Open and Closed World Types in NLP Systems. Dale Gerdemann. Seminar fur Sprachwissenschaft, Universitat Tubingen. Kl. Wilhelmstr.

Similar documents
Compiling HPSG type constraints into definite clause programs

SORT INFERENCE \coregular" signatures, they derive an algorithm for computing a most general typing for expressions e which is only slightly more comp

cmp-lg/ Jul 94

such internal data dependencies can be formally specied. A possible approach to specify

Lecture 2 - Graph Theory Fundamentals - Reachability and Exploration 1

Computational Linguistics (INF2820 TFSs)

[Ch 6] Set Theory. 1. Basic Concepts and Definitions. 400 lecture note #4. 1) Basics

P Is Not Equal to NP. ScholarlyCommons. University of Pennsylvania. Jon Freeman University of Pennsylvania. October 1989

A Simplied NP-complete MAXSAT Problem. Abstract. It is shown that the MAX2SAT problem is NP-complete even if every variable

Operational Semantics

ACHIEVING FLEXIBILITY IN UNIFICATION FORMALISMS

[LCL 94-2 ] J. van Eijck and N. Francez: Procedural Dynamic Semantics, Verb-Phrase

User-Defined Nonmonotonicity in Unification-Based Formalisms

SAMOS: an Active Object{Oriented Database System. Stella Gatziu, Klaus R. Dittrich. Database Technology Research Group

Binary Decision Diagrams

PROPOSITIONAL LOGIC (2)

Methods for Lexicon Maintenance

CSCI.6962/4962 Software Verification Fundamental Proof Methods in Computer Science (Arkoudas and Musser) Chapter p. 1/27

SOFTWARE ENGINEERING DESIGN I

Defining an Abstract Core Production Rule System

Principles of Programming Languages

Chapter 3. Set Theory. 3.1 What is a Set?

Algebraic Properties of CSP Model Operators? Y.C. Law and J.H.M. Lee. The Chinese University of Hong Kong.

H1 Spring B. Programmers need to learn the SOAP schema so as to offer and use Web services.

The Resolution Algorithm

(low) Code Area. feature symbol table type symbol table addtype table. unification table feature_val table HEAP STACK

Programming Languages Third Edition

Finding a winning strategy in variations of Kayles

A Short Summary of Javali

3.4 Deduction and Evaluation: Tools Conditional-Equational Logic

Computing least common subsumers for FLE +

Propositional Logic. Part I

Module 3. Requirements Analysis and Specification. Version 2 CSE IIT, Kharagpur

EFLUF - an Implementation of a FLexible Unification Formalism

Lecture 9 - Matrix Multiplication Equivalences and Spectral Graph Theory 1

Indexing Methods for Efficient Parsing with Typed Feature Structure Grammars

Two trees which are self-intersecting when drawn simultaneously

This example uses the or-operator ' '. Strings which are enclosed in angle brackets (like <N>, <sg>, and <pl> in the example) are multi-character symb

COMP 410 Lecture 1. Kyle Dewey

Topology Homework 3. Section Section 3.3. Samuel Otten

Computability and Complexity

9.1 Cook-Levin Theorem

The Encoding Complexity of Network Coding

Alloy: A Lightweight Object Modelling Notation

Dependent Object Types - A foundation for Scala's type system

5. Lecture notes on matroid intersection

sketchy and presupposes knowledge of semantic trees. This makes that proof harder to understand than the proof we will give here, which only needs the

Indexing Methods for Efficient Parsing

Logic Programming and Resolution Lecture notes for INF3170/4171

Unification in Maude. Steven Eker

Parameterized Complexity of Independence and Domination on Geometric Graphs

More Theories, Formal semantics

6.001 Notes: Section 8.1

A Correctness Proof for a Practical Byzantine-Fault-Tolerant Replication Algorithm

Advanced Algorithms Class Notes for Monday, October 23, 2012 Min Ye, Mingfu Shao, and Bernard Moret

Computational Linguistics (INF2820 TFSs)

A New Statement for Selection and Exception Handling in Imperative Languages arxiv: v3 [cs.pl] 24 Aug 2015

Formal semantics of loosely typed languages. Joep Verkoelen Vincent Driessen

Tidying up the Mess around the Subsumption Theorem in Inductive Logic Programming Shan-Hwei Nienhuys-Cheng Ronald de Wolf bidewolf

Hyperplane Ranking in. Simple Genetic Algorithms. D. Whitley, K. Mathias, and L. Pyeatt. Department of Computer Science. Colorado State University

CSE 20 DISCRETE MATH. Fall

A taxonomy of race. D. P. Helmbold, C. E. McDowell. September 28, University of California, Santa Cruz. Santa Cruz, CA

Constructive Coherent Translation of Propositional Logic

Exact Algorithms Lecture 7: FPT Hardness and the ETH

Taxonomic Syntax for First Order Inference. Abstract: We identify a new polynomial time decidable fragment of rst order

Higher-Order Conditional Term Rewriting. In this paper, we extend the notions of rst-order conditional rewrite systems

MA651 Topology. Lecture 4. Topological spaces 2

Lecture 15: The subspace topology, Closed sets

Module 11. Directed Graphs. Contents

Automatic Strengthening of Graph-Structured Knowledge Bases. Vinay K. Chaudhri Nikhil Dinesh Stijn Heymans Michael A. Wessel

CS323 Lecture - Specifying Syntax and Semantics Last revised 1/16/09

Detecting Logical Errors in SQL Queries

Enumeration of Full Graphs: Onset of the Asymptotic Region. Department of Mathematics. Massachusetts Institute of Technology. Cambridge, MA 02139

CS Bootcamp Boolean Logic Autumn 2015 A B A B T T T T F F F T F F F F T T T T F T F T T F F F

Localization in Graphs. Richardson, TX Azriel Rosenfeld. Center for Automation Research. College Park, MD

Lexical Considerations

The Programming Language Core

Computing intersections in a set of line segments: the Bentley-Ottmann algorithm

Flexible Structural Analysis of Near-Meet-Semilattices for Typed Unification-based Grammar Design

SETS. Sets are of two sorts: finite infinite A system of sets is a set, whose elements are again sets.

Outline. Computer Science 331. Information Hiding. What This Lecture is About. Data Structures, Abstract Data Types, and Their Implementations

Module 6. Knowledge Representation and Logic (First Order Logic) Version 2 CSE IIT, Kharagpur

Topic 1: Introduction

CMSC 330: Organization of Programming Languages

Bipartite Roots of Graphs

Mathematically Rigorous Software Design Review of mathematical prerequisites

Core Membership Computation for Succinct Representations of Coalitional Games

Liveness and Fairness Properties in Multi-Agent Systems

Combinatorial Optimization

Foundations of AI. 9. Predicate Logic. Syntax and Semantics, Normal Forms, Herbrand Expansion, Resolution

arxiv:cs/ v1 [cs.cl] 7 Jul 2000

MAX-PLANCK-INSTITUT F UR. Ordered Semantic Hyper-Linking. David A. Plaisted. MPI{I{94{235 August 1994 INFORMATIK. Im Stadtwald. D Saarbrucken

Reconciling Dierent Semantics for Concept Denition (Extended Abstract) Giuseppe De Giacomo Dipartimento di Informatica e Sistemistica Universita di Ro

Ramsey s Theorem on Graphs

AXIOMS FOR THE INTEGERS

Well-formed Default Unification in Non-deterministic Multiple Inheritance Hierarchies

From Types to Sets in Isabelle/HOL

Implementation of Axiomatic Language

Propositional Logic. Andreas Klappenecker

Complexity Theory. Compiled By : Hari Prasad Pokhrel Page 1 of 20. ioenotes.edu.np

Transcription:

Open and Closed World Types in NLP Systems Dale Gerdemann Seminar fur Sprachwissenschaft, Universitat Tubingen Kl. Wilhelmstr. 113 72074 Tubingen, Germany Email: dg@sfs.nphil.uni-tuebingen.de 1 Introduction The use of an open-world versus a closed-world type system has often been considered as a criterion upon which to evaluate NLP systems. I will argue in this paper that this is a false criterion of evaluation. The essential dierence between the two kinds of systems is that in a closedworld approach all types are declared, whereas in an open-world approach some types may be left implicit. The open-world approach may be simulated in a closed-world system such as troll [6] simply by lling in all of these implicit types. Open-World systems cannot, however, always be used to simulate a closed-world system. In some open-world systems (ale [2] and T DL [10]), it is simply impossible to rule out all of the unwanted implicit types. Thus a closedworld system such as troll is strictly more expressive than these systems. But this does not mean that closed-world systems in general are better. It simply means that open-world systems need to designed so that implicit types can be ruled out when the grammar writer so desires. The cuf system, for example, is open world. But it uses type axioms which are quite sucient to rule out any implicit types. Thus the open-world type system of cuf and the closed-world types of troll are equally expressive. The organization is as follows. In x2, I identify two kinds of closed-world assumptions, the partition condition and the disjoint species condition. These are discussed with respect to the systems ale and T DL in x3 and x4 respectively. In x5, I show that the type axioms of cuf can be translated into a closed-world type hierarchy and vice versa. Conclusions are in x6. 2 Closed vs. Open World The terms closed world and open world are most familiar from Prolog and logic programming but as applied to type hierarchies they have a somewhat extended meaning. In Prolog, the closed-world hypothesis means essentially that the Prolog data base is complete. Given the following data base we can conclude that John, Fred and Mary are the only students in our domain: student(john). student(mary). student(fred).

Similarly, this data base could be expressed as a type hierarchy: 1 john fred mary " " "" " "" student For the type hierarchy representation, however, we may identify two dierent kinds of closed-world assumptions [7] [8]. The partition condition asserts that if an object is of type t then the object is of at least one species 2 subsumed by t and the disjoint species condition asserts that the object is of no more than one species subsumed by t. So the partition condition rules out the possibility of some person other than John, Fred or Mary being a student. And the disjoint species condition rules out the possibility of John and Fred being the same person. The term \open-world assumption" is loosely applied to a system that violates either one of these assumptions. 3 The Exhaustive Species Condition In ale, an appropriateness partial function Approp: Type Feat * Type must be specied and a feature structure 3 is deemed to be well-typed i for all nodes q in the feature structure, if (q) = t and (q; f) is dened, then approp(t; f) is also dened and approp(t; f) v ((q; f)). The problem with the well-typedness test is that it allows feature structures that are (in Carpenter's terms) \ungroundable". 4 Suppose, for example, there is a type a[f : bool; g : bool], where the appropriateness specication is stated in the square brackets. And suppose further that a has subtypes a 0 [f : +; g :?] and a 00 [f :?; g : +]. With these types, the feature structure a[f : 1 ; g : 1 ] is well typed. But this is a peculiar result. What kind of object would this feature structure describe? Clearly, it describes an object of type a which is neither an a 0 nor an a 00. So there must be a third species subsumed by a, which was not included in the hierarchy. It should be noted that maintaining well-typedness represents an attempt at maintaining groundedness. Why would any system perform well-typing otherwise? The well-typing test works, however, only as a sucient test for excluding ungroundable feature structures, i.e., if a feature structure is not well-typable, then it is ungroundable. Thus it is possible to nd cases where ale would exclude an ungroundable feature structure where a system like T DL, which does 1 The hierarchy here follows the \information ordering" as in [1] and [12]. In terms of a \types-as-sets" denotational semantics, increasing information content decreases the denotation, i.e., if t v t 0 then [t] [t 0 ]. Not all systems explicitly employ this \types-as-sets" semantics. However, such a semantics can provide a useful common ground for comparison. 2 A species is a maximally specic type. 3 A feature structure is a Moore machine, i.e., a nite state automata with a type labelling each node. I will use as the transition function and as the labelling function. 4 In fact [8] introduced the closed-world assumptions into the troll system precisely in order to ensure that troll feature structures would be groundable. For a proof that troll feature structures are groundable, see [9].

not enforce run-time well-typedness would not. Suppose, for example, that the type t[f : bool; g : bool] has the subtypes t 0 [f : +; g : +] and t 00 [f : +; g :?]. Furthermore, assume that there is a type s that subsumes t 00 and u (with no features appropriate for s and u). In ale, it is not possible for t[f : 1 bool; g : 1 ] to unify with s. In T DL, on the other hand, this unication is allowed. Thus the T DL system assumes that there is another implicit type subsumed by both t and s, which is impossible to exclude (Krieger, p.c.). It is, in fact, possible to translate an ale hierarchy to a closed-world hierarchy by lling in all implicit types. First dene a depth-one feature structure of type t as a well-typed feature structure for which (q; f) is dened just in case: q = q (the root), (q) = t, and approp(t; f) is dened. A depth-one feature structure F of type t is minimal if there exists no distinct F 0, also a depth-one feature structure of type t, such that F 0 v F. 5 Similarly, a depth-one feature structure F of type t is maximal if there exists no distinct F 0, also a depth-one feature structure of type t, such that F v F 0. 6 So, for an example based on the above hierarchy and appropriateness specication, a[f : bool; g : bool] is a minimal depth-one feature structure of type a and a[f : +; g :?] is a maximal depth-one feature structure of type a. The feature structure a[f : bool], on the other hand, fails to be depth-one since the feature g is missing. Now, if there exists a maximal depth-one feature structure of type t which is not uniable with any minimal depth-one feature structure of type t 0 such that t v t 0, then add a new species s (a previously unused type name), such that t (and only t) directly subsumes s. Dene the appropriateness specication of s identically to that of t, i.e., for each f appropriate on t, let approp(s; f) = approp(t; f) and let all other features be inappropriate for s. 7 It is clear that all implicit types can be lled in by this method. So there is a translation from an ale hierarchy to the troll closed-world hierarchy. However, it is also clear that there is no general transformation in the reverse direction. If we wanted the type a to have just a 0 and a 00 as subtypes as in the above hierarchy, then there is no way to express this in ale. 8 4 The Disjoint Species Condition In addition to the exhaustive species condition, some systems, such as T DL [10], violate the disjoint species condition as well. The sort of open-worldness that 5 I assume that feature-structure subsumption is dened in terms of a homomorphism between typed feature structures as in [1] and [11]. 6 Note that a maximal depth-one feature structure may contain reentrancies. 7 This is somewhat complicated since the ale hierarchy is only open world when the appropriateness specication forces it to be. As a simpler example, suppose we give the student hierarchy above a non-exhaustive interpretation. Then to make a closed-world hierarchy, we would simply make one more species with a name like other-students and we allow this new species to be directly subsumed only by the type student. 8 See [7] [8] for discussion of how this limitation of ale aects the ability to express feature cooccurrence restrictions. Copestake et al [3] also encountered this limitation in a large scale linguistic application.

this violation produces does not, however, aect the expressivity of the type hierarchy. Such a type hierarchy can be translated into a closed-world hierarchy simply by lling in implicit types. First, let us call the maximal types in a such a hierarchy pseudo-species. The species in the closed-world equivalent of such a hierarchy will be the consistent sets of pseudo-species, where \consistent" means that there are no disjointness statements between any two pseudo-species. 9 A pseudo-species will subsume one of these newly constructed species just in case the pseudo-species is a member of the species. Then we need to ensure that the closed-world type hierarchy is bounded complete. For any two non-maximal types t and t 0, let I = fs j t v sg \ fs j t 0 v sg. If I = ;, then t and t 0 are inconsistent. If I = fsg, a singleton set, then the least upper bound of t and t 0 already exists. And if I = S, a non-singleton, non-empty set, and if there is no non-maximal type t 000 which subsumes just the species in S, then introduce t 000, a previously unused type name, and let t 000 subsumes just the species in S. This provides a translation into a closed-world hierarchy. Translation in the reverse direction is even easier: simply provide a disjointness statement between each pair of species. So the question of whether or not a system should maintain the disjoint species condition is really a question of whether consistency of types should be stated positively or negatively. The way in which user species the type hierarchy, however, in no way aects the underlying logic of the system. It is simply a user interface issue. 5 Boolean Type Axioms and Type Hierarchies Finally, we come to the method of using boolean expression as type axioms. This approach was used by [4] and incorporated into the cuf system [5]. The idea is that the interpretation of a type is determined by the denotation of the type axioms, which is given as follows: [[T ]] I := T I (T is a primitive type term) [[t ^ s]] I := [[t]] I \ [[s]] I [[t _ s]] I := [[t]] I [ [[s]] I [[:s]] I := D I? [[s]] I (D is the domain) The use of such type axioms might well be considered open-world. Dorre and Eisele [5], themselves emphasize this point: In cuf any interpretation of the types which is not invalidated by the axioms is possible. Hence, in the absence of other information, intersections of types are note considered to be empty. The Boolean axioms of cuf are, however, quite expressive. So they can be used to rule out unwanted unications or implicit types that may be impossible to rule out in other open-world systems. Thus, it will be possible to translate in both directions between type axioms and closed-world type hierarchies. 9 Typically, systems that violate the disjoint species condition allow the user to make such disjointness statements.

Translation from a closed-world type hierarchy into a conjunctive set of boolean axioms is trivial. First translate each direct subsumption relation t v t 0 into the formula :t 0 _ t. If the set of species is fs 1 ; : : :; s n g then the exhaustive species condition can be translated as? = : s 1 _ : : : _ s n, where s = : t is an abbreviation for (:s _ t) ^ (s _ :t), and the disjoint species condition can be translated as f:(s ^ s 0 ) j s; s 0 2 fs 1 ; : : :; s n gg. To translate in the reverse direction, rst convert the type axioms into disjunctive normal form. Then ensure that this DNF is maximal, i.e., if for some type term s, a disjunct d = (t 1 ^: : :^t n ) contains neither the conjunct s nor :s, then replace d by the two disjuncts (t 1 ^: : :^t n ^s) and (t 1 ^: : :^t n ^:s). Each of these disjuncts is now as fully specied as possible given the type terms used in the type axioms. The set of such disjuncts represents the set of species allowed by the type axioms. The actually allowable species are, however, specied by interaction between the type axioms and the feature declarations, so some of these potential species must be ruled out because of inconsistent feature declarations (see [5]). The set of species so constructed, however, does not constitute a full set of species since cuf also allows undeclared afs (atomic feature structure) types, which should also be considered species. A list of these undeclared types could be obtained, for example, by searching through the grammar to see what undeclared types are actually being used. 10 We could then form a power-set lattice of these species as the full type hierarchy. However, this would produce more non-maximal types than necessary given the type terms used in the type axioms. It would be more desirable to be able to reinterpret the input type terms as types in the type hierarchy. 11 Let us say that if a type term occurs as a positive literal in a species, then that species contains that type term. Let spec be the function from type terms to the set of species that contain that type term. If for some type terms t, t 0 : spec(t) = ;, then exclude t from the hierarchy. spec(t) spec(t 0 ), then let t 0 subsume t in the type hierarchy. spec(t) = spec(t 0 ), then arbitrarily let t subsume t 0 or t 0 subsume t. 12 Finally, the type hierarchy needs to be closed under least upper bounds. If, for any two types t, t 0 in the type hierarchy, it is the case that there is no type t 00 such that spec(t 00 ) = spec(t) \spec(t 0 ), then let t 000 be a new type name, let both t and t 0 subsume t 000 in the type hierarchy and let spec(t 000 ) = spec(t) \ spec(t 0 ). 6 Conclusion I have shown in this paper that open-world and closed-world type hierarchies are potentially interchangeable. It is possible to translate in both directions between open-world cuf-style type axioms and a troll-style closed-world type hierarchy. So the choice between these two approaches is arbitrary. One could even imagine a system that allowed both forms of grammar specication. 10 The few built-in types of cuf can also be translated on an ad hoc basis. 11 This correspondence will be necessary if we want to be able to translate the feature terms used in a cuf grammar into another formalism using a type hierarchy. 12 This is an exclusive or. As long as t and t 0 are on a unary branch, they will have the same denotation, regardless of which one subsumes the other [1].

On the other hand, some open-world systems (ale, T DL) have been shown to be seriously lacking in expressiveness. But the problem with these systems is not the open-world assumptions per se. It is the fact that typing system for these programs is not strong enough to exclude an open-world interpretation in those cases where this interpretation is not desirable. A system using closed-world types, on the other hand, can always simulate an open-world interpretation simply by lling in all implicit types. References [1] Bob Carpenter. The Logic of Typed Feature Structures. Cambridge Tracts in Theoretical Computer Science 32. Cambridge University Press, 1992. [2] Bob Carpenter and Gerald Penn. ALE The Attribute Logic Engine, User's Guide, 1993. [3] Ann Copestake, Antonio Sanlippo, Ted Briscoe, and Valeria de Paiva. The ACUILEX LKB: An introduction. In Inheritance, Defaults, and the Lexicon, pages 148{163. Cambridge University Press, 1993. [4] Michael Dorna. Erweiterung der Constraint-Logiksprache cuf um ein Typsystem. Master's thesis, Institut fur Informatik, Universitat Stuttgart, 1992. Diplomarbeit Nr. 896. [5] Jochen Dorre and Michael Dorna. cuf a formalism for linguistic knowledge representation. In Jochen Dorre, editor, Computational Aspects of Constraint-Based Linguistic Description, Dyana Deliverable R1.2.A, DYANA-2 ESPRIT Basic Research Project 6852, pages 8{22, 1991. [6] Dale Gerdemann and Thilo Gotz. Troll: Type resolution system, user's guide. Technical report, University of Tubingen, 1994. [7] Dale Gerdemann and Paul John King. Typed feature structures for expressing and computationally implementing feature cooccurrence restrictions. In Proceedings of 4. Fachtagung der Sektion Computerlinguistik der Deutschen Gesellschaft fur Sprachwissenschaft, pages 33{39, 1993. [8] Dale Gerdemann and Paul John King. The correct and ecient implementation of appropriateness specications for typed feature structures. In COLING 94, Proceedings, pages 956{960, 1994. [9] Paul John King. Typed feature structures as descriptions. In COLING 94, Proceedings, pages 1250{1254, 1994. [10] Hans-Ulrich Krieger and Ulrich Schafer. T DL a type description language for constraint-based grammars. In COLING-94, pages 893{899, 1994. [11] Michael Moshier. Extensions to Unication Grammar for the Description of Programming Languages. PhD thesis, University of Michigan, 1988. [12] Carl Pollard and Ivan Sag. Head-Driven Phrase Structure Grammar. CSLI Lecture Notes. Chicago University Press, Chicago, 1994.