Cross-lingual projection of Semantics

1
Cross-lingual projection of Semantics

• Sebastian Pado
• IGK Colloquium
• Dec 16th 2004

2
Overview

1. Background Role Semantics
2. Semantic Projection
3. Current and Future Work

3
Framework Role semantics

• Predicate-argument structure,
• Theta roles, who did what to whom

Agent
Recipient
Theme
Peter gives Mary a book
NB. No treatment of discourse relations,
modality, negation, etc.
4
Flavours of role semantics

• Top-down approach common, intuitively defined
  roleset for all verbs
• give is Mary Recipient or Goal or Patient?
• resemble Subj vs. Obj

• Bottom-up approach Frame Semantics
• Frames Conceptual rep of a situation Statement,
  Giving, Transaction
• Each frame is introduced by a target say, give,
  buy
• Roles are frame-specific

5
Frame Semantics

• An Example Frame Giving
• Targets give, hand out, receive
• Roles Donor, Recipient, Theme
• The Berkeley FrameNet Project
• English Frame Lexicon
• 200 Frames, 2.500 words (V/N/Adj)
• Typically 3-6 roles per frame
• Corpus of 60.000 annotated instances

6
Frame Semantics An Example
7
What do Role Semantics buy us?

• Surface-independent representation
• Solves the paraphrase problem
• Peter gives the book to Mary
• Mary receives the book from Peter
• Flexible basis for QA, Inference etc.
• Aljoscha Burchardts PhD
• Common cross-lingual semantic rep

8
Semantic Role Assignment

• Task Automatic tagging of roles on free text
• Important for NLP applications
• Linking (syntax-semantics interface)
• Statistical modelling (as classification)
• Frame semantically coherent targets
• Targets show linking idiosyncrasies
• Give Sub - Donor, Dobj - Theme, To-PP/Iobj - Rec
• Get Sub - Rec, Dobj - Theme, From-PP - Donor
• Needs lots of training data

9
Moving to another language

• SALSA Manual creation and use of a German corpus
  with semantic annotation
• Basis TIGER newspaper corpus, 1.5m words
• English frames (mostly) work for German
• Frame concept language-independent
• But Annotation slow and error-prone
• Total effort gt 10 person years
• Can we use the English data for German?

10
Overview

1. Background Role Semantics
2. Semantic Projection
3. Current and Future Work

11
Central idea Semantic Projection

• Find a large, parallel bilingual corpus
• E/G part of EUROPARL (25m words)
• Assign semantic roles on English side
• Train automatic tagger on English data
• Project semantics over to German
• Step 1 Find semantic equivalences via word
  alignment
• Step 2 Project frame
• Step 3 Project roles
• Result Large German annotated corpus

12
Projection Example
Arriving
Arriving
Peter comes home
Peter kommt nach Hause

• Three assumptions to make this work

13
Assumption 1

• Semantic representation is parallel

Arriving
Arriving
Peter comes home
Peter kommt nach Hause
14
Semantic (im-)parallelism

• Frame definition based on realisable roles
• German and English typologically similar
• Mostly, same frames evoked
• Aspect is problematic
• Proper differences
• We finish by 12 oclock Activity_finish
• Wir sind um 12 Uhr fertig Activity_done_state
• Same aspect, lexicalised differently
• I finish by saying
• Abschliessend sage ich

15
Assumption 2

• There is always parallel lexical material that is
  semantically equivalent

Arriving
Arriving
Peter comes home
Peter kommt nach Hause
16
(Im)parallelism of lexical material

• We only need semantic parallelism, only for
  targets and roles
• Dont care about discourse, modality, etc.
• Dont care about exact wording
• Insights from translation science
• Translation Recreation of text based on content
  and target language norms
• Frame structures propositional content
• Specific register
• Specific domain (no cultural differences)

17
Assumption 3

• Word Alignment provides semantic equivalence

Arriving
Arriving
Peter comes home
Peter kommt nach Hause
18
Word Alignment as Semantic Equivalence

• Current Word Alignment models use co-occurrence
  to determine alignment
• But co-occurrence ! semantic equivalence
• decide entscheiden Entscheidung treffen
• insist bestehen darauf
• Problems Phrasal verbs, Idioms, Support Verbs
  (Funktionsverbgefuege), Noise proper

19
Overview

1. Background Role Semantics
2. Semantic Projection
3. Current and Future Work

20
Current Work (1)

• Empirical assessment of assumptions
• Manual annotation of parallel corpus sample
• Independent annotation of German / English
• Evaluation of semantic parallelism
• Evaluation of lexical parallelism
• Evaluation of automatic word alignment

21
Current Work (2)

• Token-wise word alignment too noisy
• decide - treffen Deciding?
• Instead Find reliable type equivalences
• Statistics over complete corpus, filtering
• Removal of German collocations
• Result German frame lexicon
• Target x can evoke frames a,b,c
• Project frame only if licensed by German lexicon

22
Current Work (3)

• Projection of roles Find equivalences between
  constituents
• Define pairwise similarities
• Efficiently identify best match
• Graph matching
• Probabilistic model
• Choice points
• Definition of similarities
• Bijective correspondence, yes or no?
• Implementation

23
Future Work

• Thorough Evaluation
• Filtering
• Projection will be noisy
• Training a German semantic tagger
• Evaluation wrt coverage, accuracy
• Combination with manually annotated data (SALSA)
• Using another language
• English/French part of EUROPARL

24
Conclusion

• Automatic creation of semantically annotated data
  for a new language
• Projection of annotation from known
  languageusing a word-aligned parallel corpus
• Theory in place
• Potential Problems
• Semantics may diverge
• Lexical material may diverge
• Word Alignment noisy
• Empirical evaluation underway