Neural Domain Adaptation for Biomedical Question Answering

Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 281–289, Vancouver, Canada, August 3 - August 4, 2017. c⃝2017 Association for Computational Linguistics Neural Domain Adaptation for Biomedical Question Answering Georg Wiese1,2, Dirk Weissenborn2 and Mariana Neves1 1 Hasso Plattner Institute, August Bebel Strasse 88, Potsdam 14482 Germany 2 Language Technology Lab, DFKI, Alt-Moabit 91c, Berlin, Germany georg.wiese@student.hpi.de, dewe01@dfki.de, mariana.neves@hpi.de Abstract Factoid question answering (QA) has re- cently benefited from the development of deep learning (DL) systems. Neural net- work models outperform traditional ap- proaches in domains where large datasets exist, such as SQuAD (≈ 100, 000 ques- tions) for Wikipedia articles. However, these systems have not yet been applied to QA in more specific domains, such as biomedicine, because datasets are gen- erally too small to train a DL system from scratch. For example, the BioASQ dataset for biomedical QA comprises less then 900 factoid (single answer) and list (multiple answers) QA instances. In this work, we adapt a neural QA system trained on a large open-domain dataset (SQuAD, source) to a biomedical dataset (BioASQ, target) by employing various transfer learning techniques. Our network architecture is based on a state-of-the- art QA system, extended with biomedical word embeddings and a novel mechanism to answer list questions. In contrast to ex- isting biomedical QA systems, our system does not rely on domain-specific ontolo- gies, parsers or entity taggers, which are expensive to create. Despite this fact, our systems achieve state-of-the-art results on factoid questions and competitive results on list questions. 1 Introduction Question answering (QA) is the task of retriev- ing answers to a question given one or more con- texts. It has been explored both in the open- domain setting (Voorhees et al., 1999) as well as domain-specific settings, such as BioASQ for the biomedical domain (Tsatsaronis et al., 2015). The BioASQ challenge provides ≈ 900 factoid and list questions, i.e., questions with one and several an- swers, respectively. This work focuses on answer- ing these questions, for example: Which drugs are included in the FEC-75 regimen? → fluorouracil, epirubicin, and cyclophosphamide. We further restrict our focus to extractive QA, i.e., QA instances where the correct answers can be represented as spans in the contexts. Contexts are relevant documents which are provided by an information retrieval (IR) system. Traditionally, a QA pipeline consists of named- entity recognition, question classification, and an- swer processing steps (Jurafsky, 2000). These methods have been applied to biomedical datasets, with moderate success (Zi et al., 2016). The cre- ation of large-scale, open-domain datasets such as SQuAD (Rajpurkar et al., 2016) have recently en- abled the development of neural QA systems, e.g., Wang and Jiang (2016), Xiong et al. (2016), Seo et al. (2016), Weissenborn et al. (2017), leading to impressive performance gains over more tradi- tional systems. However, creating large-scale QA datasets for more specific domains, such as the biomedical, would be very expensive because of the need for domain experts, and therefore not desirable. The recent success of deep …

Content truncated to 500 words