Robust mention detection is an essential component of any coreference resolution system. BART sup­ports different pipelines for mention detection. The

Figure 1 : BART architecture choice of a pipeline depends crucially on the avail­ability of linguistic resources for a given language.

For English and German, we use the Parsing Pipeline and Mention Factory to extract mentions. The parse trees are used to identify minimal and maximal noun projections, as well as additional fea­tures such as number, gender, and semantic class.

For English, we use parses from a state-of-the-art constituent parser (Petrov et al, 2006) and extract all base noun phrases as mentions. For German, the SemEval dependency tree is transformed to a constituent representation and minimal and maxi­mal phrases are extracted for all nominal elements (pronouns, common nouns, names), except when the noun phrase is in a non-referring syntactic position (for example, expletive "es", predicates in copula constructions).

For Italian, we use the EMD Pipeline and Men­tion Factory. The Typhoon (Zanoli et al., 2009) and DEMention (Biggio et al, 2009) systems were used to recognize mentions in the test set. For each mention, its head and extension were considered. The extension was learned by using the mention an­notation provided in the training set (13th column) whereas the head annotation was learned by exploit­ing the information produced by MaltParser (Nivre et al., 2007). In addition to the features extracted from the training set, such as prefixes and suffixes (1-4 characters) and orthographic information (capi­talization and hyphenation), a number of features ex­tracted by using external resources were used: men­tions recognized by TextPro (http://textpro.fbk.eu), gazetteers of generic proper nouns extracted from the Italian phone-book and Wikipedia, and other fea­tures derived from WordNet. Each of these features was extracted in a local context of ±2 words. 3.2 Features

We view coreference resolution as a binary classifi­cation problem. Each classification instance consists of two markables, i.e. an anaphor and potential an­tecedent. Instances are modeled as feature vectors (cf. Table 1) and are handed over to a binary clas­sifier that decides, given the features, whether the anaphor and the candidate are coreferent or not. All the feature values are computed automatically, with­out any manual intervention.

Basic feature set. We use the same set of rela­tively language-independent features as a backbone of our system, extending it with a few language-specific features for each subtask. Most of them are used by virtually all the state-of-the-art coreference resolution systems. A detailed description can be found, for example, in (Soon et al., 2001).

English. Our English system is based on a novel model of coreference. The key concept of our model is a Semantic Tree - a filecard associated with each discourse entity containing the following fields:

• Types: the list of types for mentions of a given entity. For example, if an entity contains the mention "software from India", the shallow predicate "software" is added to the types.

• Attributes: this field collects the premodifiers. For instance, if one of the mentions is "the ex­pensive software" the shallow attribute "expen­sive" is added to the list of attributes.

• Relations: this field collects the prepositional postmodifiers. If an entity contains the men­tion "software from India", the shallow relation "from(India)" is added to the list of relations.

The BART toolkit supports several models of coref­erence (pairwise modeling, rankers, semantic trees), as well as different machine learning algorithms. Our final setting relies on a pairwise maximum en­tropy classifier for Italian and German.

Our English system is based on an entity-mention model of coreference. The key concept of our model is a Semantic Tree - a filecard associated to each dis­course entity (cf. Section 3.2). Semantic trees are used for both computing feature values and guiding the resolution process.

We start by creating a Semantic Tree for each mention. We process the document from left to right, trying to find an antecedent for each men­tion (candidate anaphor). When the antecedent is found, we extend its Semantic Tree with the types, attributes and relations of the anaphor, provided they are mutually compatible. Consider, for example, a list of mentions, containing, among oth­ers, "software from India", "the software" and "soft­ware from China". Initially, BART creates the fol­lowing semantic trees: "(type: software) (relation: from(India))", "(type: software)" and "(type: soft­ware) (relation: from(China))" When the second mention gets resolved to the first one, their seman­tic trees are merged to "(type: software) (relation: from(India)". Therefore, when we attempt to resolve the third mention, both candidate antecedents are re­jected, as their relation attributes are incompatible with "from(China)" This approach helps us avoid erroneous links (such as the link between the second and the third mentions in our example) by leveraging entity-level information.