Language Models Learn POS First

Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 328–330 Brussels, Belgium, November 1, 2018. c⃝2018 Association for Computational Linguistics 328 Language Models Learn POS First Naomi Saphra and Adam Lopez n.saphra@ed.ac.uk alopez@ed.ac.uk Institute for Language, Cognition, and Computation University of Edinburgh 1 Introduction A glut of recent research shows that language models capture linguistic structure. Linzen et al. (2016) found that LSTM-based language models may encode syntactic information sufficient to fa- vor verbs which match the number of their subject nouns. Liu et al. (2018) suggested that the high performance of LSTMs may depend on the lin- guistic structure of the input data, as performance on several artificial tasks was higher with natural language data than with artificial sequential data. Such work answers the question of whether a model represents linguistic structure. But how and when are these structures acquired? Rather than treating the training process itself as a black box, we investigate how representations of linguis- tic structure are learned over time. In particular, we demonstrate that different aspects of linguistic structure are learned at different rates, with part of speech tagging acquired early and global topic information learned continuously. 2 Methods 2.1 Concentration We measure the degree to which a neural network has “structured” its representation x of a particular word in a sequence through concentration. c(x) = ∥x∥2 ∥x∥1 (1) The more similar in value the cells of x are, the smaller its l2/l1 ratio is. Thus if a neural network relies heavily on a small number of cells in an acti- vation pattern, the activation is very concentrated. Likewise, a concentrated gradient is mainly mod- ifying a few specific pathways. For example, it might modify a neuron associated with particular inputs like parentheses (Karpathy et al., 2015), or properties like sentiment (Radford et al., 2017). Figure 1: Correlation between mean concentration of a word gradient and word frequency. Vertical dashes mark when the optimizer rescales step size. 2.2 SVCCA Existing work investigates how language model layers encode tags by training taggers on the ac- tivations produced by each layer (Belinkov et al., 2018). We use an alternative technique, SVCCA (Raghu et al., 2017), which interprets an arbitrary selection of neurons in terms of how they relate to another selection of neurons from any network run on the same input data. This method treats a se- lection of neurons as a subspace, spanned by their activations. Given any 2 sets of neurons, SVCCA projects the 2 distinct views of the same data onto a shared subspace which maximizes correlation between the 2 views. Intuitively, if both views encode the same se- mantic information, the correlation in the shared subspace will be high. If the 2 views are encoding disjoint properties, the correlation will be low. 3 Experiments All experiments are conducted on 1.6GB of En- glish Wikipedia (70/10/20 train/dev/test split) with a 2-layer LSTM language model featuring tied weights in the softmax and embedding layers. 3.1 Gradient Concentration During Training …

Content truncated to 500 words