Arbeitsgebiet "Textmining"

Textmining plays an important role in bioinformatics, given that a significant part of biological knowledge is available only as free text. For extracting this knowledge, it is required to recognize biological objects in texts, this is especially true for genes and proteins.

• Information extraction
• Generation of networks from texts for advanced analyses
• Analysing information derived from Text-Mining together with data from other sources (e.g. gene expression data)

Named Entity Recognition of objects in biomedical texts

The scientific literature ist the most comprehensive source for information on molecular biology objects and their interactions. In free texts the biological objects (genes, proteins, diseases, cells, organisms) are unfortunately referenced by rather unsystematic and partially ambiguous terms. Automatic text mining procedures are required to identify the objects (Named Entity Recognition, NER) and to extract their interactions. Various spelling variants and synonyms need to be recognized in order to assign unique identifiers to ambiguous free text names. Frequently, texts are matched against so-called synonym lists. Such lists for genes and proteins of higher organisms (e.g. humans) comprise 30k-50k objects per organism that can be referenced via up to 10-15 different synonyms. The size of the data is huge, e.g. multiple synonym lists could be simultaneously searched of a total of 170k objects with 1.6M synonyms against 19M PubMed abstracts (66Gb XML/26 Gb text). On the other hand, the results of the searches are highly sensitive to the quality and curation of the used synonym lists. We developed the tool syngrep that is more memory efficient and more than 100 times faster than well known Unix-tools such as fgrep. High quality matches of synonyms in the biological literature has important applications, for instance the assignment of annotation to the identified objects and augmentation of knowledge bases with literature derived facts.

Extraktion of relationships between objects from the biomedical literature

Two approaches for relation extraction are currently employed. Methods of machine learning assign relations to particular sentences after training on suitable manually annotated training corpora. Linguistic methods, on the other hand, analyze the grammatical structure of sentences and recognize relations from partial sentences by appropriate rules. Usually, very high numbers of rules have to be defined to extract relations with sufficient precision and recall. The RelEx approach uses the Stanford Lexicalized Parser to construct semantic dependency trees from the sentences. Due to the dependency trees only few tree rules are required to harvest a wide range of complex relationship phrases. 2007 Papers Katrin Fundel, R. Küffner, Ralf Zimmer. RelEx - Relation extraction using dependency parse trees. Bioinformatics, vol 23, no. 3, pp. 365-371, 2007. DOI BibTex You can find the manually curated test set of relations from the HPRD here