The Paradox of the Compositionality of Natural Language: A Neural Machine Translation Case Study

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4154 - 4175 May 22-27, 2022 c⃝2022 Association for Computational Linguistics The Paradox of the Compositionality of Natural Language: A Neural Machine Translation Case Study Verna Dankers ILCC, University of Edinburgh vernadankers@gmail.com Elia Bruni University of Osnabr¨uck elia.bruni@gmail.com Dieuwke Hupkes Facebook AI Research dieuwkehupkes@fb.com Abstract Obtaining human-like performance in NLP is often argued to require compositional general- isation. Whether neural networks exhibit this ability is usually studied by training models on highly compositional synthetic data. How- ever, compositionality in natural language is much more complex than the rigid, arithmetic- like version such data adheres to, and artificial compositionality tests thus do not allow us to determine how neural models deal with more realistic forms of compositionality. In this work, we re-instantiate three compositionality tests from the literature and reformulate them for neural machine translation (NMT). Our re- sults highlight that: i) unfavourably, models trained on more data are more compositional; ii) models are sometimes less compositional than expected, but sometimes more, exempli- fying that different levels of compositionality are required, and models are not always able to modulate between them correctly; iii) some of the non-compositional behaviours are mis- takes, whereas others reflect the natural vari- ation in data. Apart from an empirical study, our work is a call to action: we should re- think the evaluation of compositionality in neu- ral networks and develop benchmarks using real data to evaluate compositionality on natu- ral language, where composing meaning is not as straightforward as doing the math.1 1 Introduction Although the successes of deep neural networks in natural language processing (NLP) are astounding and undeniable, they are still regularly criticised for lacking the powerful generalisation capacities that characterise human intelligence. A frequently men- tioned concept in such critiques is compositionality: the ability to build up the meaning of a complex expression by combining the meanings of its parts (e.g. Partee, 1984). Compositionality is assumed 1The data and code are available at https://github. com/i-machine-think/compositionality paradox mt. We present details concerning reproducibility in Appendix E. to play an essential role in how humans understand language, but whether neural networks also exhibit this property has since long been a topic of vivid debate (e.g. Fodor and Pylyshyn, 1988; Smolensky, 1990; Marcus, 2003; Nefdt, 2020). Studies about the compositional abilities of neu- ral networks consider almost exclusively models trained on synthetic datasets, in which composi- tionality can be ensured and isolated (e.g. Lake and Baroni, 2018; Hupkes et al., 2020).2 In such tests, the interpretation of expressions is computed com- pletely locally: every subpart is evaluated indepen- dently – without taking into account any external context – and the meaning of the whole expression is then formed by combining the meanings of its parts in a bottom-up fashion. This protocol matches the type of compositionality observed in arithmetic: the meaning of (3 + 5) is always 8, independent of the context it occurs in. However, as exemplified …

Content truncated to 500 words