in a certain embodiment, please refer to the relevant descriptions of other embodiments.

The above-described embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still implement the above-mentioned implementations. The technical solutions described in the examples are modified, or some of the technical features are equivalently replaced; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions in the embodiments of this application, and should be included in within the protection scope of this application.

Claims

A text-based emotion recognition method, including:

Extract the characteristic keywords used to characterize the text to be recognized from the text to be recognized, and compare the characteristic keywords with the preset pre-recognition rules to perform first-level emotion recognition on the text to be recognized. Processing to determine whether the feature keyword hits the pre-identification rule;

If not, use the first emotion recognition model obtained by training based on the FastText model to perform secondary emotion recognition processing on the text to be recognized, and determine whether the text to be recognized is a non-negative emotion text;

If not, use the second emotion recognition model obtained based on Bert model training to perform three-level emotion recognition processing on the text to be recognized, identify the emotion classification category corresponding to the text to be recognized, and generate the emotion classification category according to the emotion classification category. Describe the emotion recognition results of the text to be recognized.
The text-based emotion recognition method according to claim 1, wherein the preset pre-recognition rules include short text matching rules and/or regular matching rules, and the short text matching rules are expressed as short text and emotion tags The corresponding relationship between the short text is the applicable condition for verifying whether the emotion label is applicable to the text to be recognized;

The regular matching rule is expressed as a correspondence between a regular expression and an emotion tag, and the regular expression is an applicable condition for verifying whether the emotion tag is applicable to the text to be recognized.
The text-based emotion recognition method according to claim 2, wherein if the pre-identification rule is a short text matching rule, then comparing the characteristic keywords with a pre-set pre-identification rule, Perform first-level emotion recognition processing on the text to be recognized to determine whether the characteristic keywords hit the pre-recognition rule, including:

Calculate the text correlation between the characteristic keywords and the short text recorded in the short text matching rule, and obtain the text correlation value between the characteristic keywords and the short text;

The text relevance value is compared with a preset relevance threshold. If the text relevance value is less than the relevance threshold, it is determined that the characteristic keyword does not hit the short text matching rule.
The text-based emotion recognition method according to claim 2, wherein if the pre-identification rule is a regular matching rule, then the characteristic keywords are compared with a pre-set pre-identification rule to determine Perform first-level emotion recognition processing on the text to be recognized, and determine whether the characteristic keywords hit the pre-recognition rules, including:

Compare the characteristic keyword with the regular expression recorded in the regular matching rule to determine whether the characteristic keyword is consistent with the regular expression. If not, determine that the characteristic keyword does not hit the target. The regular matching rules described above.
The text-based emotion recognition method according to claim 1, wherein the first emotion recognition model obtained by training based on the FastText model is used to perform secondary emotion recognition processing on the text to be recognized, and determine whether the text to be recognized is For non-negative sentiment text, include:

Using the first emotion recognition model to perform first word vector representation processing on each word in the text to be recognized, and obtaining a first word vector set corresponding to the text to be recognized;

Input the first word vector set into the sentence representation layer of the first emotion recognition model for sentence vector representation, and obtain the first sentence vector used to characterize the text to be recognized;

The first sentence vector is input to the linear layer of the first emotion recognition model for linear transformation processing to obtain the text to be recognized and the non-negative emotion category and the negative emotion category preset in the first emotion recognition model. The matching probability value between;

If the matching probability value between the text to be recognized and the non-negative emotion category is greater than the matching probability value between the text to be recognized and the negative emotion category, the text to be recognized is determined to be a non-negative emotion text.
The text-based emotion recognition method according to claim 1, wherein the second emotion recognition model obtained by training based on the Bert model is used to perform three-level emotion recognition processing on the text to be recognized, and the text to be recognized is recognized Corresponding emotion classification categories include:

Using the second emotion recognition model to perform second sub-vector representation processing on each word in the text to be recognized, and obtaining a second set of word vectors corresponding to the text to be recognized;

Input the second set of word vectors into the Transformer layer of the second emotion recognition model for bidirectional encoding and characterization processing to obtain a second sentence vector used to characterize the text to be recognized;

Input the first sentence vector into the linear layer of the second emotion recognition model for linear transformation processing to obtain probability distribution data of the text to be recognized in each emotion classification category preset by the second emotion recognition model;

According to the probability distribution data, the emotion classification category corresponding to the maximum probability value in the probability distribution data is selected as the emotion classification category of the text to be recognized.
The text-based emotion recognition method according to claim 6, wherein the second set of word vectors is input into the Transformer layer of the second emotion recognition model to perform bidirectional encoding and characterization processing to obtain a representation of the second word vector set. Describe the second sentence vector of the text to be recognized, including:

Perform self-attention calculation on each word vector in the second word vector set respectively, and obtain the self-attention data value corresponding to each word vector in the second word vector set;

The self-attention data value corresponding to each word vector is normalized to obtain a second sentence vector used to characterize the text to be recognized.
A text-based emotion recognition device, wherein the text-based emotion recognition device includes:

The first-level emotion recognition module is used to extract characteristic keywords used to characterize the text to be recognized from the text to be recognized, and compare the characteristic keywords with the preset pre-recognition rules to compare the characteristic keywords. The text to be recognized undergoes first-level emotion recognition processing to determine whether the characteristic keyword hits the pre-identification rule;

A secondary emotion recognition module, configured to perform secondary emotion recognition processing on the text to be recognized using the first emotion recognition model obtained based on FastText model training when the text to be recognized does not hit the pre-recognition rule. , determine whether the text to be recognized is a non-negative emotional text;

A three-level emotion recognition module, used to perform three-level emotion recognition processing on the text to be recognized using a second emotion recognition model based on the Bert algorithm when the text to be recognized is not judged to be a non-negative emotion text, Identify the emotion classification category corresponding to the text to be recognized, and generate an emotion recognition result of the text to be recognized according to the emotion classification category.
An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the following steps when executing the computer program:

Extract the characteristic keywords used to characterize the text to be recognized from the text to be recognized, and compare the characteristic keywords with the preset pre-recognition rules to perform first-level emotion recognition on the text to be recognized. Processing to determine whether the feature keyword hits the pre-identification rule;

If not, use the first emotion recognition model obtained by training based on the FastText model to perform secondary emotion recognition processing on the text to be recognized, and determine whether the text to be recognized is a non-negative emotion text;

If not, use the second emotion recognition model obtained based on Bert model training to perform three-level emotion recognition processing on the text to be recognized, identify the emotion classification category corresponding to the text to be recognized, and generate the emotion classification category according to the emotion classification category. Describe the emotion recognition results of the text to be recognized.
The electronic device according to claim 9, wherein the preset pre-identification rules include short text matching rules and/or regular matching rules, and the short text matching rules are represented as correspondences between short texts and emotion tags. Relationship, the short text is an applicable condition used to verify whether the emotion label is applicable to the text to be recognized;

The regular matching rule is expressed as a correspondence between a regular expression and an emotion tag, and the regular expression is an applicable condition for verifying whether the emotion tag is applicable to the text to be recognized.
The electronic device according to claim 10, wherein if the prefix identification rule is a short text matching rule, the characteristic keyword is compared with a preset prefix identification rule to match the prefix identification rule. When the text to be recognized is subjected to first-level emotion recognition processing and it is judged whether the characteristic keyword hits the pre-recognition rule, the processor implements the following steps when executing the computer program:

Calculate the text correlation between the characteristic keywords and the short text recorded in the short text matching rule, and obtain the text correlation value between the characteristic keywords and the short text;

The text relevance value is compared with a preset relevance threshold. If the text relevance value is less than the relevance threshold, it is determined that the characteristic keyword does not hit the short text matching rule.
The electronic device according to claim 10, wherein if the prefix identification rule is a regular matching rule, the characteristic keyword is compared with a preset prefix identification rule to compare the When the text to be recognized is subjected to first-level emotion recognition processing and it is judged whether the characteristic keyword hits the pre-recognition rule, the processor implements the following steps when executing the computer program:

Compare the characteristic keyword with the regular expression recorded in the regular matching rule to determine whether the characteristic keyword is consistent with the regular expression. If not, determine that the characteristic keyword does not hit the target. The regular matching rules described above.
The electronic device according to claim 9, wherein the first emotion recognition model obtained by training based on the FastText model is used to perform secondary emotion recognition processing on the text to be recognized, and it is determined whether the text to be recognized is non-negative. When writing emotional text, the processor implements the following steps when executing the computer program:

Using the first emotion recognition model to perform first word vector representation processing on each word in the text to be recognized, and obtaining a first word vector set corresponding to the text to be recognized;

Input the first word vector set into the sentence representation layer of the first emotion recognition model for sentence vector representation, and obtain the first sentence vector used to characterize the text to be recognized;

The first sentence vector is input to the linear layer of the first emotion recognition model for linear transformation processing to obtain the text to be recognized and the non-negative emotion category and the negative emotion category preset in the first emotion recognition model. The matching probability value between;

If the matching probability value between the text to be recognized and the non-negative emotion category is greater than the matching probability value between the text to be recognized and the negative emotion category, the text to be recognized is determined to be a non-negative emotion text.
The electronic device according to claim 9, wherein the second emotion recognition model obtained by training based on the Bert model is used to perform a three-level emotion recognition process on the text to be recognized, and the emotion corresponding to the text to be recognized is identified. When classifying categories, the processor implements the following steps when executing the computer program:

Using the second emotion recognition model to perform second sub-vector representation processing on each word in the text to be recognized, and obtaining a second set of word vectors corresponding to the text to be recognized;

Input the second set of word vectors into the Transformer layer of the second emotion recognition model for bidirectional encoding and characterization processing to obtain a second sentence vector used to characterize the text to be recognized;

Input the first sentence vector into the linear layer of the second emotion recognition model for linear transformation processing to obtain probability distribution data of the text to be recognized in each emotion classification category preset by the second emotion recognition model;

According to the probability distribution data, the emotion classification category corresponding to the maximum probability value in the probability distribution data is selected as the emotion classification category of the text to be recognized.
A computer-readable storage medium, the computer-readable storage medium stores a computer program, wherein the following steps are implemented when the computer program is executed by a processor:

Extract the characteristic keywords used to characterize the text to be recognized from the text to be recognized, and compare the characteristic keywords with the preset pre-recognition rules to perform first-level emotion recognition on the text to be recognized. Processing to determine whether the feature keyword hits the pre-identification rule;

If not, use the first emotion recognition model obtained by training based on the FastText model to perform secondary emotion recognition processing on the text to be recognized, and determine whether the text to be recognized is a non-negative emotion text;

If not, use the second emotion recognition model obtained based on Bert model training to perform three-level emotion recognition processing on the text to be recognized, identify the emotion classification category corresponding to the text to be recognized, and generate the emotion classification category according to the emotion classification category. Describe the emotion recognition results of the text to be recognized.
The storage medium according to claim 15, wherein the preset pre-identification rules include short text matching rules and/or regular matching rules, and the short text matching rules are represented as correspondences between short texts and emotion tags. Relationship, the short text is an applicable condition used to verify whether the emotion label is applicable to the text to be recognized;

The regular matching rule is expressed as a correspondence between a regular expression and an emotion tag, and the regular expression is an applicable condition for verifying whether the emotion tag is applicable to the text to be recognized.
The storage medium according to claim 16, wherein if the prefix identification rule is a short text matching rule, the characteristic keyword is compared with a preset prefix identification rule to match the prefix identification rule. When the text to be recognized is subjected to first-level emotion recognition processing and it is judged whether the characteristic keyword hits the pre-recognition rule, the following steps are implemented when the computer program is executed by the processor:

Calculate the text correlation between the characteristic keywords and the short text recorded in the short text matching rule, and obtain the text correlation value between the characteristic keywords and the short text;

The text relevance value is compared with a preset relevance threshold. If the text relevance value is less than the relevance threshold, it is determined that the characteristic keyword does not hit the short text matching rule.
The storage medium according to claim 16, wherein if the prefix identification rule is a regular matching rule, the characteristic keyword is compared with a preset prefix identification rule to compare the When the text to be recognized undergoes first-level emotion recognition processing and it is judged whether the characteristic keyword hits the pre-recognition rule, the computer program implements the following steps when executed by the processor:

Compare the characteristic keyword with the regular expression recorded in the regular matching rule to determine whether the characteristic keyword is consistent with the regular expression. If not, determine that the characteristic keyword does not hit the target. The regular matching rules described above.
The storage medium according to claim 15, wherein the first emotion recognition model obtained by training based on the FastText model is used to perform secondary emotion recognition processing on the text to be recognized, and it is determined whether the text to be recognized is non-negative. When writing emotional text, the computer program implements the following steps when executed by the processor:

Using the first emotion recognition model to perform first word vector representation processing on each word in the text to be recognized, and obtaining a first word vector set corresponding to the text to be recognized;

Input the first word vector set into the sentence representation layer of the first emotion recognition model for sentence vector representation, and obtain the first sentence vector used to characterize the text to be recognized;

The first sentence vector is input to the linear layer of the first emotion recognition model for linear transformation processing to obtain the text to be recognized and the non-negative emotion category and the negative emotion category preset in the first emotion recognition model. The matching probability value between;

If the matching probability value between the text to be recognized and the non-negative emotion category is greater than the matching probability value between the text to be recognized and the negative emotion category, the text to be recognized is determined to be a non-negative emotion text.
The storage medium according to claim 15, wherein the second emotion recognition model obtained by training based on the Bert model is used to perform three-level emotion recognition processing on the text to be recognized, and the emotion corresponding to the text to be recognized is identified. When classifying categories, the computer program implements the following steps when executed by the processor:

Using the second emotion recognition model to perform second sub-vector representation processing on each word in the text to be recognized, and obtaining a second set of word vectors corresponding to the text to be recognized;

Input the second set of word vectors into the Transformer layer of the second emotion recognition model for bidirectional encoding and characterization processing to obtain a second sentence vector used to characterize the text to be recognized;

Input the first sentence vector into the linear layer of the second emotion recognition model for linear transformation processing to obtain probability distribution data of the text to be recognized in each emotion classification category preset by the second emotion recognition model;

According to the probability distribution data, the emotion classification category corresponding to the maximum probability value in the probability distribution data is selected as the emotion classification category of the text to be recognized.