Event-Based Information Extraction for the Biomedical Domain: The Caderige Project

This paper gives an overview of the Ca erige project. This project involves teams from ifferent areas (biology, machine learning, natural language processing) in or er to evelop highlevel analysis tools for extracting structure information from biological bibliographical atabases, especially Me line. The paper gives an overview of the approach an compares it to the state of the art. Introd uction Developments in biology an biome icine are reporte in large bibliographical atabases either focuse on a specific species (e.g. Flybase, specialize on Drosophilia Menogaster) or not (e.g. Medline). This type of information sources is crucial for biologists but there is a lack of tools to explore them and extract relevant information. While recent named entity recognition tools have gained a certain success on these domains, event-based Information Extraction (IE) is still a challenge. The Caderige project aims at designing and integrating Natural Language Processing (NLP) and Machine Learning (ML) techniques to explore, analyze and extract targeted information in biological textual databases. We promote a corpus-based approach focusing on text pre-analysis and normalization: it is intended to drain out the linguistic variation dimension, as most as possible. Actually, the MUC (1995) conferences have demonstrated that extraction is more efficient when performed on normalized texts. The extraction patterns are thus easier to acquire or learn, more abstract and easier to maintain Beyond extraction patterns, it is also possible to acquire from the corpus, via ML methods, a part of the knowledge necessary for text normalization as shown here. This paper gives an overview of current research activities and achievements of the Caderige project. The paper first presents our ap..