The Case of Imperfect Negation Cues: A Two-Step Approach for Automatic Negation Scope Resolution

Negation is a complex grammatical phenomenon that has received considerable attention in the biomedical natural language processing domain. While neural network-based methods are the state-of-the-art in negation scope resolution, they often use the unrealistic assumption that negation cue information is completely accurate. Even if this assumption holds, there remains a dependency on engineered features from state-of-the-art machine learning methods. To tackle this issue, in this study, we adopted a two-step negation resolving approach to assess whether a neural network-based model, here a bidirectional long short-term memory, can be a an alternative for cue detection. Furthermore, we investigate how inaccurate cue predictions would affect the scope resolution performance. We ran various experiments on the open access Bio-Scope corpus. Experimental results suggest that word embeddings alone can detect cues reasonably well, but there still exist better alternatives for this task. As expected, scope resolution performance suffers from imperfect cue information, but remains acceptable on the Abstracts subcorpus. We also found that the scope resolution performance is most robust against inaccurate information for models with a recurrent layer only, compared to extensions with a conditional random field layer and extensions with a post-processing algorithm. We advocate for more research into the application of automated deep learning on the effect of imperfect information on scope resolution.