A dialog-level weakly supervised sentiment recognition method based on progressive learning

Abstract

A dialog-level weakly supervised sentiment recognition method based on progressive learning belongs to the field of artificial intelligence, comprising: obtaining a DERC training data set, containing a plurality of dialogs and corresponding dialog-level sentiment sets; converting the dialog-level sentiment set into a pseudo-sentence-level sentiment distribution through pseudo-sentence-level sentiment label initialization to construct a pseudo-ERC data set; dividing the training data into a plurality of training subsets according to the size of the dialog-level sentiment set; using a progressive learning strategy to gradually transition from a low-noise subset to a high-noise subset for model training, and in each round of training iteration, jointly updating the pseudo-sentence-level sentiment label and the ERC predictor parameter; using the trained ERC predictor to predict the sentiment category of each sentence in the new dialog, and outputting the sentence-level sentiment prediction result. The present application maintains a high labeling accuracy while reducing the labeling cost, and improves the overall prediction performance.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a dialogue-level weakly supervised emotion recognition method based on progressive learning. Background Technology

[0002] Emotion Recognition in Conversations (ERC) has significant value in various practical applications, including medical dialogue analysis, social media analysis, and dialogue system construction. ERC technology aims to train a deep learning model to automatically identify the emotion category of each utterance in a conversation, such as sadness, happiness, or anger—a fundamental task in the natural language processing community.

[0003] Current methods for solving the problem of dialogue sentiment recognition can be divided into three categories: content-based methods, knowledge-assisted methods, and relation-based methods. Content-based methods integrate utterances with historical content within the same dialogue and input them into various text encoders; knowledge-assisted methods utilize external knowledge bases and auxiliary information (such as speaker background and utterance roles) to enhance utterance representation; and relation-based methods transform dialogues into graph structures and capture contextual dependencies between utterances through graph neural networks.

[0004] The closest existing technologies to this invention are Multi-Instance Multi-Label Learning (MIML) (Zhou ZH, Zhang ML, Huang SJ, et al. Multi-instance multi-label learning[J]. Artificial Intelligence, 2012, 176(1): 2291-2320.) and Learning from Label Proportions (LLP) (Liu J, Wang B, Shen X, et al. Two-stage Training for Learning from Label Proportions[C] / / International Joint Conference on Artificial Intelligence. 2021: 2737-2743.). The MIML method associates instance bags with bag-level label sets, rather than instance-level labels, and trains the model through techniques such as self-training and contrastive learning. The LLP method, on the other hand, utilizes the label proportion information of instance bags and aggregates instance-level prediction results into bag-level outputs through objective functions such as proportion loss and consistency regularization.

[0005] Existing technologies suffer from the following drawbacks and shortcomings: First, existing dialogue sentiment recognition methods are all built on fine-grained labeled training datasets, making them entirely dependent on utterance-level fine-grained annotation. This means that each utterance in the dialogue needs to be manually labeled with a sentiment tag, resulting in high annotation costs and low efficiency. Unlike traditional sentiment recognition tasks, the sentiment of utterances in dialogue sentiment recognition is highly dependent on the dialogue context. The same utterance may express completely different sentiments in different dialogue contexts. This characteristic requires annotators to repeatedly review the entire dialogue to assign accurate sentiment tags to each utterance, leading to extremely high costs for fine-grained annotation. According to existing annotation statistics and preliminary experiments, the average time to annotate a single utterance is approximately 10 seconds, which is unacceptable for practical applications. Second, while existing weakly supervised learning methods (such as MIML and LLP) can handle bag-level labels, they fail to fully consider the special structure of dialogue. In dialogue sentiment recognition tasks, utterances in the same dialogue are interdependent coupled sequences, rather than independent bags of instances. Existing weakly supervised learning methods cannot effectively utilize the dependencies between utterances in a dialogue. Third, existing dialogue sentiment recognition methods lack effective strategies when dealing with training samples with different noise densities. When training with a dialogue-level sentiment set, the varying number of sentiments contained in the set can lead to different levels of label noise. Existing dialogue sentiment recognition methods cannot learn specifically based on the differences in noise density. Summary of the Invention

[0006] To address the problems of high annotation costs, inability to adapt to dialogue characteristics, and lack of noise-aware learning strategies in existing technologies, this invention provides a dialogue-level weakly supervised emotion recognition method based on progressive learning.

[0007] The technical solution adopted by this invention to solve the technical problem is as follows: This invention provides a dialogue-level weakly supervised emotion recognition method based on progressive learning, comprising the following steps: Step 1: Obtain the DERC training dataset, which contains multiple dialogues and their corresponding dialogue-level sentiment sets; Step 2: Initialize the dialogue-level sentiment set by using pseudo-discourse-level sentiment tags to convert it into a pseudo-discourse-level sentiment distribution, and construct a pseudo-ERC dataset; Step 3: Divide the training data into multiple training subsets according to the size of the dialogue-level sentiment set; Step 4: Employ a progressive learning strategy to gradually transition from low-noise subsets to high-noise subsets for model training. In each training iteration, jointly update the pseudo-discourse-level sentiment tags and ERC predictor parameters. Step 5: Use the trained ERC predictor to predict the sentiment category of each utterance in the new dialogue and output the utterance-level sentiment prediction results.

[0008] Furthermore, the DERC training dataset consists of N coarse-grained labeled samples, denoted as... , Indicates dialogue, Indicates dialogue The corresponding dialogue-level sentiment set, Represents a conversational emotional space; each conversation Depend on The rounds of speech are denoted as follows: , The original text content of the discourse.

[0009] Furthermore, in step two, the sentiment probabilities in the dialogue-level sentiment set are evenly distributed to all utterances in the dialogue to construct a pseudo-utterance-level sentiment distribution.

[0010] Furthermore, for each utterance Use the dialogue-level sentiment set of its respective dialogue. Initialize its pseudo-discourse-level sentiment tags : ; in, Represents a collection of conversational sentiments The number of emotions contained therein, and the pseudo-discourse-level emotion tags satisfy the normalization constraints: , Indicates the first In the first dialogue The first statement The pseudo-utterance-level sentiment probabilities for each sentiment category are M, where M represents the number of utterance turns in each dialogue, and N represents the number of coarse-grained labeled samples in the DERC training dataset.

[0011] Furthermore, in step three, a dialogue-level sentiment aggregation is used. The number of emotions included To represent from dialogue The noise density of the speech is used to divide the pseudo-ERC dataset into... A non-overlapping training subset.

[0012] Furthermore, in step four, the initial training pool contains only the lowest noise subset, and after each iteration... After each training round, a subset of the next noise level is added to the training pool until all data is included.

[0013] Furthermore, in step four, the ERC predictor is optimized using the focus loss function. The parameters are set, and the pseudo-discourse-level sentiment labels are updated for samples whose confidence exceeds the threshold and whose sentiment prediction results are covered by the dialogue-level sentiment set.

[0014] Furthermore, in step four, given any training pool... The following training objective function is used for optimization:

[0015] in, Represents the cross-entropy loss function. Indicates ERC predictor Dialogue The sentiment prediction results This represents the focus loss hyperparameter used to mitigate the data imbalance problem. Indicates ERC predictor parameter, This represents pseudo-discourse-level sentiment tags, where M represents the number of discourse turns in each dialogue. Expression of speech Pseudo-discourse level emotional tags.

[0016] Furthermore, in step four, regarding pseudo-discourse-level sentiment labels... Based on the current sentiment prediction results Detailed updates will be made:

[0017] Among them, symbols This represents element-wise multiplication. express Norm, This represents the operation of maximizing a vector. Indicates the confidence threshold. Indicates the sentiment prediction result The sharpened version, This represents a collection of conversational emotions.

[0018] Furthermore, in step four, after updating through pseudo-discourse-level sentiment tags, each pseudo-discourse-level sentiment tag is covered by the dialogue-level sentiment set, and only sentiment prediction results with high confidence are used to update pseudo-discourse-level sentiment tags.

[0019] The beneficial effects of this invention are: (1) This invention proposes an efficient coarse-grained annotation strategy, which significantly reduces annotation costs while maintaining high annotation accuracy by assigning a set of emotions to each dialogue instead of assigning a single emotion label to each utterance.

[0020] (2) This invention proposes a new weakly supervised learning task, namely Dataless ERC (DERC), which uses dialogue-level sentiment sets to train an ERC predictor that can accurately identify discourse-level sentiment, significantly reducing annotation costs.

[0021] (3) This invention proposes a training strategy based on pseudo-discourse-level sentiment labels, which initializes the pseudo-discourse-level sentiment distribution through a dialogue-level sentiment set and updates the pseudo-discourse-level sentiment labels based on high-confidence prediction results during the training process.

[0022] (4) This invention proposes a progressive learning strategy (DERC-PL), which obtains more accurate pseudo-narrative-level sentiment labels in the early stage of training by dividing the training subset according to noise density and gradually transitioning from low noise subset to high noise subset, thereby improving the overall prediction performance of this invention. Attached Figure Description

[0023] Figure 1 The flowchart illustrates a dialogue-level weakly supervised emotion recognition method based on progressive learning, as provided in this invention. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to the accompanying drawings.

[0025] The present invention provides a dialogue-level weakly supervised emotion recognition method based on progressive learning, which mainly includes two stages: a training stage and a testing stage.

[0026] During the training phase, the model's input is a DERC training dataset containing multiple dialogues and their corresponding dialogue-level sentiment sets. First, the dialogue-level sentiment sets are converted into pseudo-dialogue-level sentiment distributions using pseudo-dialogue-level sentiment tags, thus constructing a pseudo-ERC dataset. Then, the training data is divided into multiple training subsets based on the size (noise density) of the dialogue-level sentiment sets, and a progressive learning strategy is employed to gradually transition from low-noise subsets to high-noise subsets for model training. In each training iteration, the pseudo-dialogue-level sentiment tags and ERC predictor parameters are jointly updated. During the testing phase, the trained ERC predictor is used to predict the sentiment category of each utterance in the new dialogue, and the utterance-level sentiment prediction results are output.

[0027] This invention provides a dialogue-level weakly supervised emotion recognition method based on progressive learning, the specific implementation process of which is as follows: Step 1: Data Acquisition; Obtain the DERC training dataset, which contains multiple dialogues and their corresponding dialogue-level sentiment sets. Each dialogue is associated with a dialogue-level sentiment set rather than a utterance-level sentiment label.

[0028] Formally, the DERC training dataset consists of N coarse-grained labeled samples, denoted as... Each of them This indicates a dialogue. Indicates dialogue The corresponding dialogue-level sentiment set, This represents a conversational emotional space. For example, a conversational emotional space can include seven emotions: neutral, happy, sad, fearful, angry, disgusted, and surprised.

[0029] Specifically, each dialogue Depend on The rounds of speech are denoted as follows: ,in It represents the raw text content of a utterance, but its true sentiment label is unknown. Dataless ERC (DERC) aims to utilize the DERC training dataset. Training an ERC predictor This enables it to accurately predict the emotion of each utterance in future conversations.

[0030] Step 2: Initialize pseudo-discourse-level sentiment tags; This invention trains the model based on a pseudo-discourse-level sentiment labeling strategy, and converts the dialogue-level sentiment set into a pseudo-discourse-level sentiment distribution through pseudo-discourse-level sentiment labeling initialization.

[0031] For each utterance in the dialogue, pseudo-utterance-level sentiment labels are initialized: the sentiment probabilities from the dialogue-level sentiment set are evenly distributed among all utterances in the dialogue, constructing a pseudo-utterance-level sentiment distribution. Specifically, for each utterance... Use the dialogue-level sentiment set of its respective dialogue. Initialize its pseudo-discourse-level sentiment tags Its mathematical expression is as follows:

[0032] in, Represents a collection of conversational sentiments The number of emotions contained therein, and the pseudo-discourse-level emotion tags satisfy the normalization constraints: , Indicates the first In the first dialogue The first statement Pseudo-discourse-level sentiment probabilities for each sentiment category.

[0033] By initializing the pseudo-narrative-level sentiment tags as described above, a pseudo-ERC dataset can be constructed. .

[0034] Step 3: Generate a training subset; The training data is divided into multiple training subsets based on the size of the dialogue-level sentiment set, i.e., the noise density.

[0035] Because of each word pseudo-discourse-level emotional tags It must be based on its corresponding dialogue-level sentiment set It is covered, so conversational sentiment sets can be used. The number of emotions included To represent from dialogue The noise density of the speech. Based on this, the pseudo-ERC dataset... Divided into The mathematical expression for the disjoint training subsets is as follows:

[0036] Each subset Defined as: The conditions are met. .in, This represents the number of emotions contained in the dialogue-level emotion set, i.e., the noise density level corresponding to the current training subset. Representing a subset The number of dialogues between them.

[0037] Noise density corresponds to learning difficulty to some extent: the fewer emotions contained in the dialogue-level emotion set, the lower the uncertainty of the pseudo-discourse-level emotion label, and the easier the learning difficulty.

[0038] Step 4: Progressive learning strategy; A progressive learning strategy is employed to gradually transition the model training from a low-noise subset to a high-noise subset: the initial training pool contains only the lowest noise subset, and each subsequent training pool... After each training round, a subset of the next noise level is added to the training pool until all data is included. In each training iteration, pseudo-narrative-level sentiment labels are jointly updated. and ERC predictor Parameters: Optimize the ERC predictor using the focus loss function. The parameters are set, and the pseudo-discourse-level sentiment labels are updated for samples whose confidence exceeds the threshold and whose sentiment prediction results are covered by the dialogue-level sentiment set.

[0039] Given a training subset, a progressive learning strategy is employed to learn from the subset. To subset Gradually update pseudo-discourse-level emotional tags and ERC predictor The specific learning process is as follows: Step S401: Use subset Initialize the training pool Using the training pool Jointly update pseudo-discourse-level sentiment tags and ERC predictor conduct Rotational training.

[0040] Step S402: For : (a) subset Add to training pool: ; (b) Continue to utilize the expanded training pool Jointly update pseudo-discourse-level sentiment tags and ERC predictor conduct Rotational training.

[0041] Step S403: Pseudo-discourse-level sentiment tag update strategy; Given any training pool The following training objective function is used for optimization:

[0042] in, Represents the cross-entropy loss function. Indicates ERC predictor Dialogue The sentiment prediction results This represents the focus loss hyperparameter used to mitigate the data imbalance problem. Indicates ERC predictor parameter.

[0043] For parameters The update is performed using a gradient-based method.

[0044] Regarding words pseudo-discourse-level emotional tags Based on the current sentiment prediction results Detailed updates will be made:

[0045] Among them, symbols This represents element-wise multiplication. express Norm, This represents the operation of maximizing a vector. Indicates the confidence threshold; Indicates the sentiment prediction result The sharpened version, through temperature parameters Perform the calculation:

[0046] in, Indicates the ERC predictor for the first... The discourse sample belongs to the first The predicted probability of each sentiment category is determined by the temperature parameter. The result after sharpening Indicates the ERC predictor for the first... The discourse sample belongs to the first The predicted probability of each sentiment category is determined by the temperature parameter. The result after sharpening.

[0047] The above pseudo-discourse-level sentiment tag update strategy ensures that: (1) each pseudo-discourse-level sentiment tag Must be used in conversational emotional collections (2) Only high-confidence sentiment prediction results are covered; (Based on confidence threshold) Only metrics will be used to update pseudo-discourse-level sentiment tags. .

[0048] The optimized discourse-level sentiment prediction predictor, or ERC predictor, is obtained after training. .

[0049] Step 5: Testing Phase; The new dialogue to be predicted is input into the trained ERC predictor. In the process, sentiment category prediction is performed for each utterance in the new dialogue, and the utterance-level sentiment prediction results are output.

[0050] The dialogue-level weakly supervised sentiment recognition method based on progressive learning provided by this invention was validated on three benchmark datasets (MELD, IEMOCAP, and EmoryNLP), and the following technical results were achieved: (1) Significantly improved annotation efficiency and significantly reduced annotation cost: By adopting the dialogue-level sentiment aggregation annotation strategy, it only takes 10-16 seconds to annotate a dialogue containing 5-10 utterances, which is several times more efficient than traditional utterance-level annotation (about 10 seconds per utterance). At the same time, the annotation accuracy rate remains above 94.5%, which can significantly reduce the annotation cost.

[0051] (2) Improved performance of weakly supervised learning: Under the DERC setting, the method proposed in this invention is significantly better than the existing weakly supervised learning methods PLOT and ItS2CLR on all datasets, with an average improvement of about 3%-6% in Macro-F1 index.

[0052] (3) Strong versatility: The progressive learning strategy (DERC-PL) proposed in this invention is a flexible strategy that can directly integrate various existing ERC methods (such as BERTbase+MLP, RGAT, SACL, DualGATs) as basic predictors, and has good scalability and practicality.

[0053] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A dialogue-level weakly supervised emotion recognition method based on progressive learning, characterized in that, Includes the following steps: Step 1: Obtain the DERC training dataset, which contains multiple dialogues and their corresponding dialogue-level sentiment sets; Step 2: Initialize the dialogue-level sentiment set by using pseudo-discourse-level sentiment tags to convert it into a pseudo-discourse-level sentiment distribution, and construct a pseudo-ERC dataset; Step 3: Divide the training data into multiple training subsets according to the size of the dialogue-level sentiment set; Step 4: Employ a progressive learning strategy to gradually transition from low-noise subsets to high-noise subsets for model training. In each training iteration, jointly update the pseudo-discourse-level sentiment tags and ERC predictor parameters. Step 5: Use the trained ERC predictor to predict the sentiment category of each utterance in the new dialogue and output the utterance-level sentiment prediction results.

2. The dialogue-level weakly supervised emotion recognition method based on progressive learning according to claim 1, characterized in that, The DERC training dataset consists of N coarse-grained labeled samples, denoted as N_N ... , Indicates dialogue, Indicates dialogue The corresponding dialogue-level sentiment set, Represents a conversational emotional space; each conversation Depend on The rounds of speech are denoted as follows: , The original text content of the discourse.

3. The dialogue-level weakly supervised emotion recognition method based on progressive learning according to claim 1, characterized in that, In step two, the sentiment probabilities in the dialogue-level sentiment set are evenly distributed to all utterances in the dialogue to construct a pseudo-utterance-level sentiment distribution.

4. The dialogue-level weakly supervised emotion recognition method based on progressive learning according to claim 3, characterized in that, For each utterance Use the dialogue-level sentiment set of its respective dialogue. Initialize its pseudo-discourse-level sentiment tags : ； in, Represents a collection of conversational sentiments The number of emotions contained therein, and the pseudo-discourse-level emotion tags satisfy the normalization constraints: , Indicates the first In the first dialogue The first statement The pseudo-utterance-level sentiment probabilities for each sentiment category are M, where M represents the number of utterance turns in each dialogue, and N represents the number of coarse-grained labeled samples in the DERC training dataset.

5. The dialogue-level weakly supervised emotion recognition method based on progressive learning according to claim 1, characterized in that, In step three, a conversational sentiment set is used. The number of emotions included To represent from dialogue The noise density of the speech is used to divide the pseudo-ERC dataset into... A non-overlapping training subset.

6. The dialogue-level weakly supervised emotion recognition method based on progressive learning according to claim 1, characterized in that, In step four, the initial training pool contains only the lowest noise subset, and after each iteration... After each training round, a subset of the next noise level is added to the training pool until all data is included.

7. The dialogue-level weakly supervised emotion recognition method based on progressive learning according to claim 1, characterized in that, In step four, the ERC predictor is optimized using the focus loss function. The parameters are set, and the pseudo-discourse-level sentiment labels are updated for samples whose confidence exceeds the threshold and whose sentiment prediction results are covered by the dialogue-level sentiment set.

8. The dialogue-level weakly supervised emotion recognition method based on progressive learning according to claim 1, characterized in that, In step four, given any training pool The following training objective function is used for optimization: in, Represents the cross-entropy loss function. Indicates ERC predictor Dialogue The sentiment prediction results This represents the focus loss hyperparameter used to mitigate the data imbalance problem. Indicates ERC predictor parameter, This represents pseudo-discourse-level sentiment tags, where M represents the number of discourse turns in each dialogue. Expression of speech Pseudo-discourse level emotional tags.

9. A dialogue-level weakly supervised emotion recognition method based on progressive learning according to claim 1, characterized in that, In step four, regarding pseudo-discourse-level sentiment labels... Based on the current sentiment prediction results Detailed updates will be made: Among them, symbols This represents element-wise multiplication. express Norm, This represents the operation of maximizing a vector. Indicates the confidence threshold. Indicates the sentiment prediction result The sharpened version, This represents a collection of conversational emotions.

10. A dialogue-level weakly supervised emotion recognition method based on progressive learning according to claim 1, characterized in that, In step four, after updating the pseudo-discourse-level sentiment tags, each pseudo-discourse-level sentiment tag is covered by the dialogue-level sentiment set, and only the sentiment prediction results with high confidence are used to update the pseudo-discourse-level sentiment tags.

Images