Character-level Representations Improve DRS-based Semantic Parsing Even in the Age of BERT

We combine character-level and contextual language model representations to\nimprove performance on Discourse Representation Structure parsing. Character\nrepresentations can easily be added in a sequence-to-sequence model in either\none encoder or as a fully separate encoder, with improvements that are robust\nto different language models, languages and data sets. For English, these\nimprovements are larger than adding individual sources of linguistic\ninformation or adding non-contextual embeddings. A new method of analysis based\non semantic tags demonstrates that the character-level representations improve\nperformance across a subset of selected semantic phenomena.\n