Incorporating Domain Knowledge into Medical NLI using Knowledge Graphs

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 6092–6097, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 6092 Incorporating Domain Knowledge into Medical NLI using Knowledge Graphs Soumya Sharma∗ soumyasharma20† Bishal Santra∗ bsantraigi† Abhik Jana abhikjana1† T Y S S Santosh santoshtyss† Niloy Ganguly niloy‡ Pawan Goyal pawang‡ Indian Institute of Technology Kharagpur, India {†}@gmail.com, {‡}@cse.iitkgp.ac.in Abstract Recently, biomedical version of embeddings obtained from language models such as BioELMo have shown state-of-the-art results for the textual inference task in the medical do- main. In this paper, we explore how to incor- porate structured domain knowledge, available in the form of a knowledge graph (UMLS), for the Medical NLI task. Specifically, we experiment with fusing embeddings obtained from knowledge graph with the state-of-the- art approaches for NLI task, which mainly rely on contextual word embeddings. We also experiment with fusing the domain-specific sentiment information for the task. Experi- ments conducted on MedNLI dataset clearly show that this strategy improves the baseline BioELMo architecture for the Medical NLI task1. 1 Introduction Natural language inference (NLI) is one of the basic natural language understanding tasks which deals with detecting inferential relationship such as entailment or contradiction, between a given premise and a hypothesis. In recent years, with the availability of large annotated datasets like SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018), researchers have come up with sev- eral neural network based models which could be trained with these large annotated datasets and are able to produce state-of-the-art perfor- mances (Bowman et al., 2015, 2016; Munkhdalai and Yu, 2017; Sha et al., 2016; Chen et al., 2017; ∗equal contribution 1https://github.com/soummyaah/ KGMedNLI/ Tay et al., 2018). With these attempts, even though NLI in domains like fiction, travel etc. has pro- gressed a lot, NLI in medical domain is yet to be explored extensively. With the introduction of MedNLI (Romanov and Shivade, 2018), an expert annotated dataset for NLI in the clinical domain, researchers have started pursuing the problem of clinical NLI. Modeling informal inference is one of the basic tasks towards achieving natural lan- guage understanding, and is considered very chal- lenging. MedNLI is a dataset that assists in as- sessing how good a sentence or word embedding method is for downstream uses in medical domain. Recently, with the emergence of advanced con- textual word embedding methods like ELMo (Pe- ters et al., 2018) and BERT (Devlin et al., 2018), performances of many NLP tasks have improved, setting state-of-the-art performances. Following this stream of literature, Lee et al. (2019) in- troduce BioBERT, which is a BERT model pre- trained on English Wikipedia, BooksCorpus and fine-tuned on PubMed (7.8B tokens in total) cor- pus, PMC full-text articles. Jin et al. (2019) pro- pose BioELMo which is a domain-specific version of ELMo trained on 10M PubMed abstracts, and attempt to solve medical NLI problem with these domain specific embeddings, leading to state-of- the-art performance. These two …

Content truncated to 500 words