Recognizing Mentions of Adverse Drug Reaction in Social Media Using Knowledge-Infused Recurrent Models

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 142–151, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics Recognizing Mentions of Adverse Drug Reaction in Social Media Using Knowledge-Infused Recurrent Models Gabriel Stanovsky∗ Bar-Ilan University Ramat Gan, Israel gabriel.satanovsky@gmail.com Daniel Gruhl IBM Research Almaden San Jose, California, USA dgruhl@us.ibm.com Pablo N. Mendes∗ Lattice Data, Inc. Menlo Park, California, USA pablomendes@gmail.com Abstract Recognizing mentions of Adverse Drug Reactions (ADR) in social media is chal- lenging: ADR mentions are context- dependent and include long, varied and unconventional descriptions as compared to more formal medical symptom ter- minology. We use the CADEC cor- pus to train a recurrent neural network (RNN) transducer, integrated with knowl- edge graph embeddings of DBpedia, and show the resulting model to be highly accurate (93.4 F1). Furthermore, even when lacking high quality expert annota- tions, we show that by employing an active learning technique and using purpose built annotation tools, we can train the RNN to perform well (83.9 F1). 1 Introduction Identifying medical concepts in social media nar- ratives is the task of recognizing certain phrases in the context of a user’s post. Each phrase is also assigned a label from a set of predefined medi- cal types. For instance, given the sentence “As- pirin cured my terrible headache, but made me sleepy”, the following medical concepts can be identified: “Aspirin” is identified as Drug, “ter- rible headache” as a Symptom and “made me sleepy” should be spotted as a Adverse Drug Re- action. Having an automatic identification process can help domain experts examine large quantities of unstructured data, and quickly identify emerging trends. For example, associating previously un- known side effects with a given drug, or identify- ing an unforeseen impact to a change in the man- ufacturing process. ∗Work performed while at IBM Research Almaden. There are several challenges in addressing this task. First, context is crucial to type assign- ment. Compare the previous example with “As- pirin cured my sleepiness but gave me a terrible headache”, while the medical concepts are simi- lar, their context determines their particularly as- sociated type label. The social media domain poses additional chal- lenges. User narratives on social platforms tend to be non-grammatical, use colloquialisms, slang, and generally informal language. For example, a user may express sleepiness as “hard time get- ting some Z’s”. This hinders the use of pre-trained statistical parsers or simple string matching tech- niques. In this work we focus on the identification of Adverse Drug Reactions (ADR). These are un- wanted side effects which the user clearly iden- tifies as caused by the intake of a drug. ADRs are particularly challenging to spot, as they can be articulated in a variety of ways and can often be confounded with the symptoms addressed by the drug. Previous work in this field has mainly used care- fully built lexicons and hand-coded rule based sys- tems (Iqbal et al., 2015). While each individual system …

Content truncated to 500 words