Upload
jody
View
35
Download
0
Tags:
Embed Size (px)
DESCRIPTION
Combining Contexts in Lexicon Learning for Semantic Parsing. May 25, 2007 NODALIDA 2007, Tartu, Estonia. Richard Socher Saarland University Germany. Chris Biemann University of Leipzig Germany. Rainer Osswald FernUniversität Hagen Germany. Outline. - PowerPoint PPT Presentation
Citation preview
1
Combining Contexts in Lexicon Learning for Semantic Parsing
May 25, 2007
NODALIDA 2007, Tartu, Estonia
Chris BiemannUniversity of Leipzig
Germany
Rainer OsswaldFernUniversität Hagen
Germany
Richard SocherSaarland UniversityGermany
2
Outline• Motivation: lexicon extension for semantic parsing
• The semantic lexicon HaGenLex
• Binary features and complex sorts
• Method: bootstrapping via syntactic contexts
• Results
• Discussion
3
Motivation• Semantic parsing aims at finding a semantic
representation for a sentence
• Semantic parsing needs as a prerequisite semantic features of words.
• Semantic features are obtained by manually creating lexicon entries (expensive in terms of time and money)
• Given a certain amount of manually created lexicon entries, it might be possible to train a classifier in order to find more entries
• Objective is Precision, Recall is secondary
4
HaGenLex: Semantic Lexicon for German
complex sort
size: 22,700 entries of these: 13,000 nouns, 6,700 verbs
WORD SEMANTIC CLASSAggressivität nonment-dyn-abs-situationAgonie nonment-stat-abs-situationAgrarprodukt nat-discreteÄgypter human-objectAhn human-objectAhndung nonment-dyn-abs-situationÄhnlichkeit relationAirbag nonax-mov-art-discreteAirbus mov-nonanimate-con-potagAirport art-con-geogrAjatollah human-objectAkademiker human-objectAkademisierung nonment-dyn-abs-situationAkkordeon nonax-mov-art-discreteAkkreditierung nonment-dyn-abs-situationAkku ax-mov-art-discreteAkquisition nonment-dyn-abs-situationAkrobat human-object... ...
5
Characteristics of complex sorts in HaGenLex
In total, 50 complex sorts for nouns are constructed from allowed combinations of:
• 16 semantic features (binary), e.g. HUMAN+, ARTIFICIAL- • 17 sorts (binary), e.g. concrete, abstract-situation...
sort (hierarchy)
semantic features
complex sorts
6
Application: WOCADI-Parser
„Welche Bücher von Peter Jackson über Expertensysteme wurden bei Addison-Wesley seit 1985 veröffentlicht?“
7
General Methodology
Distributional Hypothesis projected on syntactic-semantic contexts for nouns: nouns of similar complex sort are found in similar contexts
We use three kinds of context elements• Adjective Modifier• Verb-Subject (deep)• Verb-Object (deep)
as assigned by the WOCADI parser for training 33 binary classifiers.
8
DataCorpus:• 3,068,945 sentences obtained from the Leipzig Corpora
Collection• parser coverage: 42%• verb-deep-subject relations: 430,916• verb-deep-object relations: 408,699• adjective-noun relations: 450,184
Lexicon• 11,100 noun entries• lexicon extension: 10-fold cross validation on known nouns• Also unknown nouns will be classified
9
Algorithm:
Initialize the training set;As long as new nouns get classified { calculate class probabilities for each context element; for all yet unclassified nouns n { Multiply class probs of context elements class-wise; Assign the class with highest probabilities to noun n; }}
Class probabilities per context element:a) count number of per classb) normalize on total number of class wrt. noun classesc) normalize to row sum=1
A threshold regulates the minimum number of different context elements a noun co-occurs with in order to be classified
Bootstrapping Mechanism
10
From binary classes to complex sorts• Binary classifiers for single features for all three context
element types are combined into one feature assignment:– Lenient: voting– Strict: all classifiers for different context types agree
• Combining the outcome: safe choices
ANIMAL +/-ANIMATE +/-ARTIF +/-AXIAL +/-... (16 features)
... (17 sorts)
ab +/-abs +/-ad +/-as +/-
Selection:compatible complex
sorts that are minimal w.r.t hierarchy and unambiguous.
result classor
reject
11
Results: binary classes for different context types
=5
=1
most of the binary features are highly biased
12
Combination of context types =1
13
Results for complex sorts=5 =1
Complex sorts with highest
training frequency
14
Typical mistakesPflanze (plant) animal-object instead of plant-objectzart, fleischfressend, fressend, verändert, genmanipuliert, transgen, exotisch, selten, giftig, stinkend,
wachsend...
Nachwuchs (offspring) human-object instead of animal-objectwissenschaftlich, qualifiziert, akademisch, eigen, talentiert, weiblich, hoffnungsvoll, geeignet, begabt,
journalistisch...
Café (café) art-con-geogr instead of nonmov-art-discrete (cf. Restaurant)Wiener, klein, türkisch, kurdisch, romanisch, cyber, philosophisch, besucht, traditionsreich, schnieke,
gutbesucht, ...
Neger (negro) animal-object instead of human-objectweiß, dreckig, gefangen, faul, alt, schwarz, nackt, lieb, gut, brav
but:
Skinhead (skinhead) human-object (ok){16,17,18,19,20,21,22,23,30}ährig, gleichaltrig, zusammengeprügelt, rechtsradikal, brutal
In most cases the wrong class is semantically close. Evaluation metrics did not account for that.
15
Discussion of ResultsBinary features:• Precision >98% for most binary features• Assigning the smaller class is hard for bias>0.9
Context types• verb-subject and verb-object are better than adjective• verb-subject is best single context for complex sorts • combination always helps for binary features
Complex sorts• Todo: more lenient combination procedure to increase
recall
16
Conclusion
• Method for semantic lexicon extension• High precision for binary semantic features• Unknown nouns:
– For 3,755 nouns not in the lexicon, a total of 125,491 binary features was assigned.
– For 1,041 unknown nouns, a complex sort was assigned
• Combination to complex sorts yet to be improved• Combination of different context types improves
results
17
Any Questions?
Thank you very much!