Identifying attack and support argumentative relations using deep learning

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1374–1379 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Identifying attack and support argumentative relations using deep learning Oana Cocarascu and Francesca Toni Department of Computing Imperial College London Abstract We propose a deep learning architecture to capture argumentative relations of attack and support from one piece of text to an- other, of the kind that naturally occur in a debate. The architecture uses two (uni- directional or bidirectional) Long Short- Term Memory networks and (trained or non-trained) word embeddings, and al- lows to considerably improve upon exist- ing techniques that use syntactic features and supervised classifiers for the same form of (relation-based) argument mining. 1 Introduction Argument Mining (AM) is a relatively new re- search area which involves, amongst others, the automatic detection in text of arguments, ar- gument components, and relations between ar- guments (see (Lippi and Torroni, 2016) for an overview). We focus on a specific type of AM, referred to as Relation-based AM (Carstens and Toni, 2015), which has recently received atten- tion by several researchers (e.g. see (Bosc et al., 2016; Carstens and Toni, 2017)). This type of AM aims at identifying argumentative relations of at- tack and support between natural language argu- ments in text, by classifying pairs of pieces of text as belonging to attack, support or neither attack nor support relations. For example, consider the three texts taken from Carstens and Toni (2015): t1: ‘We should grant politicians immunity from prosecution’ t2: ‘Giving politicians immunity allows them to focus on performing their duties’ t3: ‘The ability to prosecute politicians is the ul- timate protection against abuse of power’ Here t2 supports t1, t3 attacks t1, and t2 and t3 neither attack nor support one another. Relation-based AM is useful, for example, to pave the way towards identifying accepted opin- ions (Bosc et al., 2016) or divisive issues (Konat et al., 2016) within debates. We propose a deep learning architecture for Relation-based AM based on Long-Short Term Memory (LSTM) networks (Hochreiter and Schmidhuber, 1997; Schuster and Paliwal, 1997). Within the architecture, each input text is fed, as a (trained or non-trained) 100-dimensional GloVe embedding (Pennington et al., 2014), into a (unidi- rectional or bidirectional) LSTM which produces a vector representation of the text independently of the other text being analysed. The two vectors are then merged (using element-wise sum or concate- nation) and the resulting vector is fed to a softmax classifier which predicts whether the pair of input texts belongs to the attack, support or neither re- lations. The input texts may be at most 50 words long, but are not restricted to single sentences. We experimented with several instances of the architecture and achieved 89.53% accuracy and 89.07% F1 using unidirectional LSTMs and con- catenation as the merge layer, considerably outper- forming feature-based supervised classifiers used in the studies which presented the corpus we also use (Carstens and Toni, 2015, 2017). The remainder of the paper …

Content truncated to 500 words