We present OCMiner, a high-performance text processing system for large document collections of scienti c publications. Several linguistic options allow adjusting the quality of annotation results which can be specialized and ne-tuned for the recognition of Life Science terms. Recognized terms are mapped to semantic concepts which are ontologically located within their respective domain taxonomies. Relying on a correct identi cation and semantic interpretation of mentions of domain concepts, relations between entities are extracted. The annotated text, as well as extracted knowledge triples, can be visualized on a web-based front-end at http://www.ocminer.com/, permitting an explorative information retrieval.
Life Science knowledge mining methods rely on a correct annotation of terms and phrases with concepts from di erent knowledge domains { in particular chemistry, proteins and diseases { followed by the application of suitable semantic relation extraction algorithms. We present the implementation of a high quality context sensitive annotation of named entities in text documents that makes use of an exchangeable set of chemistry, protein and disease ontologies.
A particular challenge in recognizing Life Science terms in free text is chemical named entity recognition. The di culty of correctly annotating chemical terms lies in the large number of chemical terms and chemicals as well as in the great variability of chemical expressions: There are trivial and systematic names for chemical compounds and classes, as well as formulas and trade names for drugs. Chemical names can be extremely long and may contain variations of meaningful punctuation symbols and parentheses. Moreover, di erent chemistry name types can even be mixed within one chemical expression. Similarly, the recognition of protein terms in texts and the correct mapping to protein concepts is a non-trivial issue. Protein terms are often abbreviated and appear in various spelling variants (with or without hyphens, spaces etc., e.g. FLT1, FLT-1, FLT 1) and may be confused with other terms (e.g. ASK protein). Likewise, frequent disease terms are often homonymous to other concepts, e.g. a \ ash" might be a physiological circumstance only in certain contexts. In sum, the precise identi cation of named entities is an important prerequisite for the extraction of correct and relevant relations between annotated concepts, e.g. metabolic pathways or relations between chemicals and diseases. 2
OCMiner is a modular processing pipeline for unstructured information based on the Apache UIMA framework. The system architecture is depicted in Fig. 1.
XML detagger language detector normalizer toke&nizer Annotation pipeline
Dictionary annotation chemistry dictionary proteins dictionary diseases dictiootnhaerry domain dictionaries Ontological domain knowledge chemistry proteins diseases anatomy celllines ...
Pre-dictionary modules aanbnborteavto.r dsotrcuucmtuernet
Post-dictionary modules
Chemistry-specific modules sntarumcetu-2re- crelacsosg/gnritoiounp afnonrmotualtaor cleanup annotations coordinated entity resolution
NE rule combiner phrase tagger
Relation extraction modules mrealatctihoenr rfleitleartiinogn
index Search and look-up search back-end web front-end web server annotated
XML documents extracted data formats web service client browser input via file system or web service picture PDF full text PDF full text XML other formats
OCR PDF reader XML reader custom reader Documents are read from a variety of sources (text and picture PDF, XML, etc.) and standardized for further analysis. Then, preparatory processes such as language detection, sentence splitting, tokenization, document structuring, etc. take place. As the core of the annotation process, we have a dictionary-based named entity recognition module which uses a high performance dictionary lookup technology with support for very large dictionaries (our chemical dictionary has about 34 million entries). It implements speci c language and dictionary dependent treatment options, e.g. spelling variations, spaces/hyphens, diacritics, Greek letters, plural forms. This context-sensitive ne-tuning is especially important in the annotation of protein and chemistry terms.
Importantly, recognized terms are semantically interpreted as mentions of
concepts that are ontologically located within domain-speci c taxonomies. Our
dictionaries are generated from ne-grained domain ontologies in form of
conceptual taxonomies. This semantic mapping provides the basis for ontological
search methods and knowledge extraction technologies. Particular importance
is given to the chemical dictionary. It is generated from a compound database
built from various publicly available sources such as PubChem, MeSH,
DrugBank, ChEMBL, among others. Our system is able to automatically arrange
compounds into a single chemical ontology according to their structure or their
functional properties [
Additional components handle speci c scenarios. For instance, the abbreviation annotator nds expansions of acronyms and abbreviated terms. Another module recognizes expressions like \vitamin A and B" as a coordinated entity and annotates \vitamin A" as such and \B" as \vitamin B". A chemistry-speci c module tries to recognize whether a given chemical expression refers to a speci c compound, a compound class, or a substituent group/fragment.
A processing step which serves as a prerequisite for relation extraction is the combination of annotated concepts to complex entities. Thus, post-dictionary modules combine sequences of named entities. For instance, the text phrase \human raf kinase inhibitor", initially annotated as a sequence of named entities [organism human] [protein raf] [protein kinase] [chemistry inhibitor], is combined to a single { though internally complex { entity referring to a chemical compound class: [chemistry[protein[organism human] raf kinase] inhibitor]. This is especially useful for the recognition of higher-level combined entities made up of constituents of the domains chemistry, proteins, species, anatomy and celllines.
Our system applies a shallow pattern-based approach to the extraction of relations between annotated concepts from di erent domains, e.g. between chemicals and diseases, or metabolic pathways, physico-chemical properties of compounds, etc. First, the annotated input text is tokenized into phrase tokens, where named entities, including higher-level combined entities, constitute single tokens. Note that the system does not rely on part-of-speech tagging, parsing or other sophisticated but time-consuming natural language processing techniques. Instead, extraction rules work on phrase tokens and take speci c attributes of involved named entities into account, such as the type of a chemical entity. A dedicated relation ontology de nes a taxonomy of relations to be extracted. Examples for relation concepts are \[compound] treats [disease]" or \[compound] metabolizes to [compound]". For each relation concept, speci c mappings from natural language syntax patterns to semantic normalizations are de ned. Pattern de nitions have a syntax and a complexity similar to regular expressions, allowing for nested grouping and variable order of tokens. The relation matcher module matches the tokenized input text against these rules and generates normalized relation representations in form of triples of concept identi ers: hentity1, relation concept, entity2i.
Processed documents and extracted information can be stored in various ways. First, XML documents with inline annotations of recognized concept mentions can be generated out of heterogeneous input documents. Second, extracted knowledge such as keyword lists and relations between mentioned concepts (i.e. knowledge triples) can be stored in various formats (custom XML formats, RDF triples, SBML, CML, etc.) and accessed as a web service. Third, annotated entities and relation triples can be stored in an index (triple store or Lucene index), which is used for a web-based retrieval and visualization of the data. In particular, the index is accessed by a search back-end, which provides the indexed data to the web front-end. The OCMiner front-end displays annotated documents and provides a user interface for the navigation along relation chains, e.g. from chemicals over proteins to diseases, permitting an explorative information retrieval on multitudes of scienti c publications (Fig. 2).
The system was evaluated as part of the BioCreative IV challenge. In the
CHEMDNER task for evaluating chemical NER, we obtained a precision of 85% at a recall
of 71% (F-score 78%) [