Automatic Response Generation For Microblog Dialogue

In natural language processing,a human-machine dialogue system that exploits text,speech,graphics and so on,concentrates on how to response to a real human's oral utterance.Major of traditional dialogue systems are based on specified domains and have confirmed architecture and implementation theory.Though part of them have been commercially developed,they all need artificial and multiple repository collections and dialogue rule designations,which is time-consuming,inflexible and with strong domain limitations.Rest of current dialogue systems are domain-opened.Their approaches are various,have no standard implementations and their results are not well.Taken another route,our research concentrates on how to response to a microblog rationally,uses statistical approach to try to capture conversational characteristics,and aims at accomplishing a domain-opened human-machine dialogue system.This thesis mainly proposes an automatic response generation for twitter dialogue based on statistical machine translation(SMT)which learns English conversational characteristics from twitter status-response parallel corpus.We stimulate the response generation process as a machine translation process from status to response.We divide the process into two states.First state is corpus refinement which utilize alignment model to remove misaligned components in a status-response's sentences and remain components have been rationally aligned.In second state,with SMT technologies,a translation model is trained in the refined parallel corpus;given a new status,a response can be generated by the translation model.Moreover,in order to generalize the model,we tag the original status-response parallel corpus with Named Entity taggers and then run the above two states.As comparisons,we implement two others domain-opened dialogue systems: vector space searching and automatic knowledge learning with zero prior knowledge,whose corpora are from microblog forum and movie scripts.We use human evaluation to evaluate the response output's fluency and rationality and incidentally give the translation model's BLEU score.