UniMelb at SemEval-2018 Task 12: Generative Implication using LSTMs, Siamese Networks and Semantic Representations with Synonym Fuzzing

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 1124–1128 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics UniMelb at SemEval-2018 Task 12: Generative Implication using LSTMs, Siamese Networks and Semantic Representations with Synonym Fuzzing Anirudh Joshi1,2 Timothy Baldwin1 Richard O. Sinnott1 Cecile Paris2 1 The University of Melbourne 2 CSIRO Data61 anirudhj@student.unimelb.edu.au, tb@ldwin.net rsinnott@unimelb.edu.au, cecile.paris@data61.csiro.au Abstract This paper describes a warrant classification system for SemEval 2018 Task 12, that at- tempts to learn semantic representations of reasons, claims and warrants. The system consists of 3 stacked LSTMs: one for the reason, one for the claim, and one shared Siamese Network for the 2 candidate warrants. Our main contribution is to force the embed- dings into a shared feature space using vec- tor operations, semantic similarity classifica- tion, Siamese networks, and multi-task learn- ing. In doing so, we learn a form of genera- tive implication, in encoding implication inter- relationships between reasons, claims, and the associated correct and incorrect warrants. We augment the limited data in the task further by utilizing WordNet synonym “fuzzing”. When applied to SemEval 2018 Task 12, our system performs well on the development data, and officially ranked 8th among 21 teams. 1 Introduction This paper describes our system for the Argument Reasoning Comprehension Task of SemEval 2018 (Habernal et al., 2018). The main goal of our sys- tem is to learn semantic representations of reasons, claims and warrants upon which vector operations can be applied which encode their interrelation- ships, whilst sharing encodings. We train our system over the 1274 candidate reason, claim, first and second warrant examples from the edited Room for Debate New York Times dataset provided by the task organizers (Haber- nal et al., 2018), which we augment to generate a dataset that is used with a multi-task based ob- jective function to learn semantically meaningful representations. Our system combines these vectors to make a determination, via vector operations and semantic similarity classification, to determine which of the two warrants best fits the reason and claim repre- sentations, and best encodes the relationships be- tween them. We replicate this for each sentence type, and mirror it for each warrant. In doing so we were able to achieve strong results on the de- velopment set, and ranked 8th overall in the com- petition. 2 Approach 2.1 Model Overview As detailed in Figure 1, our system is made up of 3 stacked LSTMs, with one being essentially a Siamese network shared between the two war- rants, and the other two encoding the reason and the claim. We take an average pool over the top- layer hidden representations and use them as the semantic feature vectors for each respective sen- tence. We take these semantic vectors and ap- ply vector operations to them to generate embed- dings for each reason, claim and warrant, which we found in practice to perform well. We then do semantic comparisons between the generated and actual encodings, marking as the same those be- tween generated …

Content truncated to 500 words