- Keywords: semantic role labeling, domain adaptation
- TL;DR: The paper explores simple but effective ways of adapting neural models for domain adaptation for biological process.
- Abstract: Domain adaptation remains one of the most challenging aspects in the wide-spread use of Semantic Role Labeling (SRL) systems. Current state-of-the-art methods are typically trained on large-scale datasets, but their performances do not directly transfer to low-resource domain-specific settings. In this paper, we propose two approaches for domain adaptation that involve pre-training on existing large-scale datasets and adapting it for a low-resource corpus. Our first approach defines a mapping between the source labels and the target labels, and the other approach modifies the final CRF layer in sequence-labeling neural network architecture. We adapt LSTM-CRF based models for each of these proposed domain adaptation strategies which perform significantly better than the baselines. To simulate a low resource setting with an entirely different annotation scheme, we perform our experiments on ProcessBank dataset which contains less than 200 paragraphs on biological processes. Our results show that our best system improves the state-of-the-art system on the dataset by 21.7 F1 points, validating the strength of our domain adaptation strategies. We also show that, by incorporating event-event relationship in ProcessBank, we are able to achieve an additional 2.6 F1 gain, giving us possible insights into how to improve SRL systems for biological process using richer annotations.
- Archival status: Non-Archival
- Subject areas: Natural Language Processing, Applications: Science