Multiparty Dialogue Discourse Parsing And Application

Since Alan Turing proposed the famous "Turing Test," enabling machines to understand natural language has been a crucial research objective in the field of artificial intelligence.Existing natural language understanding tasks primarily focus on studying documents as their target.In recent years,with the development of the internet,conversational text has become increasingly prevalent,leading to significant advancements in dialogue research.Based on the number of participating members,dialogues can be divided into two-party dialogues and multiparty dialogues.Based on the different types of dialogue participants,multiparty dialogues can be further categorized into multi-person dialogues,multiparty human-machine dialogues,and group chats involving multiple machines.A discourse is a textual unit composed of one or more sentences or utterances.Discourse parsing aims to analyze the overall discourse to identify the semantic and logical relationships between basic discourse units.The goal of discourse parsing for multiparty dialogues is to parse a multiparty conversational text into a discourse dependency graph.Therefore,this task can also be referred to as multiparty dialogue discourse dependency parsing,which includes two subtasks: predicting discourse dependency links and recognizing discourse relation types.Traditional discourse parsing techniques have primarily focused on news texts.However,the discourse structure of multiparty dialogue texts often exhibits complex graph structures,which poses challenges to understanding multiparty dialogues.Therefore,further research on discourse parsing for multiparty dialogues is crucial to better serve downstream tasks in dialogue understanding.This thesis aims to investigate the techniques and applications of discourse parsing for multiparty dialogues.Starting with the construction of a large-scale dataset for multiparty dialogue discourse structure,this thesis proposes two models for analyzing the discourse structure of multiparty dialogues: a model based on speaker modeling and a model based on adapters.Furthermore,this thesis applies discourse parsing to the task of multiparty dialogue machine reading comprehension.This thesis focuses on the research of multiparty dialogue discourse parsing and its application and carries out the following four aspects of research.1.Build a large-scale multiparty dialogue discourse structure and reading comprehension dataset Molweni.The existing STAC dataset for multiparty dialogue discourse parsing is relatively small,which limits the application of current state-of-the-art representation learning models in this task.To further advance the task of discourse parsing for multiparty dialogues,this thesis introduces the Molweni dataset,which serves as a foundation for training complex neural network models for this task.In order to validate the role of discourse structure in downstream tasks,the Molweni dataset includes annotations for questions and answers specifically tailored to multiparty dialogues.As a result,this dataset can be used for training and evaluating multiparty dialogue machine reading comprehension tasks,thereby promoting advancements in this field.Molweni represents the first multiparty dialogue machine reading comprehension dataset that incorporates discourse structure information.2.A speaker-aware multiparty dialogue discourse parser is proposed.Compared to traditional discourse parsing tasks focused on news articles,multiparty dialogue discourse parsing has a distinctive feature: the input dialogue text consists of utterances from different speakers.In the task of dialogue discourse parsing,there are certain relation types that only occur between utterances from different speakers,while other relation types occur only between different utterances from the same speaker.Therefore,the speaker information in the dialogue plays a crucial role in identifying the discourse structure.To address this,this thesis employs heterogeneous graph neural networks to explicitly model the dialogue speakers and achieve significant performance improvement in the task of multiparty dialogue discourse parsing.3.An adapter-based multiparty discourse parser is proposed.In the task of multiparty dialogue discourse parsing,existing models have mainly focused on the overall performance of the dataset without considering the severe class imbalance in the distribution of different categories.This leads to underfitting issues in training certain class categories.To address this problem,this thesis proposes a multiparty dialogue discourse parsing model based on adapters.Our model addresses the issue of class imbalance by using different adapters to learn balanced representations of discourse dependency edges and discourse relation types.This thesis validates the effectiveness of our model on existing publicly available datasets.4.A discourse-aware model for multiparty dialogue reading comprehension is proposed.The effectiveness of discourse structure in traditional machine reading comprehension tasks has been experimentally validated.However,existing machine reading comprehension models for multiparty dialogues,which typically have more complex discourse structures,have not yet incorporated discourse structure information.This thesis proposes a modeling approach based on discourse structure for multiparty dialogues,using graph neural networks to model the discourse dependency graph and utterance representations.Experimental results demonstrate that incorporating discourse structure into the modeling process significantly benefits the task of multiparty dialogue machine reading comprehension.In summary,this thesis addresses the issue of data scarcity by constructing a largescale dataset for multiparty dialogue discourse parsing.In light of the task characteristics of discourse parsing for multiparty dialogue,this study delves into the significance of speaker modeling in the context of multi-party discourse parsing tasks.In response to the issue of imbalanced category distribution in multiparty discourse parsing datasets,this thesis investigates the performance enhancement brought about by the adapter-based approach.Finally,the thesis integrates the discourse structure information of multiparty dialogues into the task of multiparty dialogue machine reading comprehension,resulting in performance enhancement.The research conducted in this thesis validates that multiparty dialogue discourse structure can provide effective assistance in other natural language processing tasks.Looking forward to extending the findings of this thesis to more natural language processing tasks and further advancing the field.