| 
What is the expected
gain, if we process scientific text automatically?
Scientific publications form the major tool for the exchange
of new scientific facts. Therefore, scientists have to process
large amounts of publications, without being able to read
everything. This explains, why automatic methods for the analysis
of scientific text become very popular. If an automatic method
could extract the details needed with good quality, the scientist
would safe lot of time reading. If an automatic method could
retrieve and pre-select at 100% certainty the relevant documents
from the complete set of documents available, then the scientist
would safe lot of time searching for the best publications,
but he still would have to read them to find the facts.
Text mining methods have been developed
and applied in all fields, where large amounts of data were
available: analysis of large archives of news documents, internet
documents and in recent years scientific publication in the
biomedical field. The strong increase in publications in the
biomedical field is the consequence of the strong increase
of information gathered throughout numerous sequencing projects.
What is text
mining (TM)?
The term ‘text
mining’ refers to a variety of techniques which fulfill
different tasks mainly:
- Information retrieval (IR): Such automatic methods
extract the features of a document (words, parts of sentences),
classify them and then compare them to a profile, which consists
in general of a list of terms. If the document features match
to the profile, then the document is selected.
- Information extraction (IE): Such automatic methods
seek written evidence of facts in the text, which are in general
represented by combinations of words (patterns). Such patterns
select a sequence of words and take their morphological changes
into consideration. Other techniques include grammar rules
to ‘understand’ the sentence (Natural Language
Processing (NLP)) or complex classification techniques, e.g.
Support Vector Machines (SVM) or Bayesian Networks, to identify
patterns of high complexity.
Other approaches use similar techniques to archive documents
for better retrieval or to answer questions on the basis of
existing documents.
What data
is used for TM?
Scientific abstracts from Medline (PubMed) form the standard
data resource. In addition scientific groups analyze complete
scientific publications. Patent text from the European Patent
Office is an additional source.
The importance of correct terminology
The use of correct terminology is necessary for the identification
of the object behind the described fact and for the correct
link between the literature and scientific databases. Unfortunately
natural language is often ambiguous in the use of terminology.
The biological community is addressing these issues with work
on nomenclature and ontologies. The following ambiguities
are most prominent.
- Syntactic ambiguity occurs when a word has a noun or a verb
interpretation.
- Semantic ambiguity occurs when two identical terms have
different meanings. Disambiguation has to take the context
of the term into consideration.
Which advantages
result from natural language processing (NLP) ?
A variety of different
systems is available to do natural language processing, but
no system can cope with the complexity of natural language
in a spoken language. In addition, the time for the analysis
of a sentence increases with its complexity. As a conclusion,
NLP based systems provide solutions on chunks of sentences.
The main advantage in NLP based approaches can be found in
the fact that sentence analysis can be shaped following a
hand-crafted linguistical model and thus they do not rely
on the correct annotation of a training data set.
Which TM
tasks are important in bioinformatics ?
The most prominent
information extraction task is the extraction of protein-protein
interactions from Medline (numerous publications). Other tasks
are concerned with the association of a gene to a phenotype,
e.g. a disease or a symptome, or the identification of a gene
in conjunction with a mutation. In addition, research groups
are working on the correct identification of protein and gene
names from Medline, which is again a non-trivial task due
to the complexity of the applied names.
Evaluation of extracted results
The evaluation quantifies how many correct facts were selected
from all facts available (= recall) and how many correct facts
are contained in all selected facts (= precision). The sweat
part of this task is to generate a random sample from the
complete set of documents and to count how many facts are
contained in the sample. Other evaluation methods compare
the extracted information by the systems with independent
information stored in databases and large experimental repositories.
by Dr. Dietrich Rebholz-Schuhmann
(February, 2004).
|
