Investigating Questioner’s Explicit Information Influences in Transformer-based Community Question Answering

Web Q&A forums provide a collaborative environment where users can post subjective questions and answers based on their life experiences or knowledge of some specific domains or topics. The questions and answers relations help Learning-to-Rank models for Community Question Answering (CQA) tasks to sort user comments to answer questions by relevance. These learning models automatically rank comments as potential answer candidates for new questions. However, finding the proper models for CQA models also depends on the information used to match question and answer pairs. This paper aims to summarise a study about Community Question-Answering (CQA) models in two specific domains: Personal Finance &Money (PF&M) and Home Improvements (HI). We built CQA datasets, where all the data are scraped from user posts in a web community. This dataset provides further explicit user information to build and train our CQA baselines, such as question descriptions and multiple questioner tags to expand the information in short user questions. We offer different evaluation scenarios to assess the baseline performance associated with distinct text inputs by considering two measures: MAP and MRR. CQA models propose answer candidates ranked list based on the QA pairs relevance. We also offer an automatic tag extractor to obtain tags from question descriptions to evaluate the importance of user tags for CQA model performance. In web forums like StackExchange, questioners use multiple tags as keywords to refer to the question context (e.g., investment, tax, health insurance and so on). Our experiments show that incorporating multiple tags and question descriptions to expand user question information improves the performance of CQA for different transformer-based baselines. Our results show that our tag extraction method, which condenses question descriptions into tags, delivered a compelling outcome when we compared the effectiveness of user- and machine-generated tags as model inputs in CQA baselines.