MAF: Multimodal Alignment Framework for Weakly-Supervised Phrase Grounding

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 197–216, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 197 Intrinsic Probing through Dimension Selection Lucas Torroba Hennigen Adina Williams Ryan Cotterell , Québec Artificial Intelligence Institute (Mila) University of Cambridge Facebook AI Research lucas.torroba-hennigen@mila.quebec,ETH Zürich adinawilliams@fb.com, ryan.cotterell@inf.ethz.ch Abstract Most modern NLP systems make use of pre- trained contextual representations that attain astonishingly high performance on a variety of tasks. Such high performance should not be possible unless some form of linguistic struc- ture inheres in these representations, and a wealth of research has sprung up on probing for it. In this paper, we draw a distinction be- tween intrinsic probing, which examines how linguistic information is structured within a representation, and the extrinsic probing pop- ular in prior work, which only argues for the presence of such information by showing that it can be successfully extracted. To enable intrinsic probing, we propose a novel frame- work based on a decomposable multivariate Gaussian probe that allows us to determine whether the linguistic information in word em- beddings is dispersed or focal. We then probe fastText and BERT for various morphosyntac- tic attributes across 36 languages. We find that most attributes are reliably encoded by only a few neurons, with fastText concentrating its linguistic structure more than BERT.1 1 Introduction Natural language processing (NLP) is enamored of contextual word representations—and for good rea- son! Contextual word-embedders, e.g. BERT (De- vlin et al., 2019) and ELMo (Peters et al., 2018), have bolstered NLP model performance on myr- iad tasks, such as syntactic parsing (Kitaev et al., 2019), coreference resolution (Joshi et al., 2019), morphological tagging (Kondratyuk, 2019) and text generation (Zellers et al., 2019). Given the large empirical gains observed when they are employed, it is all but certain that word representations derived from neural networks encode some continuous ana- logue of linguistic structures. 1Code and data are available at https://github. com/rycolab/intrinsic-probing. −1.5 −1 −0.5 0 0.5 1 −2 −1.5 −1 −0.5 0 0.5 1 1.5 Dimension 477 Dimension 179 Figure 1: Scatter plot of the two most informative BERT dimensions for English present and past tense. The contours belong to our probe. Exactly what these representations encode about linguistic structure, however, remains little under- stood. Researchers have studied this question by attributing function to specific network cells with visualization methods (Karpathy et al., 2015; Li et al., 2016) and by probing (Alain and Bengio, 2017; Belinkov and Glass, 2019), which seeks to extract structure from the representations. Recent work has probed various representations for cor- relates of morphological (Belinkov et al., 2017; Giulianelli et al., 2018), syntactic (Hupkes et al., 2018; Zhang and Bowman, 2018; Hewitt and Man- ning, 2019; Lin et al., 2019), and semantic (Kim et al., 2019) structure. Most current probing efforts focus on what we term extrinsic probing, where the goal is to de- termine whether the posited linguistic structure is predictable from the learned representation. Gen- erally, extrinsic probing works argue for the …

Content truncated to 500 words