Bio-AnswerFinder

Bio-AnswerFinder retrieves and ranks literature-derived answer sentences for biomedical questions using PubMed-derived embeddings, iterative keyword-based retrieval, a supervised LSTM for keyword selection, and a fine-tuned BERT classifier.

Key Features:

• Weighted-Relaxed Word Mover's Distance: Ranks candidate answers by computing a weighted-relaxed Word Mover's Distance on word/phrase embeddings derived from PubMed abstracts to assess similarity to the question focus entity type.
• Iterative Document Retrieval: Enhances document retrieval via iterative searches using enhanced keyword queries sourced from a traditional search engine to broaden and refine coverage of relevant literature.
• Supervised LSTM Neural Network: Uses a supervised Long Short-Term Memory (LSTM) network to analyze questions and select the most pertinent keywords for iterative retrieval.
• Fine-tuned BERT Classifier: Applies a fine-tuned bidirectional encoder representations from transformers (BERT) classifier trained on 100 candidate sentences per question for 492 BioASQ questions to rerank answer sentences.

Scientific Applications:

• Biomedical question answering from PubMed: Extracts and ranks sentence-level answers from the PubMed literature corpus for question answering tasks.
• Entity-focused information retrieval: Targets retrieval and ranking according to the question's focus entity type to identify entity-relevant answer sentences.
• Benchmarking on BioASQ: Evaluates and benchmarks answer retrieval and ranking performance using BioASQ question sets.

Methodology:

Computationally, Bio-AnswerFinder employs a two-tiered approach: initial document retrieval via iterative enhanced keyword queries (keywords selected by a supervised LSTM) followed by answer sentence ranking using a weighted-relaxed Word Mover's Distance on PubMed-derived embeddings and a fine-tuned BERT classifier trained on 100 candidate sentences per question for 492 BioASQ questions.