Non-factoid Question Answering Sample Clauses

Non-factoid Question Answering. ‌ Non-factoid question answering is an umbrella term for almost all questions that cannot be classified as factoid. Several tasks have emerged in this area: community question answering [98], real-time question answering [81], vi- sual question answering [2], math word and logic questions (arithmetic) [76], event-based question answering [154], and more. This dissertation presents two approaches that are evaluated in two areas of non-factoid question an- swering: arithmetic questions and event-based questions.

Non-factoid Question Answering. During the earlier days of QA research, non-factoid questions received rela- tively little attention. The TREC QA tasks started to incorporate certain categories of non-factoid questions, such as definition questions, during the last 4 years of the challenge. One of the first non-factoid question answering systems was described by X. Xxxxxxx and X. Xxxxx [180] and was based on web search using chunks extracted from the original question. The ranking of extracted answer candidates was done using a translation model, which showed better results than the n-gram based match score. The growth of the popularity of community question answering (CQA) websites, such as Yahoo! Answers, Xxxxxxx.xxx, etc., contributed to an increased interest towards non-factoid QA. Some questions on CQA websites are repeated very often and answers can easily be reused to answer new questions, X.Xxx et al. [131] studied different types of CQA questions and answers and analyzes them with respect to answer re-usability. A number of methods for similar question retrieval have been proposed [26, 63, 105, 174]. The candidate answer passages ranking problem becomes even more diffi- cult in non-factoid questions answering as systems have to deal with a larger piece of text and need to “understand” what kind of information is expressed there. WebAP is a dataset for non-factoid answer sentence retrieval, which was developed in [232]. Experiments conducted in this work demonstrated, that classical retrieval methods do not work well for this task, and multiple additional semantic (ESA, entity links) and context (adjacent text) features have been proposed to improve the retrieval quality. One of the first ex- tensive studies of different features for non-factoid answer ranking can be found in Xxxxx Xxxxxxxx et al. [187], who explored information retrieval scores, translation models, tree kernel and other features using tokens and semantic annotations (dependency tree, semantic role labeling, etc.) of text paragraphs. Alignment between question and answer terms can serve as a good indicator of their semantic similarity. Such an alignment can be pro- duced using a machine learning model with a set of features, representing the quality of the match [217]. Alignment and translation models are usually based on term-term similarities, which are often computed from a monolin- gual alignment corpus. This data can be very sparse, and to overcome this issue [75] proposed higher-order lexical semantic model...