Research On Relation Extraction For Dialogue Text

Information extraction is the basic task of natural language processing,and relation extraction is one of the key sub-topics of information extraction.Previous researches on relation extraction mainly focus on the relatively standardized texts such as news,encyclopedia and abstract of papers,while ignoring the common dialogue scenes in daily life.Compared with traditional relation extraction,dialogue data has the characteristics of low information density,high proportion of cross-sentence relation,speaker aware and rich personal pronoun,which brings new challenges to the research of conversation-oriented relation extraction.Based on this,a variety of relation extraction models oriented to dialogue text are proposed in this paper,which integrate relationship categories,speaker information,personal pronoun co-reference information and other features to assist the model to capture more conversation-level information.The research content of this paper mainly includes three aspects:(1)In view of the characteristics of low information density and high proportion of cross-sentence relation in dialogue data,we proposed a graph convolution relation extraction model based on relation-guided attention mechanism.First of all,the whole dialogue is input into the context encoder,and the interaction between different words and different relations is evaluated through the multi-head attention mechanism,and the context representation of the dialogue is updated.Then it is input into the graph convolutional network module for document level reasoning.Finally,the entity representation obtained is input into the bilinear layer,and obtain the predicted relation category.Through the attention mechanism of relationship guidance,the model makes the context representation integrate the category characteristics of the relationship,and pay attention to different words according to different relation types.The results show that the performance of the traditional document-level relation extraction model can be improved effectively by introducing the relation type feature into the model.(2)In view of the fact that the conversation data contains a lot of speaker information and personal pronouns,a relation extraction model based on the feature perception of speaker and co-reference information is proposed.Firstly,the model introduces the speaker's feature representation in the context coding stage,and makes each word in the context match the corresponding speaker by adding the speaker information in the word representation.Secondly,in the relation prediction stage,the model uses heuristic rules and reference resolution tools to obtain the co-referance information between personal pronouns and entity mentions,and integrates the co-referance information into the entity representation to improve the representational ability of the entity.The results show that the characteristics of the speaker and co-reference can improve the ability of the model to capture the characteristics of the interlocutor,and thus improve the relation extraction effect of the dialogue text.(3)In view of the fact that dialogue context contains many kinds of information,such as the speaker,co-reference and utterance,we propose a heterogeneous graph model with multiple types of nodes,which further integrates the speaker,co-reference and utterance information on the basis of feature perception.Specifically,the model introduces speaker node,co-reference information node and utterance node in the inference layer of heterogeneous graph to assist the dialogue-level reasoning of text,and constructs eight different types of edges according to different logical relations between nodes.The edges include entity-mention edge,mention-utterance edge,speaker-utterance edge,speaker-utterance edge,mention-meta dependency path node edge,utterance-meta dependency path node edge,utterance-utterance and self-loop edge.Finally,the heterogeneous graph is inferred through relational graph convolution to capture the characteristics of nodes.The experimental results show that the method of integrating speaker information,personal pronoun information and utterance information by using heterogeneous graph is further improved than the method of feature perception on the speaker and co-reference information in dialogue-level relation extraction.