1. Concept Grounding to Multiple Knowledge Bases via Indirect Supervision
Chen-Tse Tsai and Dan Roth
University of Illinois at Urbana-Champaign

2. Concept Grounding Grounding concepts and entities
Wikipedia is not the single ideal resource some domains
Multiple ontologies in biological domain
E.g., Gene Ontology, Sequence Ontology, Protein Ontology …
We study the problem of grounding concepts to multiple knowledge bases (KBs), and use biomedical domain as our application
Mubarak, wife of Egyptian President Hosni Mubarak and …

3. The Task BRCA2 and homologous recombination. PR:000004804 EG:675
id: PR:
name: breast cancer type 2
susceptibility protein
def: A protein that is a translation
product of the human BRCA2
gene or a 1:1 ortholog thereof
synonyms: BRCA2, FACD,…
is_a: PR:
Protein Ontology
id: EG:675
symbol: BRCA2
description:
protein-coding BRCA2
breast cancer 2, early
onset
synonyms: BRCC2,
BROVCA2, …
Entrez Gene
Unlike Wikipedia entries, there is no full text article with hyperlink structure

4. Challenges Ambiguity Variability Supervision
A phrase can be used to express many different concepts
BRCA2 is used by 177 concepts
Variability
A concept may have many synonyms
EG:675 has synonyms BRCC2, FACD, FAD, FANCD, …
Supervision
Wikipedia has nice hyperlink structure which other doesn’t have
It is difficult to obtain human annotations for scientific domain
We explore the relationship between KBs to construct training examples without any document, the ranking model trained on these examples outperforms all unsupervised methods in our experiments.

5. System Overview … Concept Candidate Generation Mention
Concept Candidate Ranking
Global Inference with Knowledge
Ranked Candidates
Indirect Supervision
KB1
KB2
KBl …
Example:
Candidates
EG:675
PR:04804
EG:77244
GO:02111
Candidates
Score
PR:04804
0.8
EG:77244
0.6
EG:675
0.2
GO:02111
0.1
Candidates
Score
PR:04804
0.8
EG:77244
0.6
EG:675
0.2
GO:02111
0.1
BRCA2 and homologous recombination

6. Indirect Supervision
KB1
KBl …
Candidate Ranking
Global Inference
Candidate Generation
Candidate Generation
Given a mention, produce a small set of possible concepts
Synonym Matching
Construct a dictionary from all synonyms and names across all KBs
Phrase  possible concept IDs
Word Matching
Splitting phrases to words, and combining concepts by words
Word  possible concept IDs
Only keep top k concepts
Words are normalized by the SPECIALIST Lexical Tools

7. Candidate Ranking Relevance score between (mention, concept candidate)
Indirect Supervision
KB1
KBl …
Candidate Ranking
Global Inference
Candidate Generation
Candidate Ranking
Relevance score between (mention, concept candidate)
Representations of the mention m
context-word(m): neighboring words in the document
context-concept(m): concept candidates of other mentions in the document
Representations of the candidate c
def(c): definition of c
neighbor(c): concepts have a relation with c in all KBs
Ranking features
Common words in context-word(m) and def(c)
Common concepts in context-concept(m) and neighbor(c)

8. Indirect Supervision
KB1
KBl …
Candidate Ranking
Global Inference
Candidate Generation
Indirect Supervision
We explore the redundancy and relationship between KBs to construct training examples
Discovering positive examples
Cross reference
Chromosome
Make one as the “mention”, annotated by another
Has participant relationship
GO:
SO:
Wikipdeia:Chromosome
xref
GO:
SO:
GO:
fructose metabolic process
fructose
has_participant

9. Indirect Supervision Generating other candidates
KB1
KBl …
Candidate Ranking
Global Inference
Candidate Generation
Indirect Supervision
Generating other candidates
Apply candidate generation on the name of the concept
Uniformly sample 200 concepts from all KBs
Take the number of common ancestors between a candidate and the positive candidate as the relevance score
Extracting features from pairs of concepts
There is no context for GO:
Using def(m) instead of context-word(m)
Using neighbor(m) instead of context-concept(m)
GO:
SO:

10. Indirect Supervision
KB1
KBl …
Candidate Ranking
Global Inference
Candidate Generation
Global Inference
Enforcing a coherent global solution of all mentions in a document by constraints
Hard constraints
If a gene is from a species which is not mentioned anywhere in the document, it is removed from the final list.
Entries in Entrez Gene Database and Protein Ontology have relations to NCBI Taxonomy
Constraints
Ranking score of the j-th candidate of the i-th mention

11. Dataset Colorado Richly Annotated Full-Text corpus [Bada et al., 2012]
67 full text of biomedical journal articles
7 ontologies
Ontology
# Concepts
# Annotations
# Unique annot. PR
26,879
15,593
889
NCBITaxon
789,509
7,449
149 GO
25,471
29,443
1,235
CHEBI
19,633
8,137
553 EG
17,097,474
12,266
1,021 SO
1,704
21,284
259 CL
857
5,760
155
Total
17,961,527
99,138
4,261

12. Evaluation Comparing to 5 unsupervised methods
AUC: area under PR curve
hAUC: hierarchical version, considering common ancestors
Approach
Mean AUC
Mean hAUC
TF-IDF
40.44
48.50
PageRank
42.78
50.04
Zheng et al. (2014)
35.67
42.93
Agirre and Soroa (2009)
43.39
51.88
Agirre and Soroa (2009) w2w
46.51
55.46
Our Approach
48.58
57.37
Direct Supervision
58.98
62.59

13. Using KBs Individually v.s. Jointly
Joint: grounding a mention to all KBs simultaneously
Individual: focusing on a single KB each time
Approach
Individual
Joint
PageRank
49.85
55.74
Zheng et al. (2014)
49.46
55.42
Agirre and Soroa (2009)
52.12
54.88
Agirre and Soroa (2009) w2w
52.23
56.18
Our Approach
49.93
57.65

14. Conclusions
Concept grounding to multiple KBs without hyperlink structure
We propose an approach to construct training examples without using any document. It enables us to apply well-studied statistical models and outperforms unsupervised methods
We show that considering multiple KBs together has advantage over using each KB individually