Using Priming to Uncover the Organization of Syntactic Representations in Neural Language Models

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 66–76 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 66 Using Priming to Uncover the Organization of Syntactic Representations in Neural Language Models Grusha Prasad Johns Hopkins University grusha.prasad@jhu.edu Marten van Schijndel Cornell University mv443@cornell.edu Tal Linzen Johns Hopkins University tal.linzen@jhu.edu Abstract Neural language models (LMs) perform well on tasks that require sensitivity to syntactic structure. Drawing on the syntactic priming paradigm from psycholinguistics, we propose a novel technique to analyze the representa- tions that enable such success. By establish- ing a gradient similarity metric between struc- tures, this technique allows us to reconstruct the organization of the LMs’ syntactic rep- resentational space. We use this technique to demonstrate that LSTM LMs’ representa- tions of different types of sentences with rel- ative clauses are organized hierarchically in a linguistically interpretable manner, suggesting that the LMs track abstract properties of the sentence. 1 Introduction Neural networks trained on text alone, without explicit syntactic supervision, have been surpris- ingly successful in tasks that require sensitivity to sentence structure. The difficulty of interpret- ing the learned neural representations that under- lie this success has motivated a range of analysis techniques, including diagnostic classifiers (Giu- lianelli et al., 2018; Conneau et al., 2018; Shi et al., 2016), visualization of individual neuron activa- tions (K´ad´ar et al., 2017; Qian et al., 2016), ab- lation of individual neurons or sets of neurons (Lakretz et al., 2019) and behavioral tests of gener- alization to infrequent or held out syntactic struc- tures (Linzen et al., 2016; Weber et al., 2018; Mc- Coy et al., 2018); for reviews, see Belinkov and Glass (2019) and Alishahi et al. (2019). This paper expands the toolkit of neural net- work analysis techniques by drawing on the syn- tactic priming paradigm, a central tool in psy- cholinguistics for analyzing human syntactic rep- resentations (Bock, 1986). This paradigm is based on the empirical finding that people tend to reuse syntactic structures that they have recently pro- duced or encountered. For example, English pro- vides two roughly equivalent ways to express a transfer event: (1) a. The boy threw the ball to the dog. b. The boy threw the dog the ball. When readers encounter one of these variants in the text more frequently than the other, they ex- pect that future transfer events will more likely be expressed using the frequent construction than the infrequent one. For example, after reading sen- tences like (1a) (the prime), readers expect sen- tences like (2a), which shares syntactic structure with the prime, to occur with a greater likelihood than the alternative variant like (2b) which does not (Wells et al., 2009).1 (2) a. The lawyer sent the letter to the client. b. The lawyer sent the client the letter. We use the priming paradigm to analyze neu- ral network language models (LMs), systems that define a probability distribution over the nth word of a sentence given its first n − 1 words. Building …

Content truncated to 500 words