Transcript Mining External Resources for Biomedical IE

BioLINK Talks
Linking Literature, Information
and Knowledge for Biology
BioLINK,Detroit, June 24 (Edinburgh July 11)
• Corpora and Corpus design (2)
• NER and Term Normalisation (3)
• Annotation and Zoning (2)
• Relation Extraction (2)
• Other
Corpora and corpus design
Corpus Design for Biomedical
Natural Language Processing
K. Bretonnel Cohen et al (U of Colorado)
Main Question: why are some (bio-)corpora more used than
others? What makes them attractive?
Crucial points:
• format: XML
• code several layers of information
• publicity: write specific papers about corpus, publicise
its availability
Take home message: if you want people to use your corpus, use
XML, publish annotation guidelines, publicise corpus with
dedicated papers, use it for competitions
Corpora and corpus design
MedTag: a collection of
biomedical annotations
L. Smith et al. (National Center for Biotechnology
Information, Bethesda, Maryland)
Main Point: MedTag is a database that combines three corpora:
• MedPost (modified to include 1000 extra sentences)
• ABGene
• GENETAG (modified to reflect new defs of genes and prots)
The data is available in flat files + software to facilitate loading
data into SQL database
Take home message: integrated data, more accessible, you
should try it.
Corpora and corpus design
• MedPost
• 6700 sentences
• annotated for POS and gerund arguments
• POS tagger trained on it (97.4% accuracy)
• GENETAG
• 15000 sentences currently released
• tagged for gene/protein identification
• used in Biocreative
• ABGene
• over 4000 sentences
• annotated for gene/protein names
• NER tagger trained on it (lower 70s)
Corpora and corpus design
GOOD
Recommended Uses
• training and evaluating
POS taggers
• training and evaluating
NER taggers
• developing and evaluating
a chunker (for PubMed
phrase indexing)
• analysis of grammatical
usage in medical text
• feature extraction for ML
• entity annotation guidelines
BAD
• tokenisation!
(white spaces were deleted)
NER and TN
Weakly Supervised Learning
Methods for Improving the Quality
of Gene Name Normalization
Ben Wellner (MITRE)
Main points
1. presenting method of improving quality of training data
from BioCreative task1b. System’s performance on
improved data is better than on original data
2. weakly supervised methods can be successfully applied
for re-labeling noisy training data
(next week)
NER and TN
Unsupervised gene/protein
normalization using automatically
extracted dictionaries
A. Cohen (Oregon Health & Science U., Portland, Oregon)
Main point: dictionary-based gene and protein NER and
normalisation system; no supervised training; no human
intervention.
•
•
•
•
what curated databases are the best collections of names?
are simple rules sufficient for generating ortographic variants?
can common English words be used to decrease false positives?
what is the normalization performance of a dictionary-based
approach?
Results: near state-of-the-art; saving on annotation
METHOD
NER and TN
1. Building the dictionary
Automatically extracted from 5 databases: official symbol,
Unique identifiers, name, symbol, synonym, alias fields
2. Generating orthographic variants
Set of 7 simple rules applied iteratively
3. Separating common English words
Dictionary split in two parts: confusion and main dictionary
4. Screening out most common English words
5. Searching the text
6. Disambiguation
Note: 5% ambiguous intra-species; 85% across species.
Exploit non-ambiguous synonyms; exploit context
NER and TN
A machine learning approach to
acronym generation
Tsuruoka et al (Tokyo (Tsujii group), Japan and Salford, UK)
Task: system generates possible acronyms from a
given expanded form
Main point: acronym generation as sequence tagging problem
Method: ML approach (MaxEnt Markov Model)
Experiments:
- 1901 definition/acronym pairs
- several ranked options as output
- 75.4% coverage when including top 5 candidates
- baseline: take first letters and capitalise them
NER and TN
Classes (tags)
1.
2.
3.
4.
5.
SKIP (generator skips the letter)
UPPER (generator upper-cases letter)
LOWER (generator lower-cases letter)
SPACE (generator converts letter into space)
HYPHEN (generator converts letter into hyphen)
Features
-
letter unigram
letter bigram
letter trigam
action history (preceding action)
orthographic (uppercase or not)
length (#words in definition)
letter sequence
distance (between target letter and beginning/tail of word)
Annotation/Zoning
Searching for High-Utility
Text in the Biomedical Lit.
Shatkay et al. (Queens,Ontario and NYU and NCBI,Maryland)
Task: identify text regions that are rich in scientific content,
and retrieve docs that have many such regions
(Main idea + annotation guidelines)
High Utility Regions = regions in the text that we identify as focusing
on scientific findings, stated with a high confidence, and preferably
supported by experimental evidence.
Annotation/Zoning
[assertion = sentence or fragment]
Focus = type of information conveyed by assertion
- scientific
K=.83 - generic
- methodology
Polarity of assertion (positive/negative)
K=.81
Certainty
- complete uncertainty (0)
K=.70 - complete certainty (3)
Evidence =
- E0
K=.73 - E1
- E2
- E3
whether assertion is supported by exp evidence
= lack of evidence
= evidence exists but not reported (“it was shown..”)
= evidence not given directly but reference provided
= evidence provided
Direction/Trend = whether assertion reports increase/decrease
in specific phenomenon
K=.81
Annotation/Zoning
Automatic Highlighting of
Bioscience Literature
H. Wang et al (CS Department, University of Iowa - M. Light group)
Task: automatic highlighting of relevant passages
Approach: IR task
- sentence is passage unit
- each sentence treated as document
- user provides a query
- query box for keywords
- example passage highlighting
- system ranks sentences as to relevance to query
(* query expansion system is web-based)
Annotation/Zoning
- Corpus: 13 journal articles each highlighted by a bio
graduate student before the request for annotation
- Queries: constructed in retrospect. The annotators created
the queries for the articles they had selected.
The first highlighted region also used as query
- Processing: tokenisation (LingPipe), indexing (Zettair), ranking
of retrieved sentences (Zettair)
- Query Expansion: definitions were used. Google “define”
for each word (excluding stopwords).
Over 80% of query words had Google defs.
• poor results
• first highlighted passage works better than keywords
• Google expansion helps
Rel Extr
Using biomedical literature
mining to consolidate the set of
known human PPIs
A. Ramani et al (U of Texas at Austin - Bunescu/Mooney group)
Task: construct a database of known human PPIs by:
- combining and linking interactions from existing DBs
- mine additional interactions from 750000 Medline abs
Results:
- quality of automatically extracted interactions comparable
to that of those extracted manually
- overall network of 31609 interactions between 7748 prots
Rel Extr
1. Identify proteins in text: CRF tagger
2. Filter out less confident entities
3. Try to detect which pairs of remaining ones are interactions
- use co-citation analysis
- train model on existing set
Trained model: a sentence containing 2 protein names is classified
as correct/wrong. If a sentence has n prots (n ≥ 2), the
sentence is replicated n times
- ELCS = Extraction w Longest Common Subsequences (learned rules)
- ERK = Extraction using a Relation Kernel
Rel Extr
IntEx: A syntactic role driven PPI
extractor for biomedical text
S. Ahmed et al (Arizona State University)
Task: detect PPIs by reducing complex sentences to simple
clauses and then exploiting syntactic relations
-
pronoun resolution (third person and reflexives; simple heuristics)
entity tagging (dictionary lookup + heuristics)
parsing (Link Grammar, dependency based, CMU?)
complex sentence splitting (verb-based approach to extract
simple clauses)
- interaction extraction (from simple clauses exploiting syntactic
roles)