Font Size: a A A

Implicit Discourse Relation Inference Based On Frame Semantics

Posted on:2016-06-22Degree:MasterType:Thesis
Country:ChinaCandidate:W R YanFull Text:PDF
GTID:2308330464953053Subject:Software engineering
Abstract/Summary:PDF Full Text Request
Discourse relation analysis is an important research direction of natural language processing. In the field of discourse analysis, discourse refers to natural language texts structured by arguments which are semantic coherence and structural cohesion. Thereinto, argument is an independent semantic span in the discourse. Discourse relation is a semantic relation(such as “Cause”, “Comparison”, etc.) between adjacent arguments or arguments within a certain extent(called “argument pair”) in the same discourse. Discourse relation analysis aims to automatically recognizing and detecting the semantic relation between two arguments.Discourse relations is divided into explicit and implicit discourse relations in Penn Discourse Tree Bank according to whether there is an explicit connective(called “clue word”, such as “because”, “however”, etc.) between two arguments. Explicit discourse relation can be recognized directly according the explicit connection. On the contrary, implicit discourse relation is still unable to be detected effectively due to the lack of explicit connective.In this paper, we propose an implicit discourse relation inference method based on frame semantics for difficult implicit discourse relation. It is based on parallel inference hypothesis that “parallel arguments have parallel discourse relations”(namely explicit and implicit argument pair are semantic similarity, and then their semantic relations are the same).We mine the “explicit argument pairs” parallel to “implicit argument pairs” from the large-scale static corpus according to information retrieval technology. We build implicit relation inference system based on unsupervised learning. The main contents of the research include the following aspects: 1) Implicit discourse relation inference based on frame semantics pairsExisting methods of implicit discourse relation inference did not analyze semantic information of argument pairs, but analyze the relevance feature of argument pairs. For this, we propose a method of implicit discourse relation inference based on frame semantics knowledge. It is to automatically recognize frame semantics of arguments through Frame Net and its related identification technology. On this basis, we detect the semantic relation of arguments by the probability distributions of relation between frame semantics pairs in large-scale text data to improve the final performance of implicit relation. 2) Implicit relation recognition based on frame semantic vectorUsing probability distributions of relation between frame semantics pairs only to reasoning implicit relation can’t express the whole semantic of argument pairs. We propose a method to detect implicit discourse relation based on frame semantic vector. The compression of description is realized by frame semantic vector that abstract argument as conceptual semantic description, and then mine the comparable argument pairs through semantic vector from the large-scale static data. 3) Implicit discourse relation optimization based on local semantic forestExisting methods haven’t take into account the context information of arguments. However, the impacts of context can’t be ignored for discourse relation, especially implicit relation. For this, we propose implicit relation optimization method based on local semantic forest. We build local semantic forest for arguments and the context. And then calculate the similarity of local semantic forest that combine semantic vector similarity and context similarity between testing and comparable argument pairs, to reasoning the implicit discourse relation type for testing argument pairs.
Keywords/Search Tags:Implicit discourse relation, parallel argument, frame semantics, semantic vector, semantic forest
PDF Full Text Request
Related items