TL;DR
This paper tackles the challenge of automatically building comprehensive corpora of scientific literature on food-drug interactions (FDIs). They propose and evaluate methods for selecting effective search queries, demonstrating that feature selection techniques like gain ratio, applied to index terms, outperform simpler frequency-based approaches for identifying relevant documents.
What Problem Does This Paper Address?
Identifying and understanding food-drug interactions (FDIs) is crucial for patient safety, as these interactions can lead to treatment failures and adverse effects. While drug-drug interactions are well-studied, FDIs are often discovered post-market and are less systematically documented. Building large, high-quality corpora of scientific articles describing FDIs is essential for information extraction and knowledge discovery, but it's a non-trivial task. Existing resources like online databases require constant updates, and manually curating comprehensive corpora from scientific literature is time-consuming and expensive. The authors aim to automate the process of constructing such corpora by intelligently selecting search queries.
Methods
The core of this work lies in automating the query selection process. The authors frame the problem as a binary classification task: identifying relevant documents (those describing FDIs) versus irrelevant ones.
- Document Classification: They experiment with various classifiers (Decision Tree, SVM, Naive Bayes, Logistic Regression, RandomForest) and feature representations (term frequency, TF-IDF, one-hot encoding) to classify scientific abstracts.
- Query Selection: The primary focus is on selecting effective queries. They explore two main types of scoring functions for candidate queries:
- Frequency-based: Simple counts of how often a term appears in relevant documents.
- Feature Selection-inspired: Methods borrowed from supervised classification, such as Information Gain and Gain Ratio, which measure how well a term discriminates between relevant and irrelevant documents.
- Candidate Queries: They consider two main sources for candidate queries:
- Index Terms: Keywords assigned to scientific articles (e.g., MeSH terms).
- Document N-grams: Sequences of words within documents.
- Feature Engineering: They also incorporate domain-specific knowledge by creating generalized index terms like "Foods," "Drugs," and "Foods and Drugs" by exploiting the hierarchical structure of MeSH.
Datasets
The research utilizes several datasets for training and evaluation:
- Stockley compendium: A reference compendium of drug interactions, used as a source for identifying relevant documents.
- MEDLINE abstracts: The primary source of scientific literature.
- POMELO corpus: A previously created corpus of MEDLINE abstracts on FDIs, noted as having limited coverage.
- DBpedia: Used to extract a large list of food names (15,686) to create drug-food query candidates.
They construct several experimental subsets of abstracts, including:
Stockley2008: References cited in the Stockley compendium.FDI-HDI: Results from MeSH term queries for "Food-Drug Interaction" and "Herb-Drug Interaction."DRUGFOOD: Abstracts generated by combining drug names with food names extracted from DBpedia.MEDLINE2008: All MEDLINE abstracts published before 2008.jrnlAbstracts(): Abstracts from journals that published articles in theStockley2008andDRUGFOODsubsets.
Key Insights and Ideas
- Index terms are powerful: While word-based features (n-grams) can be noisy, index terms (like MeSH) provide more precise signals for identifying relevant FDI literature.
- Feature selection beats frequency: Methods that assess a term's discriminative power (like Gain Ratio) are more effective for query selection than simple term frequency. This is because they can identify terms that are highly indicative of FDIs, even if they aren't the absolute most frequent.
- Informed negative sampling is crucial: When dealing with unbalanced datasets (many irrelevant documents for few relevant ones), simply random sampling for negative examples is less effective. Using more informed strategies, such as considering publication venue and year, leads to better classification performance.
- Domain knowledge enhances queries: Incorporating hierarchical information from vocabularies like MeSH to create broader categories (e.g., "Foods") can improve query effectiveness.
Key Results
- Document Classification: The best document classification performance was achieved using TF-IDF features with a Support Vector Machine (SVM) classifier, particularly on datasets with more restrictive negative example selection (Experiment 3).
- Query Selection: The Gain Ratio scoring function, applied to index terms, consistently outperformed other methods (Frequency, TF-IDF, Information Gain, Pearson's Correlation) in predicting query performance.
- Best Performing Queries: Queries selected using Gain Ratio included terms like "Biological Availability," "Drug Interactions," and "Intestinal Absorption." These are noted as being more precise than broad word-based candidates.
Limitations and Future Work
- Query Complexity: The current work primarily focuses on single index term queries. Future work should investigate more complex queries that combine multiple terms.
- Keyphrase Extraction: Comparing their approach to keyphrase extraction methods for generating queries is suggested.
- Background Knowledge Integration: Further leveraging background knowledge about drugs and foods could enhance query generation.
- Dataset Freshness: The datasets used were based on older versions of compendia (e.g., Stockley 2008). Given the dynamic nature of FDI research, using more recent data would be beneficial.
Summary updated