s, in order to present a richer view of the existing literature.
The Ontogene system is based on a pipeline architecture (see gure 1), which includes, among others, modules for entity recognition and relation extraction. Some of the modules are rule-based (e.g. lexical lookup with variants) while others use machine-learning approaches (e.g. maximum entropy techniques). The initial step consists in the annotation of names of relevant domain entities in biomedical literature (currently the system considers proteins, genes, species, experimental methods, cell lines, chemicals, drugs and diseases). These names are sourced from reference databases and are associated with their unique identi ers in those databases, thus allowing resolution of synonyms and cross-linking among di erent resources.
One of the problems with sourcing resources from several databases is the possible inconsistencies among them. The fact that domain knowledge BioC XML input is scattered across dozens of data sources, occa# sionally also with some incompatibilities among (0) Wait request / validate BioC them, is a severe problem in the life sciences. Ide# ally these resources should be integrated in a sin(1) Read input with PyBioC reader gle repository, as some projects are attempting (2) Fetch Pubmed s#ource (optional) twoitdhoin (aen.gu.nOi peednPphlaatcftosrm[1.6H])o,waelvloewr,inagdqeeuperiynitneggration of the information provided by the scien# ti c literature and the content of the databases is (3) Convert to OGXML still missing.
# We train our system using the knowledge
pro(4) Sentence splitting + tokenization vided by life sciences databases as our gold
stan# dard, instead of hand-labeled corpora, since we
(5) Term annotation believe that the scope and size of manually
anno# tated corpora, however much e ort has been
in(6) Extract terms vested in creating them, is not su cient to capture
# the wide variety of linguistic phenomena that can
(7) Merge tokens be encountered in the full corpus of biomedical
lit# erature, let alone other types of documents, such
(8) Entity disambiguation as internal scienti c reports in the pharma
indus# try, which are not represented at all in annotated
(9) Compute concept relevance corpora. For example, PubMed currently contains
# more than 23 million records, while the entire
(10) Filter concepts by score set of all annotated publications probably barely
# reaches a few thousands, most of them sparsely
(11) Compute relation relevance annotated for very speci c purposes.
We generate interaction candidates using
co(12) Filter relati#ons by score occurence of entities within selected syntactic
units (typically sentences). An additional step of
# syntactic parsing using a state-of-the-art
depen(13) Annotate OGXML for visualization dency parser allows us to derive specialized
fea# tures in order to increase precision. The details of
(14) Add annotations to PyBioC writer the algorithm are presented in [
BioC XML output connects the two participating entities. All
availFig. 1. Schema of the OntoGene pipeline able lexical and syntactic information is used in
order to provide an optimized ranking for
candidate interactions. The ranking of relation
candidates is further optimized by a supervised machine learning method described in detail in [
Semi-Automated Semantic Annotation of the Biomedical Literature
The OntoGene annotator o ers an open architecture allowing for a considerable level of
customization so that it is possible to plug in in-house terminologies. We additional provide access to some of
our text mining services through a RESTful interface.2 Users can submit arbitrary documents to
the OntoGene mining service by embedding the text to be mined within a simple XML wrapper.
Both input and output of the system are de ned according to the BioC standard [
The OntoGene pipeline identi es all relevant entities mentioned in the paper, and their
interactions, and reports them back to the user as a ranked list, where the ranking criteria is the system's
own con dence for the speci c result. The con dence value is computed taking into account
several factors, including the relative frequency of the term in the article, its general frequency in
PubMed, the context in which the term is mentioned, and the syntactic con guration among two
interacting entities (for relationships). A detailed description of the factors that contribute to the
computation of the con dence score can be found in [
The user can choose to either inspect the results, using the ODIN web interface, or to have them delivered back via the RESTful web service in BioC XML format, for further local processing. ODIN (OntoGene Document Inspector) is a exible browser-based client application which interfaces with the OntoGene server. The curator can use the features provided by ODIN to visualize selected annotations, together with the statements from which they were derived, and, if necessary, add, remove or modify them. Once the curator has validated a set of candidate annotations, they can be exported, using a standard format (e.g. CSV, RDF), for further processing by other tools, or for inclusion in a reference database, after a suitable format conversion. In case of ambiguity, the curator is o ered the opportunity to correct the choices made by the system, at any of the di erent levels of processing: entity identi cation and disambiguation, organism selection, interaction candidates. The curator can access all the possible readings given by the system and select the most accurate.
As a way to verify the quality of the core text mining functionalities of the OntoGene
system, we have participated in a number of text mining evaluation campaigns [
Since internally the original database identi ers are used to represent the entities and interactions detected by the system, the annotations can be easily converted into a semantic web format, by using a reference URI for each domain entity, and using RDF statements to express interactions. While it is possible to access the automatically generated annotations for further processing by a reasoner or integrator tool, we strongly believe that at present a process of semi-automated validation is preferable and would lead to better data consistency.
Acknowledgments. The OntoGene group is partially supported by the Swiss National Science Foundation (grant 105315 130558=1 to Fabio Rinaldi) and by the Data Science Group at Ho mann-La Roche, Basel, Switzerland. 2 http://www.ontogene.org/webservices/