What Context Features Can Transformer Language Models Use?

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 851–864 August 1–6, 2021. ©2021 Association for Computational Linguistics 851 What Context Features Can Transformer Language Models Use? Joe O’Connor Jacob Andreas Massachusetts Institute of Technology {joeoc,jda}@mit.edu Abstract Transformer-based language models benefit from conditioning on contexts of hundreds to thousands of previous tokens. What aspects of these contexts contribute to accurate model prediction? We describe a series of experi- ments that measure usable information by se- lectively ablating lexical and structural infor- mation in transformer language models trained on English Wikipedia. In both mid- and long- range contexts, we find that several extremely destructive context manipulations—including shuffling word order within sentences and deleting all words other than nouns—remove less than 15% of the usable information. Our results suggest that long contexts, but not their detailed syntactic and propositional content, are important for the low perplexity of current transformer language models.1 1 Introduction Recent years have seen a significant improvement in the predictive accuracy of neural language mod- els (LMs), owing to a combination of improve- ments in model architecture (especially transform- ers; Vaswani et al. 2017) and training infrastructure (Wolf et al., 2020). The most striking change, rela- tive to both recurrent neural LMs (Mikolov et al., 2010) and count-based models (Kneser and Ney, 1995), is the length of the context that these models can effectively condition on. While count-based LMs in production speech recognition and machine translation systems typically used 10–20 tokens at a maximum (e.g., Brown, 2011), and recurrent LMs have an effective context size of 200 (Khandelwal et al., 2018), the predictive accuracy of transformer LMs appears to improve when conditioning on as many as a thousand previous tokens (Beltagy et al., 2020). A significant amount of recent work has 1Code for all experiments in this paper is available at https://github.com/lingo-mit/context-ablations. focused on making use of even longer contexts computationally feasible (Rae et al., 2019; Wang et al., 2020; Child et al., 2019; Dai et al., 2019; Kitaev et al., 2020). But despite empirical evidence that long contexts are helpful, little is understood about why. If the future of language modeling will include a focus on contexts of increasing size, it is important to first understand what contextual information con- tributes to accurate prediction in current models. This paper offers an answer to that question via the V-information framework of Xu et al. (2020). V- information, discussed more in Section 2, provides a formal framework for reasoning about how much usable information a computationally constrained predictor (like a neural LM) can extract from an input. Our experiments measure the amount of us- able information that is added when increasing LM context size, then attempt to pinpoint the source of this information by ablating features of the added context (via controlled shuffling and word deletion) and measuring the resulting loss of model predic- tive power. While this framework is general, we focus on transformer LMs. …

Content truncated to 500 words