Phrasing a text in terms the user can understand - CiteSeerX

Phrasing a text in terms the user can understand* J o h n A . B a t e m a n and C e c i l e L . P a r i s USC/Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292-6695 Abstract W h e n humans use language, they show an essential, i n b u i l t responsiveness to their hearers/readers. W h e n language is generated by machine, it is s i m i l a r l y necessary to ensure t h a t t h a t language is a p p r o p r i a t e for its intended audience. M u c h of previous research on text generation and user m o d e l l i n g has focused on b u i l d i n g a user m o d e l and selecting appropriate i n f o r m a t i o n f r o m t h e knowledge base to present to the user. It is i m p o r t a n t , however, t h a t the phrasing of a t e x t be also t a i l o r e d to the hearer - otherwise it m a y be j u s t as ineffective as texts w h i c h w r o n g l y direct a t t e n t i o n or which rely on knowledge t h a t the hearer does n o t have. T h i s research proposes a new mechanism which allows t h e t e x t p l a n n i n g process to specifically t a i l o r s y n t a c t i c phrasing to the hearer t y p e . T h i s is done in the context of an expert system e x p l a n a t i o n f a c i l i t y t h a t needs to produce e x p l a n a t i o n s of the expert system's behavior for a v a r i e t y of different users - users w h o differ in goals, expectations, and expertise concerning b o t h t h e e x p e r t system and its domain.

1

T a i l o r i n g - t h e i m p o r t a n c e of m a k i n g language a p p r o p r i a t e f o r i t s audience

Humans show an essential, i n b u i l t responsiveness to their hearers/readers in t h e i r use of language. It is s i m i l a r l y necessary to ensure t h a t the language generated by machine is a p p r o p r i a t e for its intended audience. M u c h t e x t generation research in the past has focused on the selection of t e x t content a n d organization in order to accomplish speakers' goals (e.g., [ M c K e o w n , 1985, I l o v y , 1988]). T h e presentation of t h a t content is generally also made responsive to hearers' states of focus of a t t e n t i o n (e.g., [Grosz a n d Sidner, 1986]). More recently, it has been recognized t h a t it is necessary in a d d i t i o n *The research described in this paper was supported in part by the Defense Advanced Research Projects Agency (DARPA) under a N A S A Ames cooperative agreement number NCC 2-520, by A F O S R contract F49620-87-C-0005, and by DARPA contract MDA903-87-C-641.

to make generated t e x t sensitive to the hearers' goals and knowledge a b o u t domains, t h a t is to take a user m o d e l i n t o consideration (e.g., [ M c K e o w n et a/., 1985, A p p e l t , 1985, Jameson, 1987, I l o v y , 1988, Paris, 1988, C a r b e r r y , 1988]). These are all i m p o r t a n t factors if the t e x t generated is to be b o t h i n f o r m a t i v e a n d understandable to the user. T h e language used by specific groups of people, however, often possesses syntactic patterns and lexical features t h a t are d i s t i n c t i v e to those groups; question answering systems can o n l y be effective, therefore, if they a p p r o p r i a t e l y customize their phrasing as well as their content and t e x t u a l o r g a n i z a t i o n according to each dist i n c t i v e g r o u p of users - otherwise generated texts may be j u s t as ineffective as texts w h i c h w r o n g l y direct att e n t i o n or w h i c h rely on knowledge t h a t the hearer does n o t have. In contrast to most previous research, w h i c h has focused on the selection and organization of inform a t i o n f r o m the knowledge base for presentation to the user, t h i s research addresses the issue of expressing the selected i n f o r m a t i o n in language specifically tailored to the hearer. 1 T h i s is done in the context of an expert system e x p l a n a t i o n facility.

2

W h a t is involved in ' t a i l o r i n g '

In t a i l o r i n g a response, whether it be to a user's goals for asking a question or to t h a t user's level of expertise in a d o m a i n , a generation system first has to choose w h i c h i n f o r m a t i o n f r o m the knowledge base at h a n d is most a p p r o p r i a t e and organize its overall text struct u r e . T h e result of this phase is an organized collect i o n of the p a r t i c u l a r propositions to be expressed in E n g l i s h . M o s t generation systems take these propositions as the i n p u t s for the realization components of g r a m m a t i c a l a n d lexical selection (e.g., [ M c K e o w n , 1985, M o o r e a n d S w a r t o u t , 1989]). However, the o u t p u t of this phase is t y p i c a l l y not detailed enough to c o n t r o l the many possibilities for expression t h a t current g r a m m a r components provide. There is a large gap between the 1

This issue was partially addressed in the HAM-ANS system [Morik, 1985], where, based on a user model, the system would decide whether to produce an anaphora. The work presented here is different as we are more concerned about systematic linguistic differences that exist in the language used by various groups of users.

Bateman and Paris

1511

level of detail of the o u t p u t of the text planner and the i n p u t required by a grammar. T h i s is problematic for n a t u r a l text generation for two reasons. First, generated texts could unintentionally be inappropriate in many contexts; second, a text planner's lack of control over the g r a m m a r could prevent it f r o m producing the text it intends to generate. Because different classes of users have different ways of speaking, involving systematic differences in the phrasal and grammatical patterns t h a t they employ, a generator needs to have theoretical control of phrasal organization j u s t as it does of text organization. T h e generation of natural text requires a second phase of planning t h a t bridges this gap. T h i s paper presents a mechanism t h a t achieves this. Control of the phrasing task has mostly remained below the level of detail t h a t the text planning process has under theoretical control. Phrasing involves choosing appropriate syntactic and lexical structures f r o m all the available possibilities t h a t express the same propositional content. The few attempts to gain control of phrasing in accordance w i t h general text planning goals (e.g., [Appelt, 1985, M o r i k , 1985, Hovy, 1988]) have either been restricted in the areas of phrasing they have examined, or have not maintained sufficient separation between text planning and realization. 2 In our work, we are constructing a general phrasing control component t h a t interfaces between the o u t p u t of the first phase of planning and the input to the grammar. ( T h i s is illustrated in Figure 1.) T h i s component decides b o t h which 2

I t is generally accepted that a text planner should not maintain detailed knowledge of the grammatical possibilities offered by the grammar; to do so complicates the planning process considerably by requiring the text planner to concern itself with details from an inappropriately low level of abstraction. Similarly, the grammar should not include detail at an inappropriately 'high' level of abstraction.

1512

Speech and Natural Language

aspects of the o u t p u t of phase one are most appropriate for each user type and how they are to be phrased for t h a t user. It is only the result of this second phase of selection/organization t h a t provides sufficient guidance to the g r a m m a r and lexical selection components. We use the n o t i o n of register [Halliday, 1978, Patten, 1988], which specifies the linguistic consequences of using language in particular situations. Registers may be seen as describing the 'argots' used by different classes of users. Our approach has been to restrict our attent i o n to the particular kinds of language required in our domain. We have worked back f r o m these to construct sets of 'terms' to be employed when generating language for particular users. T h i s is equivalent to a linguistic description of specifically the situation-specific aspects of some instances of language use. We can then define a phrasing a l g o r i t h m which uses these sets to effectively constrain the second phase of the generation process.

3 3.1

T h e e x p e r t system explanation task The expert system

The system for which we are generating language is an expert system constructed using the Explainable Expert System ( E E S ) framework [Swartout and Smoliar, 1987]. In this framework, an expert system includes support knowledge for explanation. T h e support knowledge contains, for example, terminology definitions, so t h a t a user can ask for the definition of specific terms used by the system and conditions of evaluation for states of affairs w i t h i n the d o m a i n (i.e., whether a particular state is good or bad). We are using a particular instantiation of this framework, an expert system designed to diagnose digital circuits. T h i s system includes in its domain knowledge a variety of facts concerning objects in the domain of dig-

i t a l c i r c u i t s , how they m a y be related, a n d w h a t can go w r o n g w i t h t h e m . W h i l e e x p l a i n i n g i t s reasoning, the system sometimes needs to provide a d e f i n i t i o n of some d o m a i n o b j e c t , such as a f a u l t y - s y s t e m . To do so, the e x p l a n a t i o n r o u t i n e f i r s t examines a l l t h e i n f o r m a t i o n contained in the knowledge base about f a u l t y systems to d e t e r m i n e w h a t to include in the t e x t ; this is phase one of t h e generation process (cf. Figure 1). Suppose the gene r a t i o n r o u t i n e decides to select o n l y the conditions u n der w h i c h a system is f a u l t y (called defining conditions). T h i s i n f o r m a t i o n (shown in Figure 2) is contained in one slot of the o b j e c t f r a m e . It is represented as predicate calculus a n d constitutes the basic p r o p o s i t i o n a l content t h a t is to be expressed in the response. T h i s i n f o r m a t i o n is highly structured. T h e structure is defined by the EES g r a m m a r w h i c h specifies the permissible predicate calculus f o r m u l a e . Such definitions are parsed according to t h i s g r a m m a r so t h a t specific tools, i n c l u d i n g a syntact i c m a t c h e r , can be used to traverse the parse tree and p r o v i d e access to its constituents. G i v e n t h i s p r o p o s i t i o n a l content, however, it is s t i l l possible to phrase this d e f i n i t i o n in several ways, each b e i n g a p p r o p r i a t e for some different t y p e of user. 3.2

T h e t y p e s of language a n d users r e q u i r e d

T h i s expert system is expected to s u p p o r t interactions w i t h at least t h e f o l l o w i n g groups of users: G r o u p 1 : System developers w h o w a n t to make sure t h a t t h e knowledge base is correctly represented and t h a t t h e system is w o r k i n g properly. G r o u p 2 : End-users w h o w a n t to f o l l o w the system's reasoning, b u t w h o do not k n o w m u c h a b o u t exp e r t system technology (or even a b o u t computer science). Each of these groups demands rather different language to be e m p l o y e d in their interactions w i t h the system. We call these groups interaction groups.3 3

These two user types are being considered for the initial study. As the system is used further, it is likely that we w i l l have to handle more interaction groups. The question of extensibility therefore becomes quite important.

For example, in response to the user question ' W h a t is a f a u l t y system?', w h i l e t h e first phase of the generat i o n process gives rise to t h e i n t e r n a l predicate calculus f o r m shown in Figure 2, a v a r i e t y of significantly different phrasings are required. For users in g r o u p 1, a definition in an English f o r m as close as possible to the exact defin i t i o n expressed in predicate calculus is needed, as these users employ the e x p l a n a t i o n f a c i l i t y for debugging purposes. There is therefore a need to be very precise and l i t e r a l . T h e t e x t generated by o u r current a l g o r i t h m for group 1 is shown in F i g u r e 3. For users in group 2, however, this t e x t is likely to be confusing at best. Figure 4 shows the more a p p r o p r i a t e t e x t t h a t is generated by our a l g o r i t h m when t a i l o r i n g p h r a s i n g for users in group 2. T h e next section describes how our generation component is able to generate these t w o very different texts f r o m the same i n t e r n a l content.

4

D e s c r i p t i o n of t h e a l g o r i t h m

4.1

T h e g r a m m a t i c a l resources available and methods for controlling t h e m

T h e g r a m m a r used in this w o r k is Nigel [ M a n n and Matthiessen, 1983], a s y s t e m i c - f u n c t i o n a l g r a m m a r [Halliday, 1985] of E n g l i s h . C o n t r o l of the selection of gramm a t i c a l features in Nigel is achieved by e m p l o y i n g a set of semantic inquiries t h a t access g r a m m a r external sources of i n f o r m a t i o n , such as the knowledge base and the i n p u t t o the g r a m m a r . An interface language ( t h e Sentence Plan Language or S P L ) t h a t provides for very flexible c o n t r o l over the g r a m m a r has recently been developed for Nigel [Kasper, 1989]. T h i s language provides for the convenient cont r o l of all aspects of Nigel at a variety of levels of abstraction, i n c l u d i n g several t h a t provide for direct responses to Nigel's inquiries, S P L can be seen as providing a set of constraints t h a t m u s t be met by the language generated. 4 Given a specification of propositional con4

This interpretation is compatible w i t h general functional unification-based schemes of control and so is not particularly bound to the Nigel grammar as currently implemented.

Bateman and Paris

1513

tent (such as t h a t shown in Figure 2), our phrasing c o m ponent constructs SPL expressions t h a t are t h e n passed t o the g r a m m a r . SPL i n p u t specifications consist of a recursive struct u r e of entities and t h e i r features to be expressed in E n glish. Each e n t i t y needs to be allocated a type w h i c h is i n t e r p r e t e d w i t h respect to a knowledge base of general conceptual categories called the upper model. T h e upper m o d e l is t y p i c a l l y used to mediate between t h e o r g a n i z a t i o n of knowledge f o u n d in an a p p l i c a t i o n dom a i n a n d t h e k i n d of o r g a n i z a t i o n t h a t is m o s t convenient for i m p l e m e n t i n g the g r a m m a r ' s inquiries. Upper m o d e l concepts possess specified roles t h a t define the possible semantic relations t h a t may appear in the S P L . An example of the SPL representation of a sentence is given in F i g u r e 5. T h i s specification states t h a t there is a r e l a t i o n of t y p e give to be expressed as an assertion, h o l d i n g over three p a r t i c i p a n t s , an actor, an actee, a n d a beneficiary. These p a r t i c i p a n t roles are d r a w n f r o m those defined in the upper model for relations of t y p e give. T h e fillers of these roles are themselves SPL expressions. A n i m p o r t a n t source o f v a r i a t i o n i n phrasing occurs i n the c o n s t r u c t i o n of the SPL specification. For e x a m p l e , in F i g u r e 6, t h e existence of the book has been made salient. T h e SPL t e r m for t h a t book now occurs at a higher level w i t h i n the s t r u c t u r e t h a n the process of giving it p a r t i p a t e s i n ; we say t h a t it is of a higher rank. T h e central t y p e d terms at each level of the SPL struct u r e (e.g., gl, p1, bl, a n d p2 in Figure 5) correspond to the heads of t h e l i n g u i s t i c expressions t h a t realize t h e m . 5 A n y p a r t i c u l a r S P L expression c o m m i t s the language t h a t The head of a linguistic expression is the 'independent

1514

Speech and Natural Language

w i l l be generated to expressing a single head-modifier organization. T h e allocation of heads a n d t h e i r respective rankings have a significant effect on p h r a s i n g . O u r mechanism for c o n t r o l l i n g this allocation relies on the n o t i o n of register. A register defines the e n t i t y a n d r e l a t i o n types t h a t are allowed to become heads in the language of t h a t register; t h i s is described in detail in Section 4.2. Once the SPL expression is c o n s t r u c t e d , however, a n u m b e r of distinct sentences satisfy its constraints, i.e., the expression is underspecified. Consider for example the SPL expression shown in F i g u r e 5. At least three dist i n c t sentences satisfy its constraints: Cecile gives John the book] Cecile is giving a book to John; to John, Cecile gave a book. If the t e x t planner is to c o n t r o l these alt e r n a t i v e phrasings, then a d d i t i o n a l guidance is required; our mechanism for this is also defined using registers and is described in Section 4.3. 4.2

Constraint on head selection

D e f i n i t i o n s 6 of the available heads a n d their relative rankings for the t w o registers used to generate the examples t a i l o r e d to interaction groups 1 a n d 2 are shown in variable' of that expression; it is often considered to be more salient than the other constituents of the expression. The head places constraints on those constituents (the head's modifiers) and thus constrains the expression's overall syntactic structure. The head-modifier organization of a linguistic expression is also recursive. The definitions divide heads among three types, the field, mode, and tenor. This imposes a further dimension of ranking in the search process drawn from the linguistic theory of register. We will not discuss these details here however.

Figure 7. T h e terms used in these definitions are linked into Nigel's upper m o d e l ; this guarantees their expressibility. They are also linked to the input specification language so t h a t instances of their occurrence can be recognized. The first definition establishes the relative importance of entities and relations according to the language used by system developers, the second according to the language of other end-users. T h e definitions state that the terms labeled RHETORICAL-RELATIONS are to be preferred for higher r a n k i n g t h a n those labeled R E L A T I O N S , which are in t u r n to be preferred to those labeled O B J E C T S , etc. I m p o r t a n t l y , the groupings are labeled in terms of their linguistic realization, and not in terms of their status in the knowledge base. For example the predicate calculus t e r m exist gets realized in English as a relation (process) in the register for system developers (giving rise to 'there exists a < e n t i t y > ...') while it gets expressed as an object modifier in the register for end users (giving rise to '... each < e n t i t y > ' ) . Thus, whereas in the register for system developers the relations exist and forall are given high r a n k i n g , they are given the lowest ranking in the register for end-users. Moreover, whereas the t e r m variables is available in the register for system developers, it does not occur in the register for end-users. The same propositional content is phrased in various ways, depending on the entities and relations available in the register and their relative rankings. By choosing a register the text planner can systematically control the phrasing of a given propositional content. Phrasing is decided upon by successively finding the highest ranking head offered by the register at each level of structural decomposition of the i n p u t , as is illustrated below. Given the i n p u t specification of Figure 2, for example, the program searches the list of register terms for the highest ranking available head whose pattern matches the structure of the i n p u t . In the register for system developers, one of the highest allowable heads is the rhetorical relation and. T h e recognition of terms in the input specification language is defined using structural patterns t h a t may be matched against the input. Patterns are expressed in terms of the input syntactic categories (specified by the EES g r a m m a r ) t h a t define the constituents of the i n p u t at t h a t level; the pattern for the rhetorical relation and is shown in Figure 8. Continuing w i t h our example, this p a t t e r n matches the top level of structure shown in Figure 2 and so offers an appropriate

head for describing the input. Accordingly, an SPL term is constructed t h a t : 1. specifies the register term found, namely and, as an appropriately ranking head, 2. places the remaining input constituents in appropriate modifier relationships. These relationships are drawn f r o m the upper model definition of the type of the head which, in this case, are conjunctl and conjunct 2. T h i s gives rise to the basic SPL fragment shown in Figure 9. T h i s process is then applied recursively for each constituent, namely assertionl and assertion2, searching for permissible lower ranking heads. B o t h assertions in the definition of a faulty system are quantifier expressions. (The pattern for exist is shown in Figure 8.) These expressions have a structural decomposition in terms of the quantifier type, the variable to which the quantifier applies, the variable range and the assertion about that variable. Therefore, for both assertionl and asseriion2, patterns for predicate calculus quantifiers (there-exists for assertionl and for-all for assertion2) match. The SPL is grown accordingly, and the process recurses. If no applicable head from the active register matches the i n p u t constituent at that level, the intrinsic struct u r i n g of the input is used to select where to search next for an applicable head. Elements of the input that are not allocated to heads at that level are retained for possible consumption later in the process. Examples of this can be seen in the differential treatment of variables and relation-types in the two registers. Whereas variables may become heads in the system developer register and are expressed as noun phrases, there are no heads corresponding to variables in the end-user register and so they do not appear in the English of t h a t register. Similarly, some relation types, such as s i g n a l - p a r t , do not appear in the end-user register and thus are not consumable in the construction of SPL for the generation of the English in t h a t register. 4.3

Constraint on grammatical feature selection

As we saw in Figure 5, the straightforward propositional content specified in SPL head and modifier terms does not restrict the possible result to a single sentence. A variety of aspects of the phrasing w i l l not be controlled and need to be constrained. We also provide this type of constraint in a register-driven way by associating particular

Bateman and Paris

1515

sets of g r a m m a t i c a l features w i t h p a r t i c u l a r selections of register. T h i s is i m p l e m e n t e d in t e r m s of SPL 'default e n v i r o n m e n t s ' w i t h i n w h i c h t h e t h e o r e t i c a l l y available g r a m m a t i c a l alternatives are restricted to some specified subset. An example of a default e n v i r o n m e n t activated when the system developer register is in force is shown in Figure 10. D e f a u l t environments consist of a set of Nigel inquiries (cf. Section 4.1) and the responses those inquiries are to be given if they are needed d u r i n g generation. By predisposing i n q u i r y responses, areas of the g r a m m a r as a whole can effectively be removed f r o m consideration so t h a t the alternatives they offer no longer appear relevant. T h i s is equivalent to the d y n a m i c c o n s t r u c t i o n of a 'subgrammar'. T h e s u b - g r a m m a r created by F i g u r e 10 has the f o l l o w i n g properties: processes of existence w i l l be lexicalized by the w o r d exist rather t h a n some more n e u t r a l w o r d , such as be, as in there is a book ( : p o s t u r e - o f - e x i s t e n c e - q ) ; w h e n possible, selection of a member of a set w i l l be made using specific expressions, e.g., one X r a t h e r t h a n some X ( : s e l e c t i o n - p a r t i c u l a r i t y - q ) ; since variables are generally refered to in the singular, t h e d i s t i n c t i o n between all and every is n e u t r a l i z e d for variables ( : r e g i s t e r - q ) ; 'de-emphasizing' expressions such as the possessive m o d i f i c a t i o n of 'X's Y' are avoided in favor of the more equal emphasis of ' Y of X' ( : d e i c t i c - p a r t - q and : p o s s e s s o r - m o d i f i c a t i o n - q ) ; pronominalization is not attempted (: empty-number-q and e m p t y - g e n d e r - m u l t i p l i c i t y - q ) ; a n d the order o f conjuncts and disjuncts in the i n p u t w i l l n o t be altered in the o u t p u t by t h e m a t i c p l a n n i n g of any k i n d (:extension-precedence-q). This sub-grammar is therefore oriented to language t h a t is precise and ex-

1516

Speech and Natural Language

p l i c i t , and w h i c h is n a t u r a l for system developers. T h i s aspect of our approach is s i m i l a r to the m e t h o d employed by P a t t e n [1988]. 4.4

S u m m a r y of the constraints

F i g u r e 11 summarizes the i n p u t s to the generation process and the l o c a t i o n of situation-specific constraints in that process. In a d d i t i o n to the basic p r o p o s i t i o n a l content i n p u t specification, we have experimented w i t h t w o f u r t h e r sources of c o n t r o l . Feature constraints t h a t h o l d sway whenever a t e x t of a p a r t i c u l a r t y p e is being generated, a n d head selection, which is specific for each specification of p r o p o s i t i o n a l content provided as i n p u t .

5

G e n e r a l i z e a b i l i t y of t h i s a p p r o a c h

G i v e n t h e need to be able to expand the use of an expert system to users of different types, it is i m p o r t a n t t h a t the t a i l o r i n g techniques employed are readily extensible. We have shown here how the phrasing phase of t a i l o r i n g can be controlled in terms of a set of entities t h a t the t y p e of language required uses and a set of restrictions on the general capabilities of the g r a m m a r . B o t h aspects w o u l d need to be addressed in any case when setting up an expert system to interact w i t h a new user group since, otherwise, the language best suited to t h a t group w o u l d not have been d e t e r m i n e d . We i l l u s t r a t e the ease of e x t e n s i b i l i t y of the technique shown here by s e t t i n g up a new user type and showing the language t h a t is generated for t h a t user t y p e given the i n p u t of F i g u r e 2. I m a g i n e a class of users w h o are not expert i n the d o m a i n o f d i g i t a l circuit analysis and w h o j u s t need to be presented w i t h responses in as s i m ple terms as possible. F i r s t , we do n o t need to invoke the g r a m m a t i c a l feature restrictions t h a t lead to pre-

ciseness shown in Figure 10. Second, we provide a set of significant entities, shown in Figure 12, in which quantifiers and variables do not appear, the space of available objects t h a t may be referred to is reduced, and 'evaluations' of domain states of affairs are p e r m i t t e d . T h e text generated when our algorithm runs w i t h this register in force is also shown in the figure.

6

Conclusion and future work

In this paper we have described a mechanism which controls the fine detail phrasing of generated language according to a definition of the type of language t h a t is required. In the short term, this work has already i m proved the control of phrasing in a text planner; the mechanisms we have designed allow for a far more flexible, and yet systematic, expression in English of single specifications of propositional content t h a n has previously been possible. In the longer t e r m , we are aiming towards a single, unified component module of the text planning/generation process t h a t adds flexible tailoring to the user during both the content selection and phrasing phases of text planning.

Acknowledgments The authors would like to thank Eduard Hovy, Johanna Moore and W i l l i a m Swartout for comments on earlier versions of this paper.

References [Appelt, 1985] Douglas E. A p p e l t . Planning Natural Language Utterances. Cambridge University Press, Cambridge, England, 1985.

[Jameson, 1987] Anthony Jameson. How to appear to be conforming to the ' m a x i m s ' even if you prefer to violate them. In G. Kempen, editor, Natural Language Generation: Recent Advances in Artificial Intelligence, Psychology, and Linguistics. Kluwer Academic Publishers, Boston/Dordrecht, 1987. [Kasper, 1989] Robert Kasper. SPL: A sentence plan language for text generation. Technical report, U S C / I S I , 1989. [Mann and Matthiessen, 1983] W i l l i a m C. Mann and Christian Matthiessen. Nigel: A systemic grammar for text generation. Technical Report ISI/RR-85-105, U S C / I S I , February 1983. [McKeown et ai, 1985] Kathleen R. McKeown, Michael W i s h , and Kevin Matthews. Tailoring explanations for the user. In Proceedings of IJCAI-85, Los Angeles, Ca., 1985. International Joint Conference on Artificial Intelligence. [McKeown, 1985] Kathleen R McKeown. Text Generation: Using Discourse Strategies and Focus Constraints to Generate Natural Language Text. Cambridge University Press, Cambridge, England, 1985. [Moore and Swartout, 1989] Johanna D. Moore and W i l l i a m R. Swartout. A reactive approach to explanation. In Proceedings of the Eleventh International Joint Conference on Artificial Intelligence, Detroit, M I , August 20-25 1989. I J C A I . [Morik, 1985] K a t h a r i n a M o r i k . User modelling, dialog structure, and dialog strategy in H A M - A N S . In Proceedings of EACL-85, Geneva, Switzerland, 1985. European Association of C o m p u t a t i o n a l Linguistics.

[Carberry, 1988] S andra Carberry. Plan recognition and user modelling. Computational Linguistics, 14 (3), September 1988.

[Paris, 1988] Cecile L. Paris. Tailoring object descriptions to the user's level of expertise. Computational Linguistics, 14 (3), September 1988.

[Grosz and Sidner, 1986] Barbara J. Grosz and Candace L. Sidner. A t t e n t i o n , intentions and the structure of discourse. Computational Linguistics Journal, 12 (3), 1986.

[Patten, 1988] Terry Patten. C o m p i l i n g the interface between text planning and realization, August 1988. Workshop on text planning and Natural Language Generation, Sponsored by A A A I .

[Halliday, 1978] Michael A. K. Halliday. Language as social semiotic. Edward A r n o l d , L o n d o n , 1978.

[Swartout and Smoliar, 1987] W i l l i a m R. Swartout and Steve W. Smoliar. Explaining the link between causal reasoning and expert behavior. In Proceedings of the Symposium on Computer Applications in Medical Care, November 1987.

[Halliday, 1985] Michael A. K. Halliday. An introduction to functional grammar. Edward A r n o l d , London, E n g l a n d , 1985. [Hovy, 1988] Eduard H. Hovy. Generating natural language under pragmatic constraints. Lawrence E r l b a u m , Hillsdale, New Jersey, 1988.

Bateman and Paris

1517