1. Using BabelNet in Bridging the Gap
Between Natural Language Queries and
Linked Data Concepts
Khadija Elbedweihy, Stuart N. Wrigley, Fabio Ciravegna and and Ziqi Zhang
OAK Research Group,
Department of Computer Science,
University of Sheffield, UK
2. Outline
• Motivation and Problem Statement
• Natural Language Query Approach
• Approach Steps
• Evaluation
• Results and Discussion
3. Motivation – Semantic Search
• Wikipedia states that Semantic Search:
“seeks to improve search accuracy by understanding
searcher intent and the contextual meaning of terms as
they appear in the searchable dataspace, whether on the
Web or within a closed system, to generate more
relevant results”
• Semantic search evaluations reported user preference for
free natural language as a query approach (simple, fast &
flexible) as opposed to controlled or view-based inputs.
4. Problem Statement
• Complete freedom increases difficulty of matching query
terms with the underlying data and ontologies.
• Word sense disambiguation (WSD) is core to the solution.
Question: “How tall is ..... ?”: property height
– tall is polysemous, should be first disambiguated:
– great in vertical dimension; tall people; tall buildings, etc.
– too improbable to admit of belief; a tall story, …
• Another difficulty: Named Entity (NE) recognition and
disambiguation.
5. Approach
• Free-NL semantic search approach, matching user query
terms with the underlying ontology using:
1) An extended-Lesk WSD approach.
2) A NE recogniser.
3) A set of advanced string similarity algorithms and
ontology-based heuristics to match disambiguated
query terms to ontology concepts and properties.
6. Extended-Lesk WSD approach
• WordNet is predominant, however its granularity is a
problem for achieving high performance in WSD.
• BabelNet is a very large multilingual ontology with widecoverage obtained from both WordNet and Wikipedia.
• For disambiguation, bags are extended with senses’
glosses and different lexical and semantic relations.
• Include synonyms, hyponyms , hypernyms , attribute, see
also and similar to relations.
7. Extended-Lesk WSD approach
• Information added from a Wikipedia page (W), mapped
to a WordNet synset includes:
1.labels; page “Play (theatre)” add play and theatre
2. set of pages redirecting to W; Playlet redirects to Play
3. set of pages linked from W; links in the page Play (theatre)
include literature, comedy, etc.
• Synonyms of synset S, associated with Wikipedia page W:
WordNet synonyms of S in addition to lemmas of
wikipedia information of W".
8. Extended-Lesk WSD approach
Feature
P
R
F1
Baseline
Synonyms
Syn + hypo
Syn + gloss examples (WN)
Syn + gloss examples (Wiki)
Syn + gloss examples (WN + Wiki)
Syn + hyper
Syn + semRel
Syn + hypo + gloss(WN)
Syn + hypo + gloss(WN) + hyper
Syn + hypo + gloss(WN) + hyper + semRel
Syn+hypo+gloss(WN)+hyper+semRel+relGlosses
58.09
59.14
62.16
61.97
61.14
60.21
60.36
59.65
64.92
65.28
65.45
69.76
57.98
59.03
62.07
61.86
61.02
60.10
60.26
59.54
64.81
65.18
65.33
69.66
58.03
59.09
62.12
61.92
61.08
60.16
60.31
59.59
64.86
65.23
65.39
69.71
• Sentences with less than seven words: f-measure of 81.34%
9. Approach – Steps
1. Recognition and disambiguation of Named Entities.
2. Parsing and Disambiguation of the NL query.
3. Matching query terms with ontology concepts and
properties.
4. Generation of candidate triples.
5. Integration of triples and generation of SPARQL queries.
10. 1.Recognition and disambiguation of Named Entities
• Named entities recognised using AlchemyAPI.
• AlchemyAPI had the best recognition performance in
NERD evaluation of SOA NE recognizers.
• AlchemyAPI exhibits poor disambiguation performance
• Each NE is disambiguated using our BabelNet-based WSD
approach.
11. 1.Recognition and disambiguation of Named Entities
• Example: “In which country does the Nile start?”
• Matches of Nile in BabelNet include:
–
–
–
–
http://dbpedia.org/resource/Nile (singer)
http://dbpedia.org/resource/Nile (TV series)
http://dbpedia.org/resource/Nile (band)
http://dbpedia.org/resource/Nile
• Match selected (Nile: river): overlapping terms between
sense and query (geography, area, culture, continent)
more than other senses.
12. 2.Parsing and Disambiguation of the NL query
• Stanford Parser used to gather lemmas and POS tags.
• Proper nouns identified by the parser and not recognized
by AlchemyAPI are disambiguated and added to the
recognized entities.
• Example: “In which country does the Nile start?”
– The algorithm does not miss the entity Nile, although it
was not recognized by AlchemyAPI.
13. 2.Parsing and Disambiguation of the NL query
• Example: “Which software has been developed by
organizations founded in California?”
Output:
Word
software
POS
NP
position
1
developed
organizations
founded
develop
organize
find
VBN
NNS
VBN
2
3
4
California
•
Lemma
software
California
NP
5
Equivalent output generated using keywords or phrases.
14. 3.Matching Query Terms with Ontology Concepts & Properties
• Noun phrases, nouns and adjectives are matched with
concepts and properties.
• Verbs are matched only with properties.
• Candidate ontology matches ordered using: Jaro-Winkler
and Double Metaphone string similarity algorithms.
• Jaro-Winkler threshold to accept a match is set to 0.791,
shown in literature to be the best threshold value.
15. 3.Matching Query Terms with Ontology Concepts & Properties
• Matching process uses the following in order:
1. query term (e.g., created)
2. lemma (e.g., create)
3. derivationally related forms (creator)
• If no matches, disambiguate query term and use
expansion terms in order:
1. synonyms
2. hyponyms
3. hypernyms
4. semantic relations (e.g., height as an attribute for tall)
16. 4. Generation of Candidate Query Triples
• Structure of the ontology (taxonomy of classes and domain
and range of properties) used to link matched concepts and
properties and recognized entities to generate query triples.
Three-Terms Rule
• Each three consecutive terms matched with set of templates.
E.g., “Which television shows were created by Walt Disney?”
• Template (concept-property-instance) generates triples:
?television_show <dbo:creator> <res:Walt_Disney>
?television_show <dbp:creator> <res:Walt_Disney>
?television_show <dbo:creativeDirector> <res:Walt_Disney>
17. Three-Terms Rule
Examples of templates used in three-terms rule:
• concept-property-instance
– airports located in California
– actors born in Germany
• instance-property-instance
– Was Natalie Portman born in the United States?
• property-concept-instance
– birthdays of actors of television show Charmed
18. Two-Terms Rule
Two-Terms Rule, used when:
1) There is fewer than three derived terms
2) No match between query terms and three-term template
3) Matched template did not generate candidate triples
E.g., “In which films directed by Garry Marshall was Julia
Roberts starring?”
<Garry Marshall, Julia Roberts, starring> : matched to a
three-terms template but does not generate triples.
19. Two-Terms Rule
Two-Terms Rule
Question: “what is the area code of Berlin?”
• Template (property-instance) generates the triples:
<res:Berlin> <dbp:areaCode> ?area_code
<res:Berlin> <dbo:areaCode> ?area_code
20. Comparatives
Comparatives Scenarios:
1) Comparative used with a numeric datatype property:
e.g., “companies with more than 500,000 employees”
?company <dbp:numEmployees> ?employee
?company <dbp:numberOfEmployees> ?employee
?company a <dboCompany>
FILTER (?employee > 500000)
21. Comparatives
2) Comparative is used with a concept:
e.g., “places with more than 2 caves”
• Generate the same triples for places with caves:
?place a <http://dbpedia.org/ontology/Place>.
?cave a <http://dbpedia.org/ontology/Cave>.
?place ?rel1 ?cave.
?cave ?rel1 ?place.
• Add the aggregate restriction:
GROUP BY ?place
HAVING (COUNT(?cave)>2).
22. Comparatives
3) Comparative is used with an object property
e.g., “countries with more than 2 official languages”
• Similarly, generate the same triples for country and
official language and add the restriction:
GROUP BY ?country
HAVING (COUNT(?official_language) > 2)
4) Generic Comparatives
e.g., “Which mountains are higher than the Nanga Parbat?”
23. Generic Comparatives
• Difficulty: identify the property referred to by the
comparative term.
1) Select best relation according to query context.
– Identify all numeric datatype properties associated
with the concept “mountain”, include:
“latS, longD, prominence, firstAscent, elevation, longM, …”
2) Disambiguate synsets of all properties and use WSD
approach to identify the most related synset to the query.
– property elevation is correctly selected
24. 5. Integration of Triples and Generation of SPARQL Queries
• Generated triples integrated to produce SPARQL query.
• Query term positions used to order the generated triples.
• Triples originating from the same query term are
executed in order until an answer is found.
• Duplicates are removed while merging the triples.
• SELECT and WHERE clauses added in addition to any
aggregate restrictions or solution modifiers.
25. Evaluation
• Test data from 2nd Open Challenge at QALD-2.
• Results produced by QALD-2 evaluation tool.
• Very promising results: 76% of questions answered correct.
Approach
Answered Correct
Precision
Recall
F1
BELA
QAKiS
Alexandria
31
35
25
17
11
5
0.62
0.39
0.43
0.73
0.37
0.46
0.67
0.38
0.45
SenseAware
SemSeK
MHE
54
80
97
41
32
30
0.51
0.44
0.36
0.53
0.48
0.4
0.52
0.46
0.38
26. Discussion
• Design choices affected by priority for precision or recall:
1. Query Relaxation
e.g., “Give me all actors starring in Last Action Hero”
– Restricting results to actors harms recall
– Not all entities in LD are typed, let alone correctly typed
– Query relaxation favors recall but affects precision
e.g. “How many films did Leonardo DiCaprio star in?”
– Return TV series rather than only films such as
res:Parenthood (1990 TV series).
• Decision: favor precision; keep restriction when specified.
27. Discussion
2. Best or All Matches
e.g., “software by organizations founded in California”
– Properties matched: foundation and foundationPlace
– Using only best match (foundation ) does not generate
all results affects recall.
– Using all properties (may not be relevant to the query)
would harm precision.
• Decision: use all matches; with high value for the
similarity threshold; perform checks against the ontology
structure to assure relevant matches are only used.
28. Discussion
3. Query Expansion
• Can be useful for recall, when the query term is not
sufficient to return all answers.
• Example: use “website” and “homepage” if any of them
used in a query and both have matches in the ontology.
• Quality of expansion terms influenced by WSD approach;
wrong sense identification will lead to noisy list of terms.
• Decision: perform query expansion only when no
matches found in the ontology for a term; or no results
generated using the identified matches.