Author ORCID Identifier

Date Available


Year of Publication


Degree Name

Doctor of Philosophy (PhD)

Document Type

Doctoral Dissertation


Arts and Sciences



First Advisor

Dr. Ramakanth Kavuluru


Biomedical text is being generated at a high rate in scientific literature publications and electronic health records. Within these documents lies a wealth of potentially useful information in biomedicine. Relation extraction (RE), the process of automating the identification of structured relationships between entities within text, represents a highly sought-after goal in biomedical informatics, offering the potential to unlock deeper insights and connections from this vast corpus of data. In this dissertation, we tackle this problem with a variety of approaches.

We review the recent history of the field of document-level RE. Several themes emerge. First, graph neural networks dominate the methods for constructing entity and relation representations. Second, clever uses of attention allow for the these constructions to focus on particularly relevant tokens and object (such as mentions and entities) representations. Third, aggregation of signal across mentions in entity-level RE is a key focus of research. Fourth, the injection of additional signal by adding tokens to the text prior to encoding via language model (LM) or through additional learning tasks boosts performance. Last, we explore an assortment of strategies for the challenging task of end-to-end entity-level RE.

Of particular note are sequence-to-sequence (seq2seq) methods that have become particularly popular in the past few years. With the success of general-domain generative LMs, biomedical NLP researchers have trained a variety of these models on biomedical text under the assumption that they would be superior for biomedical tasks. As training such models is computationally expensive, we investigate whether they outperform generic models. We test this assumption rigorously by comparing performance of all major biomedical generative language models to the performances of their generic counterparts across multiple biomedical RE datasets, in the traditional finetuning setting as well as in the few-shot setting. Surprisingly, we found that biomedical models tended to underperform compared to their generic counterparts. However, we found that small-scale biomedical instruction finetuning improved performance to a similar degree as larger-scale generic instruction finetuning.

Zero-shot natural language processing (NLP) offers savings on the expenses associated with annotating datasets and the specialized knowledge required for applying NLP methods. Large, generative LMs trained to align with human objectives have demonstrated impressive zero-shot capabilities over a broad range of tasks. However, the effectiveness of these models in biomedical RE remains uncertain. To bridge this gap in understanding, we investigate how GPT-4 performs across several RE datasets. We experiment with the recent JSON generation features to generate structured output, which we use alternately by defining an explicit schema describing the relation structure, and inferring the structure from the prompt itself. Our work is the first to study zero-shot biomedical RE across a variety of datasets. Overall, performance was lower than that of fully-finetuned methods. Recall suffered in examples with more than a few relations. Entity mention boundaries were a major source of error, which future work could fruitfully address.

In our previous work with generative LMs, we noted that RE performance decreased with the number of gold relations in an example. This observation aligns with the general pattern that recurrent neural network and transformer-based model performance tends to decrease with sequence length. Generative LMs also do not identify textual mentions or group them into entities, which are valuable information extraction tasks unto themselves. Therefore, in this age of generative methods, we revisit non-seq2seq methodology for biomedical RE. We adopt a sequential framework of named entity recognition (NER), clustering mentions into entities, followed by relation classification (RC). As errors early in the pipeline necessarily cause downstream errors, and NER performance is near its ceiling, we focus on improving clustering. We match state-of-the-art (SOTA) performance in NER, and substantially improve mention clustering performance by incorporating dependency parsing and gating string dissimilarity embeddings.

Overall, we advance the field of biomedical RE in a few ways. In our experiments of finetuned LMs, we show that biomedicine-specific models are unnecessary, freeing researchers to make use of SOTA generic LMs. The relatively high few-shot performance in these experiments also suggests that biomedical RE can be reasonably accessible, as it is not so difficult to construct small datasets. Our investigation into zero-shot RE shows that SOTA LMs can compete with fully finetuned smaller LMs. Together these studies also demonstrate weaknesses of generative RE. Last, we show that non-generative RE methods still outperform generative methods in the fully-finetuned setting.

Digital Object Identifier (DOI)

Funding Information

This dissertation was supported by the United States National Library of Medicine via grant R01LM013240.