Czech Question Answer Selection using Recurrent Neural Networks

Radoslav Sabol

Bachelor's thesis

Abstract:

Práca sa zaoberá automatizovaným výberom viet ktoré odpovedajú na otázky v českom jazyku. Modul je vyvíjaný pre systém Automatic Question Answering (AQA) pomocou databázy Simple Question Answering Dataset (SQAD). Riešenie je založené na rekurentnej neurónovej sieti, ktorá spracuje vstupnú otázku a vety zo vstupného dokumentu, a vráti z nich najpravdedpodobnejšiu odpoveď. Text práce predstavuje teoretický základ a detaily architektúry neurónovej siete, a zároveň ponúka podrobné vyhodnotenie a analýzu chýb. Vyhodnotenie poukazuje na nárast v presnosti modelu v porovnaní s jeho staršími iteráciami vďaka vylepšenej metóde trénovania, a zároveň poukazuje na spôsoby ako túto presnosť vylepšiť.

Abstract:

This thesis describes an answer selection module optimized for the Czech language. The module is being developed for open-domain question answering system Automatic Question Answering (AQA) using the Simple Question Answering Dataset (SQAD). The solution is based on recurrent neural networks for processing the input question and answer sentences from a document, returning the most probable answer. The following text presents the theoretical background and details of the model architecture and offers its detailed evaluation and error analysis. The evaluation shows a slight increase in accuracy when compared to the older iterations due to improved training techniques, and also reveals ways of improving the current setup.

Keywords

artificial intelligence, deep learning, experimental, implementation, machine learning, NLP

Language used: English

Date on which the thesis was submitted / produced: 26. 5. 2020

Thesis defence

• Date of defence: 24. 6. 2020
• Supervisor: doc. RNDr. Aleš Horák, Ph.D.
• Reader: doc. Mgr. Pavel Rychlý, Ph.D.

Masaryk University

Faculty of Informatics

Bachelor programme / field:
Applied Informatics / Applied Informatics