Research On Emotion-Cause Pair Extraction Based On Comparative Learning

Emotion-Cause Pair Extraction(ECPE)is an important sub-task in the field of NLP,which aims to identify and extract the Emotion-Cause Pairs composed of emotional expression clauses and emotional cause clauses in the text.Identifying and extracting emotion-cause pairs can find emotional changes in hot events,and then grasp the development trend of events.This task can provide technical support for public opinion control and business decision-making.The existing ECPE methods are mainly based on pipeline structure and end-to-end structure.However,these two methods do not consider the data noise in the real scene will affect the extraction results.The extraction precision of these two methods decreases when they encounter adversarial sample attacks.The existing extraction methods have a spurious correlation between the extracted emotion-cause pairs,which is intuitively expressed as the distortion of extraction precision.To address the above problems,this thesis focuses on ECPE based on comparative learning.The specific research contents are as follows:(1)This thesis firstly studies the emotion-cause pair extraction method based on adversarial training(AT-ECPE).By introducing Chinese adversarial samples for adversarial training,AT-ECPE increases the sensitivity of the extraction method to data.This method improves the extraction accuracy by enhancing the robustness of the model.Firstly,the Chinese adversarial sample generation method is used to generate adversarial samples of the original event document.The comparison set is constructed by pairing the adversarial samples with the original event document.BERT is used to dynamically obtain the clause representation of the event document and the comparison set.Then,the positive and negative pairs in the comparison set are transferred for adversarial training to enhance the robustness of the model.Experiments are conducted on the EMNLP2016 Chinese sentiment cause discovery dataset.The results show that AT-ECPE has the characteristics of high extraction precision and strong robustness.(2)Based on the emotion-cause pair extraction method based on adversarial training,this thesis proposes the emotion-cause pair extraction model based on adversarial training and context enhancement,named ATCE-ECPE.ATCE solves the problem of precision distortion caused by spurious correlation between emotion-cause pairs in existing model extraction results.Firstly,the contrast set is constructed by using the adversarial sample generation method.The attention mechanism and Bi-LSTM are used to obtain shallow information about the original event document and the contrast set,which is convenient for the preliminary extraction of the emotion-cause pair.Then,the loss function is optimized to determine the existence of spurious correlation emotion-cause pairs in the extraction results.Finally,adversarial training is used to enhance context information.Adversarial training reduces the spurious correlation caused by the confusion factors of full matching calculation,thus improving the extraction precision of emotion-cause pairs.Experiments demonstrate the superiority of ATCE-ECPE over state-of-the-art methods in emotional detection,cause detection,and other evaluation indicators.The precision,recall,and F1 value of the emotion-cause pair extraction reached 83.36%,79.62%,and 81.45%.ATCE-ECPE achieves the known highest performance for ECPE tasks,which can provide technical support for public opinion management and business decision-making.Table [13] Figure [23] Reference [87]...