DFKI-LT - Answering Non-Factoid questions of Online Users using Text-based End-to-end Trainable models

Stalin Varanasi, GŁnter Neumann
Answering Non-Factoid questions of Online Users using Text-based End-to-end Trainable models
2 The Twenty-Fifth Text REtrieval Conference (TREC 2016) Notebook, Gaithersburg, Maryland, USA, Trec, TREC, 11/2016
 
Answering Non-Factoid questions automatially requires more anlaysis and design than answering factoid questions. While supervised methods using linguistic features or curated databases help to achieve decent results in domain specific questions, they are not easy to scale . Using end-to-end trainable methods on large corpora is more attractive as the process is not domain specific. This paper deals with end-to-end trainable models for answering non-factoid questions from Yahoo! Answers using deep learning methods. The questions from Yahoo!Answers1 pose an extra challenge as they sometimes contain a narrative by the online user (see Table 1). Our method propose a heuristic in summarizing these questions and further use two models to train the question and answer pairs. We manually evaluate our results for hundred questions. We used our system to participate in the TREC LiveQA 2016 competition.
 
Files: BibTeX