Besides the well established problem of pronoun resolution, pronoun generation is now attract­ing renewed attention. In the past, generation systems generated pronouns without attaching much importance to the problem, one notable exception being the classical algorithm of Dale (1990), loosely based on centering theory. With the emergence of corpus based studies in compu­tational linguistics, the question arises whether it is possible to refine known standard algo­rithms, or whether an improvement is only to be achieved with the help of world knowledge reasoning - a matter too complex to be dealt with reliably at this time. The former direction is represented by the pioneering work of McCoy and Strube (1999). They propose a refined algo­rithm for the choice between definite description on the one hand and pronoun on the other for

* The work reported in this paper has been carried out with the financial support of UK ESPRC grant L51126.

animate referents , which is based on distance, time structure and ambiguity constraints.

Here we introduce a more general algorithm for the pronominalization decision that is valid not only for animate but for inanimate referents as well. In conformity with McCoy and Strube, we group noun phrases with definite determiner and proper names together under the term "def­inite description". The algorithm proposes a new pronominalization strategy, which beyond McCoy and Strube (1999)'s criteria makes use of the discourse status of the antecedent and parallelism effects.

The algorithm has been implemented as the reusable module gnome-np. It has been re-used in the web hypertext generation system ILEX (see Oberlander et al. (1998)). It shows an accuracy over 87% with respect to two corpora (each 5000 words) of different genres.

In previous accounts pronominalization has been related to the idea of a local focus of at­tention: a set of discourse referents who/which is in the center of attention of the speaker (e.g. Sidner (1979), givenness hierarchy (Gundel et al., 1993), centering theory (Grosz et al., 1995), RAFT/RAPR (Suri, 1993)). Whereas (Gundel et al., 1993) do not attempt to make their fo­cus notion operationalizable, this has been at­tempted by further developments of centering. However these have mostly been applied to the pronoun resolution problem. In the following we discuss three versions of centering and show that their application to the pronoun generation problem is nevertheless limited.

Centering. Centering was developed to ex­plain local discourse coherence; the extent to which it benefits pronoun generation is how­ever not immediately clear. In centering, the discourse entities evoked in a certain ut­terance ui are called forward-looking centers (Cfs). It is assumed that they are partially or­dered. As a major determinant of the ordering, the grammatical function hierarchy (roughly: subj>obj>others) has been proposed. Be­cause other factors affecting the order have not been elaborated in detail, this ranking (as the only operationalizable handle) has become the standard ranking in several computational ap­plications of centering. The backward-looking center (henceforth Cb) is a distinguished mem­ber of the Cfs, which is defined as the most highly ranked member of the Cfs of the previous utterance ui-1 which is realized in ui. The Cb is considered as the local focus of attention. Cen­tering states two rules. Only the first rule makes a claim about pronominalization: If any element of the utterance ui-1 is realized in ui as pro­noun, then the Cb must be pronominalized in Ui as well. As noted by McCoy and Strube (1999), this rule applies only in the case that two subse­quent utterances share more than one referent, and that the non-Cb referent is pronominalized in the second utterance. But why this non-Cb referent is realized as a pronoun is not given by the theory.

However, following more the spirit of center­ing than the actual definition, one can under­stand the Cb as the referent which is preferably pronominalized. General pronominalization of the backward-looking center was in fact a claim of early centering, but had to be abandoned be­cause of counter-evidence from real discourse. But the idea that pronominalization of the Cb could be a means of establishing local discourse coherence is still prevalent. It has accordingly been used by some generation systems to con­trol pronominalization - e.g. in the ilex sys­tem (Oberlander et al., 1998), the Cb is always realized as a pronoun.

Semantic centering. Centering is also found in Dale (1990) as the method of pronominaliza­tion control. However, Dale's center definition differs from standard centering theory in that it is defined semantically and not on the basis of a syntactic ranking. This approach has some appeal, especially for generation, because it sup­ports the natural modularity between strategic generation - which would determine the se­mantic center for each utterance - and tactical generation - which decides about grammatical functions.

Functional centering. Finally, the centering version suggested by Strube and Hahn (1999) appears to reveal an underlying discourse mech­anism responsible for centering: the information structure of an utterance (roughly the given-new pattern) is the deeper reason for the rank­ing of the forward-looking centers. This per­mits a generalization of standard centering into a language-independent theory covering both free and fixed word-order languages. It is how­ever then surprising that this result is not made maximal use of in the subsequent generation-oriented work of McCoy and Strube (1999).

Beyond centering. The questions which re­main open with all three approaches - stan­dard centering, semantic centering and func­tional centering - are:

P1; I Why are in real texts a large number of _ Cb's not pronominalized?

P2; I Why are non-Cb referents pronominalized? or expressed independently of centering:

Why are in real texts a large number of discourse entities with an antecedent in the previous utterance not pronominalized? Why can more than one entity be pronom-inalized in one utterance?

McCoy and Strube (1999) were the first to suggest an algorithm for generation which solves these problems. It was motivated by the ob­servation that a large percentage of NPs which would have been realized by pronouns using known algorithms, are in fact not realized as pronouns in real text. They suggest that such NPs serve to mark 'time changes' in the dis­course. Their algorithm accordingly makes use of distance, context ambiguity and temporal discourse structure to decide about pronomi-nalization. In our work, we have considered a corpus of a different genre in which temporal change does not play a determining role: de­scriptive texts. We propose a new algorithm

From a corpus-driven view, question larger problem.

The algorithm we will present below has been developed in close relation to the muse corpus -a corpus of museum exhibit labels. The corpus is a collection of web pages of the Paul Getty Museum, pages from an exhibition catalogue, and pages from a jewellery book. Typical char­acteristics are the central role of inanimate ref­erents in these texts, and the lack of temporal change - thus providing an interesting counter­part to the newspaper genre investigated by Mc­Coy and Strube.

With an overall set of around 5000 words, the corpus contains 1450 NPs. Each NP has been annotated with respect to, among others, gram­matical function, discourse status, gender, num­ber, countability, and antecedent relationships. 23% of the NPs form reference chains (i.e. at least two mentions of one and the same referent in one text), the other 77% are only mentioned once. We have 101 different reference chains; the chain-forming NPs fall into 101 discourse-new and 213 anaphoric NPs. In the following, we will only discuss the anaphoric NPs. 50% of the anaphoric NPs are realized as definite de­scriptions, 50% as pronouns. We distinguish be­tween locally bound pronouns, which are deter­mined syntactically (Binding Theory, (Chom­sky, 1981)), and which we expect the tacti­cal generator to handle correctly, and pronouns which are not locally bound - so-called dis­course pronouns. We investigated possible cor­relations between the discourse pronouns and semantic/pragmatic features of their context.

The basic notions that we found were dis­tance, discourse status of the antecedent, and grammatical function of the antecedent. All three notions need a precise definition.

Distance. To be able to determine the dis­tance between a discourse entity and its an­tecedent, a precise determination of what counts as utterance unit is necessary. Following Kameyama (1998), we take as utterance unit the finite clause. Relative clauses and com­plement clauses are not counted as utterances on their own. This means that we count clauses containing complement clauses or rel­ative clauses as single utterances.' The pre­vious utterance is the preceding utterance at the same level of embedding.

Note that we allow the treatment of clauses with VP coordination (subject ellipsis) as com­plex coordinated clauses as done in Kameyama (1998), thus handling subject ellipsis as a dis­course pronoun; our algorithm does not insist on this view however.

The following correlation between pronoun use and distance was found in our corpus: 97% of the pronouns have an antecedent in the same or the previous utterance.

Discourse status. The information status of a discourse entity in an utterance is either given or new. We use these terms with an identi­cal meaning as ground and focus in Vallduvi (1993). Discourse status, as introduced by Prince (1992), is a similar but different notion: A discourse entity is discourse-old, ifithas been mentioned before in the same discourse; it is discourse-new otherwise. All cases of givenness by indirect means like part-whole, set-member relationships, other bridging rela­tions, inferences (Prince's inferrables, anchored and situationally evoked entities) are judged as discourse-new, thus taking into account only the identity antecedent relationship. We share Prince's opinion that pronominalization has to do with discourse status, whereas definiteness has to do with information status.

66% of all short-distance discourse pronouns in the muse corpus refer to an antecedent which is in itself discourse-old.

Subjecthood. The third strong correlation is the relation between pronoun use and the gram­matical function of the antecedent. 63% of dis­course pronouns have a subject as antecedent. The following table shows the overall distribu­tion of antecedent properties for short-distance discourse pronouns and (shown in brackets) for short-distance definite descriptions.

Based on these corpus study results, we define a new notion of the local focus - the set of refer­ents which are available for pronominalization. The local focus is updated at each utterance boundary, and is defined as the set of referents of the previous utterance which are:

(a) discourse-old, or (b) realized as subject.

This set can theoretically contain more than one referent, but in most cases, (a) and (b) are one and the same singleton set, which could be seen as the well-known Cb. Thus standard center­ing appears as a special case of our approach. This account means that newly introduced ref­erents are not immediately pronominalized in the following utterance, unless they have been introduced as subject - an observation made by Brennan (1998) and now confirmed with respect to our data also.

The proposed definition of the local focus gen­eralizes the focusing mechanisms assumed in centering and introduces the discourse status of the antecedent as one main criterion behind the pronominalization decision. It is interesting to note that McCoy and Strube (1999) also make use of the discourse status of the antecedent without mentioning it explicitly. For a certain subset of intrasentential anaphoric relations in ambiguous contexts they propose pronominal-ization in case the antecedent would be the pre­ferred one in Strube (1998)'s pronoun resolu­tion algorithm. Because the set of antecedents is ranked there with respect to information sta­tus, this is identical with our proposal. Why they do not use the discourse status as a gen­eral criterion is not clear. We believe that the discourse status of the antecedent as pronomi-nalization trigger is a general rule in discourse semantics.

The central role of discourse status and sub-jecthood are in our opinion not accidental. The two notions reflect two typical strategies to introduce a new referent into the discourse. We will assume here the unmarked information structure of an utterance: given — new. The subject usually is part of (or identical to) the given. Let X be a certain referent which is newly introduced in utterance (u1), and referred to again in the following utterance (u2). In the first strategy, X is introduced in the new non-subject part of (u1). And in this pattern the sec­ond mention of X in (u2) is not pronominalized. In example (1) given in Figure 1 the local focus for utterance (u2) has one element: {he}; "the main rooms" is new in (u1) and not pronominal-ized in (u2). The other typical strategy is where the referent is first mentioned in a subject posi­tion. This is typical for a segment onset, or the beginning of a text. Often this referent is given by other means - for example, by reference to a picture, or to a related object. In example (2) of Figure 1, the second mention is pronominal-ized. Thus the subject position seems to func­tion as creating a givenness allocation for the denoted referent. These two strategies roughly correspond with two types of thematic develop­ment identified in Danes (1974).

Parallelism. Our definition of the local fo­cus licenses 91% (62 of 68 pronouns) of all short-distance discourse pronouns in our corpus. Looking at the pronouns violating the proposed account, we made an interesting observation: most of them occur in contexts of strong par­allelism. We call an anphoric NP np2 parallel if it has an antecedent np1 in the previous utter­ance, and np1 and np2 have the same grammati­cal function. For work with real text, it is useful to include cases where np2 is a possessive or gen­itive NP inside a certain np3, and np1 and np3have the same grammatical function. Depend­ing on the concrete function, we distinguish sub­ject and object parallelism. Strong parallelism is a simultaneous subject and object parallelism in two consecutive clauses. Strong parallelism always overrides the local focus criterion, and allows for pronominalization of referents with discourse-new antecedents in nonsubject posi­tion.

but most cases of

P1.1 ambiguous context,

The local focus definition refined by the parallelism effect is an explanation for question and a small portion of remain open. Two reasons for not pronominalizing a referent which is a member of the local focus need to be considered:

A comparison of the performance of our algo­rithm with the annotated muse corpus and Mc­Coy and Strube's newspaper corpus is given in Table 1. The evaluation has been carried out for the algorithm gnome-np without employing the repetition blocking rule and without a fine­grained discourse segmentation. Layout seg­ments were used for the muse corpus. Because the number of annotated segment onsets for the newspaper corpus is not easy to re-establish, we give here two figures for this corpus: first with­out any segment onset signalling (lower bound), and second with the assumption that 15 short-distance definite descriptions mark segment on­sets. The figures include locally-bound pro­nouns to yield better comparability with McCoy and Strube. The figures in the columns 'gnome-np' represent those NPs whose form is predicted correctly by the new algorithm when evaluated against the annotated corpora.

The figures in Table 1 show that our al­gorithm performs very well in both domains, even without using a finer discourse segmen­tation such as temporal structure. Moreover, it performs better on McCoy and Strube's cor­pus than their own algorithm, which success­fully predicted the choice between realization by pronoun and realization by definite description in 84.7% of all cases. The disagreements oc­cur first for long distance pronouns (in our ter­minology: pronouns more than one clause dis­tant) and, second, in longer referent chains with well established focus. For the latter, whereas gnome-np would always suggest a pronoun, the real discourse swaps between pronoun and defi­nite description. Thus a finer segmentation or a repetition blocking rule could still improve the result further.

This paper has presented a new algorithm for the pronominalization of third person discourse entities. The algorithm, first, is implemented as a reusable module for generation systems and, second, provides a theoretical account of pronominalization in general.

The proposed algorithm provides a solution for question P2 above by widening the defini­tion of local focus to be a set with possibly more than one referent. The algorithm also offers a new solution for problem P1.1 above, ambigu­ous pronoun generation. Discourse structuring ( P1.2 ) is assumed as given. A sufficiently fine­grained discourse structuring has been explored, for example, by McCoy and Strube for their do­main of newspaper articles, but remains an issue for future research for other domains. We have shown that next to proximity, the discourse sta­tus of the antecedent is a main criterion for trig­gering pronominalization.

The suggested algorithm generalizes known focusing accounts. Gundel et al. (1993)'s cog­nitive status of being "in focus" is now approxi­mated by the set of all discourse-old entities and the subject of the previous utterance. The new focus determination is also a generalization of centering's Cb. The focus so defined serves two functions simultaneously: to trigger pronomi-nalization, and to provide the set of competitors for pronoun generation in ambiguous contexts. Although our training corpus is too small to jus­tify general claims, the evaluation with respect to the newspaper genre provides evidence that this finding is valid for planned discourse in gen­eral, independent of the concrete genre.