Recurrent Networks in AQA Answer Selection

Publikace nespadá pod Filozofickou fakultu, ale pod Fakultu informatiky. Oficiální stránka publikace je na webu muni.cz.

Autoři

SABOL Radoslav MEDVEĎ Marek HORÁK Aleš

Druh Článek ve sborníku
Konference Proceedings of the Twelfth Workshop on Recent Advances in Slavonic Natural Languages Processing, RASLAN 2018
Fakulta / Pracoviště MU

Fakulta informatiky

Citace
WWW https://nlp.fi.muni.cz/raslan/2018/paper12-Sabol_Medved_Horak.pdf
Klíčová slova question answering; answer selection; QA dataset; SQAD; AQA
Popis Unlimited, or open domain, question answering system AQA is being developed and tested with the Simple Question Answering Data-base (SQAD) for the Czech language. AQA is optimized for work with morphologically rich languages and makes use of syntactic cues provided by the morphosyntactic analysis. In this paper, we introduce a new answer selection module being developed for the AQA system. The module is based on recurrent neural networks processing the question and answer sentences to derive the most probable answer sentence. We present the details of the module architecture and offer a detailed evaluation of various hyperparameter setups. The module is trained and tested with 8,500 question-answer pairs using the SQAD v2.1 benchmark dataset.
Související projekty: