Collecting Cognitively Salient Semantic Relations
Gerhard Kremer
[email protected]
Introduction
Preliminary Focus
Motivation
Electronic learner’s dictionary for German & Italian (ELDIT) – among
others, shows semantically related words for a target word.
Composed part relations (adj modifier + part-noun), e. g.
rabbit: has long ears
dog: has a wet nose
Method
Goal
Assuming that salient instances of part relations for a given concept
have been identified already:
Find the most salient modifiers.
Production experiment data was used both as input (concept, part)
and for evaluation (modifier-part pairs produced for a concept)
Approach
Based on occurrence frequencies of concept/part/modifier combinations in the WaCKy Webcorpus, create ranked list and select the 5
highest ranked modifier candidates.
Concept dog:
DE
1.0
has paws, a tail, barks ©
has a heart, breathes §
IT
1.0
Examples
Evaluated Rank Lists (DE)
modifier [...] part within 4 words
(and within 20 sentences of concept)
same, but not considering concept context
combination: multiplication of frequencies from both sources
reranking: pull up those modifiers which are similar to those at
higher ranks (by calculating their cosine distance based on
nouns they co-occurr with)
without considering concept
in concept context
combined: freq(contextless) x freq(in context)
reranked combi list (acc. to cos distance)
0.2
0.4
Precision
0.2
0.4
Precision
0.6
0.8
without considering concept
in concept context
combined: freq(contextless) x freq(in context)
reranked combi list (acc. to cos distance)
0.8
Goal
Semi-automatically harvest instances of cognitively salient semantic
relations from text resources from the web.
0.6
. . . but entries were selected manually, on an arbitrary basis.
Webcorpus Excerpt Example
. . . Die mittelgroßen Affen leben in Gruppen von etwa 15 Tieren
auf Bäumen im Regenwald. [. . . ] Die Kipunji sind verwandt mit
anderen Mangaben, doch sie weisen einige Besonderheiten auf. Sie
haben braunes oder hellbraunes Fell und geben Töne von sich. . .
Task: Describe the given concept in short phrases
Stimuli: 50 concrete concepts from 10 classes (DE & IT)
Data: Categorised into types of relations (cf. McRae et al.)
Analysis: Deviations of the top 6 types from overall production
frequency distribution (for each concept class)
go
cate
mammal
ry
German
Italian
type
type
part
ty aviour ction ation
i
l
a
qu beh fun
loc
ry
go
cate
Pearson
residuals:
22.9
4.0
2.0
0.0
−2.0
−4.0
class
class
0.4
0.6
0.8
1.0
0.0
Recall
0.2
0.4
0.6
0.8
1.0
Recall
Production vs. Perception (DE)
Follow-up judgement experiment:
“The part of a concept is modifier.” – plausible or not?
Result for the acceptance rate of 0.75:
Small overlap of modifiers both produced and accepted (46), compared to modifiers only produced (53) or only accepted (42).
Conclusion
vegetable
bodypart
clothing
implement
implement
vehicle
vehicle
building
0.2
Pearson
residuals:
23.3
fruit
vegetable
furniture
0.0
bird
fruit
clothing
ty aviour ction ation
i
l
a
qu beh fun loc
mammal
bird
bodypart
part
0.0
0.0
Experiment: Feature Production
−14.2
p−value =
< 2.22e−16
furniture
building
4.0
2.0
0.0
−2.0
−4.0
−12.6
p−value =
< 2.22e−16
Results
Preferred relation types depending on which (super-) class the
concept belongs to
Similar patterns for German and Italian data
→ General Procedure:
For a given concept
select those relation types
which are prominent within its concept class,
and for each of these
find those relations which are cognitively most salient.
Best method combines in-context and contextless information
(with similar performance for German and Italian), and
yields both produced and perceived modifiers
Reranking improvable?
Adaptable to part-noun collection and other relation types?
References
Kremer, G., Abel, A., and Baroni, M. (2008).
Cognitively salient relations for multilingual lexicography.
In Proceedings of the Workshop on Cognitive Aspects of the Lexicon ( COGALEX 2008),
pages 94–101, Manchester, United Kingdom. Coling 2008 Organizing Committee.
McRae, K., Cree, G. S., Seidenberg, M. S., and McNorgan, C. (2005).
Semantic feature production norms for a large set of living and nonliving things.
Behaviour Research Methods, Instruments & Computers, 37(4):547–559.
Spence, D. P. and Owens, K. C. (1990).
Lexical co-occurrence and association strength.
Journal of Psycholinguistic Research, 19(5):317–330.
Advisors: Marco Baroni (CIMeC), Andrea Abel (EURAC)