Analysing Neural Language Models: Contextual Decomposition Reveals Default Reasoning in Number and Gender Assignment

TL;DR

This paper introduces a generalized method called GCD to analyze neural language models, revealing that they rely heavily on default reasoning rather than purely syntactic or semantic cues for tasks like agreement and co-reference.

Contribution

The paper presents GCD, a novel technique to dissect model decisions, uncovering the reliance on default reasoning in neural language models.

Findings

Models depend on default reasoning for syntactic agreement.

GCD effectively distinguishes semantic, syntactic, and bias influences.

Neural models show strong reliance on heuristics rather than true understanding.

Abstract

Extensive research has recently shown that recurrent neural language models are able to process a wide range of grammatical phenomena. How these models are able to perform these remarkable feats so well, however, is still an open question. To gain more insight into what information LSTMs base their decisions on, we propose a generalisation of Contextual Decomposition (GCD). In particular, this setup enables us to accurately distil which part of a prediction stems from semantic heuristics, which part truly emanates from syntactic cues and which part arise from the model biases themselves instead. We investigate this technique on tasks pertaining to syntactic agreement and co-reference resolution and discover that the model strongly relies on a default reasoning effect to perform these tasks.