Correlating Neural and Symbolic Representations of Language

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2952–2962 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 2952 Correlating neural and symbolic representations of language Grzegorz Chrupała Tilburg University g.chrupala@uvt.nl Afra Alishahi Tilburg University a.alishahi@uvt.nl Abstract Analysis methods which enable us to better understand the representations and functioning of neural models of language are increasingly needed as deep learning becomes the domi- nant approach in NLP. Here we present two methods based on Representational Similarity Analysis (RSA) and Tree Kernels (TK) which allow us to directly quantify how strongly the information encoded in neural activation pat- terns corresponds to information represented by symbolic structures such as syntax trees. We first validate our methods on the case of a simple synthetic language for arithmetic ex- pressions with clearly defined syntax and se- mantics, and show that they exhibit the ex- pected pattern of results. We then apply our methods to correlate neural representations of English sentences with their constituency parse trees. 1 Introduction Analysis methods which allow us to better under- stand the representations and functioning of neu- ral models of language are increasingly needed as deep learning becomes the dominant approach to natural language processing. A popular technique for analyzing neural representations involves pre- dicting information of interest from the activation patterns, typically using a simple predictive model such as a linear classifier or regressor. If the model is able to predict this information with high accu- racy, the inference is that the neural representation encodes it. We refer to these as diagnostic models. One important limitation of this method of anal- ysis is that it is only easily applicable to rela- tively simple types of target information, which are amenable to be predicted via linear regression or classification. Should we wish to decode acti- vation patterns into a structured target such as a syntax tree, we would need to resort to complex structure prediction algorithms, running the risk that the analytic method becomes no simpler than the actual neural model. Here we introduce an alternative approach based on correlating neural representations of sen- tences and structured symbolic representations commonly used in linguistics. Crucially, the cor- relation is in similarity space rather than in the original representation space, removing most con- straints on the types of representations we can use. Our approach is an extension of the Representa- tional Similarity Analysis (RSA) method, initially introduced by Kriegeskorte et al. (2008) in the context of understanding neural activation patterns in human brains. In this work we propose to apply RSA to neural representations of strings from a language on one side, and to structured symbolic representations of these strings on the other side. To capture the sim- ilarities between these symbolic representations, we use a tree kernel, a metric to compute the pro- portion of common substructures between trees. This approach enables straightforward compari- son of neural and symbolic-linguistic representa- tions. Furthermore, we introduce RSAREGRESS, a similarity-based analytic method which combines …

Content truncated to 500 words