Research On Implicit Discourse Relation Recognition Integrating Different Levels Of Clues

Recognizing implicit discourse relation is a challenging sub-task in discourse analysis,which aims to understand and annotate the latent relations between the adjacent text spans in one discourse,such as temporal relation,comparison relation.This task is beneficial to improving the performance of many downstream natural language processing(NLP)tasks,such as semantic recognition,question answering systems,machine translation and text mining.It also can facilitate the basic researches and applications at discourse-level of NLP.Currently,most of the previous studies focus on the semantic representation of discourse,ignoring the interactions between words,phrases,sentences out of discourse and even the relevant external knowledge.They could not effectively integrate multi-level features of discourse relation,which reduces the performance of implicit discourse relation identification obviously.This thesis is inspired by the relevant theories of cognitive psychology and linguistics.From the shallow linguistic features of discourse itself to the relevant knowledge,this thesis gradually investigates different levels of discourse clues,and explores how to effectively utilize these clues for improving implicit discourse relation recognition.The main contributions of our work are summarized as follows:(1)How to appropriately represent two discourse arguments is a key issue in implicit discourse relation recognition.Considering the different semantic information contained in different granularities of discourse,we propose a novel Dynamic Chunk-based Max Pooling Bi LSTM-CNN model(DC-BCNN)for identifying implicit discourse relation,which could obtain richer argument representations.Macroscopically,we encode the global contextual information of discourse,while capturing the local information in different ranges of discourse.Microscopically,according to the length of each argument and the structure of network,we dynamically divide the argument and capture crucial information from a wider range of n-grams by max pooling operation,which could retain the sequence cues.Thus,we obtain multi-granularity shallow linguistic features to improve the performance of our task.The experiments on English and Chinese corpora show the effectiveness of our proposed model.(2)How to effectively extract the interactive information is the core factor of this task.Because the shallow linguistic clues obtained in research(1)are independent,we present a novel Mutual Attentive Neural Tensor Network framework(MATS)to further explore the interactions between two discourse arguments,which is based on contextual interaction perception and pattern filtering.Inspired by the human-like reading strategy,we model the asymmetric connections by constructing a static interactive attention matrix,and utilize the neural tensor network with sparse constraint to mine the indicative discourse relation patterns.Due to the limitations of the static interactive attention,we design a dynamic mutual attention mechanism to fully obtain the asymmetric and bidirectional connections between the arguments,which could improve the ability of identifying implicit discourse relation.Finally,the experimental results on the PDTB corpus also demonstrate that our MATS model improves the performance of recognizing implicit discourse relation significantly.(3)Implicit discourse relations are rooted in the semantic understanding of discourse,which could not exist without relevant external knowledge according to the association theory in cognitive psychology.Therefore,we imitate the working memory strategy of the brain in the process of speech acquisition or discourse comprehension,and propose a working memory-driven neural network framework.This model devises a novel knowledge enhancement paradigm by the implicit and explicit fusion forms to model the instant and long-term memories,which promotes the deeply semantic representation and the deep understanding of the semantic connection between two arguments.It breaks the inherent model of existing knowledge enhancement.Thus,they can integrate the deeper knowledge clues associated with the discourse,which could promote the semantic representation of the arguments and the in-depth understanding of implicit discourse relation.External experiments show that compared with other models,our proposed model has improved the performance of implicit discourse relation classification,while enriching the discourse argument representations.In summary,this thesis first proposes a hybrid neural network based on dynamic chunk-based pooling to model the global and local features of discourse,and obtains the multi-granularity shallow linguistic clues.On this basis,it presents a novel neural network model based on contextual interaction perception and pattern selection,which considers the interactions between two discourse arguments and the specific patterns clues.In additional,the discourse cannot exist independently of its associated external knowledge,we design a new knowledge enhancement paradigm,effectively integrating the deeper knowledge clues to enrich the argument representations and improve the ability of recognizing implicit discourse relation.These three models are homologous to the same task,and gradually explore the relevant clues of implicit discourse relation.They have a great reference value to the relevant researches in discourse analysis,which have both certain theoretical significance and application value.