Parallel Inference For Detecting Implicit Discourse Relation

Discourse relation analysis is a core issue in natural language processing. In the fieldof discourse analysis, Discourse refers to a series of consecutive clauses, sentences or par-agraphs which are constituted as a whole, and Discourse Relation is the semantic relation-ship (such as Contingency and Comparison) between the two different arguments occurredin the same discourse. In addition, Discourse Relation Analysis just means recognizing anddetecting the specific class of discourse relation. According to whether there is an explicitconnective (also known as the clue word, such as because, but) between the two relatedarguments, discourse relation can be divided into explicit relation and implicit relation. Atpresent, explicit relation has been recognized very well with the method of machine learn-ing and based on explicit connective itself. On the other hand, implicit relation is still una-ble to effectively detect because the complicated syntax and semantic information betweenimplicit arguments are difficult to extract and formalize.This paper presents a method based on parallel inference for implicit discourse rela-tion detection. For an adjacent implicit argument pair, we assume that if there exists an ex-plicit argument pair which is parallel with the implicit one, then the relations between im-plicit and explicit argument pairs are similar. With the help of information retrieval andlarge scale of data, we can mine and extract the parallel arguments and explicit clues, inorder to establish an unsupervised inference system for predicting the implicit relation. Themain research content includes the following three parts:1) Retrieval-driven for parallel argument inferenceWith large-scale web information, the parallel argument pairs can be automaticallymined. On one hand, we construct high-quality queries for extracting explicit parallel ar-gument pairs which are similar to implicit argument pair on syntactic and semantic levels.On the other hand, three inference models (Similarity, Confidence, Relevance) are pre-sented to evaluate the quality of queries and parallel argument pairs. Based on the TopN high-quality parallel argument pairs and the explicit clues, we realize the Explic-it-to-Implicit relation mapping among parallel argument pairs and the inference of implicitdiscourse relation.2) Relation disambiguation based on explicit connectivesIn the process of detecting implicit discourse relation based on parallel inference, thereare two factors affect the final detection performance such as ambiguous connectives andexisting pseudo cues. To resolve those two issues and improve the performance respective-ly, we correspondingly proposed the methods of relation sense disambiguation based onlocal rigorous connectives and pseudo cues filtering based on recognizing latent real con-nectives. The former is to disambiguate the relation senses of explicit connectives in themapping process (e.g. discrimination between Contingency and Temporal senses of theconnective since); and the latter is to shield the inference from the misleading cuescaused by the unbalanced distribution of discourse relations (averagely42.45%Expansionrelations but only12.89%Temporal in PDTB).3) Relation optimization based on bilingual collaborationIn the process of mining parallel argument pairs, due to the limit of information ex-tracted from implicit argument pair and retrieval performance of search engine, the parallelargument pairs are lower quality, which are not suitable to infer the corresponding implicitrelation. Therefore, with the help of cross-language retrieval, we attempt to mine additionalparallel argument pairs in the other language, in order to expand the scale of parallel re-sources, further optimize the performance of retrieval-driven for parallel argument infer-ence.