Extracting Biological Keywords from Scientific Text

1
Extracting Biological Keywords from Scientific
Text

• Wen-Juan Hou and Hsin-Hsi Chen (2004). Enhancing
  Performance of Protein and Gene Name Recognizers
  Using Filtering and Integration Strategies.
  Journal of Biomedical Informatics, 37(6),
  December 2004, 448-460.

2
Mining Protein Collocates

• collocate
• keywords/key phrases that co-occur with protein
  names

Tag the raw material

• Preprocessing
• 1. Remove stop words
• . Stem

Calculate collocation value
Extract significant biological keywords
3
Step 1 Tagging the Corpus

• to calculate the collocation values of words with
  proteins from a corpus ? recognize protein
  names
• to improve the performance of protein name
  tagging ? application of protein collocates
• Preparing a protein name tagged corpus and
  developing a high performance protein name tagger
  ? a chicken-egg problem
• Dictionary-based approach ? a partial-tagged
  corpus

4
Step 1 Tagging the Corpus (continued)

• ltNAME TYPE"PROTEIN"gt Chloroperoxidase lt/NAMEgt
  (CPO) is a versatile heme-containing enzyme that
  exhibits ltNAME TYPE"PROTEIN"gt peroxidase lt/NAMEgt
  , ltNAME TYPE"PROTEIN"gt catalase lt/NAMEgt and
  ltNAME TYPE"PROTEIN"gt cytochrome P450 lt/NAMEgt
  -like activities in addition to catalyzing
  halogenation reactions.

5
Step 2 Preprocessing

• Step 2.1 Exclusion of Stopwords
• the stoplists (Fox, 1992) was adopted, but the
  words also appearing in the protein lexicon were
  removed.
• capsid of the lumazine ? of is excluded
• 387 stopwords were used
• Step 2.2 Stemming
• Stemming can group the same word semantics and
  reflect more information around the proteins.
• inhibited and inhibition will be mapped into
  the root form inhibit after stemming.

6
Step 3 Computing Collocation Statistics
Frequency

• define a collocation window to be five words on
  each side of protein names
• capture collocation bigrams at a distance
• both bind and signal are good collocates with
  proteins, but the frequencies of bind and
  signal are 365 and 9, respectively
• A low threshold strategy will keep both of these
  two words, but many false candidates will pass
  the threshold at the same time.

7
Step 3 Computing Collocation Statistics Mean
and Variance

• is the average distance for word i in the
  collocation windows
• is the distance of the j-th occurrence of
  word i away from proteins in the collocation
  windows
• is the total occurrences of
  word I
• is the standard deviation of

8
Step 3 Computing Collocation Statistics Mean
and Variance (continued)

• standard deviation of value zero
• the collocates and the protein names always occur
  at exactly the same distance equal to the mean
  value
• low standard deviation
• two words usually occur at about the same
  distance, i.e., near the mean value
• high standard deviation
• the collocates and the protein names occur at
  random
• phosphoinositide owns zero standard deviation,
  but it only occurs twice in the corpus

9
Step 3 Computing Collocation Statistics t-test
Model

• is the probability of protein
• When ? (confidence level) is equal to 0.005, the
  value of t is 2.576.
• If the t-value is larger than 2.576, the word is
  regarded as a good collocate of protein with
  99.5 confidence.

N 4n - 15
10
Step 4 Extraction of Collocation Keywords

• PASTA website in Sheffield University with 1,514
  MEDLINE abstracts
• 4,782 different stemmed words appear in the
  collocation windows
• 541 collocations generated in Step 3
• the output may contain nouns, prepositions,
  numbers, verbs, etc.
• assign parts of speech to these words
• there remained 154 collocation keywords with
  verbal part of speech

11
The Collocates with the Highest 15 t-value in
Step 3
12
The Collocates with the Highest 15 t-value in
Step 4
13
Interpretation of interact

• the verb interact is located at average
  -0.388 position with the value of standard
  deviation 3.256.
• The minus sign denotes that the word is on the
  left side of the protein.
• The absolute value of distance minus 1 indicates
  how many words are inserted in between the
  collocate and the protein class.

14
interact appears at two different locations

• As ltNAME TYPE"PROTEIN"gt ubiquitin
  lt/NAMEgt-conjugating enzymes interact with
  different substrates or other accessory proteins
  in the ubiquitination pathway, these variable
  surface regions may confer distinct specificity
  to individual enzymes.

• The modelremain free to interact with ltNAME
  TYPE"PEOTEIN"gt ferredoxin lt/NAMEgt and ltNAME
  TYPE"PROTEIN"gt flavodoxin lt/NAMEgt, the
  physiological partners of ltNAME TYPE"PROTEIN"gt
  ferredoxin lt/NAMEgt NADP() reductase.

15
Collocates mined from corpus

• act (-, -ed, -ing, -ion, -ive, -ivities, -ivity,
  -s), activat (-e, -ed, -es, -ing, -ion, -or) ,
  adopt (-,ed, -s), affect (-, -ed, -ing, -s),
  allow (-, -ed, -s), analys (-sed, -ses, -sis,
  -zed, -zing), appear (-, -s), arrange (-d,
  -ment), assembl (-ing, -y), associat (-e, -ed,
  -ion), bas (-e, -ed, -is), belong (-, -ing, -s),
  bind (-, -ing, -s) / bound, bond (-, -ed, -ing,
  -s), bridge (-, -d, -s), calculat (-ed, -ion),
  called, carr (-ied, -ier, -ies), cataly (-sed,
  -ses, -stic, -ze, -zed, -zes, -zing), cause,

16
Evaluation

• ask an expert to examine the resultant keywords
• make sure if our keyword set contains significant
  information that most experts talked about in
  their papers
• apply the keyword set to some application, e.g.,
  protein name recognition, and to confirm if the
  performance of this application is improved when
  the keyword set is introduced

17
(No Transcript)
18
Extracting Key Phrases

• apply the proposed method with window size of one
  word on each side of the original single keywords
  
• select candidates of standard deviations less
  than 0.4
• There are 99 compound collocation words with
  99.5 confidence and the values of standard
  deviations are less than 0.4

19
Examples of Compound Collocates with Prepositions
20
Examples of Compound Collocates with Nouns
21
Enhancing Performance of Protein Name Recognizers
Using Collocation
22
Protein Name Recognizers

• Challenging issues
• variant structural characteristics
• resemblance to regular noun phrases
• similarity with other kinds of biological
  substances
• Approaches
• rule-based
• Kex (Fukuda, et al., 1998)
• Yapex (Olsson, et al., 2002)
• corpus-based
• HMM model (Collier, et al., 2000)

23
Comparisons Kex vs. Yapex

• Kex
• 30 abstracts on SH3 domain and 50 abstracts on
  signal transduction
• 94.70 precision and 98.84 recall
• Yapex
• 48 documents were queried from protein binding
  and interaction
• 53 documents were randomly chosen from GENIA
  corpus
• 67.8 precision and 66.4 recall
• Changing the domain may result in the variant
  performance
• 40.4 precision and 41.1 recall of Kex on the
  test data of Yapex

24
Filtering Strategies

• M0 (baseline model)
• the output generated by Yapex
• M1
• For each candidate in M0, check if a collocate is
  found in its collocation window.
• If yes, tag the candidate as a protein name.
  Otherwise, discard it.
• M2
• If a collocate appears in the candidate or in the
  collocation window of the candidate, then tag the
  candidate as a protein name.
• Otherwise, discard it.

25
Filtering Strategies (Continued)

• M3
• During checking if there exists a collocate
  co-occurring with a protein candidate, the
  candidate without any collocate is kept
  undecidable instead of definite no.
• After all the protein names are examined, those
  undecidable candidates may be considered as
  protein names when one of their co-occurrences
  containing any collocate.
• M31 and M32

26
Experimental Results

• to enhance the precision rate, but not to reduce
  the significant recall rate

27
Integration Strategies

• how to improve the recall rates
• different protein name taggers have their own
  specific features
• a protein name recognizer may tag a protein name
  that another recognizer cannot identify, or
• both of them may accept certain common proteins
• The integration strategies are used to select
  correct protein names proposed by multiple
  recognizers

28
Candidates Proposed by Two Systems
29
Integration Strategies

• When the protein names produced from two
  recognizers are totally separated (i.e., type A),
  retain them as the protein candidates.
• When the protein names produced from two
  recognizers are exactly the same (i.e., type B),
  retain them as the protein candidates.
• When the protein names tagged by two taggers have
  partial overlap (i.e., types C, D and E), two
  additional integration strategies are employed
• Yapex-based strategies
• Kex-based strategies

30
Experiment Designs

• YA and KA
• Use the collocates automatically extracted to
  filter out the candidates
• YB and KB
• Use the terms suggested by human experts for the
  filtering strategies
• YA-C and KA-C
• If Yapex and Kex recommend the same protein names
  (i.e., type B), regard them as protein names
  without consideration of collocates.
• Otherwise, use the collocates proposed in this
  study to make filtering.
• YB-C and KB-C
• Similar to (3) except that the collocates are
  replaced by the terms suggested by human experts

31
Results

• Yapex-based Integration Strategy

• Kex-based Integration Strategy

32
Discussions

• The strategy of delaying the decision until clear
  evidence is found is workable
• The tendency M32gtM31gtM2gtM1 is still kept in the
  new experiments
• The set of collocates proposed by our system is
  more complete than the set of terms suggested by
  human expert
• The performances of YA, YA-C, KA, and KA-C are
  better than the performances of the corresponding
  models (i.e., YB, YB-C, KB, and KB-C)

33
Discussions

• The recall rates of both Yapex- and Kex-based
  integration are increased
• The precision rates are decreased more than the
  increase of recall rates
• Kex performed not well in this test set
• Yapex performed better in this test corpus