RNNs as psycholinguistic subjects: Syntactic state and grammatical dependency
This addresses the problem of understanding what grammatical features RNNs implicitly learn, which is important for researchers in NLP and cognitive science, but it is incremental as it builds on existing psycholinguistic methods.
The study investigated whether recurrent neural networks (RNNs) learn grammatical structures like humans by testing LSTMs on English and Japanese, finding they represent incremental syntactic state but fail to generalize like humans and do not learn correct grammatical dependencies for reflexives or negative polarity items.
Recurrent neural networks (RNNs) are the state of the art in sequence modeling for natural language. However, it remains poorly understood what grammatical characteristics of natural language they implicitly learn and represent as a consequence of optimizing the language modeling objective. Here we deploy the methods of controlled psycholinguistic experimentation to shed light on to what extent RNN behavior reflects incremental syntactic state and grammatical dependency representations known to characterize human linguistic behavior. We broadly test two publicly available long short-term memory (LSTM) English sequence models, and learn and test a new Japanese LSTM. We demonstrate that these models represent and maintain incremental syntactic state, but that they do not always generalize in the same way as humans. Furthermore, none of our models learn the appropriate grammatical dependency configurations licensing reflexive pronouns or negative polarity items.