Human Inspired Progressive Alignment and Comparative Learning for Grounded Word Acquisition

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 15475–15493 July 9-14, 2023 ©2023 Association for Computational Linguistics Human Inspired Progressive Alignment and Comparative Learning for Grounded Word Acquisition Yuwei Bao† Barrett Martin Lattimer§∗ Joyce Chai† †Computer Science and Engineering, University of Michigan §ASAPP {yuweibao, lattimer, chaijy}@umich.edu Abstract Human language acquisition is an efficient, su- pervised, and continual process. In this work, we took inspiration from how human babies ac- quire their first language, and developed a com- putational process for word acquisition through comparative learning. Motivated by cognitive findings, we generated a small dataset that en- ables the computation models to compare the similarities and differences of various attributes, learn to filter out and extract the common infor- mation for each shared linguistic label. We frame the acquisition of words as not only the information filtration process, but also as representation-symbol mapping. This proce- dure does not involve a fixed vocabulary size, nor a discriminative objective, and allows the models to continually learn more concepts effi- ciently. Our results in controlled experiments have shown the potential of this approach for efficient continual learning of grounded words. 1 Introduction Two of the important word acquisition problems are: 1) what must be learned to acquire a word, and 2) how to learn the word? To the first question, cog- nitive studies have shown that several critical steps in learning language naturally comes from joint at- tention establishment (Tomasello and Farrar, 1986), and symbol grounding (Harnad, 1990). Children’s attention are usually redirected through a mother or teacher’s guidance, and they learn to map these at- tended sensory inputs (e.g. color, sound, heat) with their corresponding words or sentences. Living in a rich and diverse world enabled by our multiple body sensors, we learned to filter out the noise and pay attention to specific aspects of an input which we assign linguistic labels to. This attention estab- lishment and information filtration process is the first step of word acquisition. After filtering out the noise, we are left with a mental representation of ∗ Work done during master study at the University of Michigan. what a word entails. Just as the word “car” triggers certain impressions of a common transportation in head, we store these representations as they could come in handy later when we use them to reason, imagine, and express ourselves. To acquire a word, humans learn to filter out noise to focus on key information from sensory inputs that contributes to its meaning (Gentner, 1983; Tomasello and Farrar, 1986), and store that meaning representation for future use (Harnad, 1990; Kuehne et al., 2000). As for the second question, one of the common but under-explored methods is implicit or explicit comparisons. Caretakers may lay out stuffed ani- mals around a baby and name them one by one to differentiate them. In school, teachers may compare different components of the learning ma- terial, e.g. “Today we learn ‘colors’. This is red/blue/yellow...”. Comparison is the process …

Content truncated to 500 words