Research On The Approaches Of Textual Emotion-cause Pair Extraction

As a research hotspot in the field of natural language processing,text emotion cause extraction aims to identify the potential causes that induce a certain emotional expression in texts,which has expansive application prospects in public opinion analysis,commodity comment mining,human-computer intelligent dialogue,and so on.After nearly two decades of development,the technology of text emotion cause extraction has achieved great improvements,from the initial rule-based and machine learning-based methods to the current deep learning-based methods.At the same time,task selection has shifted from Emotion Cause Extraction(ECE),which aims at identifying the corresponding causes of a given emotion,to Emotion-Cause Pair Extraction(ECPE)that is a more challenging and aims at simultaneously identifying emotions and their corresponding causes.Nevertheless,on the one hand,ECPE has never been defined by a unified formal paradigm,and only a small number of benchmark datasets based on the formal text are available.On the other hand,existing ECPE methods still have shortcomings in achieving the mutual enhancement between emotions and causes,fully modeling the semantic information of text,handling the sample conflicts between emotions and causes,and alleviating the data sparsity in emotion-cause pairs,which limits the further improvement of performance for ECPE,so as to seriously hinders the application of ECPE technology in real scenarios.Towards the above issues,this article proposes a series of ECPE methods based on deep learning.The specific research work is unfolded around different formal definitions of ECPE,mainly focusing on the following aspects:Firstly,towards the problem that only the unidirectional enhancement is achieved between emotion extraction and cause extraction in the first stage of the pipelined methods,an ECPE method based on mutually auxiliary model and Self-Distillation is studied.According to the idea of multitask learning based on auxiliary tasks and the mechanism of sharing shallow network,a joint emotion-cause extraction model based on mutual assistance is proposed,which achieves bidirectional enhancement between two subtasks while avoiding a dead cycle in the calculation process.Then,a self-distillation method based on Born Again Networks(BAN)is designed to train the proposed multitask model,which further improve the extraction performance of emotions and causes,so as to reduce the negative impact of cross-stage error propagation on pairing the emotions and causes in the second stage.Secondly,towards the problem that semantic information in a document is not sufficient modeling,an end-to-end ECPE method based on hierarchical heterogeneous graph attention is proposed.The designed heterogeneous graph in this paper contains a variety of semantic elements involved in ECPE and the rich semantic relationships between them,not only the local semantic relationships between the homogeneous nodes corresponding to clauses or clause pairs.The proposed hierarchical model first adopts a node-level heterogeneous graph attention network to model the dependency between the clause nodes and word nodes,and then employs a meta-path based heterogeneous graph attention network to model the correlation between the clause pair nodes,which can obtain the higher-order representation of clauses and clause pairs,so as to extract the emotion-cause pairs more effectively.Thirdly,towards the data sparsity problem of emotion-cause pairs,an end-to-end ECPE method based on neural transition system is studied,which formalizes the ECPE as a directed graph construction task to alleviate the data sparsity problem caused by using Cartesian product to generate candidate clause pairs.A transition system for the unified construction of directed graphs is proposed,and the newly designed self-loop arcs(edges)is used to handle sample conflicts between emotions and causes,so as to provide the possibility for the global optimization of ECPE.On the basis of our transition system,a graph state-aware generative model is proposed,which adopts a graph neural network to encode the sub-graph states during the construction process.This strengthens the state representation of the transition system,so as to extract the emotion-cause pairs more effectively.Finally,towards the high computation complexity of emotion-cause pair extraction,an end-to-end ECPE method based on unified sequence tagging is studied.Formalizing ECPE as a clause-level sequence tagging task,which extracts the emotion cause pairs through one pass clause tagging,can further alleviate the high computation complexity problem caused by using Cartesian product to generate candidate clause pairs.A semantic block oriented unified sequence tagging scheme is proposed,and the newly introduced boundary tags are responsible for segmenting a document into semantic blocks containing a whole causality,which is conducive to simultaneously identifying multiple causes for an emotion and distinguishing the causes of different emotions.On the basis of the new tagging scheme,a dual-channel based sequence tagging model is proposed.The two channels respectively handle the dependency relationship between clauses and the dependency relationship between predicted tags,which can improve the performance of our model in ECPE.