Research On Multilingual Semantic Textual Similarity Computation And Applications

Semantic textual similarity(STS)assesses the degree of semantic equivalence between a pair of sentences.The study of STS is an essential basis towards natural language understanding.Semantic textual similarity has been widely used in many natural language processing(NLP)applications,for example,similar question retrieval in question answering(QA)and quality estimation in machine translation(MT).There exist two problems in semantic textual similarity;(1)previous methods adopted feature engineering with machine learning algorithms,which heavily rely on expert domain knowledge;(2)most studies focused on English,but research on other languages lag behind,especially in low-resource language.Low-resource language STS suffers from lack of annotated data and previous work employ off-the-shelf MT to translate them into resource-rich language like English.However,due to error propogation,the errors in MT will deteriorate the overall performance.To address the above issues,this thesis describes:1.Study of universal multilingual STS.To improve the performance of traditional NLP methods,we combine these methods with deep learning methods to build an ensemble model,and conduct experiments on multi-lingual datasets using MT.We won the first place in SemEval 2017 STS task and this work has been published in SemEval 2017 workshop.2.Study of low-resource language STS.To decouple the dependency of MT while still taking advantage the annotated data in resource-rich languages,we propose a deep multi-task learning model,which consists of two parameter sharing networks.One neural network used in the low-resource language is driven to learn from the other neural network in the resource-rich language.This work has been published in ECIR 2018 conference.3.Application of semantic matching in argument reasoning task.Semantic matching is the basis of semantic textual similarity semantic,and we apply it in the argument reasoning task.In order to bridge the gap between the given reasons,claims and the candidate warrants,we propose an attention-based neural network.We won the third place in SemEval 2018 Argument Reasoning Comprehension task and this work has been published in SemEval 2018 workshop.In order to examine the effectiveness of the above proposed models,we conducted a series of quantitative and qualitative experiments.The experimental results show that the universal multilingual STS model and the deep multi-task learning model can effectively measure the multilingual and low-resource language STS,respectively.Besides,semantic matching is effective in argument reasoning task.