A method and device for identifying deep reasons of text emotion based on common sense reasoning

Abstract

The application relates to the technical field of artificial intelligence, and discloses a text emotion deep reason recognition method and device based on common sense reasoning, which comprises the following steps: first, potential emotional clauses and potential reason clauses in the text are recognized, and corresponding emotional clause graphs and reason clause graphs are constructed; then, emotional and causal common sense related to the clauses is searched to expand the graph structure, and two graph neural networks are obtained to represent the context semantics and common sense knowledge of the emotional and reason graphs; and an interactive attention mechanism is used to find out the emotional-reason combination clause pair with the largest correlation, and further decode the deep reason with a strong causal relationship and the corresponding emotion. Since the text grammar structure is fully depicted and the related external common sense is fused, the application can better decode the deep emotional reason, realize the mining of the potential deep emotional reason behind the surface emotional reason event of the text, and effectively improve the accuracy and reliability of the emotional reason inference.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a text emotion deep reason recognition method and device based on common sense reasoning. BACKGROUND

[0002] With the rapid development of the Internet, more and more text data with subjective emotion are generated, such as review data in the e-commerce field, dialogue data in the psychological counseling field, etc. These emotions are a series of physical and mental states produced by a combination of factors such as feelings, cognition, thought and behavior. Such text with emotion contains rich subjective knowledge and reflects people's preferences and feelings. Mining these preference knowledge has important value. Decision makers see the essence from the phenomenon and realize from "knowing that" to "knowing why" to solve the root cause problem. For example, in the dialogue system, the machine can generate more appropriate replies by analyzing the emotions and reasons implied in the user's dialogue expression; for product or service operation, enterprises can accurately grasp the user's real experience and demand by analyzing the emotions and internal reasons induced in the review data. This is conducive to clarifying the improvement and optimization direction and launching products and services that meet the expectations of consumers.

[0003] For the problem of emotion reason discovery, existing works mostly only focus on the shallow reasons that trigger emotions, and the research on deep reasons is very weak. Here, "shallow" and "deep" refer to the degree of semantic cognition. Shallow reasons generally point out the emotional reasons or events that appear in the literal text, and people can directly extract the sentence where the reason event is located, or extract some fragments from the text as reasons through "cause" "because" and other clue words. For example, for the text "Xiaoming is very sad because he failed the exam.", the negative emotional reason can be extracted as "failed the exam". These methods can identify simple reasons on the surface. However, the expression of emotion is diverse and complex, and some emotional reasons are ambiguous and implicit. Figure 1As shown, the comment "Recently bought this brand mobile phone really annoying, call 'please ask... you... is... etc... etc...'. The customer service said that it can try to put the phone in a higher position, or put a signal amplifier next to the window. Can't stand it, the next step is to change Huawei." contains the emotion of "anger". Through literal matching, the shallow reason event is "buying this mobile phone", or "calling". However, these neutral actions do not have any emotional tendency and are not the root cause of the anger emotion, that is, "poor mobile phone signal". This reason, which is not directly present in the text but can be easily obtained by understanding the semantic meaning of the text and combining common sense, is called deep reason. Identifying deep reasons can better reveal the nature of things and more accurately guide decision-making. Therefore, the present application focuses on the discovery of deep emotional reasons, which has important commercial value. On the other hand, since the deep reason often does not directly appear in the literal text, the traditional extraction-based method is difficult to identify. Moreover, deep reasons usually involve semantic understanding and common sense reasoning, but machines are not good at abstract knowledge understanding and reasoning ability. In this challenging task, there is currently a lack of research, and the academic research value is great.

[0004] In the academic field, traditional methods study the independent extraction of emotional reasons or the joint extraction of emotional-reason pairs. Among them, the emotional reason extraction task refers to extracting emotion-related fragments or word groups as reasons from the given text; the emotion-reason pair extraction task is generally a clause-level extraction, aiming to extract a series of sentences with causal relationship from the given document as emotion-reason pairs. However, both of these two types of work belong to shallow emotion recognition. It is understood that there is currently no research on deep emotional reason discovery.

[0005] For emotion cause discovery, traditional techniques can be summarized into three categories, namely rule-based method based on cause trigger words, statistical-based machine learning method and deep learning-based method. The first category of method, people construct rule set by observing the language features of the data under the guidance of experience, and then extract emotional expression and its cause. The statistical method mainly extracts the features related to emotion and reason, and extracts emotion and reason by regarding emotion and reason extraction as a text classification or sequence labeling problem. This method generally assumes that the cause of triggering emotion is an event or a series of events, and the emotion cause is near the emotion word. The general process is as follows: first, find meaningful real words in a paragraph, then determine the classification features such as event features, linguistic features, distance features, and lexical features, and finally use these features to complete the classification or sequence labeling of emotion causes. With the wide application of deep learning in natural language processing, neural network-based methods have been applied to text emotion cause extraction since 2017. The general process is as follows: first, map the words to the vector space; second, automatically extract text features through a neural network model; and finally, use the softmax function to calculate the maximum probability to extract emotion causes. Existing deep extraction models include convolutional neural networks, recurrent neural networks, long short-term memory networks, gated recurrent networks, Transformers, and graph convolution networks, which are combined, transformed, and fused with attention mechanisms to construct more complex and efficient models. In addition, there are studies that use a multi-task learning framework to jointly learn emotion classification and emotion cause extraction tasks, or introduce sub-clause relative position information, external knowledge, and external emotion vocabulary to improve the performance of emotion cause recognition. In the above methods, the rule method generally identifies a phrase in the text, and the rules usually contain lexical features or semantic features. However, due to the limitations of rules, the coverage and completeness are insufficient. The statistical method generally relies on feature engineering and large-scale training data, so the disadvantage is the over-reliance on feature selection. These features usually need to be designed manually, and the scalability is weak, making it difficult to handle complex semantic information and the logic between sentences. The deep learning method is still in the stage of identifying and extracting surface cause events, and lacks the mechanism to discover implicit or deep reasons, and the reasoning ability is still weak.

[0006] In the patent field, there is currently a lack of technology for identifying the underlying causes of emotion. Existing work has primarily focused on extracting superficial causes. CN113486657A incorporates external knowledge into the word encoding representation stage to assist in the extraction of emotion-cause pairs. However, due to insufficient knowledge utilization, the ability to discern causal relationships between clauses remains weak. CN113505583A and CN114065769B respectively use semantic decision graph neural networks and graph attention networks to extract emotion-cause clauses. CN114357164A uses an attention module integrated with a Gaussian prior to extract emotion-cause pairs. However, this method combines the sets of emotion clauses and cause clauses using a Cartesian product, resulting in high computational cost and low efficiency. CN113988048A proposes using a multi-round machine reading comprehension method with a reflection mechanism to verify extracted emotion-cause pairs and reduce noise. However, these methods are only suitable for identifying superficial causes. Through methods such as sequence labeling or classification, fragments or sentences are extracted from a given text as causes. The identification of emotional causes only stays at the surface level of the text, ignoring the fact that these surface causes contain deeper causes that have a more causal logical relationship with the generation of emotions. Summary of the Invention

[0007] The purpose of the present invention is to provide a method and device for identifying the underlying causes of text emotions based on common sense reasoning, so as to discover the underlying causes of emotions contained in the semantics of the text.

[0008] To achieve the above objectives, the present invention provides a method for identifying the underlying causes of text emotions based on common sense reasoning, comprising:

[0009] S1. First, extract potential sentiment clauses and potential cause clauses from the sentences in the text. Then, construct a sentiment clause graph and a cause clause graph by treating the clauses as nodes and the dependencies between the clauses as edges. For each node in the sentiment clause graph and the cause clause graph, retrieve common sense first. Then, expand the original graph by adding the retrieved common sense as an adjacent knowledge node to the corresponding node, thus obtaining the expanded sentiment clause graph and cause clause graph.

[0010] S2. For the expanded sentiment clause graph and cause clause graph, the nodes are updated through the calculation of multiple layers of graph attention. During the calculation process, the node representation incorporates the information of the clause context and the corresponding external common sense knowledge, and the updated sentiment clause graph and cause clause graph are obtained.

[0011] S3. For the updated emotion clause graph and reason clause graph, first filter out the nodes for combination and pairing, then pair the filtered nodes of the emotion clause graph with the nodes of the reason clause graph to obtain multiple candidate emotion-reason combinations; after obtaining a candidate emotion-reason combination, concatenate the hidden vector representations of the emotion clause graph nodes and the reason clause graph nodes of the candidate combination in the graph structure to obtain a joint representation vector of the candidate combination; then perform multi-head attention interaction calculation on the candidate emotion-reason combination to calculate the emotion-reason vector combination with the highest correlation; decode the vector of the emotion-reason vector combination with the highest correlation to obtain the deep emotion cause.

[0012] As a preferred solution, in step S1, the sentences in the text are preprocessed and segmented, and the obtained words are searched in the sentiment dictionary SentiWordNet, and the clauses in which the words appear in the sentiment dictionary SentiWordNet are regarded as potential sentiment clauses to obtain the potential sentiment clause set. ; and regard all sentences in the text as potential cause clauses to obtain the set of potential cause clauses .

[0013] As a preferred solution, the pre-trained model RoBERTa is used to encode all clauses in the text, with the goal of obtaining each clause Vector representation of , in order to better understand and process natural language. Before inputting into RoBERTa, special characters [CLS] and [SEP] are added to the front and back of the clause to mark the beginning and end of the sentence respectively. Each clause is represented as ,in, Indicates a word in a clause.

[0014] As a preferred solution, when generating common sense knowledge for a clause, the clause is replaced with a placeholder and used as a prompt to input the GPT large model. By sampling, a set of common sense descriptions of the clause can be obtained. ; Considering that these descriptions may contain noise, they are filtered by point mutual information PMI, which can calculate the vector representation of the clause And the correlation between each description k, refer to formula (1),

[0015]

[0016] because and k are sentences, use PrLM to estimate the probability, and use the clause As a condition to calculate the mutual information between the common sense description and the clause:

[0017]

[0018] PMI measures the dependency between clauses and knowledge, the greater the correlation, the greater the PMI value; that is, when the PMI score is positive, there is a positive correlation between the two; when PMI is zero, there is no correlation; when PMI is negative, the clause and the knowledge may be mutually contradictory; by only retaining the knowledge with a positive PMI score, and expanding it to the adjacent knowledge node of the corresponding clause node.

[0019] As a preferred solution, in step S1, COMET model is used to retrieve common sense related to each clause from the knowledge base; for the reason clause graph, 4 types related to causality are used: xIntend, xWant, xNeed, xEffect, wherein xIntend describes the intention or purpose behind the behavior; xEffect describes the impact caused by the event or behavior; xWant represents a goal or desire that a person wants to achieve; xNeed represents the needs or requirements of a person; for the emotion clause graph, 3 types related to emotional expression are used: xReact, oReact, xAttr, wherein xReact represents the feelings and reactions of others to the event; oReact represents the feelings and reactions of oneself to the event; xAttr represents attributes or traits.

[0020] As a preferred solution, the knowledge generated by COMET in the new knowledge node is subjected to emotion recognition, the knowledge is divided according to the emotion category, and then the knowledge consistent with the emotion category in the document is selected, the emotion is mapped to the sentiment polarity, and the sentiment dictionary SentiWordNet is used to calculate the sentiment score,

[0021] For a piece of common sense knowledge, the positive score, negative score and objective score of each word in the sentiment dictionary SentiWordNet are queried, and then the average value of the three scores is calculated, respectively, to obtain and . Then according to formula (3):

[0022]

[0023] the sentiment score of this piece of common sense knowledge is calculated if , it represents that the common sense knowledge is positive, negative, otherwise neutral, after calculating the sentiment of all common sense knowledge, the common sense knowledge is divided into three groups according to the emotion, and the common sense matching the emotion contained in the input text is selected.

[0024] As a preferred solution, in step S2, for the cause clause graph, the number of nodes n in the expanded cause clause graph is the sum of the number of clauses in the original document and the newly added adjacent knowledge nodes of all nodes.

[0025] Assume { } represents the input node representation of the t-th layer graph attention, For nodes No. The output of the layer represents that the input of the first layer of graph attention is { }. Each node in the graph is updated using the following aggregation method, as shown in Formula 4:

[0026]

[0027] in, For output representation, and are the learning parameters, Representation and Node Connected neighbors, then, the attention weight can be calculated , used to represent nodes and nodes Correlation

[0028]

[0029] in, It is a concatenation operation. By stacking T layers of graph attention layers to model the relationship between sentences, we get the output clause representation of the last layer of graph attention, which is denoted as .

[0030] As a preferred solution, in step S3, the node screening rules of the sentiment clause graph and the cause clause graph are as follows:

[0031] For the cause clause graph, the nodes of the original document clause are not considered, and only the expanded knowledge nodes are retained to avoid discovering some shallow causes;

[0032] For the sentiment clause graph, only the nodes of the original document clauses in the sentiment clause graph are considered, and the expanded knowledge nodes are not considered;

[0033] Consider only clauses with sentiment Relative position in the original document The extended adjacent knowledge node of the cause clause, the extended node in the cause clause graph is included in the scope of pairing with the emotion sentence as an independent reason.

[0034] As a preferred solution, in step S3, the joint representation vector of the candidate combination is ;

[0035] all candidate combinations are represented as a matrix where t is the number of all candidate emotion-reason combinations that satisfy the condition, in order to fully obtain the global information of an emotion-reason combination, a multi-head attention mechanism is used on H, as shown in the following formula (7):

[0036]

[0037] where, , , , , ; represents the number of attention heads, is the dimension of the multi-head attention hidden layer vector, in order to explicitly introduce relative position information, relative position encoding is added and spliced through the formula , where and are learned parameters;

[0038] The emotion-reason vector combination is decoded by the decoder:

[0039] First, the hidden state of the decoder is initialized sequence, then the context vector of the decoder at time step t is calculated At each time step t, the decoder GRU layer accepts the input vector to obtain the hidden state at time step t , and the attention distribution is obtained through formulas 8 and 9:

[0040]

[0041] where, , , and are learnable parameters, next, the obtained attention distribution is weighted and summed with the hidden state of the encoder to obtain the context vector , ; concatenate and and input them into two linear layers to obtain the word distribution , . Where , , , are learnable parameters, is the probability distribution of all words in the vocabulary;

[0042] For each time step t, the probability of generating a sentence is obtained through the context vector , decoder state and decoder input all three are obtained . wherein, , , , are learnable parameters, is a sigmoid activation function; next, is a soft switch that decides whether to generate a word from the vocabulary or copy a word from the source text according to the vocabulary probability The final vocabulary probability distribution is represented by equation 10:

[0043]

[0044] In this equation, if w is a word in the vocabulary, is greater than zero; if w does not appear in the source text, will be equal to zero; during the training process, the loss value of time step t is the negative log-likelihood value of the target word The overall loss of the entire sequence is: .

[0045] The application also provides a text emotion deep reason recognition device based on common sense reasoning, comprising:

[0046] A common sense causal graph construction unit is configured to extract potential emotional clauses and potential reason clauses from sentences in the text, and then construct an emotional clause graph and a reason clause graph by taking the clauses as nodes and the dependency relationships between the clauses as edges; for each node in the emotional clause graph and the reason clause graph, common sense is first searched out, and the searched common sense is extended to the original graph by being newly added as an adjacent knowledge node of the corresponding node, so as to obtain the extended emotional clause graph and the extended reason clause graph.

[0047] An emotional semantic graph representation unit is configured to update nodes by calculating a multi-layer graph attention layer for the extended emotional clause graph and the extended reason clause graph, and in the calculation process, the information of the clause context and the information of the corresponding external common sense knowledge are integrated into the node representation, so as to obtain the updated emotional clause graph and the updated reason clause graph.

[0048] ​​The graph interaction reasoning generation unit is used for screening out nodes for combination pairing of the updated sentiment clause graph and the cause clause graph, pairing the screened nodes of the sentiment clause graph with the nodes of the cause clause graph to obtain a plurality of candidate sentiment-cause combinations, splicing the hidden layer vector representations of the sentiment clause graph nodes and the cause clause graph nodes in the graph structure of a candidate combination to obtain a joint representation vector of the candidate combination after obtaining a candidate sentiment-cause combination, performing multi-head attention interaction calculation on the candidate sentiment-cause combination, and calculating a sentiment-cause vector combination with the highest correlation degree, and decoding the vector of the sentiment-cause vector combination with the highest correlation degree to obtain deep sentiment causes.

[0049] Compared with the prior art, the present application has the beneficial effects that:

[0050] The present application first identifies potential sentiment clauses and potential cause clauses in the text, constructs corresponding sentiment clause graphs and cause clause graphs, then retrieves sentiment and causal common sense related to the clauses to expand the graph structure, obtains two graph neural networks representing the context semantics and common sense knowledge of the sentiment graph and the cause graph, and finds the sentiment-cause combination clause pair with the highest correlation degree by using the interaction attention mechanism, and further decodes the deep causes with strong causal relationship and the corresponding sentiment. Since the text grammar structure is fully described and the relevant external common sense is fused, the present application can better decode the deep sentiment causes, realize the mining of the potential deep sentiment causes behind the surface sentiment cause events of the text, and effectively improve the accuracy and reliability of sentiment cause inference. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 is an example graph of deep sentiment causes;

[0052] Figure 2 is a flowchart of the text emotion deep reason recognition method based on common sense reasoning of the embodiment of the present application;

[0053] Figure 3 is a prompt example of common sense knowledge of the embodiment of the present application;

[0054] Figure 4 is a COMET generation example of the embodiment of the present application;

[0055] Figure 5 is a principle diagram of the text emotion deep reason recognition device based on common sense reasoning of the embodiment of the present application. DETAILED DESCRIPTION

[0056] The specific embodiments of the present application will be further described in detail below in combination with the drawings and examples. The following examples are used to illustrate the present application, but not to limit the scope of the present application.

[0057] Term explanation:

[0058] Common sense reasoning: an important research direction of natural language processing, which aims to help computers learn common sense knowledge shared by most people and use the knowledge to make deep understanding and reasoning.

[0059] Text sentiment: the user emotion category and related attributes contained in the text.

[0060] Deep reason of emotion: the deep reason of causing user emotion obtained by induction and summary, which is used to infer unobserved causal relationships or new causal relationship hypotheses.

[0061] Embodiment one

[0062] As shown in the figure, a text emotion deep reason recognition method based on common sense reasoning in the preferred embodiment of the application comprises: Figure 2 S1, first extract potential emotional clauses and potential reason clauses from the sentences in the text, and then construct an emotional clause graph and a reason clause graph by taking the clauses as nodes and the dependency relationships between the clauses as edges; for each node in the emotional clause graph and the reason clause graph, first retrieve common sense, and then extend the original graph by adding the retrieved common sense as an adjacent knowledge node of the corresponding node to obtain the extended emotional clause graph and reason clause graph;

[0063] S2, for the extended emotional clause graph and reason clause graph, the nodes are updated through the calculation of multiple layers of graph attention layers, and in the calculation process, the information of the clause context and the information of the corresponding external common sense knowledge are integrated into the node representation, to obtain the updated emotional clause graph and reason clause graph;

[0064] S3, for the updated emotional clause graph and reason clause graph, first select the nodes used for combination and pairing, then pair the selected nodes of the emotional clause graph with the nodes of the reason clause graph to obtain multiple candidate emotion-reason combinations; after obtaining a candidate emotion-reason combination, the hidden layer vector representation of the emotion clause graph node and the reason clause graph node in the graph structure of the candidate combination is spliced to obtain the joint representation vector of the candidate combination; then, the candidate emotion-reason combination is calculated through multi-head attention interaction, and the emotion-reason vector combination with the highest relevance is calculated; the vector of the emotion-reason vector combination with the highest relevance is decoded to obtain the deep emotional reason.

[0065]

[0066] ​The embodiment first identifies potential emotional clauses and potential reason clauses in the text, and constructs corresponding emotional clause graphs and reason clause graphs. Then, the emotional and causal common sense related to these clauses is retrieved to expand the graph structure, obtaining two graph neural networks representing the context semantics and common sense knowledge of the emotional and reason graphs. The interactive attention mechanism is used to find the emotional-reason combination clause pair with the maximum relevance, and further decode the deep reason with strong causal relationship and the corresponding emotion. Since the text grammar structure is fully described and the relevant external common sense is fused, the embodiment can better decode the deep emotional reason, realize the mining of the potential deep emotional reason behind the surface emotional reason event of the text, and effectively improve the accuracy and reliability of the emotional reason inference.

[0067] Specifically, the method can depict the context association between text sentences, better understand the text semantics and emotion, and also integrate external common sense knowledge for reasoning. The method can realize causal reasoning with the aid of graph neural networks and external common sense knowledge, and generate deep emotional reasons. The method can provide better decision support for applications such as psychological counseling, product optimization, and public opinion monitoring. The advantages of the method are as follows:

[0068] 1. Deep reason reasoning capability: The method can not only find shallow emotional reasons, but also reason out deep emotional reasons. Unlike traditional extraction methods, the generative reason discovery model in the method expresses more diversely and freely when generating reasons, and can better capture complex relationships in the text. Compared with the extraction method, the method is more suitable for complex situations. Generating deep emotional reasons can directly express the internal mechanism and nature of the event, and therefore has more explanatory and persuasive power. It can mine more implicit and subtle causal relationships, thus producing more accurate and reliable prediction results.

[0069] 2. Introduction of external common sense knowledge: The method has common sense perception and reasoning capability, and can not only use common sense to understand the deep reasons implied in literal events. These advantages can make the model better understand and explain emotional events, effectively improving the accuracy and reliability of emotional reason inference.

[0070] 3. Stronger application value: The deep emotional reason discovery device has stronger practicality and popularization value, and can be applied to various emotional analysis scenarios, such as social media monitoring, public opinion analysis, brand management, etc.

[0071] Embodiment Two

[0072] The difference between this embodiment and Embodiment One is that, based on Embodiment One, step S1 is further described in this embodiment.

[0073] S1. First, potential sentiment clauses and potential cause clauses are extracted from the sentences in the text. Then, a sentiment clause graph and a cause clause graph are constructed respectively by treating the clauses as nodes and the dependencies between clauses as edges. For each node in the sentiment clause graph and the cause clause graph, common sense is first retrieved. The retrieved common sense is then expanded by adding the adjacent knowledge nodes of the corresponding nodes to obtain the expanded sentiment clause graph and cause clause graph.

[0074] Step S1 constructs an association graph of reasons and associated sentiment clauses by analyzing the semantic relations and potential related common sense in a given text.

[0075] Step S1 constructs a sentiment clause graph and a cause clause graph. This graph-based approach depicts the logical relationships between textual sentences in a fine-grained manner and introduces commonsense knowledge that facilitates sentiment understanding and causal reasoning. Specifically, this step first analyzes the document using a grammatical parser to extract potential sentiment sentences and potential cause clauses, and then constructs the corresponding clause association graph. Subsequently, relevant commonsense is retrieved for each graph node, expanding the implicit sentiment and commonsense information through this explicit approach.

[0076] Clause graph construction: First, sentiment clauses and cause clauses are extracted from sentences in the text, and then sentiment information is used more finely to support causal reasoning.

[0077] The potential sentiment clauses are obtained by pre-processing the sentences in the input text and searching the obtained words in the sentiment dictionary SentiWordNet. SentiWordNet is a sentiment dictionary widely used in natural language processing, which contains a large number of sentiment words and the sentiment polarity scores corresponding to the words. The clauses in which the words appear in the sentiment dictionary are regarded as potential sentiment clauses, and the set of potential sentiment clauses is obtained. Since the cause clause may be the sentiment clause itself or other clauses in the document, all clauses in the text are considered as potential cause clauses, and the set of potential cause clauses is obtained. .

[0078] Subsequently, this paper conducts characterization learning for clauses. The pre-trained model RoBERTa is used to encode all clauses in the text, with the goal of obtaining each clause. Vector representation of , in order to better understand and process natural language. Before inputting into RoBERTa, special characters [CLS] and [SEP] are added to the front and back of the clause to mark the beginning and end of the sentence respectively. Each clause is represented as ,in, Represents the words in the clause. Finally, the sentiment and reason clause graphs are constructed respectively by taking the clauses as nodes and the dependencies between clauses as edges.

[0079] Then, common sense is introduced by extending the graph. Based on the clause graph, this step retrieves relevant common sense to improve causal reasoning capabilities. One of the most direct sources of common sense is the knowledge graph. However, since the graph is usually limited to fixed relationship categories, the coverage is limited. This embodiment proposes to use a large language model containing a large amount of common sense to obtain more complete knowledge. By combining these two knowledge sources, the knowledge gap between events and deep causes can be better bridged, and the clues involved in the implicit intermediate reasoning steps can be supplemented, thereby helping to discover deep causes. Based on the above construction of each graph node in the sentiment and cause clause graph, common sense is first retrieved, and the original graph is expanded by adding adjacent knowledge nodes to the corresponding clause nodes, thereby achieving the purpose of enriching general common sense information, sentiment information and causal information. The specific method is as follows:

[0080] (a) Large Language Model: Prompt learning is a technique for extracting the knowledge built into a large language model. Prompt sentences can be designed and constructed to guide the large model in generating sentences tailored to specific tasks. Large language models, pre-trained on massive amounts of text data, have already acquired a wealth of general language and world knowledge. In this step, prompt sentences are designed to enable the large language model (GPT) to flexibly generate common sense related to the events in the sentence, regardless of type.

[0081] like Figure 3 The common sense knowledge prompt example shown in the figure contains instructions, examples, and placeholders. When generating common sense knowledge for a new clause, the clause is replaced with the placeholder. This is used as a prompt to input the GPT large model, and a set of common sense descriptions of the clause can be obtained through sampling. Considering that these descriptions may contain noise, we filter them through pointwise mutual information (PMI). Pointwise mutual information PMI can calculate the vector representation of the clause The correlation between each description k is shown in Formula 1.

[0082]

[0083] because and k are sentences, use PrLM to estimate the probability, and use the clause As a condition to calculate the mutual information between the common sense description and the clause:

[0084]

[0085] PMI measures the dependency between clauses and knowledge, the greater the correlation, the greater the PMI value. That is, when the PMI score is positive, there is a positive correlation between the two; when PMI is zero, there is no correlation; when PMI is negative, the clause and the knowledge between them may be mutually contradictory. We only keep the knowledge with a positive PMI score and extend it to the adjacent knowledge nodes of the corresponding clause nodes.

[0086] (b) Knowledge graph: This embodiment uses the COMET model to retrieve knowledge related to each clause from the knowledge base. The COMET model is pre-trained on the knowledge graph and contains rich graph knowledge, which can output associated knowledge of a specified type.

[0087] Considering that social common sense knowledge is beneficial to reasoning the impact, consequences and emotions of events, and helps to discover deep reasons. This method uses social common sense knowledge to expand the clause graph. The common sense base is provided by the knowledge graph ATOMIC-2020. In order to facilitate the query of knowledge in ATOMIC-2020, researchers trained a BART-based generation model COMET using triples in the graph. COMET takes text and relationship types as input and outputs corresponding common sense knowledge in the form of short sentences.

[0088] 1. Causally related common sense knowledge: For the cause clause graph, this embodiment uses four types related to causality: xIntend, xWant, xNeed, and xEffect. Among them, xIntend describes the intention or purpose behind the behavior; xEffect describes the impact caused by the event or behavior; xWant represents a person's goal or desire to achieve; xNeed represents a person's needs or requirements.

[0089] 2. Emotion-related common sense knowledge: For the emotion clause graph, this embodiment uses three types related to emotional expression: xReact, oReact, and xAttr. Among them, xReact represents the feelings and reactions of others to events; oReact represents the feelings and reactions of oneself to events; xAttr represents attributes or traits. Figure 4 An example of generating common sense knowledge through the xReact relationship is given.

[0090] Considering that the results generated by COMET have noise and the deep reason is often contained in the input clauses with consistent emotions, the embodiment proposes to filter these noises by emotion type. Specifically, the embodiment performs emotion recognition on the knowledge generated by COMET in the newly added knowledge node, divides the knowledge according to the emotion category, and then selects the knowledge consistent with the emotion category in the document. Map the emotion to the sentiment polarity, and use SentiWordNet to calculate the sentiment score. For a piece of common sense knowledge, query the positive score, negative score and objective score of each word in SentiWordNet, and then calculate the average value of the three scores respectively to obtain and . According to the formula:

[0091]

[0092] the sentiment score of this piece of common sense knowledge is calculated if , it means that the common sense knowledge is positive, negative, otherwise neutral. After calculating the sentiment of all common sense knowledge, the common sense knowledge is divided into three groups according to the emotion, and the matching one is selected according to the emotion contained in the input document.

[0093] The other steps of the embodiment are the same as those of embodiment one, which will not be repeated here.

[0094] Embodiment three

[0095] The difference between the embodiment and embodiment two is that the embodiment further describes step S2 on the basis of embodiment two.

[0096] S2, for the expanded emotion clause graph and reason clause graph, the nodes are updated through the calculation of multiple layers of graph attention layers, and in the calculation process, the information of the clause context and the information of the corresponding external common sense knowledge are integrated into the node representation, to obtain the updated emotion clause graph and reason clause graph.

[0097] In order to find the most relevant reasons for emotions, the structural knowledge in the modeling graph and the potential common sense association are built. Specifically, the structural information of the graph is represented by the graph attention layer, the common sense knowledge and the context knowledge are integrated and represented, and then the correlation between the emotion and the reason is calculated.

[0098] Among them, the graph interaction attention is modeled by stacking multiple layers of graph attention layers to model the inter-sentence relationship. Taking the reason clause graph as an example, the number of knowledge-enhanced expanded reason graph nodes n is the sum of the number of original document clauses and all newly added adjacent knowledge nodes. Assuming that represents the input node representation of the t-th layer of graph attention, is the output representation of the node . The input of the first layer graph attention is . For each node in the graph, the update is done by the following aggregation as shown in equation 4.

[0099]

[0100] where is the output representation, and are the learned parameters, denotes the neighbors connected to the node . Then, the attention weight can be calculated to represent the relevance between the node and the node .

[0101]

[0102] where is the concatenation operation. By stacking T layers of graph attention layers to model the inter-sentence relationship, the output sentence representation of the last layer of graph attention is denoted as .

[0103] This step updates the graph nodes through the calculation of multiple layers of graph attention layers. In the calculation process, the node representation is integrated with the information of the sentence context and the corresponding external common sense knowledge. This can enhance the sentence representation ability of the model. At the same time, the external introduced common sense knowledge is fully interacted with the original document sentence, realizing the transmission of knowledge in the graph in the process of node attention calculation, and then assisting the model to understand the semantics. In step S3, the model has sufficient background common sense information and context information for reasoning the deep reason of emotion.

[0104] The other steps of this embodiment are the same as those of embodiment two, which will not be described here.

[0105] Embodiment four

[0106] The difference between this embodiment and embodiment three is that, on the basis of embodiment three, this embodiment further describes step S3.

[0107] ​S3. For the updated emotion clause graph and reason clause graph, first filter out the nodes for combination and pairing, then pair the filtered nodes of the emotion clause graph with the nodes of the reason clause graph to obtain multiple candidate emotion-reason combinations; after obtaining a candidate emotion-reason combination, concatenate the hidden vector representations of the emotion clause graph nodes and the reason clause graph nodes of the candidate combination in the graph structure to obtain a joint representation vector of the candidate combination; then perform multi-head attention interaction calculation on the candidate emotion-reason combination to calculate the emotion-reason vector combination with the highest correlation; decode the vector of the emotion-reason vector combination with the highest correlation to obtain the deep emotion cause.

[0108] Step S3 first designs rules to select candidate combinations of emotion-cause pairs. Then, based on the combination of the emotion sentence and the cause sentence represented by the graph, the underlying emotion cause is decoded. Specifically, a pairing rule for emotion and cause is first designed to filter out non-underlying causes. The interactive attention module then selects combinations with strong causal relationships. Finally, the cause node representations in these combinations are decoded to output the underlying cause.

[0109] (a) Filtering: Considering the large amount of noise in the candidate emotion reason combinations, this embodiment designs the following filtering rules:

[0110] Considering that deep-level reasons generally do not appear in the original text, for the reason clause graph, the nodes of the original document clauses are not considered, and only the expanded knowledge nodes are retained to avoid discovering some shallow reasons. Considering that the emotions corresponding to deep reasons are generally reflected in the original text. For the sentiment clause graph, only the nodes of the original document clauses in the sentiment clause graph are considered, and the expanded knowledge nodes are not considered to ensure the accuracy of the sentiment expression of the original document. In order to explore the potential reasons behind the emotions, this module starts from the sentiment clause graph nodes and combines them with the reason sentences. Considering that emotions and reasons are generally close, we calculate the relative positions of emotions and reason sentences. Only the relative position d in the original document with respect to the sentiment clause is considered. The extended adjacent knowledge node of the cause clause. The extended node in the cause graph is included in the scope of pairing with the emotion sentence as an independent reason. and reasons The hidden vector representations in the graph structure are concatenated to obtain the joint representation vector of the candidate combination .

[0111] (b) Interaction: This embodiment performs multi-head attention interaction calculation on candidate emotion-cause combinations to calculate emotion-cause combinations with strong causal relationships. All candidate combinations are represented as a matrix where t is the number of all candidate emotion-reason combinations satisfying the condition. To fully capture the global information of an emotion-reason combination, the embodiment uses multi-head attention mechanism on H, as shown in the following formula 7:

[0112]

[0113] where, , , , , .

[0114] where, represents the number of attention heads, is the multi-head attention hidden layer vector dimension. To explicitly introduce relative position information, the module adds relative position encoding and concatenates through the formula , where and are learned parameters.

[0115] (c) Generation: The embodiment is responsible for decoding the emotion-reason vector combination pair with the highest correlation calculated by the above modules to obtain deep emotional reasons. The decoder part uses a gated recurrent unit (GRU) based on attention mechanism. The GRU unit is mainly used to help the model capture context information and language structure to generate high-quality text sequences. Specifically, the embodiment first initializes the hidden state of the decoder sequence, and then calculates the context vector of the decoder at time step t At each time step t, the decoder GRU layer accepts the input vector to obtain the hidden state at time step t , and the attention distribution is obtained through formulas 8 and 9:

[0116]

[0117] where, , , and are learnable parameters. Next, the obtained attention distribution is weighted and summed with the hidden state of the encoder to obtain the context vector , . After concatenating and , the word distribution , is obtained by inputting to two linear layers. Where , , , are learnable parameters, is the probability distribution over all words in the vocabulary.

[0118] The probability of generating a sentence for each time step t is obtained by the context vector , the decoder state and the decoder input . Where, , , , are learnable parameters, is the Sigmoid activation function. Next, is a soft switch that decides whether to generate a word from the vocabulary or copy a word from the source text according to the vocabulary probability and the attention distribution . For the final vocabulary probability distribution, we use the following equation:

[0119]

[0120] In this equation, if w is a word in the vocabulary, is greater than zero; if w does not appear in the source text, will be equal to zero. During training, the loss value for time step t is the negative log-likelihood value of the target word : and the overall loss for the entire sequence is:

[0121] To avoid repetition in the generated text, we introduce a Coverage mechanism that can record the generated content and punish the repeated content by changing the attention. Specifically, a coverage vector is defined based on the sum of all past time steps' attention: where c represents the coverage degree of these words from the attention mechanism up to time t. The coverage vector is used as an additional input to the attention mechanism, and the formula for the coverage vector is replaced by , is a learnable parameter vector with the same length as v. This way, the attention will not focus on the same position repeatedly, thus avoiding the generation of the same content.

[0122] This embodiment defines a coverage loss that will give a penalty when focusing on the same position: where, is the attention distribution at time t, a hyperparameter λ is added to the coverage vector, and the final loss function is . ​

[0123] The other steps of this embodiment are the same as those of Example Three, which will not be repeated here.

[0124] Example Five

[0125] As Figure 5 shown, the embodiment provides a common sense reasoning based text emotion deep reason recognition device, which comprises:

[0126] A common sense causal graph construction unit is configured to extract potential emotion clauses and potential reason clauses from sentences in the text, and then construct an emotion clause graph and a reason clause graph by taking the clauses as nodes and the dependency relationships between the clauses as edges. For each node in the emotion clause graph and the reason clause graph, common sense is first retrieved, and the retrieved common sense is extended to the original graph by adding adjacent knowledge nodes to the corresponding nodes to obtain the extended emotion clause graph and the extended reason clause graph. The common sense causal graph construction unit is configured to implement the above step S1.

[0127] An emotion semantic graph representation unit is configured to update nodes in the extended emotion clause graph and the extended reason clause graph through the calculation of a multi-layer graph attention layer, and in the calculation process, the information of the clause context and the information of the corresponding external common sense knowledge are integrated into the node representation to obtain updated emotion clause graphs and reason clause graphs. The emotion semantic graph representation unit is configured to implement the above step S2.

[0128] A graph interaction reasoning generation unit is configured to first screen nodes for combination and pairing in the updated emotion clause graph and the updated reason clause graph, then pair the screened nodes in the emotion clause graph with the nodes in the reason clause graph to obtain a plurality of candidate emotion-reason combinations, concatenate the hidden layer vector representations of the emotion clause graph nodes and the reason clause graph nodes in the graph structure of a candidate emotion-reason combination to obtain a joint representation vector of the candidate combination after obtaining a candidate emotion-reason combination, and then perform multi-head attention interaction calculation on the candidate emotion-reason combination to obtain an emotion-reason vector combination with the highest relevance. The vector of the emotion-reason vector combination with the highest relevance is decoded to obtain a deep emotion reason. The graph interaction reasoning generation unit is configured to implement step S3.

[0129] In summary, the embodiment of the present application provides a text emotion deep reason recognition method and device based on common sense reasoning, which first recognizes potential emotional clauses and potential reason clauses in the text, and constructs corresponding emotional clause graphs and reason clause graphs; then, the emotion and causal common sense related to the clauses are retrieved to expand the graph structure, and two graph neural networks are obtained to represent the context semantics and common sense knowledge of the emotion and reason graphs; and the interactive attention mechanism is used to find the emotion-reason combination clause pair with the maximum relevance, and further decode the deep reason with a strong causal relationship and the corresponding emotion. Since the text grammar structure is fully described and the relevant external common sense is fused, the present application can better decode the deep emotional reason, realize the mining of the potential deep emotional reason behind the surface emotional reason event of the text, and effectively improve the accuracy and reliability of the emotional reason inference.

[0130] The above only describes the preferred embodiments of the present application, and it should be noted that those skilled in the art can make several improvements and replacements without departing from the technical principles of the present application, and these improvements and replacements should also be considered as the protection scope of the present application.

Claims

1. A method for identifying the underlying causes of textual emotions based on common sense reasoning, characterized by: include: S1. First, extract potential sentiment clauses and potential cause clauses from the sentences in the text. Then, construct a sentiment clause graph and a cause clause graph by treating the clauses as nodes and the dependencies between the clauses as edges. For each node in the sentiment clause graph and the cause clause graph, retrieve common sense first. Then, expand the original graph by adding the retrieved common sense as an adjacent knowledge node to the corresponding node, thus obtaining the expanded sentiment clause graph and cause clause graph. When generating common sense knowledge for a clause, the clause is replaced with a placeholder and used as a prompt to input the GPT large model. A set of common sense descriptions of the clause can be obtained through sampling. ; Considering that these descriptions may contain noise, they are filtered by point mutual information PMI, which can calculate the vector representation of the clause And the correlation between each description k, refer to formula (1), because and k are sentences, use PrLM to estimate the probability, and use the clauses as conditions to calculate the mutual information between the common sense description and the clauses: PMI measures the dependency between clauses and knowledge. The greater the correlation, the greater the PMI value. That is, when the PMI score is positive, there is a positive correlation between the two. When the PMI is zero, there is no correlation. When the PMI is negative, the clause and knowledge may be contradictory. By retaining only the knowledge with a positive PMI score and expanding it into the adjacent knowledge nodes of the corresponding clause node, S2. For the expanded sentiment clause graph and cause clause graph, the nodes are updated through the calculation of multiple layers of graph attention. During the calculation process, the node representation incorporates the information of the clause context and the corresponding external common sense knowledge, and the updated sentiment clause graph and cause clause graph are obtained. S3. For the updated emotion clause graph and reason clause graph, first filter out the nodes for combination and pairing, then pair the filtered nodes of the emotion clause graph with the nodes of the reason clause graph to obtain multiple candidate emotion-reason combinations; after obtaining a candidate emotion-reason combination, concatenate the hidden vector representations of the emotion clause graph nodes and the reason clause graph nodes of the candidate combination in the graph structure to obtain a joint representation vector of the candidate combination; then perform multi-head attention interaction calculation on the candidate emotion-reason combination to calculate the emotion-reason vector combination with the highest correlation; decode the vector of the emotion-reason vector combination with the highest correlation to obtain the deep emotion cause.

2. The method for identifying the underlying causes of text emotions based on common sense reasoning according to claim 1, characterized in that: In step S1, the sentences in the text are preprocessed and segmented, and the obtained words are searched in the sentiment dictionary SentiWordNet. The clauses in which the words appear in the sentiment dictionary SentiWordNet are regarded as potential sentiment clauses, and the potential sentiment clause set is obtained. ; and all sentences in the text are regarded as potential cause clauses to obtain the set of potential cause clauses .

3. The method for identifying the underlying causes of text emotions based on common sense reasoning according to claim 2, characterized in that: Use the pre-trained model RoBERTa to encode all clauses in the text, with the goal of obtaining each clause Vector representation of , in order to better understand and process natural language; before inputting into RoBERTa, special characters [CLS] and [SEP] are added to the front and back of the clause to mark the beginning and end of the sentence, respectively. Each clause is represented as ,in, Indicates a word in a clause.

4. The method for identifying the underlying causes of text emotions based on common sense reasoning according to claim 3 is characterized in that: In step S1, the COMET model is used to retrieve common sense related to each clause from the knowledge base; for the cause clause diagram, four types related to causality are used: xIntend, xWant, xNeed, and xEffect, where xIntend describes the intention or purpose behind the behavior; xEffect describes the impact caused by an event or behavior; xWant represents the goal or desire a person wants to achieve; xNeed represents someone's need or requirement; for the emotion clause diagram, three types related to emotional expression are used: xReact, oReact, and xAttr, where xReact represents other people's feelings and reactions to an event; oReact represents one's own feelings and reactions to an event; and xAttr represents attributes or traits.

5. The method for identifying the underlying causes of text emotions based on common sense reasoning according to claim 4 is characterized in that: Emotion recognition is performed on the knowledge generated by COMET in the newly added knowledge nodes. The knowledge is divided according to the emotion category, and then the knowledge that matches the emotion category in the document is selected. The emotion is mapped to the emotion polarity and the emotion score is calculated using the emotion dictionary SentiWordNet. For a piece of common sense knowledge, query the positive score, negative score and objective score of each word from the sentiment dictionary SentiWordNet, and then calculate the average of the three scores to get and , and then according to formula (3): Calculate the sentiment score of this common sense knowledge ,if It means that common sense knowledge is positive. If the sentiment is negative, then it is negative, otherwise it is neutral. After calculating all the common sense knowledge emotions, the common sense knowledge is divided into three groups according to the emotions, and the common sense that matches the emotion contained in the input text is selected.

6. The method for identifying the underlying causes of text emotions based on common sense reasoning according to claim 5, characterized in that: In step S2, for the reason clause graph, the number of nodes n in the expanded reason clause graph is the sum of the number of clauses in the original document and the newly added adjacent knowledge nodes of all nodes. Assume { } represents the input node representation of the t-th layer graph attention, For nodes No. The output of the layer represents that the input of the first layer of graph attention is { }; For each node in the graph, update it using the following aggregation method, as shown in Formula 4: in, For output representation, and are the learning parameters, Representation and Node Connected neighbors, then, the attention weight can be calculated , used to represent nodes and nodes Correlation in, It is a concatenation operation. By stacking T layers of graph attention layers to model the relationship between sentences, the output clause representation of the last layer of graph attention is obtained, which is recorded as .

7. The method for identifying the underlying causes of text emotions based on common sense reasoning according to claim 6, characterized in that: In step S3, the node screening rules for the sentiment clause graph and the cause clause graph are as follows: For the cause clause graph, the nodes of the original document clause are not considered, and only the expanded knowledge nodes are retained to avoid discovering some shallow causes; For the sentiment clause graph, only the nodes of the original document clauses in the sentiment clause graph are considered, and the expanded knowledge nodes are not considered; Only consider the relative position of the sentiment clause in the original document The extended adjacent knowledge node of the cause clause, the extended node in the cause clause graph is included in the scope of pairing with the emotion sentence as an independent reason.

8. The method for identifying the underlying causes of text emotions based on common sense reasoning according to claim 7, characterized in that: In step S3, the joint representation vector of the candidate combination is ; All candidate combinations are represented as matrices , where t is the number of all candidate emotion-reason combinations that meet the conditions. In order to fully obtain the global information of an emotion-reason combination, a multi-head attention mechanism is used on H, as shown in the following formula (7): in, , , , , ; represents the number of attention heads, is the multi-head attention hidden layer vector dimension. In order to explicitly introduce relative position information, relative position encoding is added. , and by the formula Splicing, where and are the learning parameters; Decode the emotion-reason vector combination through the decoder: First initialize the decoder's hidden state sequence, and then calculate the context vector of the decoder at time step t , at each time step t, the decoder GRU layer accepts the input vector and obtains the hidden state of time step t , the attention distribution is obtained by formulas 8 and 9: in, 、 、 and As a learnable parameter, the obtained attention distribution is weighted and summed with the hidden state of the encoder to obtain the context vector , ;Will and After concatenation, input into two linear layers to obtain vocabulary distribution , ,in 、 、 、 is a learnable parameter, is the probability distribution of all words in the vocabulary; For each time step t, the probability of generating a sentence Through the context vector , decoder status and decoder input The three get ,in, 、 、 、 is a learnable parameter, is the Sigmoid activation function; next, It is a soft switch, based on the vocabulary probability To decide whether to generate words from the vocabulary or based on the attention distribution Copy the words from the source text; for the final vocabulary probability distribution, use Equation 10 to express it: In this formula, if w is a word in the vocabulary, greater than zero; if w does not appear in the source text, then will be equal to zero; during training, the loss value at time step t is the target word The negative log-likelihood of : , the overall loss of the entire sequence is: .

9. A device for identifying the underlying causes of text emotions based on common sense reasoning, based on the method according to any one of claims 1 to 8, characterized in that: include: The common sense causal graph construction unit is used to extract potential sentiment clauses and potential cause clauses from sentences in the text. It then constructs a sentiment clause graph and a cause clause graph, respectively, by using clauses as nodes and dependencies between clauses as edges. For each node in the sentiment clause graph and the cause clause graph, common sense is first retrieved. This retrieved common sense is then added as an adjacent knowledge node to the corresponding node to expand the original graph, resulting in the expanded sentiment clause graph and cause clause graph. The sentiment semantic graph representation unit is used to update the nodes of the expanded sentiment clause graph and cause clause graph through calculations of multiple layers of graph attention. During the calculation process, the node representation incorporates the clause context information and the corresponding external common sense knowledge information, resulting in the updated sentiment clause graph and cause clause graph. The graph interactive reasoning generation unit is used to first screen out the nodes used for combination pairing of the updated emotion clause graph and reason clause graph, and then pair the screened nodes of the emotion clause graph with the nodes of the reason clause graph to obtain multiple candidate emotion-reason combinations; after obtaining a candidate emotion-reason combination, the hidden vector representations of the emotion clause graph nodes and the reason clause graph nodes of the candidate combination in the graph structure are spliced ​​to obtain the joint representation vector of the candidate combination; then multi-head attention interaction calculation is performed on the candidate emotion-reason combination to calculate the emotion-reason vector combination with the highest correlation; the vector of the emotion-reason vector combination with the highest correlation is decoded to obtain the deep emotion cause.

Images