Using Argument Concept And Focus For Detecting Implicit Discourse Relation

Discourse relation analysis is a natural language processing technology aboutautomatic identification and determination of semantic relationship between two arguments.In the field of NLP, Discourse refers to natural language texts which are constituted byarguments of semantic relevance and structured organization, and Discourse Relation is thesemantic correlation properties between adjacentarguments or arguments of spans within acertain range (such as “Comparison”) in the same discourse. Among them, the Argument isindependent semantic snippets in the discourse, which is the basic performance of unitform discourse relationship. One of the core tasks of Discourse Relation Analysisisrecognizing and detecting the specific class of discourse relation. That is to explainsemantic relations between the arguments. According to whether there is an explicitconnective (such as “so”) between the two related arguments, discourse relations can bedivided into two categories: explicit relation and implicit relation. At present, explicitrelation has been recognized very well because of explicit connective. On the other hand,implicit relation is still unable to be effectively detected, because the complicated syntaxand semantic information between implicit arguments are difficult to extract.This paper presents a method using argument concept and focus for detecting implicitdiscourse relation. Based on the hypothesis of semantic parallel argument has parallelism,we mine explicit parallel reference arguments set, in the aspect of sentence structure andcontent with high semantic relevance of implicit arguments with the help of connectingclue words, including functional connective. Using concept and co-referential focus cluesbetween parallel arguments, formed parallel inference mechanism of “explicit guideimplicit”, we build implicit relation inference system based on unsupervised learning, themain research content includes the following three aspects: Implicit relation detection based on conceptual model of parallel argumentsBecause of explicit connectives are difficult to provide enough clues to miningparallel reference arguments, we use functional connectives which can express internalsemantic relations of arguments and development trend of language environment andexplicit connectives as connecting clue words, contribute to automatically mine the parallelreference arguments. Specifically, we build parallel arguments concept model which candescribe argument property in the light of each cule word, and establish the mappingsystem of “argument concept-discourse relation”.Then statistical strategy is used toidentify the conceptual model of the arguments and realize the inference.Argument focus recognition oriented relation disambiguation modelThere usually exit many relation ambiguities between implicit arguments with relationambiguities. Relation ambiguities mainly include cule words ambiguity (polysemy) andmultiple implicit clues ambiguity (diverse clues metaphor different relations). Cluesambiguity appears more frequent than connectives and it is difficult to identify.To this end,we develop a method of topic-driven focus identification to determine words as argumentfocus when words adhere to topic, and it’s used for relation disambiguation.Parallel inference optimization based on co-referential focusBased on the identification of focus, we propose a focus detection model that is usedto determine the focus consistency. We first establish focus probability curve of parallelarguments, then measure the matching deviation and obtain the probability of consistency.Finally, we optimize parallel inference system based on co-referential focus to improvethe detection performance of implicit discourse relation.