Coloring the Black Box: What Synesthesia Tells Us about Character Embeddings

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 2673–2685 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 2673 Coloring the Black Box: What Synesthesia Tells Us about Character Embeddings Katharina Kann∗ University of Colorado Boulder katharina.kann@colorado.edu Mauro M. Monsalve-Mercado∗ Columbia University mauro.m.monsalve@columbia.edu Abstract In contrast to their word- or sentence-level counterparts, character embeddings are still poorly understood. We aim at closing this gap with an in-depth study of English char- acter embeddings. For this, we use resources from research on grapheme–color synesthesia – a neuropsychological phenomenon where let- ters are associated with colors –, which give us insight into which characters are similar for synesthetes and how characters are orga- nized in color space. Comparing 10 differ- ent character embeddings, we ask: How sim- ilar are character embeddings to a synesthete’s perception of characters? And how similar are character embeddings extracted from dif- ferent models? We find that LSTMs agree with humans more than transformers. Compar- ing across tasks, grapheme-to-phoneme con- version results in the most human-like charac- ter embeddings. Finally, ELMo embeddings differ from both humans and other models. 1 Introduction Neural network models have become crucial tools in natural language processing (NLP) and define the state of the art on a large variety of tasks (Wang et al., 2018). However, they are difficult to un- derstand and are often considered ”black boxes”.1 This can make their use difficult to defend in many settings, for instance in a legal context, and consti- tutes a barrier for model improvement. Therefore, a lot of research has been dedicated to investigat- ing the information encoded by neural networks. Especially word embeddings, contextualized word representations, and language representation mod- els like BERT (Devlin et al., 2019) have been ex- haustively studied (Rogers et al., 2020). ∗Equal contribution. 1This fact has led to the establishment of a workshop with the same name: https://blackboxnlp.github.io A​ ​B​ ​C​ ​D​ ​E​ ​F​ ... Figure 1: Characters as they might be seen by a grapheme–color synesthete. The colors in this exam- ple are randomly chosen. Character embeddings are used for a large set of tasks, either as a supplement to word-level in- put, e.g., for part-of-speech tagging by Plank et al. (2016), or on their own, e.g., for character-level sequence-to-sequence (seq2seq) tasks by Kann and Sch¨utze (2016). Despite this, they have not yet been explicitly analysed. One reason for this might be that identifying relevant properties to study is more challenging than for their word-level coun- terparts. However, we argue that, in order to truly shine light into black-box NLP models, it is nec- essary to understand each and every component of them. In this paper, we perform a detailed study of English character embeddings. Our first contribu- tion is a character similarity task (§3) in analogy to the word-based version: Do embeddings agree with humans on which pairs of characters are simi- lar? While e.g., cat and tiger are generally consid- ered more similar than cat and …

Content truncated to 500 words