Tiago Grego

Chemical Entity Recognition and Resolution to ChEBI

Chemical entities are ubiquitous through the biomedical literature and the development of text-mining systems that can efficiently identify those entities are required. Due to the lack of available corpora and data resources, the community has focused its efforts in the development of gene and protein named entity recognition systems, but with the release of ChEBI and the availability of an annotated corpus, this task can be addressed. We developed a machine-learning-based method for chemical entity recognition and a lexical-similarity-based method for chemical entity resolution and compared them with Whatizit, a popular-dictionary-based method. Our methods outperformed the dictionary-based method in all tasks, yielding an improvement in F-measure of 20% for the entity recognition task, 2–5% for the entity-resolution task, and 15% for combined entity recognition and resolution tasks.

Identification of Chemical Entities in Patent Documents

Biomedical literature is an important source of information for chemical compounds. However, different representations and nomenclatures for chemical entities exist, which makes the reference of chemical entities ambiguous. Many systems already exist for gene and protein entity recognition, however very few exist for chemical entities. The main reason for this is the lack of corpus to train named entity recognition systems and perform evaluation. In this paper we present a chemical entity recognizer that uses a machine learning approach based on conditional random fields (CRF) and compare the performance with dictionary-based approaches using several terminological resources. For the training and evaluation, a gold standard of manually curated patent documents was used. While the dictionary-based systems perform well in partial identification of chemical entities, the machine learning approach performs better (10% increase in F-score in comparison to the best dictionary-based system) when identifying complete entities.

Identifying bioentity recognition errors of rule-based text-mining systems

An important research topic in Bioinformatics involves the exploration of vast amounts of biological and biomedical scientific literature (BioLiterature). Over the last few decades, text-mining systems have exploited this BioLiterature to reduce the time spent by researchers in its analysis. However, state-of-the-art approaches are still far from reaching performance levels acceptable by curators, and below the performance obtained in other domains, such as personal name recognition or news text. To achieve high levels of performance, it is essential that text mining tools effectively recognize bioentities present in BioLiterature. This paper presents FIBRE (Filtering Bioentity Recognition Errors), a system for automatically filtering mis annotations generated by rule-based systems that automatically recognize bioentities in BioLiterature. FIBRE aims at using different sets of automatically generated annotations to identify the main features that characterize an annotation of being of a certain type. These features are then used to filter mis annotations using a confidence threshold. The assessment of FIBRE was performed on a set of more than 17,000 documents, previously annotated by Text Detective, a state-of-the-art rule-based name bioentity recognition system. Curators evaluated the gene annotations given by Text Detective that FIBRE classified as non-gene annotations, and we found that FIBRE was able to filter with a precision above 92% more than 600 mis annotations, requiring minimal human effort, which demonstrates the effectiveness of FIBRE in a realistic scenario.