A scalable machine-learning approach to recognize chemical names within large text databases.

Area of Science:

• Bioinformatics
• Computational Chemistry
• Natural Language Processing

Background:

• Scientific literature contains vast amounts of information on chemical compounds.
• Accurate identification of chemical names is essential for various informatics tasks, including database curation and information retrieval.

Purpose of the Study:

• To evaluate a first-order Markov Model (MM) for its effectiveness in recognizing chemical names within text.
• To assess the scalability of the MM for large-scale chemical name recognition.

Main Methods:

• A first-order Markov Model was developed and tested for distinguishing chemical terms from general words.
• The model's performance was evaluated on smaller test sets and then scaled to process 13.1 million MEDLINE records.

Main Results:

• The Markov Model achieved approximately 93% recall for chemical terms and 99% precision for non-chemical terms on smaller datasets.
• On a large-scale test of 13.1 million MEDLINE records, the model achieved an average precision of 82.7% for extracted chemical terms.
• The study found a correlation between the frequency of a chemical name and its spelling variants, impacting information retrieval.

Conclusions:

• The Markov Model is a scalable and effective method for chemical name recognition in large biomedical text corpora.
• Understanding chemical name variability is important for improving information retrieval and term mapping in databases like PubMed and Ovid.