Transcript Mining External Resources for Biomedical IE

Mining External Resources
for Biomedical IE
Why, How, What
Malvina Nissim
[email protected]
Why
• goal: Named Entity Recognition
• method: supervised learning
• feature extraction
• (text) internal features: word shape, n-grams, ...
protein-indicative features:
- of shape a0a0a0a…
- followed by /bind/
- shorter than 5 characters
• generalisations on training data might be incomplete
• acquired evidence might be absent in test instance
Getting Additional Evidence
internal features might be insufficient, but
good evidence might be somewhere else...
• small and accurate lists of proteins (gazetteers)
• use as rules
• use as features
• other texts might contain indicative n-grams
• how to use other texts
• which texts to use
Note: some systems (MaxEnt for instance) can easily and
successfully integrate a huge number of features
How
patterns
“X gene/protein/DNA”
“X sequence/motif”
A. Create patterns (aim, method, input)
B. Search corpus for patterns and obtain counts
C. Use counts as appropriate
Create Patterns (I)
1. AIM (granularity)
distinguish entities from non-entities
“X gene OR DNA OR protein”
+ bypass ambiguities and data sparseness
– less information
distinguish between entities
“X gene”
“X DNA”
“X binds”
+ more information
– ambiguities, data sparseness
1. AIM
2. METHOD
3. INPUT
Create Patterns (II)
1. AIM
2. METHOD
3. INPUT
2. METHOD
by hand (experts)
+ high precision, exact target
– time consuming, experts needed
automatically (collocations, clustering)
+ no human intervention
– lower precision, not necessarily interesting
patterns
Create Patterns (III)
3. INPUT
1. AIM
2. METHOD
3. INPUT
(“X gene”)
low frequency words (as estimated from a non-specific corpus)
words not found in standard dictionary
NP chunks
first output of classifier
increase precision but lower recall
all features
– web
prec
.813
.807
rec
.861
.864
f-score
.836
.835
What? Google vs PubMed
• PubMed: searchable collection of over 12M biomedical
abstracts, more sophisticated search options
• Everything: Google searches over 8 billion pages, raw search, API
“p53 gene”
PubMed
5,843 documents
Google
~165,000 pages
Google + PubMed
“anything you want” site:<specific_site>
“p53 gene” site:www.ncbi.nlm.nih.gov
Rob Futrelle has this function available on this webpage:
http://www.ccs.neu.edu/home/futrelle/bionlp/search.html
• comment: sometimes PubMed reports
“Quoted phrase not found” even when
Google finds the phrase.
PubMed provides phrase search only on pre-indexed phrases
PubMed > Google
• query expansion
PubMed uses the MeSH headings to match synonyms
(it will expand “Pol II” to search for “DNA Polymerase II”)
Google will only try correct misspelling
• field specific search
PubMed allows field-specific searches (eg year)
Google cannot refine its search in this respect
• timeliness
PubMed is updated daily
Google is slow in updating
PubMed > Google
(cont’d)
• ranking
Google does a ‘vote’-based ranking: not necessarily good
PubMed does not do any ranking (possibly bad too...)
• truncation and flexibility
PubMed accepts truncated entries and will look for all possible
Variations. It will try break phrases if no matches are found.
Google has a rigid search
• manual indexing
PubMed’s MeSH contain keywords not necessarily contained
in the abstract
Google cannot find something that is not mentioned
in the abstract
What
What to
to Use?
Use?
(or
(or How
How to
to Use
Use the
the Evidence)
Evidence)
• as a rule
+ sure identification of entities
– too powerful -> high risk of false positives
might be better to use PubMed: less info but precise
• as a feature
+ less false positives
+ some systems (MaxEnt) can integrate huge number of features
– might still not get used or provide enough evidence
might be OK to use Google: more info but not necessarily precise
iHOP
Nature Genetics, Vol. 36(7), July 2004
(Information Hyperlinked Over Proteins)
A gene network for navigating the literature
http://www.pdg.cnb.uam.es/UniPub/iHOP
• uses genes and proteins as hyperlinks between sentences
and abstracts http://www.pdg.cnb.uam.es/UniPub/iHOP
• each step through the network produces information about
one single gene and its interactions
• information retrieved by connecting similar concepts
• precision of gene name and synonym identification: 87-99%
• readers can still check correctness of sentences when they are
presented to them
• shortest path between any 2 genes is on average 4 steps only