Sensitivity to Input Order: Evaluation of an Incremental and Memory-Limited Bayesian Cross-Situational Word Learning Model

Proceedings of the 27th International Conference on Computational Linguistics, pages 3170–3180 Santa Fe, New Mexico, USA, August 20-26, 2018. 3170 Sensitivity to Input Order: Evaluation of an Incremental and Memory-Limited Bayesian Cross-Situational Word Learning Model Sepideh Sadeghi Tufts University Human-Robot Interaction Laboratory Medford, MA, USA sepideh.sadeghi@tufts.edu Matthias Scheutz Tufts University Human-Robot Interaction Laboratory Medford, MA, USA matthias.scheutz@tufts.edu Abstract We present a variation of the incremental and memory-limited algorithm in (Sadeghi et al., 2017) for Bayesian cross-situational word learning and evaluate the model in terms of its functional performance and its sensitivity to input order. We show that the functional performance of our sub-optimal model on corpus data is close to that of its optimal counterpart (Frank et al., 2009), while only the sub-optimal model is capable of predicting the input order effects reported in experimental studies. 1 Introduction It has been suggested that early word learning is guided by observing the referent of words across situ- ations (cross-situational word learning). Surely the co-occurrence statistics of word-object mappings is noisy (e.g., when the referent of a word is absent in the scene) and only provides ambiguous information (due to referential ambiguity) about word-object mappings, but it is the only source of information used by a true novice who does not have access to social and linguistic cues (Baldwin, 1993; Arunachalam and Waxman, 2010) to narrow down the hypothesis space of words’ meanings. In fact, development of some initial meaning interpretations is the prerequisite for the development of useful social and syntactic knowledge (cues) for bootstrapping the process of word learning. Recent experimental findings (Medina et al., 2011; Trueswell et al., 2013; Zhang and Yu, 2017) in- dicate that word learning is sensitive to the presentation order of learning trials. More specifically, their results suggest that early wrong interpretations of words’ meanings can inhibit the acquisition of correct words’ meanings and that performance is best when highly informative learning trials (e.g., with lower referential ambiguity) precede the low informative learning trials, so that the possibility of starting with wrong meaning interpretations is minimized. These results are incompatible with the prediction of the existing well-known word learning models which employ global learning and are equipped with perfect memory of past observations (Yu and Ballard, 2007; Kachergis et al., 2012; Xu and Tenenbaum, 2007; Frank et al., 2009; Alishahi and Fazly, 2010). These models are insensitive to input order due to two major reasons: (1) being purely driven by co-occurrence regularities and assuming full access to co- occurrence statistics which stays the same regardless of input order or (2) processing the data in batch. Word learning is one of a wide range of cognitive tasks which exhibit sensitivity to the presentation style of input. Inductive category learning (Carvalho and Goldstone, 2015), category generalization (Spencer et al., 2011), property projection (Lawson, 2017), etc. all demonstrate important differences under se- quential vs. simultaneous presentation of input. Therefore, the root causes of the effects of presentation order on word learning results should be understood as it may …

Content truncated to 500 words