How LSTM Encodes Syntax: Exploring Context Vectors and Semi-Quantization on Natural Text

Abstract

Long Short-Term Memory recurrent neural network (LSTM) is widely used and known to capture informative long-term syntactic dependencies. However, how such information are reflected in its internal vectors for natural text has not yet been sufficiently investigated. We analyze them by learning a language model where syntactic structures are implicitly given. We empirically show that the context update vectors, i.e. outputs of internal gates, are approximately quantized to binary or ternary values to help the language model to count the depth of nesting accurately, as Suzgun et al. ( 2019) recently showed for synthetic Dyck languages. For some dimensions in the context vector, we show that their activations are highly correlated with the depth of phrase structures, such as VP and NP. Moreover, with an L 1 regularization, we also found that it can be accurately predicted whether a word is inside a phrase structure or not from a small number of components of the context vector. Even for the case of learning from raw text, context vectors are still shown to correlate well with the phrase structures. Finally, we show that natural clusters of the functional words and the parts of speech that trigger phrases are represented in a small but principal subspace of the context-update vector of LSTM.

Effects of Parameter Norm Growth During Transformer Training: Inductive Bias from Gradient Descent

“…Experiments suggest that LSTMs trained on synthetic tasks learn to implement counter memory (Weiss et al, 2018;Suzgun et al, 2019a), and that they fail on tasks requiring stacks and other deeper models of structure (Suzgun et al, 2019b). Similarly, Shibata et al (2020) found that LSTM language models trained on natural language data acquire saturated representations approximating counters.…”

Probing Multilingual Cognate Prediction Models

“…Phonology To study whether our models learn phonologically meaningful representations, we study our high-dimensionality hidden representation for each item of our vocabulary, as suggested in Madsen et al (2021). We reduce the dimensionality of our encoded representations using PCA (Pearson, 1901) andt-SNE (der Maaten andHinton, 2008) and look at the emerging underlying organisation of the phonetic space, as was done in Jacobs and Mailhot (2019) and Shibata et al (2020) for, respectively, seq2seq phonetic and LSTM syntactic representations analysis.…”

Formal Language Theory Meets Modern NLP

“…Similarly, they cannot reliably reverse strings (Hao et al, 2018;Merrill, 2019). Shibata et al (2020) show that LSTM language models trained on natural language acquire semi-saturated representations where the gates tightly cluster around discrete values. Thus, sLSTMs appear to be a promising formal model of the counting behavior of LSTMs on both synthetic and natural tasks.…”