Decision-Theoretic Salience Interactions in Language Generation

1 Introduction

The salience of an entity, in intuitive terms, refers to its degree of prominence, importance or activation . In natural language generation (NLG), it is interpreted as a measure of how well the entity stands out from other entities. When the generator is faced with the choice of a word or a phrase, a ranking of the available choices is obtained on the basis of the degree of salience of entities in the generation context. The word or phrase is then selected for utterance on the basis of maximum salience. The role of salience in NLG thus involves a causal re­ lation between prominence of entities in a ranking, and preference of a choice among alternatives.

The first computational study of salience in an NLG system is that of [Conklin and McDonald, 1982], who used it as a heuristic for selecting topics in a scene de­ scription task and for ordering the topics in generat­ ing paragraphs . In disciplines related to NLG, such as linguistics and psycholinguistics, salience has been in­ voked in several accounts of language phenomena. For instance, the constituents of sentences are said to be or­ dered according to their relative salience, and the facility to order the constituents of a sentence is said to be used by speakers to communicate the relative salience of el­ ements in the context of language production [Osgood and Bock, 1977], [Osgood, 1980]. Guided by the premise that the choice of words and constructions on the basis of salience is pervasive in NLG, our research [Pattabhi­ raman, 1992] aims to flesh out the notion of salience in a theoretically-significant form.

A number offactors have been claimed to contribute to the salience of entities. Vividness, pervasiveness, unique­ ness, recency of mention and frequency of mention are some of the determinants of salience, several of which are at play even in the simplest of NLG settings. A ma­ jor challenge to the computational study of salience in NLG is in unravelling how the determinants of salience interact in their effects on language generation decisions. As the first step towards meeting this challenge, we de­ lineate an important pair of categories in the salience theory: canonical salience and instantial salience. The former is characterized as a built-in preference in the gen­ eral conceptual- and linguistic knowledge of the speaker. An example of such a natural tendency is the preference to express agent before object in simple action sentences. The latter refers to the salience of specific objects in the context of NLG , and may accrue through such determi­ nants as vividness and recency of mention . These cat­ egories are presented briefly in section 2 below, and in more detail, in [Pattabhiraman, 1992].

We consider the interaction between canonical salience and instantial salience as a crucial subject of study in NLG. We then formalize this interaction in terms of decision-theoretic combinations of probabilities and util­ ities. The choice of the formal framework is inspired primarily by the conclusions of psycholinguistic experi­ ments most notably, those of [Osgood and Bock, 1977]. In this paper we construe two distinct language gener­ ation tasks in terms of decision-theoretic salience in­ teractions: (1) phrase selection (section 3), (2) basic level/entry level preference in noun generation (section 4) . A third task, viz ., comparison (simile) generation, has been modelled in this framework and presented in [Pattabhiraman , 1992] and [Pattabhiraman and Cer­ cone, 1992].

2 Canonical Salience and Instantial Salience

We postulate that there exist built-in preferences or in­ ternal biases in the linguistic choice systems of human adult speakers. These biases are regarded as pre-defined relative salience configurations. By this postulate, even if (for example) sentence John gave a book to Mary has the logical, 'transformed' equivalent John gave Mary a book, in the presence of a built-in preference in the choice set

the choices will not be equally presented when the gen­ eration context involves talking about a(ny) specific X giving a(ny) specific Y to a(ny) specific Z. That is, ev­ ery time the aforementioned choice set is encountered in generation, the built-in preference will be automatically activated as a determinant of salience, as if by force of habit. If there are other determinants in the instance of generation (like vividness, etc) which define a relative salience configuration among X, Y and Z, their effects will have to be 'gated through' the built-in preference inherent in the choice system.

2.1 Canonical Salience

The inherent, built-in bias by which the generator tends to exercise preferences is termed canonical salience in our theory. The naturalness hypothesis of [Osgood, 1980] most directly addresses the origin of canonical salience from a psycholinguistic perspective. A major assump­ tion of Osgood's theory is that the cognitive structures underlying the production and comprehension of sen­ tences develop in prelinguistic, perceptual experience; these structures are 'taken over' by the linguistic sys­ tem as it develops later . The basic structure of sim­ ple cognitions is of the form relation(argl,arg2) . Os­ good discusses two very basic cognitions which are cru­ cial for the prelinguistic child: stative cognitions of the form state(figure,ground), and action cognitions of the form action( actor, recipient) . Osgood then argues that in a prelinguistic child, for stative cognitions, figure-state­ ground is the natural order, and for action cognitions , actor-action-recipient is the natural order.

2.2 Instantial Salience

Salience configurations set up in specific situations or contexts of generation by such determinants as vividness, recency of mention, speaker motivation and so on come under instantial salience. For example, in generating A UFO was spotted by the boy, the entity UFO (the object) is more salient than the entity boy (the agent) due to vividness, and consequently gets expressed earlier (as a passive sentence). Instantial salience depends on the presence of other entities in the context: UFO is vivid relative to boy . It could also be influenced by the presence of speaker-motivation or goals, and on dynamic factors which, as the text is unfolding, may make entities salient for pronominalization and other decisions. The passivization in My stereo system has been stolen and in the second clause of He started arguing loudly, and had to be told to calm down are licensed by the salience determinants in the respective instances, viz ., speaker motivation and recency of mention, respectively.

2.3 Canonical-Instantial Salience Interactions

The choice of a word or phrase is exercised on the basis of effective salience, ~hich is the net result of the interac­ tion of canonical- and instantial salience. A distinguishing feature of canonical-instantial salience interactions is that they sometimes agree in their dictates (synergistic interaction) and sometimes are opposed in their dictates (competitive interaction) . For example, in generating the sentence The Mayor was struck by a car yesterday, the natural tendency of canonical salience dictates that the instrument 'car' be expressed prior to the patient 'the Mayor'. On the other hand, the prominence (instantial salience) of the Mayor presses for it to be positioned ear­ lier in the sentence. lnstantial salience 'wins out', and the passive voice sentence is generated.

Psycholinguistic experimental results of [Osgood and Bock, 1977] suggest that the interaction between canon­ ical salience and instantial salience is multiplicative. We capture this interaction in a decision-theoretic formal­ ization which models canonical salience as probabilities, instantial salience as utilities, and the selection criterion based on effective salience as maximization of expected utility (or its variants). The details of the available evi­ dence and our arguments for the choice of decision the­ ory as the formal framework are presented in [Pattabhi­ raman, 1992}.

3 Phrase Choice

In this section, we first show how effective salience in the selection between active voice and passive voice is obtained as expected utility. Next , we illustrate the relative rarity of the passive voice by applying the de­ cision criterion to a small set of example sentences. The criterion is then stated in general terms for all bi­ nary linguistic choice systems in syntactic linearization. Psycholinguistically-significant characteristics are then presented .

3.1 Effective Salience as Expected Utility

Let us consider the verbalization of a simple agent­ action-object cognition.

(1) The environment in which the selection decision is carried out is governed by the probabilities that the agent Ag and the object Ob respectively will be ex­ pressed earlier in the utterance. These probabilities cor­ respond to the canonical saliences of Ag and Ob respec­ tively, and are denoted by p(Ag) and p(Ob) respectively. p(Ag) + p(Ob) = 1 .

(2) The instantial saliences of Ag and Ob are given as u(Ag) and u(Ob) respectively. In a specific situation, these are evaluated as functions of such determinants as vividness, recency of mention, etc.

(3) Whatever be the instantial saliences, the agent Ag may or may not be expressed sentence-initially, be­ cause of the influence of the canonical probabilities. We will have four possible outcomes in this decision, two corresponding to each relative instantial salience config­ uration . Utilities are associated with those outcomes, expressing their preferability, on the basis of instantial saliences. We expand this point more formally in ( 4)-(8) below.

( 4) We will represent the possible relative instantial salience configurations as a set of states: S = {A, 0} where A= u(Ag) ~ u(Ob) and 0 = u(Ob) > u(Ag).

1. The linguistic choice system is the set of available actions (decisions), denoted as A= {Ac, Pa} where Ac = "express Ag before Ob" , and Pa = "express Ob before Ag" .
2. The four possible outcomes are labelled in terms of the states and actions that lead to them, as (state,action) pairs: {Ac, A}, {Ac, 0), (Pa, A}, (Pa, OJ .
3. Preferences for these outcomes are expressed as utilities. Utility of action a E A in state s E S is repre­ sented as U(a, s) .
4. We assign values for these utilities on the basis of knowledge of instantial saliences, and the principle of tending to express the more salient element earlier: U{Ac, A)= u(Ag), U(Pa, A)= 0, U(Ac, OJ= 0, and U{Pa, OJ= u{Ob).
5. Given the above information on probabilities and utilities, we can now compute the expected utility of each action as follows :

Expected utility of expressing Ag sentence-initially:

Expected utility of expressing Ob sentence-initially:

(10) Decision theory recommends choosing the action with maximum expected utility. Salience principles in NLG recommend selecting the alternative with maxi­ mum effective salience. The above derivation shows how the effective salience of an entity corresponds to the ex­ pected utility of a linguistic action, thus giving salience a causal role in a linguistic effect .

3.1.1 An Illustration

Consider the set of sentences

1. Carl Sagan discovered a new comet last month .
2. A baffiing new galaxy has been discovered at the centre of the Milky Way.
3. I bought a pen .
4. The astronomers at Palomar discovered a new as­ teroid.
5. Carl Sagan bought a shirt .

For these sentences, table 1 summarizes the decisions, and the decision-theoretic quantities on which the de­ cisions are based. p(Ag) and p(Ob) are 0.87 and 0.13 respectively (approximates, from the empirical work of [Svartvik, 1966J). The instantial saliences have been sub­ jectively assigned based on the vividness of entities. 3 of the 5 sentences above have a high-salient object. 4 out of 5 have an object with instantial salience no less than that of the agent . Yet, only 1 out of the 5 cases produces a passive.

3.2 General Form for Binary Choice Contexts

The derivation in section 3.1 can be expressed easily in a general form, applicable to all binary linguistic choice contexts in linearization. Let x 1 and x2 be entities in the context of NLG whose saliences are being compared. Denote the canonical saliences of x 1 and x2 as p(xl) and p(x2 ) respectively (p(x 1 ) + p(x2) = 1) and the instantial saliences as u(xl) and u(x2) respectively. Let {X1 , X2} be a linguistic choice context, with the alternatives X1 and X2 corresponding to the earlier expression of x1 and x 2 respectively.

Then, analogous to the derivation in section 3.1 ,

and

The construct selected is the one with greater EU.

3.2.1 Psycholinguistically-Relevant Characteristics

Several important psycholinguistically-relevant as­ pects are captured straightforwardly in this model :

1. Canonically unbiased linguistic choice context: If there is no skew in the canonical probabilities, then the decision for earlier expression is governed by instantial salience. That is,

Choice here is apparently situation-controlled.

1. Equi-salient entities in the generation context: If there is no skew in the instantial saliences, then the decision for earlier expression is governed by canonical salience . That is,

Choice here is apparently grammar-controlled.

1. Extremes of lack of availability of choice: If X1 is the only available linguistic choice then x 1 will always be expressed earlier. That is,

Similarly,

·

1

1

4 Basic Level/Entry Level Preference in Noun Generation

Central to canonical salience is Osgood's conception of naturalness as arising in prelinguistic, perceptual experi­ ence, and as being cultivated in childhood and persisting in adulthood. We proposed viewing the interactions be­ tween canonical salience and instantial salience (in lin­ earization) as decision-theoretic combinations of prob­ abilities and utilities. We next asked ourselves whether we could find other language generation tasks (than syn­ tactic linearization) in which a similar, built-in expres­ sive preference played a crucial role, and whether, in such a task, a possibility exists for overriding the natural preference, and under what conditions such a possibility is admissible. If we can find such a phenomenon, we thought, we can learn more about the nature of canon­ ical salience. Moreover, we can consider applying the decision-theoretic model of salience interactions, suit­ ably instantiated, to model decision criteria in that phe­ nomenon.

Interestingly, we found several dimensions of similar­ ity between the role of basic levels of taxonomic concept knowledge in the free naming of objects, and the na­ ture of canonical salience discussed in section 3. To our knowledge, this similarity has not been observed or ex­ plained in the literature. In day-to-day speech, one refers to objects by labels which are neither too general nor too specific. For example, we generally say look at that dog! rather than look at that collie! or look at that mammal! This effect is described in cognitive psychology by [Rosch and Mervis, 1975] in terms of basic levels in human tax­ onomic knowledge. Later work by [Jolicoeur eta/. , 1984] recognizes exceptions to the norm of basic level prefer­ ence, and introduces the notion of entry level effect, by which some specific instances of certain categories are named at a level other than the basic level. For example, when pointing a 'generic' bird to a friend, you are more likely to say look at that bird; however, when you see an ostrich, you are more likely to say look at that ostrich. Basic level preference is one of several constraints on noun generation in [Reiter, l991]'s system. While Re­ iter handcodes basic level annotations in his knowledge base, our work deals with the non-linguistic basis for determining basic levels in a concept hierarchy.

The stark similarities between the role of basic lev­ els in object naming and that of canonical salience in linearization have been pointed out in detail in [Pattab­ hiraman, 1992]. In the same thesis, we also argue, using the conclusions of Jolicoeur et a/. 's experiments, that en­ try level effects are an instance of decisions based on ef­ fective salience. In this section, we present some major aspects of our results in modelling free naming of ob­ jects in terms of salience interactions. Using probabilis­ tic concept representations in a taxonomic hierarchy, we formalize entry level preference as maximization of ex­ pected utility. Three major components constitute this model (A fully detailed illustration is given in [Pattab­ hiraman, 1992]): (1) Modelling canonical salience (basic level preference) as a probability distribution. By this model, basic level is the level in the classification hierarchy of an object at which the object is most likely to be described in a free naming task. The derivation of this distribution depends on differential category utility, a measure modified and adapted from category utility proposed by [Gluck and Corter, 1985]. (2) Utilities to ex­ press the instantial salience of the objects being named. (3) Effective salience as expected utility, determining the choice of names.

4.1 Probabilistic Concepts in Taxonomic Knowledge

Formally, a concept C is a labelled set of properties Pi, and each property Pi is a pair Ai : Vi, where Ai is an attribute and vi is its probabilized value space. vi is a set of pairs V;i : Pii, with Pii E (0, 1], and I::i Pii = 1. For example, in one of our implementations, the concept mango has, among its properties, colour = {yellow:0.25, golden:O. 60, green:0.15}. These representations are a variant of the dimensional approach to probabilistic con­ cept representation, discussed in cognitive psychology literature by [Smith and Medin, 1981] . We can denote the probabilities more perspicuously so as to reflect the associated property and concept , as follows: p(Ai = V;i) is the probability that the attribute Ai (of an object) has value V;i . p(A; = V;j ICk) is the probability that an object in class Ck has value V;i for attribute Ai.

In the taxonomic knowledge base, probabilistic sum­ maries of value distributions for attributes are stored at several levels. Each category in the taxonomic hierar­ chy is thus a probabilistic concept. It is to be noted that the probabilities at successive levels of a path in the hierarchy are related to one another: for example, the probability of the red colour of apples summarizes contributions from the probabilities of the red colour of all subtypes of apples.

The root node of the hierarchy is denoted as Centity; the leaf nodes are the most specific object classes. C,up(k l is the immediate superordinate of Ck. Apart from property probabilities p(A; = V;i ICk) stored at probabilistic concept nodes , concept probabilities are also relevant to our computations. The concept probabilities we consider are: p(Centity) = l for the root node of the hierarchy, and p( ck IC.up(k )) for all other concepts ck. The concept taxonomy hierarchy encodes p(CkiC,up(k)) as a value stored at the node ck.

4.2 Canonical Salience from Differential Category Utility

The scenario we consider is a guessing game in which a participant guesses the values for various attributes of a specific object . The participant is assumed to possess probabilistic concept knowledge, and to adopt what [Gluck and Corter, 1985] call a probability match­ ing strategy: for an attribute Ai, (s )he guesses a value V;i with a probability equal to the value of p(Ai = V;i) in her own knowledge. The guess is likely to be correct with the same probability.

Assuming that the category information (Ck) is given prior to guessing (as in: "I have an apple with me; can you guess its colour?"), for an attribute Ai for which the value is guessed, the score awarded to the participant is the expected value of p(A; = Vii ICk) taken over all pos­ sible values of A;, given by L:;j~\ (p(A; = Vii ICk )j2 where n;k is the number of possible values of attribute A; in concept C~c . If guessing proceeds for every one of the I attributes of the object, the score is L::=l L:j~\(p(A; = Vii ICk )j2. This is the expected number of values of at­ tributes that can be correctly guessed for an object given that it belongs to the class C~; .

At higher (more abstract) levels of taxonomic knowl­ edge, probability distributions for attribute values tend to be more dispersed than at lower {more specific) levels. (By dispersion we mean the number of possible values, as well as the flatness of the distribution envelope). That is, when I am told, "/ have an entity on mind; can you guess its colour?", all possible values on the colour di­ mension are candidate guesses. When more specific cat­ egory information is given, as in, "/have a type of apple on mind; can you guess its colour?", the possible values are more restricted, and the probability distribution for attribute values generally tends to show more skew or prommence.

This tendency will be reflected accordingly in the val­ ues of Lj (p(A; = V;j }]2: for an attribute dimension A; having n; possible values, L:;j~ 1 (p(A; = V;i )j2 E (~;, 1]. It is minimum ( = ,;, ) when the attribute values are equiprobable, and maximum(= 1) when there is only one possible value for A;. Therefore, if more specific categorial information is given, the score in the guess­ ing game is higher . That is, for any C k. L: i (p( A; V;i)IC,up(k)F will tend to be less than Lj(p(A; = V;j IC~c)J2 °

[Mervis and Rosch, 1981] note that basic levels are where a number of inference-related abilities are maxi­ mized in humans. [Rosch and Mervis, 1975] state that the categories subordinate to the basic level are also bun­ dles of predictable attributes and functions, but contain little more information than the basic level to which they are subordinate.

Consider a traversal of the hierarchy starting at the root (Centity): as we go down, not only does the score in the guessing game increase, but the rate at which it in­ creases, increases . Soon, there comes a point where the rate of increase decreases, though the score itself con­ tinues to increase. We consider the expected increase in score, made possible by knowledge of a category ck over the score with prior knowledge of its superordinate C,up(k)· That is, compare I have an apple with me; can you guess its colour? versus I have a fruit with me; can you guess its colour? Basic level is tJielevel at which the increase of score over the score at the immediate superordinate is maximum.

4.2.1 Differential Category Utility

The differential category utility (DCU) of a concept C~:, written as DCU(C~.:) , is:

That is, DCU is defined as the expected increase in score made possible by knowledge of C~: over the score with knowledge only of C,up(k) · As a boundary case, DCU(Centity) = 0. DCU differs from category util· ity [Gluck and Corter, 1985], defined as the expected increase in score made possible with knowledge of ck over the score with no prior category information- that is, over the score with knowledge of Centity · We argue in [Pattabhiraman, 1992] that category utility will always predict the most specific level, which need not necessar­ ily be the basic level.

4.2.2 Probabilistic Definition of Canonical Salience

Given a set of concepts { C1 , C2 , .. . , Cn} constituting a path in taxonomic concept knowledge such that C~c = C,up(k+l) for k = 1 .. . (n- 1), define:

so that Pa(C~c) E [0, 1] and L:7= 1 Pa(Ck) = 1. The canonical salience of concept Ck is Pa(C~c) . The basic level is the 'most probable' concept in the Pa distribu­ tion .

4.3 lnstantial Saliences and Utilities

According to the experimental results of [Jolicoeur et a/., 1984], in the presence of atypicality, preference tends to shift down from the basic level to the subordinate level. That is, when we see an object which is in some way (in­ stantially) salient, we tend to use a more specific name. A typicality of concepts is assessed in the context of their immediate superordinate. More than one subconcept of a concept can be atypical: Consider a set of concepts { C1 , C2 , ... , Cn} grouped under C. One or more of the C1.: 's could be salient in C. For example, in

both ostrich and peacock are salient, the former because of its size, flightlessness and shape, and the latter, for its vivid feathers.

We propose that the saliences of the concepts C~c be evaluated by a function S : C ---. [0, 1], with S(Ck) being determined as a single number representing a combination of scores from determinants like atypical­ ity and vividness . In the bird example, we would ex­ pect S(peacock) and S(ostrich) to be high(> 0.5). The choice between C~: and C in a free-naming task is gov­ erned in part by a function u which apportions the num­ ber 1 between C~: and C as follows :

and

Moreover, since the context in which the salience of Ck is evaluated is provided by its superordinate C, u( c') =

0 for all concepts c' in the path, other than C~.: and C. For example, the salience of crabapple is evaluated with respect to apple. We take u( C) = 0 for all other higher concepts C (fruit, planLproduct and entity) in the crabapple path. When a specific concept is salient, the u score is 'taken away' from the superordinate, to the extent that the specific concept is salient. Thus, for the choice context {bird, robin}, owing to the low salience of robin, we will have u(robin) = 0.5 = u(bird) . For the choice context {bird, ostrich}, owing to the high salience of ostrich, we will have (say), u(ostrich) = 0.8 and u(bird) = 0.2.

4.3.1 Probabilistic Measure of Instantial Salience

We propose that the function S evaluate the instantial saliences of the concepts C~.: as a number in [0, 1] . In general, S will be evaluated as a composition of several subfunctions, one corresponding to each determinant of salience. We are interested in determining the extent to which an object classified under C~.: possesses an unusual property for its category C. The unusualness we consider is one by which we judge, for example, a crabapple, to be too sour or too small for an apple.

The index of this unusualness is computable from the probabilistic concept representations in taxonomic knowledge, as follows: given a concept C~.:, a supercon­ cept C and an attribute A;, if the most probable value A; = V;i is atypical in C~.:, the probabilistic degree of salience of C~.: because of V;i in the context of concepts constituting C is denoted by Sp(C~.;, A;), and is com­ puted as:

The subtraction of the square-bracketted term prevents double-counting of the contribution of C~.: to p(A; "ViiiC).

4.3.2 Using a Threshold to Recognize High Salience

The attributes taste, size and colour are highly salient in crabapples in the context of apples. There could, how­ ever be attributes which are not salient in the probabilis­ tic sense. For instance, the shape of crabapples does not particularly stand out when the shape of apples is con­ sidered. In fact, the distributions of shape in crabapple and apple could be identical. Such low scores do not contribute to the instantial salience of objects, and are ignored if they are below a threshold(},. We take (}, to be 0.5 in our . computations.

4.3.3 Prioritizing Property Saliences with Intrinsicness Weights

We then compute the net (instantial) salience of con­ cepts C~.: as a weighted sum of Sp 's:

where L; w(A;) = 1. The weights w(A;) serve the pur­ pose of prioritizing the contribution of the unusualness of attribute A; to the overall salience of C~.;. For exam­ ple , in judging crabapples as salient subspecies of apples, it may matter more that crabapples are very sour than that they are yellow. This could be because taste is in some sense more essential to the identity of an object, and 'less alienable' from the object than colour. This conception of attribute salience tallies with the notion of intrinsicness, discussed in cognitive linguistics by [Lan­ gacker, 1987] .

4.3.4 Evaluating the Functions S and u

The degree of atypicality of C~.: with respect to C,up(k)· is computed as

u(C~.:) and u(C.up(J.:)), which represent the instantial saliences of C~.: and Cwp(k) respectively, are now cal­ culated from S(crabapple) according to expressions (11) and (12) . These utilities will interact with the prob­ abilities Pa (expression 10) in choosing the label as a referring expression in a free-naming task.

4.4 Entry Level Effect as Expected Utility Maximization

The derivation of the expected utility model for selecting a concept label as a referring expression for an object closely parallels the derivation of the criterion for voice selection illustrated in section 3. We are given:

• an object classified under a set of concepts { C1 , C2, ... Cn} representing a path in taxonomic concept knowledge such that C~.: = C.up(k-l) for k=l ... (n-1).
• the probabilities Pa( C~.;) E [0, 1] for k = 1 ... n, rep­ resenting the canonical salience of the concepts, and
• u(C~.:) E [0, 1] for k = 1 ... n, standing for the in­ stantial salience of the concepts.

In decision theoretic terms, we have:

• a set of actions A= {Ac,, Ac,, . . . , Acn }, where Ac, stands for the selection of the label corresponding to concept C;
• a set of states representing the maximal sta­ tuses of the instantial saliences of concepts: S = { S1 , S2, ... , Sn} where S; is the state in which u(C;) =max( {u(Cl) , u(C2), ... u(Cn)})
• a set of possible outcomes in the form of (state,action) pairs. Here we have n 2 of them: (Ac;,Sj) fori= 1 .. . n,j = 1 .. . n .
• Utilities U(Ac;, Sj) associated with the outcomes. U(Ac;, Sj) is the utility of selecting the label cor­ responding to C; when Cj has maximum instan­ tial salience among all the concepts. The values of U(Ac,, Sj) are derived from the instantial saliences u(C;) and are shown in matrix form in table 2.
• An environment of decision making, characterized by the probability distribution Pa with n values, one corresponding to each state. We can transfer the distribution on to the set of states and write: p(Si) = Pa(Ci) ·

(16)

Table 2: Matrix of U(Ac,, Si)

The expected utilities of action Ac, denoted by EU(Ac.), represents the effective salience of C;, and stands for the strength of the motivational factor for se­ lecting the label corresponding to C; . Given the utilities tabulated in the matrix 2, we get:

When Cj has maximum instantial salience among all the concepts, the utility of selecting the label corresponding to C; for any i :f:. j is 0. Thus, the only nonzero contribu­ tion to the expected utility of Ac, comes from the pair (Ac,, S;), and is given by the product 'Pa(C;) · u(C;) .

5 Concluding Remarks

Decision-theoretic formalization of salience-induced phrase choice helps capture the relative frequencies of linguistic constructs, as illustrated in section 3 for the passive voice construction. Our view of salience­ based phrase choice complements as well as reinforces probabilistic themes in systemic linguistics, such as those of [Halliday, 1991] : we discuss the overlaps and complementarity in [Pattabhiraman, 1992] . With the canonical-instantial interactions and decision theory, we have been able to 'travel' to another phenomenon, viz., basic/entry-level effect in object naming, and have fleshed it out computationally.

Our research plans for the future include work on structuring the decisions on the basis of the probabil­ ities and utilities that enter into them. This will yield new NLG architectures in the long term. The multitude of numbers that one has to deal with in a quantitative approach can be a deterrent to ventures of large-scale implementations in existing AI languages and environ­ ments. However, the profusion of interacting numbers is more easily manageable in more 'linguistic' tasks like syntactic linearization than, say, in comparison genera­ tion, where heavy reliance on conceptual knowledge en­ tails many numbers coming into play. Probabilistic con­ texts in linguistic choice are mostly binary, while those in conceptual choice typically involve several possibilities.

Acknowledgments: The authors thank David McDon­ ald, Christian Matthiessen, Norman Foo, Ed Hovy and G.Sivakumar for comments, criticism and dis­ cussions . Thanks also to the three anonymous ref­ erees for their comments.