Query-Focused Medical Text Summarization
Investigating the status of QFS in the medical domain
We conduct a comprehensive comparison of several extractive and abstractive models for medical summarization like Pointer-Generator, COWTS+UMLS and BERTSumm in a query-focused setting. We also collect a dataset of 435K PubMed abstracts and their corresponding titles, to fine-tune our models on a summarization setting, to compare the effect of domain-specific finetuning on the task. Initial results indicate a better performance of extractive models compared to abstractive approaches. However, due to the large number of novel tokens in the reference summaries of abstractive summarization algorithms, a naive semantic overlap tends to favour extractive approaches due to their simple copy-and-paste approach. Hence, we modify the ROUGE score to account for similar Universal Medical Language System (UMLS) concepts (i.e. fever and flu will be considered the same), to alleviate some of this bias. The performance of extractive and abstractive approaches are much more comparable, and in line with the evaluations performed by medical professionals.