Representation of Word Meaning in the Intermediate Projection Layer of a Neural Language Model

Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 362–364 Brussels, Belgium, November 1, 2018. c⃝2018 Association for Computational Linguistics 362 Representation of Word Meaning in the Intermediate Projection Layer of a Neural Language Model Steven Derby1 Paul Miller1 Brian Murphy1,2 Barry Devereux1 1 Queen’s University Belfast, Belfast, United Kingdom 2 BrainWaveBank Ltd., Belfast, United Kingdom {sderby02, p.miller, brian.murphy, b.devereux}@qub.ac.uk Abstract Performance in language modelling has been significantly improved by training recurrent neural networks on large corpora. This progress has come at the cost of interpretabil- ity and an understanding of how these archi- tectures function, making principled develop- ment of better language models more difficult. We look inside a state-of-the-art neural lan- guage model to analyse how this model repre- sents high-level lexico-semantic information. In particular, we investigate how the model represents words by extracting activation pat- terns where they occur in the text, and com- pare these representations directly to human semantic knowledge. 1 Introduction & Related Work Language modelling involves learning to predict the next word in a sequence of words, using large text corpora as the training input. Language mod- els must therefore learn to represent information from the preceding context which is relevant for future word prediction, and, intuitively, this should include information about the syntactic structure of the context and the meanings of constituent words. Today’s state-of-the-art language models make use of Recurrent Neural Networks (RNNs) with Long Short-Term Memory cells (LSTMs) (Hochreiter and Schmidhuber, 1997) which can handle time series information by remembering salient information over latent variables (Mikolov et al., 2010). Because of their wide applicability, there has been much interest in developing a bet- ter understanding of the inner workings of RNN models, and, in particular, researchers have inves- tigated how syntactic knowledge is encoded and processed by such networks (Dyer et al., 2016; Linzen et al., 2016; Jozefowicz et al., 2016; Mc- Coy et al., 2018; Gulordava et al., 2018). Karpa- thy et al. (2015) performed an in-depth analysis of the types of errors RNN’s make, in order to un- derstand how recurrent mechanisms can encode long-term dependency information. Linzen et al. (2016) present a more direct analysis by examin- ing LSTM language models’ ability to understand difficult long-range dependencies such as the form of a verb linked to a noun subject. Recently, re- searchers have started to study the semantic em- beddings generated by these networks (Chrupała et al., 2015), especially for those focused on en- coding visual grounding (Kiela et al., 2017; Yoo et al., 2017). However, compared to syntax, there has been relatively less work on how LSTM net- works represent lexical semantic knowledge. In this work, we evaluate latent semantic knowl- edge present in the LSTM activation patterns pro- duced before and after the word of interest. We evaluate whether these activations predict human similarity ratings, human-derived property knowl- edge, and brain imaging data. In this way, we test the model’s ability to encode important semantic information relevant to word …

Content truncated to 500 words