Identification method and system for multiple emotion identification models based on DS theory

Abstract

The invention provides a DS theory-based recognition method and system for multiple emotion recognition models, and the method comprises the steps: obtaining a comment emotion data set and a dialogue data set, and carrying out the preprocessing of the obtained data; constructing and training a plurality of independent emotion recognition sub-models by utilizing the preprocessed data set, and respectively outputting probabilities; a weight coefficient alpha i of each sub-model is distributed, an uncertainty coefficient 1-alpha i is introduced, the sub-model output probability is combined, basic probability distribution is completed, and a quality function is constructed; combining the quality functions of the four sub-models into a new quality function; calculating a sub-trust function based on the new quality function, and then calculating a combined likelihood function; the trust function and the likelihood function form a trust interval which represents the confirmation degree of emotion recognition, and final probability distribution is determined according to needs; and performing emotion judgment based on probability distribution.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing and sentiment analysis, and in particular relates to a recognition method and system for multiple sentiment recognition models based on DS theory. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Sentiment Analysis is an important task in Natural Language Processing (NLP), which aims to identify and extract subjective information (such as opinions, emotions, attitudes, etc.) from text.

[0004] With the rapid development of social media and internet content, sentiment analysis technology has been widely used in areas such as public opinion monitoring and user behavior analysis. Sentiment classification models are machine learning or deep learning models used to automatically determine the emotional orientation of text. Different model architectures and methods can be used depending on task complexity, data size, and requirements.

[0005] However, existing sentiment classification models have the following limitations:

[0006] (1) A single model provides too little information for psychological monitoring and is prone to omissions, making it difficult to fully analyze emotions.

[0007] (2) In actual data, the proportion of normal emotion samples is significantly higher than that of negative emotions. During the training process of traditional models, the imbalance of samples may lead to insufficient recognition accuracy of a few negative emotions such as anger, disgust, fear, and sadness, and there is a bias in classification decisions.

[0008] (3) The outputs of different models may be contradictory, and the traditional weighted average method cannot effectively solve the problem of evidence conflict.

[0009] Furthermore, existing technologies use DS theory as an effective tool for handling uncertainty. It can provide a reasonable fusion mechanism when conflicting outputs from different sub-models exist, quantifying uncertainty through trust and likelihood functions. However, its direct application faces the following challenges: the deterministic probabilities output by the classifier lead to the degradation of trust intervals, resulting in the Zadeh paradox. Summary of the Invention

[0010] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a recognition method for multiple emotion recognition models based on DS theory, adopts an adaptive uncertainty quality allocation strategy and decision function to construct an efficient multi-model fusion framework, which is mainly suitable for single-modal model analysis of text.

[0011] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0012] In the first aspect, a recognition method of multiple emotion recognition models based on DS theory is disclosed, including:

[0013] Obtain comment sentiment datasets and conversation datasets, and preprocess the data in the obtained data;

[0014] Use the preprocessed dataset to build and train multiple independent emotion recognition sub-models and output probabilities respectively;

[0015] Assign weight coefficient α to each sub-model i , introduce the uncertainty coefficient 1-α i , combined with the output probability of the sub-model, complete the basic probability distribution and construct the quality function;

[0016] The quality functions of the four sub-models are combined into a new quality function;

[0017] Based on the new quality function, calculate the sub-trust function and then calculate the combined likelihood function;

[0018] The trust function and likelihood function form a trust interval, which indicates the degree of confirmation of the emotion recognition and determines the final probability distribution as needed;

[0019] Emotional judgment is performed based on the probability distribution, and the sum of negative emotion probabilities is calculated. If the sum of negative emotions is greater than the probability of normal emotions, the negative emotion with the highest probability is selected as the final output, otherwise it is judged as a normal emotion.

[0020] As a further technical solution, multiple independent emotion recognition sub-models include:

[0021] The XGBoost bag-of-words model is trained based on a review sentiment dataset and uses the bag-of-words model to extract lexical features of the text.

[0022] Sentiment dictionary polarity analysis model, used to extract sentiment polarity features in text and calculate sentiment scores based on the polarity value and intensity of the vocabulary;

[0023] A supervised learning hidden Markov model uses two conversation datasets as training data, organizes consecutive sentences of the same user into context sequence features, uses text as the observed state and emotion as the implicit state, and outputs the forward probability as the emotion probability distribution;

[0024] The hidden Markov model fine-tuned based on the EM algorithm, on the basis of the hidden Markov model training based on supervised learning, introduces sentiment-balanced text data, and uses the EM algorithm to optimize the model parameters.

[0025] As a further technical solution, the bag-of-words model is used to extract text features, including: using the word segmentation tool to segment the Chinese text, constructing TF-IDF feature vectors for the segmentation results, and implementing two-level classification based on the gradient boosting algorithm XGBoost for the feature vectors: the first-level classification distinguishes normal emotions from negative emotions, and the second-level classification further subdivides negative emotions into multiple categories, and finally outputs a probability distribution.

[0026] As a further technical solution, the hidden Markov model based on supervised learning also includes: extracting the implicit emotional features in the context, converting them into the model's initial probability distribution, state transition probability distribution and observation probability distribution, and outputting the forward probability as the emotional probability distribution.

[0027] As a further technical solution, the sentiment score is calculated based on the polarity and intensity of the vocabulary. The specific steps include:

[0028] For each emotion, obtain its raw score;

[0029] Get the total sentiment score based on the raw score of each sentiment;

[0030] For each emotion, its final score ratio is obtained.

[0031] As a further technical solution, the quality function m corresponding to the i-th model i (e) The calculation formula is:

[0032] m i (e) = P i (e)·α i

[0033] m i (Ω)=1-α i

[0034] Among them, Ω represents the set of all possible emotions, and the uncertainty parameter 1-α i Used to represent the uncertain part of the model output, the actual α i The value can be determined based on the accuracy of each model on the validation set or by domain experts. i (e) Output the probability of emotion e for the i-th sub-model.

[0035] Secondly, a recognition system for multiple emotion recognition models based on DS theory is disclosed, including:

[0036] The dataset acquisition module is configured to: acquire the comment sentiment dataset and the conversation dataset, and preprocess the acquired data;

[0037] The sub-model construction module is configured to: construct and train multiple independent emotion recognition sub-models using the pre-processed dataset and output probabilities respectively;

[0038] The quality function building block is configured to: assign a weight coefficient α to each sub-model i , introduce the uncertainty coefficient 1-α i , combined with the output probability of the sub-model, complete the basic probability distribution and construct the quality function;

[0039] The final probability distribution determination module is configured to: combine the quality functions of the four sub-models into a new quality function;

[0040] Based on the new quality function, calculate the sub-trust function and then calculate the combined likelihood function;

[0041] The trust function and likelihood function form a trust interval, which indicates the degree of confirmation of the emotion recognition and determines the final probability distribution as needed;

[0042] The recognition module is configured to: make emotion judgment based on probability distribution, calculate the sum of negative emotion probabilities, and if the sum of negative emotions is greater than the probability of normal emotions, select the negative emotion with the highest probability as the final output; otherwise, judge it as a normal emotion.

[0043] One or more of the above technical solutions have the following beneficial effects:

[0044] The technical solution of the present invention realizes the effective integration of multi-source emotion evidence by constructing a set of independent sub-models, designing an uncertainty quality allocation strategy and optimizing decision rules, thereby improving the accuracy and comprehensiveness of emotion recognition and the ability to recognize a small number of negative emotions.

[0045] Given the difficulty of collecting multimodal data and the difficulty of achieving universality for online textual comments, we prioritized analyzing pure textual content, particularly for comments with multiple replies. Hidden Markov models and corresponding fine-tuned models can better analyze the correlations between them. Furthermore, the entire approach utilizes machine learning algorithms, rather than deep learning, which reduces computing power and training costs. Furthermore, DS theory can better combine the strengths of each sub-model, resulting in a more comprehensive and balanced recognition of different sentiments.

[0046] To address the problem that the deterministic probability of the classifier output leads to the degradation of the trust interval, the sub-technical solution of this embodiment introduces the uncertainty coefficient α, which recognizes that the emotion recognition of the sub-model has an uncertain component, that is, there is a certain probability of recognizing any evidence, thereby preventing the following situation: if any piece of evidence denies a basic hypothesis, no matter how much confidence other evidence supports the basic hypothesis, the synthesis result will completely deny the basic hypothesis.

[0047] The technical solution of the present invention extracts the lexical features, sentiment polarity features and context sequence features of the text through heterogeneous sub-models, making up for the problem of insufficient feature coverage of a single model; the fine-tuned hidden Markov model is more accurate in identifying negative emotions, making up for the defect of low accuracy in identifying a few negative emotions caused by other models in pursuit of accuracy; the probability of five emotions being identified is output, and based on the uncertainty mass allocation and synthesis rules of DS theory, the output conflicts of different models are reasonably handled to form a trust interval containing direct evidence and potential supporting evidence, thereby improving the robustness and decision reliability of the fusion model.

[0048] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0050] Figure 1 This is a flowchart of sub-model fusion according to an embodiment of the present invention;

[0051] Figure 2 This is a comparison chart of the confusion matrices of each sub-model and the DS fusion model in the embodiment of the present invention.

[0052] Figure 2 (a)-(e) represent the confusion matrices of the XGBOOST bag-of-words model, sentiment dictionary polarity analysis model, first-order hidden Markov model, fine-tuned hidden Markov model, and DS fusion model, respectively. The horizontal axis is the predicted sentiment and the vertical axis is the true sentiment. DETAILED DESCRIPTION

[0053] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0054] It should be noted that the terms used herein are for describing particular embodiments only and are not intended to limit the exemplary embodiments according to the present invention.

[0055] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0056] Example 1

[0057] This embodiment discloses a recognition method for multiple emotion recognition models based on DS theory. The confidence intervals of various emotions can be obtained by inputting text data, including:

[0058] Step (1): Construct four heterogeneous sub-models based on different training data and feature representation methods, input text training data, and realize that the i-th sub-model outputs the emotion probability P i (e);

[0059] Step (2): Assign weight coefficient α to each sub-model i , introduce the uncertainty coefficient 1-α i , combined with the above sub-model output probability, complete the basic probability distribution, and construct the quality function m i (e);

[0060] Step (3): Based on Dempster’s synthesis rule, the quality functions of the four sub-models are combined into a new quality function m(e);

[0061] Step (4): Based on the new quality function calculated in step (3), calculate the sub-belief function Bel(e), which is calculated as the sum of direct evidence; then calculate the combined likelihood function Pl(e), which is calculated as the sum of non-negative evidence;

[0062] Step (5): The trust function and the likelihood function form a trust interval [Bel(A), Pl(A)], which represents the degree of confirmation of the emotion recognition. Depending on the required situation, if you pursue accuracy, you can choose the trust function as the final probability distribution; if you pursue precision without missing detection, you can choose the likelihood function as the final probability distribution. Usually, the average of the trust function and the likelihood function is taken as the final probability distribution;

[0063] Step (6): Emotional judgment is performed based on the probability distribution. The sum of the probabilities of the four negative emotions (anger, disgust, fear, and sadness) is calculated. If the sum of the negative emotions is greater than the probability of normal emotions, the negative emotion with the highest probability is selected as the final output. Otherwise, it is judged as a normal emotion.

[0064] Before building the four heterogeneous sub-models, both the comment sentiment dataset and the conversation dataset underwent unified data preprocessing, including removing special symbols and filtering stop words. For the conversation dataset, we further screened for sequences with more than three consecutive texts and more than two valid words per text to ensure the validity of the model parameters.

[0065] In this embodiment, step (1) specifically includes:

[0066] (1.1) XGBoost Bag-of-Words Model: We use the bag-of-words model to extract text features. After segmenting Chinese text using the Jieba word segmentation tool, we use the Python function CountVectorizer to construct word frequency and inverse text frequency (TF-IDF) feature vectors. We then implement two-level classification using the XGBoost gradient boosting algorithm: the first-level classification distinguishes between normal and negative emotions, and the second-level classification further subdivides negative emotions into anger, disgust, fear, and sadness. Finally, we use the Python function predict_proba to output the probability distribution. The training data is a dataset of annotated comment sentiment.

[0067] When training the XGBoost classification model based on the bag-of-words model, the XGBoost classification model uses the XGBClassifier of the scikit-learn library, sets the learning rate to 0.2, the maximum depth to 7, and the subsampling rate to 0.8, and performs two-level classification training based on the comment sentiment dataset.

[0068] (1.2) Sentiment Dictionary Polarity Analysis Model: Based on the Dalian University of Technology DUTIR sentiment vocabulary ontology, sentiment words are extracted from the text, sentiment scores are calculated based on the polarity and intensity of the words, and sentiment judgments are output in a quantitative manner;

[0069] When training the polarity analysis model based on the DUTIR sentiment dictionary, the DUTIR sentiment dictionary is used to construct a sentiment vocabulary mapping table, the sentiment scores are calculated according to the formula, and the normalized output is the sentiment recognition probability.

[0070] The above-mentioned DUTIR sentiment dictionary comes from the existing technology, for example, from the literature citation "Xu Linhong, Lin Hongfei, Pan Yu, et al. Construction of sentiment vocabulary ontology [J]. Journal of the China Society for Scientific and Technological Information, 2008, 27(2): 180-185.".

[0071] (1.3) Hidden Markov model based on supervised learning: Using two dialogue datasets as training corpus, the continuous sentences of the same user are organized into state sequences, with text as the observed state and emotion as the implicit state. The emotional features implicit in the context are extracted based on statistics and converted into the model's initial probability distribution, state transition probability distribution, and observation probability distribution. The statistical range of the observation probability is based on the DUTIR emotional vocabulary ontology, and the forward probability is output as the emotion probability distribution.

[0072] When extracting the implicit emotional features of a context based on statistics, we detect whether a text contains the emotional terms in the aforementioned emotional vocabulary mapping table. The probability of these terms being present in the training set is calculated as the observed probability. The statistical probability of the emotion of the first text in each text group is the initial state distribution. The emotional transition between the previous and next texts in the same text group is the state transition matrix.

[0073] After obtaining the vocabulary range of the observation probability, the extraction is the probability of this word in the text corresponding to the emotional state of the entire data set.

[0074] When training a first-order hidden Markov model based on supervised learning, the dialogue dataset is used to statistically analyze the initial state distribution, state transition distribution, and observation probability distribution, and the forward algorithm is used to calculate the sentiment probability distribution of the new input text.

[0075] The emotional vocabulary contained in the above emotional vocabulary mapping table is used to detect whether the text contains the emotional vocabulary, and the probability of the above vocabulary contained in the training set is counted as the observation probability. The statistical probability of the emotion of the first text in each group of texts is the initial state distribution. The emotional transition of the previous and next texts in the same text group is the state transition matrix. The statistical scope is the entire training set, the denominator of the observation probability is the number of all words that appear in the emotional text of the training set, and the numerator is the number of times a certain word appears (and each different emotional state is calculated separately). The state matrix is ​​one of the elements, and the denominator is the total number of emotional transitions in all training sets. If the element is normal to normal, the numerator is the number of normal to normal transitions, and the state transition matrix is ​​a 5*5 matrix. The initial state matrix is ​​the distribution of all emotional states of the first text in all sequences in the training set. The denominator is the total number of sequences, and the numerator is the total number of emotional labels for the first text in the sequence of a certain emotion.

[0076] (1.4) Hidden Markov model fine-tuned based on the EM algorithm: Based on the model training in (1.3), emotionally balanced text data is introduced in a targeted manner. User diaries in Chinese Internet communities that mainly express negative emotions are selected as the fine-tuning dataset, and the EM algorithm is used to optimize the model parameters.

[0077] When training a hidden Markov model fine-tuned using the EM algorithm, an iterative optimization algorithm based on a first-order hidden Markov model is performed using the Baum-Welch algorithm (a form of the EM algorithm). User diaries from Chinese internet communities, primarily expressing negative emotions, were selected as a fine-tuning dataset, with a focus on optimizing the model's ability to recognize negative sentiment. To avoid overfitting, a constrained update strategy was adopted: only the emission probabilities of existing words were updated (no new words were introduced), while the initial probability distribution and state transition probabilities were updated normally. Model parameters were saved with each iteration, ultimately retaining only the optimized parameter set that showed improved and stable performance.

[0078] Among them, the training data of the XGBoost bag-of-words model and the hidden Markov model are different. The former is a comment sentiment dataset, and the latter is a conversation sentiment dataset. In addition, the training data used to fine-tune the hidden Markov model, that is, the fine-tuning dataset, is also different from the former two. This shows that these four sub-models have weak independence and meet the use conditions of DS theory.

[0079] Of the four models mentioned above, the XGBOOST bag-of-words model offers high accuracy, the sentiment polarity analysis model offers high average recall, and the hidden Markov model can better identify sentiment based on text that is continuous over a period of time, such as the preceding and following text of one party in a conversation. The fine-tuned model also demonstrates high accuracy for anger.

[0080] In this implementation example, the specific emotion score judgment expression in step (1.2) is as follows:

[0081] For each emotion e∈E, its raw score S e for:

[0082]

[0083] The total sentiment score is the sum of all normalized sentiment scores:

[0084]

[0085] Where T represents the input text, E = {normal, angry, afraid, sad, disgust} represents the emotion set, and f T (w) represents the frequency of word w in text T, L e (w) represents the intensity value of word w in sentiment e, D e Represents the subset of the sentiment dictionary corresponding to sentiment e.

[0086] For each emotion e, its final score ratio F e for:

[0087]

[0088] Among them, δ e,normal is the Kronecker function, which takes the value 1 when e is a normal emotion and 0 otherwise. e The output probability of emotion e is used to map the vocabulary judgment in the emotion mapping table to the emotion recognition of the actual text through the above method.

[0089] In step (2), the weight coefficient α is assigned to the i-th sub-model i , and introduce the uncertainty parameter 1-α i The quality function m corresponding to the i-th model i (e) The calculation formula is:

[0090] m i (e) = P i (e)·α i

[0091] m i (Ω)=1-α i

[0092] Where Ω represents the full set (i.e., the set of all possible emotions). Uncertainty parameter 1-α i Used to represent the uncertain part of the model output, the actual α i The value can be determined based on the accuracy of each model on the validation set or by domain experts. This method eliminates the Zadeh paradox and makes the output an interval, providing more information.

[0093] In this embodiment, the uncertainty coefficient is introduced. For the XGBoost bag-of-words model, α1 is 0.9, for the sentiment dictionary polarity analysis model, α2 is 0.99, for the first-order hidden Markov model, α3 is 0.8, and for the fine-tuned hidden Markov model, α4 is 0.7. After adding the weight and the sentiment probability obtained in step (2), the final quality function m is generated. i (e i ).

[0094] In particular, the Dempster synthesis rule for step (3) is as follows:

[0095] On the recognition frame Θ, for the recognition element Finite number of mass functions m1(A1), m2(A2), ..., m n (A n ) is expressed as follows:

[0096]

[0097] Where K is the normalization constant, which is calculated as follows:

[0098]

[0099] This normalization constant K represents the total amount of non-conflicting evidence; the denominator is 1 minus the total amount of conflicting evidence. When conflicting evidence is generated by different sub-models, the synthesis rule uses the K factor to resolve the conflict and achieve effective fusion of evidence.

[0100] Specifically, for e∈E, the quality functions m1(e1), m2(e2), m3(e3), and m4(e4) of the four sub-models are combined based on the Dempster synthesis rule to form m(e), which is the new quality function, expressed as follows:

[0101]

[0102] Among them, K is a normalization constant used to eliminate conflicts between different evidences:

[0103]

[0104] In this embodiment, in step (4), for A, the belief function Bel(A) and the likelihood function Pl(A) are theoretically defined as follows:

[0105]

[0106] Bel(A) represents the total amount of direct evidence that fully supports A, representing the minimum confidence level for A; Pl(A) represents the total amount of evidence that does not negate the sentiment set A, representing the maximum likelihood for A. Together, the two constitute the confidence interval [Bel(A), Pl(A)], describing the range of support for the sentiment judgment.

[0107] Specifically in this implementation, A represents the proposition that the sentiment of a text is this sentiment, B represents the quality function of the evidence supporting this proposition, and Bel represents the sum. The belief function Bel(A) represents the minimum confidence that proposition A holds based on the current evidence. The likelihood function Pl(A) includes both direct evidence supporting sentiment e and uncertain evidence that may support sentiment e, and outputs a confidence interval [Bel(A), Pl(A)], which includes a certain degree of uncertainty.

[0108] In step (5), the final probability distribution and sentiment judgment are performed. Usually, the average value of the trust function and the likelihood function is taken as the final probability distribution. The final probability distribution expression P final (e) as follows:

[0109] P final (e)=(Bel(e)+Pl(e)) / 2

[0110] In order to improve the ability to identify a few negative emotions, the optimized emotion judgment rule is as follows: calculate the sum of the probabilities of four negative emotions (anger, disgust, fear, and sadness). If the sum of the probabilities of negative emotions is greater than the probability of normal emotions, then choose the negative emotion with the highest probability, otherwise it is judged as a normal emotion.

[0111] To verify the difference between the proposed method and other models, we compared the performance of different methods. We conducted comparative experiments on a public sentiment dataset and obtained the results shown in Table 1.

[0112] Experimental results show that the DS fusion model is significantly better than the single sub-model in terms of macro-average precision and macro-average F1 score. Figure 2 The confusion matrix comparison diagram of each sub-model and the DS fusion model shows that the recognition performance of a few negative emotion categories is significantly improved, which verifies the effectiveness and advancement of the method of the present invention.

[0113] Table 1 Comparison of classification performance of different sub-models and DS hybrid model

[0114]

[0115] In summary, the present invention constructs a multi-sub-model fusion framework through the DS theory, which effectively solves the problems of insufficient information of a single model, minority class recognition bias and multi-model conflict. It also retains the reliability of direct supporting evidence and considers the possibility of non-negative evidence, forming a sentiment recognition method that is both robust and sensitive, providing an efficient and reliable technical solution for sentiment analysis in complex scenarios.

[0116] Since a single model provides too little information for psychological monitoring and is prone to omissions, making it difficult to fully analyze emotions, the technical solution of the present invention can adopt multi-model output, introduce confidence to increase the amount of output information, and improve the accuracy and robustness of text sentiment classification. In view of the problem of sample imbalance in the traditional model training process, the judgment basis is optimized. When the total negative emotions in the text account for too large a proportion, it is determined that there is a problem with the emotions, thereby improving the accuracy of identifying a small number of negative emotions. In view of the problem that the outputs of different models may be contradictory, the technical solution of the present invention introduces a multi-model fusion method based on DS theory to eliminate conflicts between different models.

[0117] The sub-technical solution of this embodiment adopts an adaptive uncertainty quality allocation strategy and decision function to build an efficient multi-model fusion framework. The quality allocation strategy is determined by the probability of each emotion output of all sub-models and the uncertainty coefficient of the allocation, and finally the confidence interval output is obtained by the combination. In the decision-making part, the present invention uses the mean of the trust function and the likelihood function as the final probability output. Then, when the probability of negative emotion recognition is greater than the probability of normal emotion, the emotion recognition is judged to be negative, and then the one with the highest probability is selected from the negative emotions. Of course, in actual applications, the confidence intervals of each emotion can be used for comprehensive judgment. The confidence interval provides more information. The above decision function is only for performance comparison.

[0118] Example 2

[0119] The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0120] Example 3

[0121] The purpose of this embodiment is to provide a computer-readable storage medium.

[0122] A computer-readable storage medium stores a computer program, which, when executed by a processor, performs the steps of the above method.

[0123] Example 4

[0124] The purpose of this embodiment is to provide a recognition system for multiple emotion recognition models based on DS theory, including:

[0125] The dataset acquisition module is configured to: acquire the comment sentiment dataset and the conversation dataset, and preprocess the acquired data;

[0126] The sub-model construction module is configured to: construct and train multiple independent emotion recognition sub-models using the pre-processed dataset and output probabilities respectively;

[0127] The quality function building block is configured to: assign a weight coefficient α to each sub-model i , introduce the uncertainty coefficient 1-α i , combined with the output probability of the sub-model, complete the basic probability distribution and construct the quality function;

[0128] The final probability distribution determination module is configured to: combine the quality functions of the four sub-models into a new quality function;

[0129] Based on the new quality function, calculate the sub-trust function and then calculate the combined likelihood function;

[0130] The trust function and likelihood function form a trust interval, which indicates the degree of confirmation of the emotion recognition and determines the final probability distribution as needed;

[0131] The recognition module is configured to: make emotion judgment based on probability distribution, calculate the sum of negative emotion probabilities, and if the sum of negative emotions is greater than the probability of normal emotions, select the negative emotion with the highest probability as the final output; otherwise, judge it as a normal emotion.

[0132] Example 5

[0133] The purpose of this embodiment is to provide a computer program product containing instructions, which, when running on a computer, enables the computer to execute the methods and functions involved in any of the above embodiments.

[0134] The steps involved in the apparatus of the above embodiment correspond to those of the method embodiment 1. For detailed implementation, please refer to the relevant description of embodiment 1. The term "computer-readable storage medium" should be understood to mean a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and causing the processor to perform any method of the present invention.

[0135] Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computer device. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0136] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A recognition method for multiple emotion recognition models based on DS theory, characterized by: include: Obtain comment sentiment datasets and conversation datasets, and preprocess the data in the obtained data; Use the preprocessed dataset to build and train multiple independent emotion recognition sub-models and output probabilities respectively; Assign weight coefficient α to each sub-model i , introduce the uncertainty coefficient 1-α i , combined with the output probability of the sub-model, complete the basic probability distribution and construct the quality function; The quality functions of the four sub-models are combined into a new quality function; Based on the new quality function, calculate the sub-trust function and then calculate the combined likelihood function; The trust function and likelihood function form a trust interval, which indicates the degree of confirmation of the emotion recognition and determines the final probability distribution as needed; Emotional judgment is performed based on the probability distribution, and the sum of negative emotion probabilities is calculated. If the sum of negative emotions is greater than the probability of normal emotions, the negative emotion with the highest probability is selected as the final output, otherwise it is judged as a normal emotion.

2. The identification method of multiple emotion recognition models based on DS theory as claimed in claim 1, characterized in that: Multiple independent emotion recognition sub-models include: The XGBoost bag-of-words model is trained based on a review sentiment dataset and uses the bag-of-words model to extract lexical features of the text. Sentiment dictionary polarity analysis model, used to extract sentiment polarity features in text and calculate sentiment scores based on the polarity value and intensity of the vocabulary; A supervised learning hidden Markov model uses two conversation datasets as training data, organizes consecutive sentences of the same user into context sequence features, uses text as the observed state and emotion as the implicit state, and outputs the forward probability as the emotion probability distribution; The hidden Markov model fine-tuned based on the EM algorithm, on the basis of the hidden Markov model training based on supervised learning, introduces sentiment-balanced text data, and uses the EM algorithm to optimize the model parameters.

3. The identification method of multiple emotion recognition models based on DS theory as claimed in claim 1, characterized in that: The bag-of-words model is used to extract text features, including: using the word segmentation tool to segment the Chinese text, constructing TF-IDF feature vectors for the segmentation results, and implementing two-level classification based on the feature vectors based on the gradient boosting algorithm XGBoost: the first-level classification distinguishes normal emotions from negative emotions, and the second-level classification further subdivides negative emotions into multiple categories, and finally outputs a probability distribution.

4. The method for identifying multiple emotion recognition models based on DS theory as claimed in claim 1, wherein: The hidden Markov model based on supervised learning also includes: extracting the implicit emotional features in the context, converting them into the model's initial probability distribution, state transition probability distribution and observation probability distribution, and outputting the forward probability as the emotional probability distribution.

5. The identification method of multiple emotion recognition models based on DS theory as claimed in claim 1, characterized in that: The sentiment score is calculated based on the polarity and intensity of the words. The specific steps include: For each emotion, obtain its raw score; Get the total sentiment score based on the raw score of each sentiment; For each emotion, its final score ratio is obtained.

6. The method for identifying multiple emotion recognition models based on DS theory as claimed in claim 1, characterized in that: The quality function m corresponding to the i-th model i (e) The calculation formula is: m i (e)=P i (e)·α i m i (Ω)=1-a i Among them, Ω represents the set of all possible emotions, and the uncertainty parameter 1-α i Used to represent the uncertain part of the model output, the actual α i The value can be determined based on the accuracy of each model on the validation set or by domain experts. i (e) Output the probability of emotion e for the i-th sub-model.

7. A recognition system for multiple emotion recognition models based on DS theory, characterized by: include: The dataset acquisition module is configured to: acquire the comment sentiment dataset and the conversation dataset, and preprocess the acquired data; The sub-model construction module is configured to: construct and train multiple independent emotion recognition sub-models using the pre-processed dataset and output probabilities respectively; The quality function building block is configured to: assign a weight coefficient α to each sub-model i , introduce the uncertainty coefficient 1-α i , combined with the output probability of the sub-model, complete the basic probability distribution and construct the quality function; The final probability distribution determination module is configured to: combine the quality functions of the four sub-models into a new quality function; Based on the new quality function, calculate the sub-trust function and then calculate the combined likelihood function; The trust function and likelihood function form a trust interval, which indicates the degree of confirmation of the emotion recognition and determines the final probability distribution as needed; The recognition module is configured to: make emotion judgment based on probability distribution, calculate the sum of negative emotion probabilities, and if the sum of negative emotions is greater than the probability of normal emotions, select the negative emotion with the highest probability as the final output; otherwise, judge it as a normal emotion.

8. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method described in any one of claims 1 to 6 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 6 are performed.

Images