Utilizing Bert for Question Retrieval on Vietnameses E-commerce Sites

Utilizing BERT for Question Retrieval in Vietnamese E-commerce Sites Thi-Thanh Ha HaNoi Uni. of Science and Technology, VietNam ThaiNguyen Uni. of Information and Communication Technology, VietNam htthanh@ictu.edu.vn Van-Nha Nguyen HaNoi Uni. of Science and Technology nha282@gmail.com Kiem-Hieu Nguyen HaNoi Uni. of Science and Technology hieunk@soict.hust.vn Kim-Anh Nguyen HaNoi Uni. of Science and Technology anhnk@soict.hust.vn Tien-Thanh Nguyen HaNoi Uni. of Science and Technology 20144052@student.hust.edu.vn Abstract Question retrieval is an important task in ques- tion answering. This task is considered to be challenging due to the lexical gap issue, i.e., similar questions could be expressed in dif- ferent words or phrases. Although there are numerous researches conducted on question retrieval task in English, the corresponding problem in Vietnamese hasn’t been studied much. In this investigation, we highlight our efforts on question retrieval in Vietnamese e-commerce sites majorly in two directions: (1) Building a Vietnamese dataset for ques- tion retrieval in e-commerce domain. (2) Con- ducting experiments using recent deep learn- ing techniques including BERT-based clas- sifiers. Our results provide practical exam- ples of effectively employing these models on Vietnamese e-commerce data. Particularly, we demonstrate that a BERT model trained on e-commerce texts yields significant improve- ment on question retrieval over BERT trained on general-domain texts. 1 Introduction Community-based Question Answering (CQA) sys- tems12 have become an increasingly popular on- line platform. Community websites, where users can post their own questions or answers to other users’ questions, provide frameworks for people with dis- similar backgrounds to share their knowledge and experiences. When a user posts a new question on a community website, it usually takes a while for 1https://stackoverflow.com/ 2https://www.qatarliving.com/ other users to respond. Moreover, in a certain pe- riod of time, the number of questions and answers stored in a database gradually becomes enormous and challenging to handle, which means that the pos- sibility of finding duplicated questions increases. As a result, it is time-consuming to retrieve good an- swers to a given question in an archive of question- answer pairs. In order to reduce latency, CQA sys- tems should automatically find questions which are similar to a given new question. It is hoped that the answers of these related questions could be useful for the new question. The problem of question retrieval is defined as follows: Given a query question and a set of exist- ing questions, return the most similar questions to the query. Question retrieval has been extensively in- vestigated with the purpose of answering new ques- tions using previous answers in databases [Zhou et al.2013, Zhou et al.2015]. Previous studies delved into the lexical gap challenge in which query ques- tion might contain words and phrases different from its similar questions. Figure 1 is a typical pair of sim- ilar questions in our Vietnamese dataset. In order to deal with lexical gap challenge, pre- vious research applied soft alignment technique originated from machine translation or implicitly disambiguated word meaning using topic models [Cai et al.2011]. A huge number of research meth- ods in recent years have focused on …

Content truncated to 500 words