Cross Copy Network for Dialogue Generation

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 183–196, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 183 Information-Theoretic Probing with Minimum Description Length Elena Voita1,2 Ivan Titov1,2 1University of Edinburgh, Scotland 2University of Amsterdam, Netherlands lena-voita@hotmail.com ititov@inf.ed.ac.uk Abstract To measure how well pretrained representa- tions encode some linguistic property, it is common to use accuracy of a probe, i.e. classifier trained to predict the property froma the representations. Despite widespread adop- tion of probes, differences in their accuracy fail to adequately reflect differences in repre- sentations. For example, they do not substan- tially favour pretrained representations over randomly initialized ones. Analogously, their accuracy can be similar when probing for gen- uine linguistic labels and probing for random synthetic tasks. To see reasonable differences in accuracy with respect to these random base- lines, previous work had to constrain either the amount of probe training data or its model size. Instead, we propose an alternative to the standard probes, information-theoretic prob- ing with minimum description length (MDL). With MDL probing, training a probe to pre- dict labels is recast as teaching it to effectively transmit the data. Therefore, the measure of interest changes from probe accuracy to the de- scription length of labels given representations. In addition to probe quality, the description length evaluates ‘the amount of effort’ needed to achieve the quality. This amount of effort characterizes either (i) size of a probing model, or (ii) the amount of data needed to achieve the high quality. We consider two methods for esti- mating MDL which can be easily implemented on top of the standard probing pipelines: varia- tional coding and online coding. We show that these methods agree in results and are more in- formative and stable than the standard probes.1 1 Introduction To estimate to what extent representations (e.g., ELMo (Peters et al., 2018) or BERT (Devlin et al., 2019)) capture a linguistic property, most previous 1We release code at https://github.com/ lena-voita/description-length-probing. Figure 1: Illustration of the idea behind MDL probes. work uses ‘probing tasks’ (aka ‘probes’ and ‘diag- nostic classifiers’); see Belinkov and Glass (2019) for a comprehensive review. These classifiers are trained to predict a linguistic property from ‘frozen’ representations, and accuracy of the classifier is used to measure how well these representations encode the property. Despite widespread adoption of such probes, they fail to adequately reflect differences in repre- sentations. This is clearly seen when using them to compare pretrained representations with randomly initialized ones (Zhang and Bowman, 2018). Anal- ogously, their accuracy can be similar when prob- ing for genuine linguistic labels and probing for tags randomly associated to word types (‘control tasks’, Hewitt and Liang (2019)). To see differ- ences in the accuracy with respect to these random baselines, previous work had to reduce the amount of a probe training data (Zhang and Bowman, 2018) or use smaller models for probes (Hewitt and Liang, 2019). As an alternative to the standard probing, we take an information-theoretic view at the …

Content truncated to 500 words