Inspire at SemEval-2016 Task 2: Interpretable Semantic Textual Similarity Alignment based on Answer Set Programming

Proceedings of SemEval-2016, pages 1109–1115, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics Inspire at SemEval-2016 Task 2: Interpretable Semantic Textual Similarity Alignment based on Answer Set Programming Mishal Kazmi Sabanci University Istanbul, Turkey mishalkazmi@sabanciuniv.edu Peter Schüller Marmara University Istanbul, Turkey peter.schuller@marmara.edu.tr Abstract In this paper we present our system de- veloped for the SemEval 2016 Task 2 - Interpretable Semantic Textual Similarity along with the results obtained for our sub- mitted runs. Our system participated in the subtasks predicting chunk similarity alignments for gold chunks as well as for predicted chunks. The Inspire system ex- tends the basic ideas from last years par- ticipant NeRoSim, however we realize the rules in logic programming and obtain the result with an Answer Set Solver. To pre- pare the input for the logic program, we use the PunktTokenizer, Word2Vec, and WordNet APIs of NLTK, and the POS- and NER-taggers from Stanford CoreNLP. For chunking we use a joint POS-tagger and dependency parser and based on that determine chunks with an Answer Set Pro- gram. Our system ranked third place over- all and first place in the Headlines gold chunk subtask. 1 Introduction Semantic Textual Similarity (STS), refers to the degree of semantic equivalence between a pair of texts. This helps in explaining how some texts are related or unrelated. In Interpretable STS (iSTS) systems, further explanation is provided as to why the two texts are related or unre- lated. Finding these detailed explanations helps in gathering a meaningful representation of their similarities. The competition at SemEval 2016 was run on three different datasets: Headlines, Images and Student-Answers. Each dataset included two files containing pairs of sentences and two files containing pairs of gold-chunked sentences. Ei- ther the gold chunks provided by the organizers or chunks obtained from the given texts would be used as input to the system. The expected outputs of the system are m:n chunk-chunk alignments, corresponding similarity scores be- tween 0 (unrelated) and 5(equivalent), and a label indicating the type of semantic relation. Possible semantic relations are EQUI (seman- tically equal), OPPO (opposite), SPE1/SPE2 (chunk 1/2 is more specific than chunk 2/1), SIMI (similar but none of the relations above), REL (related but none of the relations above), and NOALI (not aligned, e.g., punctuation). All relations shall be considered in the given context. For details see the companion pa- per (Agirre et al., 2016). Similarity alignment in the Inspire system is based on ideas of previous year’s NeRoSim (Ban- jade et al., 2015) entry, however we reimple- mented the system and realize the rule engine in Answer Set Programming (ASP) (Lifschitz, 2008; Brewka et al., 2011) which gives us flex- ibility for reordering rules or applying them in parallel. To account for differences in datasets we detect the type of input with a Naive Bayes classifier and use different parameter sets for each of the datasets. Chunking in the Inspire system is based on a joint POS-tagger and dependency parser (Bohnet et al., 2013) and an …

Content truncated to 500 words