Re-solve it: simulating the acquisition of core semantic competences from small data

Proceedings of the 24th Conference on Computational Natural Language Learning, pages 344–354 Online, November 19-20, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 344 Simulating the acquisition of core semantic competences from small data Aur´elie Herbelot Center for Mind/Brain Sciences Dept. of Information Engineering and Computer Science University of Trento aurelie.herbelot@unitn.it Abstract Many tasks are considered to be ‘solved’ in the computational linguistics literature, but the corresponding algorithms operate in ways which are radically different from human cog- nition. I illustrate this by coming back to the notion of semantic competence, which includes basic linguistic skills encompassing both referential phenomena and generic knowl- edge, in particular a) the ability to denote, b) the mastery of the lexicon, or c) the ability to model one’s language use on others. Even though each of those faculties has been exten- sively tested individually, there is still no com- putational model that would account for their joint acquisition under the conditions experi- enced by a human. In this paper, I focus on one particular aspect of this problem: the amount of linguistic data available to the child or ma- chine. I show that given the first competence mentioned above (a denotation function), the other two can in fact be learned from very lim- ited data (2.8M token), reaching state-of-the- art performance. I argue that both the nature of the data and the way it is presented to the system matter to acquisition. 1 Introduction Many tasks and datasets are considered solved problems in the computational linguistics literature. However, the data, training regimes and system ar- chitectures required to obtain top performance are unrealistic from the point of view of human cogni- tion. Thus, state-of-the-art data-driven frameworks can be considered excellent engineering solutions to particular linguistic tasks, but they are not us- able as ‘models’ of language acquisition, and thus of limited applicability to test hypotheses about human language. This paper argues that core problems in compu- tational linguistics should be ‘re-solved’ – solved again – not as tasks, but as phenomena to simulate. This would involve a more careful attention to i) the type of data fed to the system; ii) the knowledge already in-built in its architecture; iii) the mode of learning implied by the training regime; iv) the specific features exploited by the learning process; and of course, v) the theoretical proposals explain- ing the phenomenon. Some of these desiderata have started being explored in the literature: the BlackBox NLP events, for instance, are currently fostering important discussions on the interpretabil- ity of artificial neural systems (Linzen et al., 2018, 2019). Still, the field remains far from satisfying all of them. The work described in the following pages is a step towards the simulation of a particular phe- nomenon: the acquisition of core semantic com- petences. Its specific focus is on data: more par- ticularly, the type and size of the corpus a system is exposed to. As we will see, talking about in- put data (desideratum i. above) naturally brings in questions …

Content truncated to 500 words