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

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 interpretability and an understanding of how these architectures function, making principled development of better language models more difficult. We look inside a state-of-the-art neural language model to analyse how this model represents high-level lexico-semantic information. In particular, we investigate how the model represents words by extracting activation patterns where they occur in the text, and compare these representations directly to human semantic knowledge.

Why is penguin more similar to polar bear than to sea gull? Analyzing conceptual knowledge in distributional models

“…Several studies investigate the relation between semantic features recorded in feature norm datasets (McRae et al, 2005;Devereux et al, 2014;Vinson and Vigliocco, 2008;Vigliocco et al, 2004) and embedding vectors (Fagarasan et al, 2015;Tsvetkov et al, 2015Tsvetkov et al, , 2016Herbelot and Vecchi, 2015;Herbelot, 2013;Riordan and Jones, 2011;Glenberg and Robertson, 2000;Derby et al, 2018;Forbes et al, 2019;Rubinstein et al, 2015). These studies indicate that (at least partial) mappings between distributional and conceptual spaces are possible and that conceptual knowledge can complement distributional representations.…”

“…For instance, in the CSLB norms (Devereux et al, 2014), has legs is listed for several birds, but not for owl, duck, and eagle. This introduces noise to already rather small datasets used to investigate property knowledge in distributional data (Derby et al, 2018). Yaghoobzadeh et al (2019) apply diagnostic classification to investigate semantic classes using a large, automatically generated dataset derived from Wikipedia, which is likely to contain noise.…”