Discourse-Givenness of Noun Phrases: theoretical and computational models

ness/SpecificityNews texts usually report on events involving concrete referents. Concrete referents tendto be described rather specifically, often resulting in longer expressions (length in chars).Also, embedded definite phrases or expressions with a possessive ‘s’ often have specificreferents: typical possessors are e.g. persons or organizations. On the other hand, certainsuffixes point to abstract common nouns, such as -ion, -ness etc. (suffix n).Specificity is a precondition for coreference in many corpora. I understand specificity asa feature representing two components:a) an expression’s descriptive content (preciseness of description), andb) relatedness to the context it occurs in.When referring to a particular object, the speaker needs to distinguish it a) from otherobjects in the world, and b) from other discourse referents in the context.A distinction from other objects in the world can be accomplished by using proper namesor precise descriptions (e.g. the blue book or the lexicon is more precise than the book).A distinction from other discourse referents can be made by using different concepts,or, when re-using lexical material, by adding descriptive content (e.g. the town thenew town, a bus another bus) or by using the indefinite plural form (e.g. Investigatorscontinued their search. At Calverton, investigators began piecing together the airplane).For an operationalisation of specificity, I assume that1) the more precise the description, the fewer documents it occurs in, and2) the more important a discourse referent is for a discourse d , the more frequently itoccurs in d .As a measure for these two components of specificity, I use1) inverse document frequency (idf) and2) term frequency (tf), combined to tfidf, a well-known measure from the field of Infor-mation Retrieval.For the purpose of discourse-givenness classification, tfidf is calculated using the fullexpression as a term on the one hand, and using a sliding window of characters as a termon the other hand (4 characters in the experiments presented here). The sliding windowwas implemented to make the method more robust against inflection, composition, andthe shortening of names (e.g. Alex for Alexander), considering in particular that it isapplied not only to English.The calculation of tf-idf-related features is sketched at the end of this section.SimilaritySynonyms are sometimes used for referring to the same entity for stylistic reasons, e.g. toavoid word repetitions. In earlier work, latent semantic analysis and variants (Hempel-mann et al., 2005) have been used, as well as semantic relatedness, measured e.g. bymeans of WordNet or GermaNet (Markert et al., 2012; Cahill and Riester, 2012). In thepresent work, semantic similarity as calculated by DISCO (Kolb, 2008) is used in thefeature maxsimilar mention. The calculation of this feature is sketched at the end ofthis section.ContextMotivated by the findings in the theoretical part (Section 2.3), I also introduce featuresexploiting an expression’s context. These contextual features are designed to complementDistributional similarity (LSA and span) between all words of the NP and the NP’s preceding contexthas been used by Hempelmann et al. (2005) in their logistic regression experiments.