<?xml version="1.0" standalone="yes"?>
<Paper uid="E93-1055">
  <Title>Lexical Choice Criteria in Language Generation</Title>
  <Section position="2" start_page="0" end_page="454" type="metho">
    <SectionTitle>
2 Word Choice Criteria
</SectionTitle>
    <Paragraph position="0"> Only few contributions have been made towards establishing word choice criteria in NLG. 1 Hovy's \[1988\] generator PAULINE selected lexical items according to pragmatic aspects of the situation (rhetorical goals of the speaker giving rise to stylistic goals, which in turn lead to certain lexical choices). Also looking at the pragmatic level, Elhadad \[1991\] examined the influence of a speaker's argumentative intent on the choice of adjectives. Wanner and Bateman \[1990\] viewed lexical choice from a situationdependent perspective: the various aspects of the message to be expressed by the generator can have different degrees of salience, which may give rise to certain thematizations and also influence lexical choice. Reiter \[1990\] demonstrated the importance of basic-level categories (as used by Rosch \[1978\]) for generation, overriding the popular heuristic of always choosing the most specific word available.</Paragraph>
    <Paragraph position="1"> Generally speaking, the point of &amp;quot;interesting&amp;quot; language generation (that is, more than merely mapping semantic elements one-to-one onto words) is to tailor the output to the situation at hand, where 'situation' is to be taken in the widest sense, including the regional setting, the topic of the discourse, the social relationships between discourse participants, etc. There is, however, no straightforward one-to-one mapping from linguistic features to the parameters that characterize a situation, as, for example, stylisticians point out \[Crystal and Davy, 1969\]. Various levels of description are needed to account for the complex relationships between the intentions of the speaker and the variety of situational parameters, which together determine the (higher-level) rhetorical means for accomplishing the speaker's goM(s) and then on lower levels their stylistic realizations.</Paragraph>
    <Paragraph position="2"> Here we are interested in the descriptional level of lexis: we want to identify linguistic features that 1 Considerable work has been done on the construction of referring expressions, but this is just one specific sub-problem of lexical choice, and moreover a context-sensitive one. In this paper, we restrict ourselves to choice criteria that apply independently of the linguistic context.</Paragraph>
    <Paragraph position="3">  serve as a basis for choosing a particular lexical item from a set of synonyms. Not all these features are equally interesting, however; as Crystal and Davy \[1969\] noted, the relation between situational features and linguistic features is on a scale from total predictability to considerable freedom of choice.</Paragraph>
    <Paragraph position="4"> Among the less interesting dimensions are dialect and genre (sub-languages pertaining to particular domains, for example legal language or sports talk), because they tend to merely fix a subset of the vocabulary instead of Mlowing for variation: the fact that what Americans call a lightning rod is a lightning conductor in British English does not imply a meaningful (in particular, not a goal-directed) choice for a speaker; one rarely switches to some dialect for a particular purpose. More interesting is the degree of semantic specificity of lexical items. An example from Cruse \[1986\]: see is a general term for having a visual experience, but there is a wide range of more specific verbs that convey additional meaning; for instance, watch is used when one pays attention to a changing or a potentially changing visual stimulus, whereas look at implies that the stimulus is static. Such subtle semantic distinctions demand a fine-grained knowledge representation if a generator is expected to make these choices \[DiMarco et ai., 1993\].</Paragraph>
    <Paragraph position="5"> An important factor in lexical choice are collocalionai constraints stating that certain words can co-occur whereas others cannot. For instance, we find rancid butter, putrid fish, and addled eggs, but no alternative combination, although the adjectives mean very much the same thing. 2 Collocations hold among lexemes, as opposed to underlying semantic concepts, and hence have to be represented as lexicai relations. They create the problem that individual lexical choices for parts of the semantic representation may not be independent: roughly speaking, the choice of word x for concept a can enforce the choice of word y for concept b.</Paragraph>
    <Paragraph position="6"> Finally, a highly influential, though not yet very well-understood, factor in lexical choice is style.</Paragraph>
  </Section>
  <Section position="3" start_page="454" end_page="455" type="metho">
    <SectionTitle>
3 Lexical Style
</SectionTitle>
    <Paragraph position="0"> The notion of style is most commonly associated with literary theory, but that perspective is not suitable for our purposes here. Style has also been investigated from a linguistic perspective (e.g., Sanders \[1973\]), and recently a computational treatment has been proposed by DiMarco and Hirst \[1993\]. What, then, is style? Like Sanders, we view it broadly as the choice between the various ways of expressing the same message. Linguists interested in style, as, for instance, Crystal and Davy \[1969\], have analyzed the relationships between situational parameters (in 2In NLG, collocation knowledge has been employed by, inter alia, Smadja and McKeown \[1991\] and Iordanskaja, Kittredge and Polgu~re \[1991\].</Paragraph>
    <Paragraph position="1"> particular, different genres) and stylistic choice, and work in artificial intelligence has added the important aspect of (indirectly) linking linguistic choices to the intentions of a speaker \[Hovy, 1988\]. Clearly, the difficult part of the definition given above is to draw the line between message and style: what parts of an utterance are to be attributed to its invariant content, and what belongs to the chosen mode of expressing that content? In order to approach this question for the level of lexis, hence to investigate iezicai style, it helps to turn the question &amp;quot;What criteria do we employ for word choice?&amp;quot; around and to start by analyzing what different words the language provides to say roughly the same thing, for example with the help of thesauri. By contrastively comparing similar words, their differences can be pinned down, and appropriate features can be chosen to characterize them. A second resource besides the thesaurus are guidebooks on &amp;quot;how to write&amp;quot; (especially in foreign-language teaching), which occasionally attempt to explain differences between similar words or propose categories of words with a certain &amp;quot;colour&amp;quot; (cf. \[DiMarco et ai., 1993\]). One problem here is to determine when different suggested categories are in fact the same (e.g., what one text calls a 'vivid' word is labelled 'concrete' in another).</Paragraph>
    <Paragraph position="2"> An investigation of lexical style should therefore look for sufficiently general features: those that can be found again and again when analyzing different sets of synonymous words. It is important to separate stylistic features from semantic ones, cf.</Paragraph>
    <Paragraph position="3"> the choice criterion of semantic specificity mentioned above. The whole range of phenomena that have been labelled as associative meaning (or as one aspect under the even more fuzzy heading connotation) has to be excluded from this search for features. For example, the different overtones of the largely synonymous words smile, grin (showing teeth), simper (silly, affected), smirk (conceit, self-satisfaction) do not qualify as recurring stylistic features. Similarly, a sentence like Be a man, my son/alludes to aspects of meaning that are clearly beyond the standard 'definition' of man (human being of male sex) but again should not be classified as stylistic. And as a final illustration, lexicM style should not be put in charge to explain the anomaly in The lady held a white lily in her delicate fist, which from a 'purely' semantic viewpoint should be all right (with fist being defined as closed hand).</Paragraph>
    <Paragraph position="4"> Stylistic features can be isolated by carefully comparing words within a set of synonyms, from which a generator is supposed to make a lexical choice. Once a feature has been selected, the words can be ranked on a corresponding numerical scale; the experiments so far have shown that a range from 0 to 3 is sufficient to represent the differences. Several features, however, have an 'opposite end' and a neutral position in the middle; here, the scale is -3... 3.</Paragraph>
    <Paragraph position="5">  Ranking words is best being done by constructing a &amp;quot;minimal&amp;quot; context for a paradigm of synonyms so that the semantic influence exerted by the surrounding words is as small as possible (e.g.: They destroyed/annihilated/ruined/razed/.., the building).</Paragraph>
    <Paragraph position="6"> Words can hardly be compared with no context at all -- when informants are asked to rate words on a particular scale, they typically respond with a question like &amp;quot;In what sentence?&amp;quot; immediately. If, on the other hand, the context is too specific, i.e., semantically loaded, it becomes more difficult to get access to the inherent qualities of the particular word in question.</Paragraph>
    <Paragraph position="7"> These are the stylistic features that have been determined by investigating various guides on good writing and by analyzing a dozen synonym-sets that were compiled from thesauri: * FORMALITY: -3...3 This is the only stylistic dimension that linguists have thoroughly investigated and that is well-known to dictionary users. Words can be rated on a scale from 'very formal' via 'colloquial' to 'vulgar' or something similar (e.g., motion picture-movie-flick).</Paragraph>
    <Paragraph position="8"> * EUPHEMISM: 0...3 The euphemism is used in order to avoid the &amp;quot;real&amp;quot; word in certain social situations. They are frequently found when the topic is strongly connected to emotions (death, for example) or social taboos (in a washroom, the indicated activity is merely a secondary function of the installation). null</Paragraph>
  </Section>
  <Section position="4" start_page="455" end_page="455" type="metho">
    <SectionTitle>
* SLANT: -3...3
</SectionTitle>
    <Paragraph position="0"> A speaker can convey a high or low opinion on the subject by using a slanted word: a favourable or a pejorative one. Often this involves metaphor: a word is used that in fact denotes a different concept, for example when an extremely disliked person is called a rat. But the distinction can also be found within sets of synonyms, e.g., gentleman vs. jerk.</Paragraph>
    <Paragraph position="1"> * ARCHAIC ... TRENDY: -3... 3 The archaic word is sometimes called 'obsolete', but it is not: old words can be exhumed on purpose to achieve specific effects, for example by calling the pharmacist apothecary. This stylistic dimension holds not only for content words: albeit is the archaic variant of even though. At the opposite end is the trendy word that has only recently been coined to denote some modern concept or to replace an existent word that is worn out.</Paragraph>
    <Paragraph position="2"> * FLOPdDITY: -3...3 This is one of the dimensions suggested by Hovy \[1988\]. A more flowery expression for consider is entertain the thought. At the opposite end of the scale is the trite word. Floridity is occasionally identified with high formality, but the two should be distinguished: The flowery word is used when the speaker wants to sound impressively &amp;quot;bookish&amp;quot;, whereas the formal word is &amp;quot;very correct&amp;quot;. Thus, the trite house can be called habitation to add sophistication, but that would not be merely 'formal'. Another reason for keeping the two distinct is the opposite end of the scale: a non-flowery word is not the same as a slang term.</Paragraph>
    <Paragraph position="3"> * ABSTRACTNESS: -3...3 Writing-guidebooks often recommend to replace the abstract with the concrete word that evokes a more vivid mental image in the hearer. But what most examples found in the literature really do is to recommend semantically more specific words (e.g., replace to fly with to float or to glide), which add traits of meaning and are therefore not always interchangeable; thus the choice is not merely stylistic. A more suitable example is to characterize an unemployed person (abstract) as out of work (concrete).</Paragraph>
    <Paragraph position="4"> * FORCE: 0...3 Some words are more forceful, or &amp;quot;stronger&amp;quot; than others, for instance destroy vs. annihilate, or big vs. monstrous.</Paragraph>
    <Paragraph position="5"> There is an interesting relationship (that should be investigated more thoroughly) between these features and the notion of core vocabulary as it is known in applied linguistics. Carter \[1987\] characterizes core words as having the following properties: they often have clear antonyms (big--small); they have a wide collocational range (fat cheque, fat salary but .corpulent cheque, .chubby salary); they often serve to define other words in the same lexical set (to beam = to smile happily, to smirk = to smile knowingly); they do not indicate the genre of discourse to which they belong; they do not carry marked connotations or associations. This last criterion, the connotational neutrality of core words could be measured using our stylistic features, with the hypothesis being that core words tend to assume the value 0 on the scales. However, the coreness of a word is not only a matter of style, but also of semantic specificity: Carter notes that they are often superordinates, and this is also the reason for their role in defining similar words, which are, of course, semantically more specific. It seems that the notion of core words corresponds with basic-level categories, which have been employed in NLG by Reiter \[1990\], but which had originated not in linguistics but in cognitive psychology \[Rosch, 1978\].</Paragraph>
  </Section>
  <Section position="5" start_page="455" end_page="456" type="metho">
    <SectionTitle>
4 Towards a Model for Lexicalization
</SectionTitle>
    <Paragraph position="0"> When the input to the generator is some sort of a semantic net (and possibly additional pragmatic parameters), lexical items are sought that express all the parts of that net and that can be combined into a grammatical sentence. The hard constraint on which  (content) words can participate in the sentence is that they have the right meaning, i.e., they correctly express some aspect of the semantic specification. The second constraint is that collocations are not to be violated, to avoid the production of a phrase like addled butter. The other factors mentioned above enter the game as preferences, because their complete achievement cannot be guaranteed -- if we want to speak 'formally', we can try to find particularly formal words for the concepts to be expressed; but if the dictionary does not offer any, we have to be content with more 'standard' words, at least for some of the concepts underlying the sentence. We can maximize the achievement of lexical-stylistic goals, but not strive to fully achieve them.</Paragraph>
    <Paragraph position="1"> To arrive at this kind of elaborate lexical choice, I first employ a iexical option finder (following ideas by Miezitis \[1988\]) that scans the input semantic net and produces all the lexical items that are semantically (or truth-conditionally) appropriate for expressing parts of the net. If the set of options contains more than one item for the same sub-net, these items can differ either semantically (be more or less specific) or connotationally (have different stylistic features associated with them).</Paragraph>
    <Paragraph position="2"> The second task is to choose from this pool a set of lexical items that together express the complete net, respect collocational constraints (if any are involved), and are maximal under a preference function that determines the degree of appropriateness of items in terms of their stylistic and other connotational features. Finally, the choice process has to be integrated with the other decisions to be made in generation (sentence scope and structure, theme control, use of conjunctions and cue words, etc.), such that syntactic constraints are respected.</Paragraph>
    <Paragraph position="3"> Two parts of the overall system have been realized so far. First, a lexical option finder was built with LOOM, a KL-ONE dialect. Lexical items correspond to configurations of concepts and roles (not just to single concepts, as it is usually done in generation), and the option finder determines the set of all items that can cover a part of the input proposition (represented as LOOM instances). Using inheritance, the most specific as well as the appropriate more general items are retrieved (e.g., if the event in the proposition is darning a sock, the items darn, mend, fix are produced for expressing the action).</Paragraph>
  </Section>
  <Section position="6" start_page="456" end_page="456" type="metho">
    <SectionTitle>
5 Stylistic Lexical Choice in
PENMAN
</SectionTitle>
    <Paragraph position="0"> At the 'front end' of the overall system, a lexical choice process based on the stylistic features listed in section 3 has been implemented using the PENMAN sentence generator \[Penman-Group, 1989\].</Paragraph>
    <Paragraph position="1"> Its systemic-functional grammar has been extended with systems that determine the desired stylistic &amp;quot;colour&amp;quot; and, with the help of a distance metric (see below), determine the most appropriate lexical items that fit the target specification.</Paragraph>
    <Paragraph position="2"> Figure 1 shows a sample run of the system, where the :lexstyle keyword is in charge of the variation; its filler (here, slang or newspaper) is being translated into a configuration of values for the stylistic features. This is handled by the standard mechanism in PENMAN that associates keyword-fillers with answers to inquiries posed by the grammatical systems. In the example, the keyword governs the selection from the synonym-sets for evict, destroy, and building (stored in Penman's lexicon with their stylistic features). The chosen transformation of the :lexstyle filler into feature values is merely a first step towards providing a link from low-level features to more abstract parameters; a thorough specification of these parameters and their correspondence with lexical features has not been done yet.</Paragraph>
    <Paragraph position="3"> More specifically, for every stylistic dimension one system is in charge to determine its numeric target value (on the scale -3 to 3). Therefore, the particular :lexstyle filler translates into a set of feature/value pairs. When all the value-inquiries have been made, the subsequent system in the grammar looks up the words associated with the concept to be expressed and determines the one that best matches the desired feature/value-specification. For every word, the distance metric adds the squares of the differences between the target feature value (tf) and the value found in the lexical entry (wf) for each of the n features: ~i~=l(tfi - wfi) 2 The fine-tuning of the distance-metric is subject to experimentation; in the version shown, the motivation for taking the square of the difference is to, for example, favour a word that differs in two dimensions by one point over another one that differs in one dimension by two points (they would otherwise be equivalent). The word with the lowest total difference is chosen; in case of conflict, a random choice is made.</Paragraph>
  </Section>
class="xml-element"></Paper>