This paper describes an interactive graphical environment for computational semantics. The system provides a teaching tool, a stand alone extendible grapher, and a library of algorithms together with test suites. The teaching tool allows users to work step by step through derivations of semantic representations, and to compare the properties of various semantic formalisms such as Intensional Logic, DRT, and Situation Semantics. The system is freely available on the Internet.
The CT,EARS tool (Computational Linguistics Education and Research Tool in Semantics) was developed as part of the FraCaS project 1 which aimed to encourage convergence between different semantic formalisms. Although formalisms such as Intensional Logic, DR?l', and Situation Semantics look different on first sight, they share many common assumptions, and provide similar treatmeats of many phenomena. The CLEARS tool allows exploration and comparison of these different formalisms, enabling the user to get an idea of the range of possibilities of semantic construction. It is intended to be used as both a research tool and a tutoriM tool.
The first part of the paper shows the potential of the system for investigating the properties of different seinantic formMisms, and for teaching students formal semantics. The next section outlines the library contents and the system architecture, which was designed to reflect convergence between theories. The result is a highly modular and, we beliew~, a highly flexible system which
1A lh'amework tor Computational Semantics, F,uropean Community LRE 62-051. s app(1,2) vp id I v la!ghs (B) )~C, laughs(C) laughs np id I pn anna \] XA.~®A Figure 1: Initial ffepresentation of anna with Aq)l{l' allows user prograrns to be integrated at various levels. The final part of the paper describes the grapher which was designed as a stand alone tool which can be used by various applications. 2 A Tutorial System for Computational Semantics As a tutorial tool, CI, PArtS allows students to investigate certain tbrmalisms and their relationship. It also provides the possibility for the teacher to provide interactive demonstrations ami to produce example slides and handouts.
In this section we show how a user can interactively explore the step-by-step construction of a semantic representation out of a syntax tree. Figures 1 and 2 show a possible initial display for the sentence "Anna laughs" in a compositional version of I)RT (Bos et al., 1994) and in 'Montague Grammar' (Dowty et al., 198:1).
The user controls the semantic construction process by moving to particular nodes in the derivation tree, and performing operations by using mouse double-clicks, or by selecting froln a pop-up menu. For example, clicking on app(2,1) s app(2,1) J C Ioves(C,J) vp id pn v f I anna laughs anna XA.laughs(A) np id man(C) woman ( J ) VC.(man(C) ~ 3J.(Ioves(C,J) ^ woman(J))) Figure 4: 'I'ranslating I)I{T to Predicate Logic li'igure 2: Initial Representation of Anna laughs with 'Montague Grammar' s laughs(anna) I .. app(2,1) /N np vp I irA.laughs (A) )n v nna i a!qhs anna kA.laughs (A) -Figure 3: Final Representation of Anna laughs in ' Montague-Grammar' in the tree shown in l?igure 2 has the effect of applying the lambda-ext)ression lA.laughs(A) to anna. The resulting display is given in t,'igure 3.
The poI)-up menu allows a user to pertbrm single derivation steps. For example, the user can first form an application term AA.hmghs(A)(anna) and then reduce this at the next step. Menu options include the possibility of cancelling intensional operators, performing lmnbda reduction, applying meaning postulates, and \[)RS merging. The glenn also allows a user to choose whether or not to perform quantifier storage or discharge, and thereby pick a particnlar reading for a sentence. Alterxlatively the user can choose to fully process a node, in which case all readings are simultaneously displayed. 3 Comparing Theories A major use of the tool is for comparison of different semantic theories and methods of semantic construction. To akl comparison of theories, there are translation routines between some semantic tbrmalisms. For example, \],'igure 4 shows a translation from a D|{S to a formula in Predicate Logic.
The user can try out various options for semantic construction by using a menu to set various parameters. An illustrative subset of the parameters and their possible va.lues is given below: semantic forntalism
l,ogic of Generalized Quantitiers,
lntensional Logic,
Compositional 1)RT (Muskens, 1993),
Aq)R'F (Bos et al., 1994),
'lbp-l)own-Dl{T (Kamp and Reyle, \[993),
Situation Semantics. granllnar
simple PSG, PSG with features,
Categorial Grammar with features. parser
top-down, incremental (for CG only). lexicon
simple lexicon, lexicon with features. syntax-semantics mai)plng
rule-to-.rule, syntactic template. syntax-semantlcs ('onstruetlon
serial, parallel. subject applied to verb phrase
yes, no. quantifier storage me(:hanism
Cooper Storage (Cooper, 1983),
Nested Cooper Storage (Keller, 1988) fl-reduction
unification based, substitution based. 4 The Library Because a tutorial system of this kind has to be based largely on standard routines and algorithms that are fundamental for the area of computational semantics, a secondary aim of the project was to provide a set of well documented programs which could form the nucleus of a larger library of reusable code for this field. Most of the library contents correspond directly to particular values of parameter settings. However there are some extra library routines, for example a very generalised form of flmction composition. The library is being expanded with routines for semantic construction driven by semantic types. It is also intended to integrate a wider range of grammars, parsing strategies and pronoun resolution strategies. For program documentation we largely have followed the approach taken in LEDA (Ngher, 1993)).
Apart from the routines concerned directly with computational semantics, there are also routines designed to aid application developers who want to provide a graphical output tbr semantic representations. These routines are mainly concerned with translating from Prolog syntax into the description string syntax used by the CLiG grapher. Currently they rely on the Tcl/Tk library package provided by Sicstus 3.
A standard approach to modularisation is to split a problem into independent black boxes, e.g. a grammar, a parser etc. This top-down modularisation is then followed by some bottom-up modularisation in the sense of supplying general utilities which each of the larger modules can use. For this application, such an approach had obvious inadequacies. For example, there are subtle differences in some steps of quantifier storage according to the formalism being used, similarly, differences even in lambda reduction (for intensional ogic it is natural to interleave the step of operator caneellation between/?-reductions). Even the parsing stage cannot be totally independent unless we generalise to the worst case (the Situation Semantics fragment requires an utterance node as well as a sentence node).
One of the aims in building the tool was to show where semantic formalisms converge. Thus there was theoretical motivation to ensure components of the system were shared wherever possible. There was also practical motivation, since there is more chance of finding errors in shared code. The solution adopted was to use parameterised modularisation. This allows differences to be located in as small pieces of code as possible (e.g. single lines Ii00 I parameterised node formation 1 I semantic construction I I parameterised extraction from nodes 1 Figure 5: Architecture of a pm't of the Syntax-Semantics lnt, erface of tile quantifier storage routine), with the parameters picking up the correct; piece of code at run time. There are some small costs due to indirection (instead of calling e.g. a /?-reducer directly, a program first calls a routine which chooses the /?-reducer according to the parameters). But with these parameterisation layers we provide natural points where the system can be extended or modified by the user. The approach also gets rid of the need to create large data structures which include information which would be relevant for one choice of parameters, but not the current choice. For example, in parsing, a parameterised level chooses how to annotate nodes so that the syntax trees only have the relevant inibrmation for the chosen syntax-semantics strategy. The architecture is illustrated in Figure 5.
The result of the parameterised approach is a system which provides several thousand possible valid combinations of semantic tbrmalism, grammar, reducer etc. using a small amount of code.
A major part of our work on the educational tool was the development of a general graphical browser or grapher for the graphical notations used in computational linguistics, especiMly those in computational semantics such as trees, Attribute-Value-Matrices, EKN (Barwise and Cooper, 1993) and 1)RSs. The grapher was written in Tcl/Tk, a programming system tbr developing graphical user interfaces (Ousterhout, 1994). Two attrilmtes of Tel/Tk which were important lbr this applieattion were the l)rowision of translation routines from graphic canvasses into Postscript (allowing generation of diagrams such as Figures 1 to d), and the ease of providing scaling routines for zooming.
The grapher was designed to be extendible for future al)plications. Graphical structures are described using a (les(:ril)tion stritlg, a. plain text hi-erarchical description of the object to be drawn without any exact positioning information, l,'or example, the following tree: S A is created by the description string: {tree {plain-text "S"}
{plain-text "NP"}
{plain-text "VP"}} CLIG Call display hale,active graphical slA'llcl;llres which aJlow tim user to perform actions by clicking on mouse-sensitive r gions ill the display are;~. The grapher and an underlying application therefore can behaw.' in a way that the grapher is not only a way to visual*st the data of t;he application, but also providc.s a real interface I)etween user and af)plication. 6 Availability of the System The system ('urrently requires Sicstus 3 plus '\['cl version 7.d and 'l'k w;rsion 4.0 (or later versions), lit, is awfilablc at the' ftp address: ftp.coli.uni-sb.de:/pub/fracas or on the WWW at the UI/J,: http ://coli. uni-sb, de/~ clears/clears, html l;urther (toeumentation of the' system is given in (l,'raCaS, 1996a) and (FraC, aS, 1996b), which are available from: http://www, cogsci.ed, ac.uk/~fracas/ 7 Conclusion Initial reactions to demonstrations of the educational tool suggest that it has the potential to become a widely used educatioual aid. We also believe that the programs iml~lemented and documented it* this work provide the nucleus of a larger library of rensab\[e programs for computational semantics. Our current plans a.re to test t;\[l(', system ii01 with a wide (:lass of users to discover areas requiring extension or modification. A longer term aim is to integrate the system with existing grammar develol)ment environments. Acknowledgements 'l'his work would not have been I)ossible without the encouragement and support of the other men> hers of the l"ra(~aS Project. We. would especially like to thank Ih)bin (,'ooper, Mass*me Poe.sio and Steven lhdman for eontril)utions to the code. References .J. Barwise and R. Cooper. 1993. Extended l(anq~
notation. In Y. I(atagiri P. A(:zel, 1). Israel and
S. Peters, editors, ,Situation 7'heorg and its Ap-
plication Vol. 3, chapter 2, t)ages 29 54. CSIA,
Stanford. J. Bos, E. Mastenbroek, S. McClashan, S. Mil-
lies, and M. l'inkal. 1994. A compositional
l)i{S-based \['orm~dism for nip applic;~tions. In
Proceedings of the International Workshop
on Computational ,5'cmanlics, pages 21 3 ,
Tilbucg. 11. (JOOl)er. 1983. Q'uanlificalion and 5'gnlactw
Tttcory. SI,AI'. l{eidel, 1)ordrecht. I). l)owty, R. Wall, and 8. Peters. 1981. Intro-
duction I.o Monla(lUC ,%:man*its. SI,A I'. t/x~ide\[,
I)ord vecht. FraCaS. 1996a. Ih,il(ling the framework, l"racas
I)eliver~d)lc l)J 5. l"ra(:aS. 1996b. I.Jsing tim framework. Fracas
Deliverable 1)16. 1l. l(amp and 11. lt.(:yle. 1993. l')zmt I)iscours(' to
Logic. Kluwer, l)ordrecht. W. Keller. 11988. Nested cooper storag< In
U. Reyle and C. Rohrer, editors, Natural Lan-
(tltag('~ Parsing and Ling.uistic Theories, pages
432 447. i/,eidel, 1)ordreeht. If. Muskens. 1993. A compositional discoul:se
representation theory. In P. l)ekker and
M. Stokhof, editors, I'rocecdings of the 9th A'mo
sterdam Colloquium, pages 467 48(5. IIAX\], Uni-.
versity of Amsterdam. S. Niher. 1993. Le(la manuM version 3.11. 'Ib.chni-
eel l~eport MP1-I-93-109, IHa.x-Planck-lnstitut
fiir Infornmtik, S~.~arbriicken, I"ebru~ry. a. Ousterhout. 1994. "Fcl and the 7'k 7'oolkii. Pro--
fessional Computing. Addison-Wesley, I{eading,
Massachusetts.