Although most pronominal anaphora resolu­tion algorithms only account for anaphors re­ferring to individual entities, anaphors re­ferring to abstract entities evoked by verbal phrases, clauses or discourse segments (hence­forth APAs) are quite common in English see i.a. (Byron and Allen, 1998) and even more in Danish (Navarretta, 2004). Recently, two algorithms for resolving APAs and individ­ual pronominal anaphors (henceforth IPAs) in specific English dialogues have been proposed Eckert and Strube's (2000) ESÜ0 and Byron's (2002) PHORA. Furthermore Strube and Müller (Strube and Müller, 2003) have presented a machine learning approach for resolving APAs in spoken dialogues. Both ESÜ0 and phora recognise IPAsand APAs on the basis of se­mantic constraints on the argument position oc­cupied by the anaphors and account for differ­ences in reference between personal and demon­strative pronouns. In ES00 demonstrative pro­nouns preferentially refer to abstract entities, while personal pronouns preferentially refer to individual ones. ES00 resolves IPAs applying Strube's (1998) algorithm. In phora the an­tecedents of personal pronouns are searched for looking at their degree of salience which is im­plemented by word order as in (Grosz et al., 1995). Demonstratives, instead, are searched for in the list of activated entities (Gundel et al., 1993) containing non NP antecedents, which are assumed to be less salient. In phora demon­stratives can also refer to Kinds.

ES00 requires that the structure of dialogues has been marked. Byron's PHORA-algorithm does not rely on predefined dialogue structure, but only searches for abstract antecedents of APAs in the sentence preceding the anaphor. Thus it does not account for APAs referring to larger discourse segments. phora relies on both semantic knowledge and a model of speech acts and accounts for more phenomena than ES00. Differing from ES00, phora has been imple­mented.

To resolve IPAsand APAs inDanishtexts and dialogues Navarretta (2004) has proposed the so-called DAR-algorithm (Navarretta, 2004). In dar APAs are resolved following the ES00 strategy, but dar accounts for the Danish ab­stract anaphors which occur in much more con­texts than the English ones. Individual anaph­ors are resolved in dar following a novel strat­egy which combines models which identify high degree of salience with high degree of givenness (topicality) of entities in the hearer's cognitive model, e.g. (Grosz et al., 1995), with Hajicova et al.'s (1990) salience account which assigns the highest degree of salience to entities in the focal part of an utterance in Information Structure terms. These entities often introduce new in­formation in discourse.

In the present paper we describe an extended version of the DAR-algorithm accounting for dif­ferences in the reference of various types of Dan­ish demonstratives which we found analysing the uses of pronouns in three text collections (computer manuals, novels and newspaper ar­ticles) and three corpora of recorded naturally-occurring dialogues (the SL (Duncker and Her­mann, 1996), the BYSOO (Gregersen and Peder-sen, 1991) and the pid corpora (Jensen, 1989)). In the following we first discuss the background for our proposal (section 2) then we describe the extended DAR-algorithm (section 3). In sec­tion 4 we evaluate it and compare its perfor­mance with that of other known algorithms. Fi­nally we make some concluding remarks.

In most applied approaches resolving pronom­inal anaphors mainly consists in the following steps: 1: determining the anaphoric antecedent domain; 2: choosing the most prominent or salient antecedent among possible candidates. Thus determining the degree of salience of dis­course elements, henceforth DEs, is essential to anaphor resolution. Although there is not al­ways an identity relation between linguistic an­tecedents and referents, we also follow this strat­egy, well aware that it is particularly problem­atic for resolving APAs (se especially (Webber, 1991)). Nearly all salience-based algorithms identify high degree of salience with high de­gree of givenness of DEs. In fact, although dif­ferent criteria are used for ranking DEs, such as linear order, hierarchy of grammatical roles, in­formation structure, Prince's Familiarity Scale (Prince, 1981), all algorithms assign the highest prominence to the DEs which are most topi­cal, known, bound, familiar and thus given,e.g. (Grosz et al., 1995; Brennan et al., 1987; Strube, 1998). Analysing the Danish data we found a restricted number of cases where high degree of salience did not correspond to high degree of topical ity, as it is the case in example (1). (1) A: hvem...hvem arbejdede [din mor]i med? (with whom... whom did [your mother]i work) B: [Hun]i arbejdede med [vores nabo]k([She]i worked with [our neighbour] k) [Hun]k var enke ... havde tre s0nner [bysoc] ([She]k was a widow... had three sons) In (1) the antecedent of the second occurrence of the pronoun hun (she) is the object vores nabo (our neighbour) which provides the information re­quested in the preceding question. This nominal is assigned lower prominence than the subject pronoun hun (she)inmostsalience models.

The only salience model which departs from the givenness assumption has been proposed by Hajicova et al. (1990). Hajicova et al., in fact, assign the highest degree of salience to DEs in the focal part of an utterance in information structure terms (Sgall et al., 1986). These enti­ties often represent new information. Hajicova et al.'s approach is original and can account for the data in (1), but it is problematic from an applied point of view. In the first place it is difficult to determine the information structure of all utterances. Secondly, focal candidate an­tecedents are ranked highest in Hajicova et al.'s model, but they still compete with given can­didate antecedents in their system. Finally the data does not confirm that all entities in the fo­cal part of an utterance have the highest degree of accessibility.

We agree with Hajicova's insight, but in or­der to operationalise the influence of focality in a reliable way propose the following. Accessi­bility by default is connected with givenness as assumed in most algorithms. However, when speakers explicitly change the degree of accessi­bility of entities in discourse by marking them as salient with information structure related de­vices, these entities get the highest degree of salience and are proposed as the preferred an­tecedents of anaphors. In cases of explicit focus marking the shift of focus of attention is as co­herent as continuing speaking about the same entities, because it is preannounced to the ad­dressee. On the basis of the data we propose a list of identifiable constructions in which ex­plicit focus marking occurs. Examples from the list are the following:

a: Entities referred to by NPs which in Danish are focally marked structurally (clefts, existen­tial and topicalised constructions). b: Entities referred to by NPs that follow fo­cusing adverbs.

c: Entities focally marked by the prosody (if this information is available) and/or entities providing information requested in questions, as in (1).

Giveness subsumes here concepts such as topicality and familiarity.

We have manually tested dar on randomly cho­sen texts and dialogues from our collections. The performance of dar on dialogues has been compared with that of Es00. The function for resolving IPAs(ResolveIpa) has similarly been tested on texts, where APAswereex-cluded. We have compared the obtained re­sults with those obtained by testing bfp (Bren-nan et al., 1987) and str98 (Strube, 1998). In all tests the intrasentential anaphors have been manually resolved and expletive and cat­aphoric uses of pronouns have been marked and excluded from the test. Dialogue act units were marked and classified by three annotators fol­lowing (Eckert and Strube, 2000). The relia­bility for the two annotation tasks (^-statistics (Carletta, 1996)) was of 0.94 and 0.90 respec­tively. Pronominal anaphors were marked, clas­sified and resolved by two annotators. The k-statistics for the pronoun classification was 0.86. In few cases (one in the texts and two in the dia­logues) where the annotators did not agree upon resolution, the pronouns were marked as am­biguous and were excluded from the test. The results obtained for bfp and str98 are given in table 1, while the results of dar's ResolveIpa are given in table 2. In the tables CR stands for "correctly resolved", HR stands for "resolved by humans", RA stands for "resolved over all", P stands for precision and R stands for recall. Because dar both classifies and resolves anaph-ors, both precision and recall (respect to hu­man resolution) are given in table 2. The re­sults indicate that ResolveIpa performs sig­nificantly better than bfp and str98 on the

1 anteced.

anteced.

and list ambig.

In this paper we presented dar, an algorithm for resolving IPAsandAPAsinDanishtexts and dialogues. In dar differences between the referential characteristics of Danish weak and strong pronouns are accounted for and a novel strategy for resolving individual anaphors is proposed. This strategy combines givenness with focality preferences to model salience and also accounts for parallelism preferences. dar performs significantly better on IPAsthan al­gorithms which only rely on givenness-based salience models.

dar extends the es00 strategy of classify­ing and resolving (some types of) APAs. The tests of dar indicate that the es00's approach of recognising APAsisalsopromisingfortexts and other languages than English.

dar has not been compared with phora which is the only abstract anaphora algorithm implemented. We find the algorithm very inter­esting because it addresses many of the same phenomena, but with different strategies. It would be useful to combine some of these strate­gies with the approaches proposed in dar and Es00 to improve the still problematic resolution of abstract anaphors.