This document summarizes an approach to automatically extract microorganisms and their habitats from free text using text mining workflows. The approach uses a named entity recognizer combining dictionaries and machine learning to identify organisms and habitats. It then employs relation mining to extract sentences expressing relationships between organisms and habitats. Evaluation shows the organism recognition achieves 84% precision but habitat recognition is lower at 68% precision, limiting the relation mining performance. Ongoing work aims to improve the habitat model and make the overall approach more robust to noise from PDF to text conversion.

1. Automa'c
extrac'on
of
microorganisms
and
their
habitats
from
free
text
using
text-­‐mining
workflows
BalaKrishna
Kolluru,
Sirintra
Nakjang,
Robert.
P.
Hirt,
Anil
Wipat
and
Sophia
Ananiadou

2. Outline
of
the
talk
• Mo'va'on
• Experiments
• Results
&
inferences
• Discussion
• Current
work

3. Mo'va'on
• In
the
study
of
symbio'c
rela'onships,
host-­‐
microbe
interac'ons
play
an
important
role
• To
date,
there
is
no
comprehensive
database
regarding
microbe—habitat
rela'on,
but
there
is
an
explosion
in
the
numbers
of
taxa
• With
this,
there
is
an
urgent
need
for
automated
host-­‐microbe
rela'on
extrac'on

4. Experiments:
relevant
work
• Iden'fica'on
of
named
en''es
such
as
microorganisms,
diseases,
genes
etc.,
has
received
sufficient
importance
from
the
scien'fic
community
at
large
[Sasaki,
Hanisch,
Chikashi]
• Researchers
have
also
used
ontology
based
approaches
to
iden'fy
concepts
such
as
public
health
rumors
etc
[Biocaster]

5. Experiments:
our
approach
Named
en'ty
recogni'on
• Free
text
• Habitats
&
ar'cles
organisms
• pdf
Text
Rela'on
processing
mining
Employ
text
mining
workflows
consis'ng
of
•
text/pdf
processor
•
Named
en'ty
recognizer
to
iden'fy
microorganisms
and
their
habitats
•
Rela'on
mining
component
to
extract
sentences
which
express
this
rela'on

6. Experiments:
our
approach
• The
named
en'ty
recognizer
used
a
hybrid
dic'onary-­‐machine
learning
based
approach
– It
combined
the
informa'on
dic'onaries
with
a
feature
set
for
a
condi'onal
random
field
(CRF)
based
classifier
[Mallet]
– The
CRFs
used
a
linear
chain
model
and
were
trained
on
a
corpus
consis'ng
of
32
full
papers

7. Experiments:
our
approach
– The
feature
set
included
• lexical
informa'on
of
the
word
e.g.,
word,
POS
tag
etc
• Orthographic
informa'on
e.g.
any
uppercase
le^ers,
numbers
• Contextual
informa'on;
informa'on
about
two
word
preceding
and
succeeding
the
word
• For
the
rela'on
mining
component,
a
linear
chain
CRF
was
trained
using
– Occurrence
of
organisms
and
habitats
– Contextual
informa'on
of
all
the
en''es
in
a
sentence

8. Results
and
inference
Performance
of
our
named
en'ty
recognizer
on
a
9-­‐fold
cross-­‐valida'on
Class
of
Precision(%)
Recall(%)
F-­‐score(%)
en**es
2PR/(P+R)
Organisms
84
79
81
Habitats
68
55
61
improved
results
from
the
'me
of
submission
•
Microorganisms
have
been
recognized
quite
well.
•
Habitat
recogni'on
is
modest
•
One
of
the
observa'ons
is
that
in
a
free
text,
the
descrip'on
of
habitats/host
is
devoid
any
salient
features
such
as
uppercase
le^ers,
hyphens
etc.
•
Instances
such
as
abscess,
lung
were
typical
misses

9. Results
and
inference
Rela'on
mining
results
• For
the
rela'on
mining
experiment,
the
CRF-­‐based
classifier
achieved
a
precision
of
~
80%
• Most
of
the
false
nega'ves
(
sentences
which
should
have
been
picked
up,
but
were
not)
due
to
the
noise
in
pdf
to
text
conversion
• Another
reason
for
false
nega'ves
is
the
modest
performance
of
habitat
recogni'on
which
affected
the
rela'on
mining
algorithm

10. Discussion
• The
workflows
we
have
developed
bring
together
pdf-­‐conversion,
machine
learning
and
dic'onaries
together
– Performance
of
individual
components
obviously
has
an
impact
its
overall
performance
– Pdf
conversion
is
not
trivial
by
any
means
and
this
component
is
the
most
limi'ng
factor
for
any
sentence-­‐based
classifica'on
task

11. Discussion
• Pdf-­‐to-­‐text
sentence
examples

These
mechanisms
may
have
evolved
in
bacterial
pathogens
to
increase
the
frequency
of
phenotypic
varia'on
in
genes
involved
in
1
100,000
200,000
300,000
1,600,00
Figure
2
Circular
representa'on
of
the
H.
pylori
26695
chromosome.
[Clearly,
data
from
a
table
and
figure
corrupted
the
sentence]

airborne
pigs
[noisy
conversion
of
table
discussing
airborne
diseases
in
pigs
]

12. Discussion
• The
CRF
model
for
habitats
is
evidently
weak
– There
is
a
need
to
augment
the
features
to
alleviate
this
weakness.
We
are
currently
enhancing
model
to
include
more
features
such
as
character-­‐level
n-­‐grams
–
Results
reflect
ini'al
success
• Rela'on
mining
is
a
hyper-­‐classifica'on
task
and
perhaps
it
is
prone
to
cascading
errors

13. Current
work
• Work
is
underway
to
improve
the
rela'on
mining
component
using
bag-­‐of-­‐words
and
character
level
n-­‐grams
to
augment
the
feature
space
• We
are
also
working
on
less
noisy
conversion
techniques
for
pdf-­‐to-­‐text
• Export
the
workflows
to
the
public
domain
so
that
scien'sts
across
the
spectrum
can
use
our
workflows

14. Snapshot
of
rela'on
miner
References
•
Hanisch,
D.
et
al.
ProMiner:
Organism
specific
protein
name
detec'on
using
approximate
string
matching.
Embo
Workshop
Granada,
Spain,
2004
• Sasaki,
Y.
et
al.
(2008).
How
to
make
the
most
of
NE
dic'onaries
in
sta's'cal
NER?
In:
BMC
Bioinforma'cs,
9(Suppl
11),
S5
•
Collier,
N.
et
al.
BioCaster:
detec'ng
public
health
rumors
with
a
Web-­‐based
text
mining
system.
Bioinforma'cs,
24(24),
2008.
•
Nobata,
C.
et
al
Mining
Metabolites:
Extrac'ng
the
Yeast
Metabolome
from
the
Literature.
Metabolomics,
2010.