Probing Linguistic Systematicity

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1958–1969 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 1958 Probing Linguistic Systematicity Emily Goodwin,∗1,5 Koustuv Sinha,2,3,5 Timothy J. O’Donnell1,4,5 1Department of Linguistics, 2School of Computer Science, McGill University, Canada 3Facebook AI Research (FAIR), Montreal 4Canada CIFAR AI Chair, Mila 5Quebec Artificial Intelligence Institute (Mila) {emily.goodwin, koustuv.sinha}@mail.mcgill.ca timothy.odonnell@mcgill.ca Abstract Recently, there has been much interest in the question of whether deep natural language understanding models exhibit systematicity— generalizing such that units like words make consistent contributions to the meaning of the sentences in which they appear. There is ac- cumulating evidence that neural models often generalize non-systematically. We examined the notion of systematicity from a linguistic perspective, defining a set of probes and a set of metrics to measure systematic behaviour. We also identified ways in which network ar- chitectures can generalize non-systematically, and discuss why such forms of generalization may be unsatisfying. As a case study, we per- formed a series of experiments in the setting of natural language inference (NLI), demonstrat- ing that some NLU systems achieve high over- all performance despite being non-systematic. 1 Introduction Language allows us to express and comprehend a vast variety of novel thoughts and ideas. This creativity is made possible by compositionality— the linguistic system builds utterances by combin- ing an inventory of primitive units such as mor- phemes, words, or idioms (the lexicon), using a small set of structure-building operations (the gram- mar; Camap, 1947; Fodor and Pylyshyn, 1988; Hodges, 2012; Janssen et al., 2012; Lake et al., 2017b; Szabó, 2012; Zadrozny, 1994; Lake et al., 2017a). One property of compositional systems, widely studied in the cognitive sciences, is the phe- nomenon of systematicity. Systematicity refers to the fact that lexical units such as words make consistent contributions to the meaning of the sen- tences in which they appear. Fodor and Pylyshyn ∗Corresponding author (1988) provided a famous example: If a compe- tent speaker of English knows the meaning of the sentence John loves the girl, they also know the meaning of The girl loves John. This is because for speakers of English knowing the meaning of the first sentence implies knowing the meaning of the individual words the, loves, girl, and John as well as grammatical principles such as how transi- tive verbs take their arguments. But knowing these words and principles of grammar implies know- ing how to compose the meaning of the second sentence. Deep learning systems now regularly exhibit very high performance on a large variety of natural language tasks, including machine translation (Wu et al., 2016; Vaswani et al., 2017), question answer- ing (Wang et al., 2018; Henaff et al., 2016), visual question answering (Hudson and Manning, 2018), and natural language inference (Devlin et al., 2018; Storks et al., 2019). Recently, however, researchers have asked whether such systems generalize sys- tematically (see §4). Systematicity is the property whereby words have consistent contributions to composed mean- ing; the alternative is the situation …

Content truncated to 500 words