Codeco: A Grammar Notation for Controlled Natural Language in Predictive Editors
0 likes999 views
The document presents Codeco, a grammar notation for controlled natural languages in predictive editors. Codeco was designed to address limitations of existing grammar frameworks for expressing concepts important for CNLs, like lookahead features and anaphoric references. The notation is formal and declarative yet implementable in different languages. Codeco grammars can express a large subset of the Attempto Controlled English language and have been evaluated through language generation and parsing tests.
![Grammar Rules in Codeco
• Grammatical categories with flat feature structures
• Category Types:
• non-terminal (e.g. vp)
• pre-terminal (e.g. noun)
• terminal (e.g. [ does not ])
• Grammar Rule Examples:
• vp num: Num
neg: Neg
:
−→ v
num: Num
neg: Neg
type: tr
np case: acc
• v neg: +
type: Type
:
−→ [ does not ] verb type: Type
• np noun: Noun
:
−→ [ a ] noun text: Noun
• noun text: woman
gender: fem
→ [ woman ]
Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 9 / 25](https://image.slidesharecdn.com/codecotalk-130811041940-phpapp01/85/Codeco-A-Grammar-Notation-for-Controlled-Natural-Language-in-Predictive-Editors-9-320.jpg)
![Scopes
• Opening of scopes:
• Scopes in (controlled) English usually open at the position of the
scope triggering structure, or nearby.
• Scope opener category “ ” in Codeco:
quant exist: –
:
−→ [ every ]
• Closing of scopes:
• Scopes in (controlled) English usually close at the end of certain
structures like verb phrases, relative clauses, and sentences.
• Scope-closing rules “
∼
−→” in Codeco:
vp num: Num
∼
−−→ v num: Num
type: tr
np case: acc
Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 11 / 25](https://image.slidesharecdn.com/codecotalk-130811041940-phpapp01/85/Codeco-A-Grammar-Notation-for-Controlled-Natural-Language-in-Predictive-Editors-11-320.jpg)
![Position Operators
• How to define reflexive pronouns like “herself ” that can only
attach to the subject?
• With the position operator “#”, position identifiers can be
assigned:
np id: Id
:
−→ # Id prop gender: G >
id: Id
gender: G
type: prop
ref subj: Subj
:
−→ [ herself ] < id: Subj
gender: fem
s
vp
np
ref
herself
v
tv
helps
np
prop
Maryp0 p1 p2 p3
#
#
Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 12 / 25](https://image.slidesharecdn.com/codecotalk-130811041940-phpapp01/85/Codeco-A-Grammar-Notation-for-Controlled-Natural-Language-in-Predictive-Editors-12-320.jpg)
![Complex Backward References
• How to define pronouns like “him” that cannot attach to the
subject?
*John helps him.
• Complex backward references “<+... −...” have one or more
positive feature structure “+” and zero or more negative
ones “−”.
• They succeed if there is an antecedent that matches one of the
positive feature structures but none of the negative ones:
ref subj: Subj
case: acc
:
−→ [ him ] <+ human: +
gender: masc
− id: Subj
• A more complicated (but probably less useful) example:
ref subj: Subj
:
−→ [ this ] <+ hasvar: –
human: –
type: relation − id: Subj type: prop
Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 14 / 25](https://image.slidesharecdn.com/codecotalk-130811041940-phpapp01/85/Codeco-A-Grammar-Notation-for-Controlled-Natural-Language-in-Predictive-Editors-14-320.jpg)
Codeco: A Grammar Notation for Controlled Natural Language in Predictive Editors
- 1. Codeco: A Grammar Notation for Controlled Natural Language in Predictive Editors Tobias Kuhn Department of Informatics University of Zurich Second Workshop on Controlled Natural Language 15 September 2010 Marettimo (Italy)
- 2. Introduction • Problem: Existing grammar frameworks do not work out particularly well for CNLs. • Reason: • CNLs have essential differences to other languages (natural and formal ones) • To solve the writability problem, CNLs have to be embedded in special tools with very specific requirements. • Error messages and suggestions • Predictive editors • Language generation • Solution: A new grammar notation that is dedicated to CNLs and predictive editors. Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 2 / 25
- 3. CNL Grammar Requirements Concreteness: CNL grammars should be fully formalized and interpretable by computers. Declarativeness: CNL grammars should not depend on a concrete algorithm or implementation. Lookahead Features: CNL grammars should allow for the retrieval of possible next tokens for a partial text. Anaphoric References: CNL grammars should allow for the definition of nonlocal structures like anaphoric references. Implementability: CNL grammars should be easy to implement in different programming languages. Expressivity: CNL grammars should be sufficiently expressive to express CNLs. Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 3 / 25
- 4. Lookahead Features Predictive editors need to know which words can follow a partial text: Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 4 / 25
- 5. Anaphoric References • Anaphoric references in CNLs are resolvable in a deterministic way: A country contains an area that is not controlled by the country. If a person X is a relative of a person Y then the person Y is a relative of the person X. John protects himself and Mary helps him. • Anaphoric references that cannot be resolved should be disallowed: * Every area is controlled by it. * The person X is a relative of the person Y. • Scopes have to be considered too: Every man protects a house from every enemy and does not destroy ... ... himself. ... the house. * ... the enemy. Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 5 / 25
- 6. Existing Grammar Frameworks • Grammar Frameworks for Natural Languages • Head-Driven Phrase Structure Grammars • Lexical-Functional Grammars • Tree-Adjoining Grammars • Combinatory Categorial Grammars • Dependency Grammars • ... and many more • Backus-Naur Form (BNF) • Parser Generators • Yacc • GNU Bison • ANTLR • Definite Clause Grammars (DCG) • Grammatical Framework (GF) Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 6 / 25
- 7. How Existing Grammar Frameworks Fulfill our Requirements for CNL Grammars NL BNF PG DCG GF Concreteness + + + + + Declarativeness +/– + – (+) + Lookahead Features – + (+) (+) + Anaphoric References (+) – – (+) – Implementability – + – – ? Expressivity + – + + + Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 7 / 25
- 8. The Codeco Notation Codeco = “Concrete and Declarative Grammar Notation for Controlled Natural Languages” • Formal and Declarative • Easy to implement in different programming languages. • Expressive enough for common CNLs. • Lookahead features can be implemented in a practical and efficient way. • Deterministic anaphoric references can be defined in an adequate and simple way. Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 8 / 25
- 9. Grammar Rules in Codeco • Grammatical categories with flat feature structures • Category Types: • non-terminal (e.g. vp) • pre-terminal (e.g. noun) • terminal (e.g. [ does not ]) • Grammar Rule Examples: • vp num: Num neg: Neg : −→ v num: Num neg: Neg type: tr np case: acc • v neg: + type: Type : −→ [ does not ] verb type: Type • np noun: Noun : −→ [ a ] noun text: Noun • noun text: woman gender: fem → [ woman ] Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 9 / 25
- 10. Forward and Backward References The special categories “>” and “<” can be used to establish nonlocal dependencies, e.g. for anaphoric references: np : −→ det def: – noun text: Noun > type: noun noun: Noun ref : −→ det def: + noun text: Noun < type: noun noun: Noun s vp vp np ref noun area det the v tv control aux does not conj and vp np noun area det an v tv contains np >noun country det A > < Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 10 / 25
- 11. Scopes • Opening of scopes: • Scopes in (controlled) English usually open at the position of the scope triggering structure, or nearby. • Scope opener category “ ” in Codeco: quant exist: – : −→ [ every ] • Closing of scopes: • Scopes in (controlled) English usually close at the end of certain structures like verb phrases, relative clauses, and sentences. • Scope-closing rules “ ∼ −→” in Codeco: vp num: Num ∼ −−→ v num: Num type: tr np case: acc Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 11 / 25
- 12. Position Operators • How to define reflexive pronouns like “herself ” that can only attach to the subject? • With the position operator “#”, position identifiers can be assigned: np id: Id : −→ # Id prop gender: G > id: Id gender: G type: prop ref subj: Subj : −→ [ herself ] < id: Subj gender: fem s vp np ref herself v tv helps np prop Maryp0 p1 p2 p3 # # Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 12 / 25
- 13. Negative Backward References • How to define that the same variable can be introduced only once? *A person X knows a person X. • Negative backward references “ ” succeed only if no matching antecedent is accessible: newvar : −→ var text: V type: var var: V > type: var var: V Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 13 / 25
- 14. Complex Backward References • How to define pronouns like “him” that cannot attach to the subject? *John helps him. • Complex backward references “<+... −...” have one or more positive feature structure “+” and zero or more negative ones “−”. • They succeed if there is an antecedent that matches one of the positive feature structures but none of the negative ones: ref subj: Subj case: acc : −→ [ him ] <+ human: + gender: masc − id: Subj • A more complicated (but probably less useful) example: ref subj: Subj : −→ [ this ] <+ hasvar: – human: – type: relation − id: Subj type: prop Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 14 / 25
- 15. Strong Forward References • How to define that propernames like “Bill” are always accessible? *Mary does not love a man. Mary hates him. Mary does not love Bill. Mary hates him. • Strong forward references “ ” are always accessible: np id: Id : −→ prop human: H gender: G id: Id human: H gender: G type: prop Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 15 / 25
- 16. Reference Resolution: Accessibility Forward references are only accessible if they are not within a scope that has already been closed before the position of the backward reference: s ∼ vp vp ∼ np ref ... v tv destroy aux does not conj and vp pp np >n enemy det every prep from np n house det a v tv protects np n man det Every ∼ ( ( () > > < Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 16 / 25
- 17. Reference Resolution: Accessibility Strong forward references are always accessible: s ∼ vp vp ∼ np ref ... v tv likes conj and vp np pp np prop Bill prep of >n friend det every v tv knows np prop Mary ∼ ( ) < Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 17 / 25
- 18. Reference Resolution: Proximity If a backward reference matches more than one forward reference then the closest one is taken: s s ... ... np ref pn it conj then s vp np n error det an v tv causes np pp np n machine det a prep of n part det a conj If > > > < Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 18 / 25
- 19. Possible Extensions • Semantics (e.g. with λ-DRSs) • General feature structures (instead of flat ones) • ... Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 19 / 25
- 20. Parsers for Codeco Two parsers with different parsing approaches exist: • Transformation into Prolog DCG • fast (1.5 ms per sentence) • no lookahead features • ideal for regression tests and parsing of large texts in batch mode • Execution in a chart parser (Earley parser) under Java • slower, but still reasonably fast (130 ms per sentence) • lookahead features • ideal for predictive editors in Java Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 20 / 25
- 21. ACE in Codeco • Large subset of ACE in Codeco • Includes: countable nouns, proper names, intransitive and transitive verbs, adjectives, adverbs, prepositions, plurals, negation, comparative and superlative adjectives and adverbs, of-phrases, relative clauses, modality, numerical quantifiers, coordination of sentences / verb phrases / relative clauses, conditional sentences, questions, and anaphoric references (simple definite noun phrases, variables, and reflexive and irreflexive pronouns) • Excludes: Mass nouns, measurement nouns, ditransitive verbs, numbers and strings as noun phrases, sentences as verb phrase complements, Saxon genitive, possessive pronouns, noun phrase coordination, and commands • 164 grammar rules • Used in the ACE Editor: http://attempto.ifi.uzh.ch/webapps/aceeditor/ Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 21 / 25
- 22. Evaluation of ACE Codeco Exhaustive Language Generation: • Evaluation subgrammar with 97 grammar rules • Minimal lexicon • 2’250’869 sentences with 3–10 tokens: sentence length number of sentences growth factor 3 6 4 87 14.50 5 385 4.43 6 1’959 5.09 7 11’803 6.03 8 64’691 5.48 9 342’863 5.30 10 1’829’075 5.33 3–10 2’250’869 • All are accepted by the ACE parser → ACE Codeco is a subset of ACE • None is generated more than once → ACE Codeco is unambiguous Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 22 / 25
- 23. Evaluation of the Codeco notation and its implementations Prolog DCG representation versus Java Earley parser: • Equivalence of the Implementations: Generate the same set of sentences up to 8 tokens → The two implementations process Codeco in the same way • Performance Tests: task grammar implementation seconds/sentence generation ACE Codeco eval. subset Prolog DCG 0.00286 generation ACE Codeco eval. subset Java Earley parser 0.0730 parsing ACE Codeco eval. subset Prolog DCG 0.000360 parsing ACE Codeco eval. subset Java Earley parser 0.0276 parsing full ACE Codeco Prolog DCG 0.00146 parsing full ACE Codeco Java Earley parser 0.134 parsing full ACE APE 0.0161 Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 23 / 25
- 24. Conclusions Codeco ... • ... fulfills our requirements for CNLs in predictive editors. • ... is suitable to describe a large subset of ACE. • ... allows for automatic tests. • ... stands for a principled and engineering focused approach to CNL. Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 24 / 25
- 25. Thank you for your attention! Questions & Discussion Tobias Kuhn, University of Zurich Codeco CNL 2010, 15 September, Marettimo 25 / 25