Chinese Dependency Parsing Of Social Media Text Based On Deep Neural Transfer Learning

Dependency parsing has been a hot topic in the natural language processing(NLP)community,and plays a vital role in practical applications such as sentiment analysis,information extraction and retrieval.In recent years,the studies of dependency parsing have made great progress,and the core methods and techniques are relatively mature.However,currently the excellent performance of dependency parsing is more restrained to a single specific domain(such as news),where the labeled data is mostly from general normative text.As new-media technology develops rapidly,social media platforms have reserved a large number of users’ spoken and informal text data.For these domains,not only are labeled corpora rare,but due to the gap of data distribution,the performance of general dependency parser would be significantly dropped.Considering the existing problems and challenges,domain adaptation technique is a feasible solution,the core idea is to migrate source-domain knowledge to a specific target domain,making the dependency parser learned from the source domain can be effectively applied to the target domain,even if the target domain has no manually-annotated data.Since dependency parsing is a word-level task and Chinese text is character sequence in nature,it is necessary to perform word segmentation before parsing.But this pipeline way may cause that word segmentation errors lead to the wrong parsing results.Thus,this paper firstly studies the character-level dependency parsing model,and proposes an end-to-end multi-task learning framework for joint word segmentation,part-of-speech tagging and dependency parsing,alleviating the potential error propagation problem.Furthermore,this work explores domain adaption for Chinese dependency parsing to improve the parsing performance of dependency model in social media texts much.In practice,it is hard to manually annotate dependency trees on texts and the annotation cost is high,thus this paper focuses on studying unsupervised cross-domain dependency parsing techniques.From model-centric and data-centric perspectives,we mainly investigate three essential domain adaptation methods.Extensive experiments validate the effectiveness of our methods,meanwhile improving the domain transfer ability of dependency parser without available labeled target-domain training examples.