A Multilingual Approach to Question Classification

A Multilingual Approach to Question Classification Aikaterini-Lida Kalouli, Katharina Kaiser, Annette Hautli-Janisz, Georg A. Kaiser, Miriam Butt Department of Linguistics, University of Konstanz firstname.lastname@uni-konstanz.de Abstract In this paper we present the Konstanz Resource of Questions (KRoQ), the first dependency-parsed, parallel multilingual corpus of information-seeking and non-information-seeking questions. In creating the corpus, we employ a linguistically motivated rule-based system that uses linguistic cues from one language to help classify and annotate questions across other languages. Our current corpus includes German, French, Spanish and Koine Greek. Based on the linguistically motivated heuristics we identify, a two-step scoring mechanism assigns intra- and inter-language scores to each question. Based on these scores, each question is classified as being either information seeking or non-information seeking. An evaluation shows that this mechanism correctly classifies questions in 79% of the cases. We release our corpus as a basis for further work in the area of question classification. It can be utilized as training and testing data for machine-learning algorithms, as corpus-data for theoretical linguistic questions or as a resource for further rule-based approaches to question identification. Keywords: (non-)information-seeking questions, parallel multilingual corpora, question classification, Bible 1. Introduction A central phenomenon in natural language as well as human-computer interaction is that of questions. Although this is a central phenomenon, it has been understudied in computational linguistics. Most of the existing work has concentrated on dealing with ‘factoid’ questions such as When was Alan Turing at Bletchley Park? This research has mostly been driven by the goal of building Question- Answering (QA) systems and finding intelligent, quick and reliable ways of matching a query to terms found in a given text collection, e.g., see Wang and Chua (2010). Com- paratively less research has focused on understanding the structure of questions per se or on distinguishing differ- ent types of questions, i.e. information-seeking vs. rhetori- cal or discourse-structuring questions, among several other types. While a few approaches explicitly focus on non- information-seeking questions (Harper et al., 2009; Paul et al., 2011; Li et al., 2011), this work is either based on big data or on information gained from crowdsourcing. It does not tend to take recent theoretical linguistic work on questions into account, as has also been observed within the CReST project (K¨ubler et al., 2012), in which the existing PennTreebank (Marcus et al., 1993) annotation scheme was amended for yes-no and back-channeling questions. This paper makes use of theoretical linguistic insights for automatically classifying questions as information-seeking or non-information-seeking. We devise a rule-based sys- tem with a heuristic scoring methodology that uses linguis- tic indicators to classify questions into information-seeking (ISQ) or non-information seeking questions (NISQs) across four different languages: German, French, Spanish and Koine Greek.1 One result of our work has been the creation of a new resource: KRoQ (Konstanz Resource of Ques- tions), a first dependency-parsed, parallel multilingual cor- pus of ISQs vs. NISQs. 1Dialect of Greek, also known as Alexandrian dialect, com- mon Attic, Hellenistic or Biblical Greek; spoken and written dur- ing the Hellenistic and …

Content truncated to 500 words