Building Question Answering systems based on Semantic Parsing with Combinatory Categorial Grammar (CCG)

1. Semantic Parsing with Combinatory Categorial
Grammar (CCG)
Sherzod Hakimov
Sherzod Hakimov

2. Publication
Sherzod Hakimov, Christina Unger, Sebastian Walter, Philipp Cimiano. Applying Semantic Parsing to Question
Answering Over Linked Data: Addressing the Lexical Gap. NLDB 2015: 103-109

3. QALD-4
Question Answering over Linked Data
● multilingual (English, German, Spanish, Italian, French, Dutch, Romanian)
● over 200 training, 50 test questions
● DBpedia 3.9

4. QALD-4
Question:
Which river does the Brooklyn Bridge cross?
Query:
PREFIX dbo: <http://dbpedia.org/ontology/>
PREFIX res: <http://dbpedia.org/resource/>
SELECT DISTINCT ?uri
WHERE {
res:Brooklyn_Bridge dbo:crosses ?uri .
}
Answer:
http://dbpedia.org/resource/East_River

5. Semantic Parsing
● ZC_05 [1] Parsing
○ Lambda Calculus
○ CCG (Combinatory Categorial Grammar)[2]
○ CYK - dynamic programming
● Training
○ Perceptron Algorithm
[1] Luke S. Zettlemoyer, Michael Collins. Learning to Map Sentences to Logical Form: Structured Classification with Probabilistic
Categorial Grammars. UAI 2005: 658-666
[2] Mark Steedman. Surface structure and interpretation. Linguistic inquiry 30, MIT Press 1997, ISBN 978-0-262-69193-2, pp. 1-126

6. Geobase880
● 880 questions paired with semantics
● geographic concepts about the U.S.
● categories : city, river, mountain ...
● functions : next_to, loc ...
● entities : colorado, texas ...
Question : what are the major cities in Texas?
Semantics : λx (major(x) ∧ city(x) ∧ location(texas, x))

7. Geobase880
● 880 questions paired with semantics
● geographic concepts about the U.S.
● categories : city, river, mountain ...
● functions : next_to, loc ...
● entities : colorado, texas ...
Question : what are the major cities in Texas?
Semantics : λx (major(x) ∧ city(x) ∧ location(texas, x))
Precision Recall F1
ZC05 96.25 79.29 86.95

8. Semantic Parsing using CCG and Lambda Calculus
CCG - Combination rules CCG - Categories

9. Semantic Parsing using CCG and Lambda Calculus
What are the major cities in Texas?

10. what are the major cities in Texas
S/(SNP) (SNP)/NP (NP/N) N/N N (NN)/NP NP
λg.λx. (g(x)) λf.f(x) λf.f(x) λg.λx.major(x) ∧ g(x) λx.city(x) λy.λg.λx.location(y, x) ∧ g(x) texas
NN
λg.λx.location(texas, x) ∧ g(x)

11. what are the major cities in Texas
S/(SNP) (SNP)/NP (NP/N) N/N N (NN)/NP NP
λg.λx. (g(x)) λf.f(x) λf.f(x) λg.λx.major(x) ∧ g(x) λx.city(x) λy.λg.λx.location(y, x) ∧ g(x) texas
NN
λg.λx.location(texas, x) ∧ g(x)
N
λx.location(texas, x) ∧ city(x)

12. what are the major cities in Texas
S/(SNP) (SNP)/NP (NP/N) N/N N (NN)/NP NP
λg.λx.(g(x)) λf.f(x) λf.f(x) λg.λx.major(x) ∧ g(x) λx.city(x) λy.λg.λx.location(y, x) ∧ g(x) texas
NN
λg.λx.location(texas, x) ∧ g(x)
N
λx.location(texas, x) ∧ city(x)
N
λx.major(x) ∧ location(texas, x) ∧ city(x)

13. what are the major cities in Texas
S/(SNP) (SNP)/NP (NP/N) N/N N (NN)/NP NP
λg.λx.(g(x)) λf.f(x) λf.f(x) λg.λx.major(x) ∧ g(x) λx.city(x) λy.λg.λx.location(y, x) ∧ g(x) texas
NN
λg.λx.location(texas, x) ∧ g(x)
N
λx.location(texas, x) ∧ city(x)
N
λx.major(x) ∧ location(texas, x) ∧ city(x)
NP
λx.major(x) ∧ location(texas, x) ∧ city(x)

14. what are the major cities in Texas
S/(SNP) (SNP)/NP (NP/N) N/N N (NN)/NP NP
λg.λx.(g(x)) λf.f(x) λf.f(x) λg.λx.major(x) ∧ g(x) λx.city(x) λy.λg.λx.location(y, x) ∧ g(x) texas
NN
λg.λx.location(texas, x) ∧ g(x)
N
λx.location(texas, x) ∧ city(x)
N
λx.major(x) ∧ location(texas, x) ∧ city(x)
NP
λx.major(x) ∧ location(texas, x) ∧ city(x)
(SNP)
λx.major(x) ∧ location(texas, x) ∧ city(x)

15. what are the major cities in Texas
S/(SNP) (SNP)/NP (NP/N) N/N N (NN)/NP NP
λg.λx.(g(x)) λf.f(x) λf.f(x) λg.λx.major(x) ∧ g(x) λx.city(x) λy.λg.λx.location(y, x) ∧ g(x) texas
NN
λg.λx.location(texas, x) ∧ g(x)
N
λx.location(texas, x) ∧ city(x)
N
λx.major(x) ∧ location(texas, x) ∧ city(x)
NP
λx.major(x) ∧ location(texas, x) ∧ city(x)
(SNP)
λx.major(x) ∧ location(texas, x) ∧ city(x)
S
λx (major(x) ∧ location(texas, x) ∧ city(x))

16. Semantic Parsing on QALD
● update GENLEX[1] rules for DBpedia predicates, categories, resources
● Convert logical forms in Lambda Calculus to SPARQL queries
[1] Luke S. Zettlemoyer, Michael Collins. Learning to Map Sentences to Logical Form: Structured Classification with Probabilistic
Categorial Grammars. UAI 2005: 658-666

17. GENLEX for QALD-4

18. Parsing Example

19. Results
- QALD-4
○ training (200 instances)
○ test (50 instances)

20. Results -1
- trained lexical entries
- DBpedia Ontology labels (predicates, categories)

21. Results -1
- trained lexical entries
- DBpedia Ontology labels (predicates, categories)
Precision Recall F1
Learned lexicon + ontology labels 0.66 0.05 0.09

22. Results -2
- lexical entries generated by M-ATOLL
- DBpedia Ontology labels (predicates, categories)

23. Results -2
- lexical entries generated by M-ATOLL[1]
- DBpedia Ontology labels (predicates, categories)
Precision Recall F1
Learned lexicon + ontology labels + M-ATOLL 0.70 0.18 0.30
[1] Sebastian Walter, Christina Unger, Philipp Cimiano. M-ATOLL: A Framework for the Lexicalization of Ontologies in Multiple
Languages. International Semantic Web Conference (1) 2014: 472-486

24. Results -3
- handcrafted lexical entries
- DBpedia Ontology labels (predicates, categories)

25. Results -3
- handcrafted lexical entries
- DBpedia Ontology (predicates, categories)

26. Results -3
- handcrafted lexical entries
- DBpedia Ontology labels (predicates, categories)

27. Comparison for QALD-4