Learning English tenses from Sentential Input: A Neural Network Approach

Anonymous

Abstract: Children are able to productively use and understand tense information, such as lexical verbs, auxiliaries and tense morphemes. How children acquire the tense information remains unclear. One controversy is whether linguistic input alone is sufficient enough for the children to learn tense information, or whether children extract the tense information using abstract syntactic knowledge and/or multimodal cognition. This study uses transformer models to understand the process of tense acquisition from sentential input. We train transformer models on English tense classification tasks with sentences in child directed speech as the input. When the transformer models successfully learn the tense, we find that 1) the models are sensitive to auxiliary verbs (e.g. was, do) but not phrases (e.g. is going to), 2) the past tense -ed form facilitates classification, and 3) temporal adverbs have limited impact in tense classification.