ANCOR-AS: Enriching the ANCOR Corpus with Syntactic Annotations

ANCOR-AS: Enriching the ANCOR Corpus with Syntactic Annotations Loïc Grobol∗ †, Isabelle Tellier∗, Éric de la Clergerie†, Marco Dinarelli∗, Frédéric Landragin∗ ∗ Lattice ; † ALMAnaCH, Inria 1 rue Maurice Arnoux, 92120 Montrouge, France ; 2 rue Simone Iff, 75589 Paris, France {loic.grobol, isabelle.tellier, marco.dinarelli, frederic.landragin}@ens.fr, eric.de_la_clergerie@inria.fr Abstract This paper presents ANCOR-AS, an enriched version of the ANCOR corpus. This version adds syntactic annotations in addition to the existing coreference and speech transcription ones. This corpus is also released in a new TEI-compliant XML format. Keywords: Coreference, Parsing, Oral Data, TEI 1. Introduction Since its inception, even as far back as (Hobbs 1986), au- tomatic detection of coreference and anaphora has made consequent use of rich syntactic knowledge. To this day, most of the state-of-the-art systems make use of at least some information inferred from syntax analyses available in coreference-annotated treebanks. Until now, however, no French coreference corpus had publicly available syntactic analysis, which led previous works on automatic coreference detection to either use ad-hoc automatic parsing and shallow syntactic analysis or to use manually annotated mentions, thus avoiding the problem of mention detection. In an effort to at least start to address this lack of resource, we present ANCOR-AS, an enriched version of ANCOR (Muzerelle et al. 2014) — the current reference corpus for coreference in French — that includes syntactic analysis ob- tained in the Universal Dependencies framework (Nivre et al. 2016) through state-of-the-art automatic parsing techniques. ANCOR-AS builds upon our previous efforts in (Grobol, Landragin, and Heiden 2017) to develop a standard format for coreference annotations and extends our proposal for TEI-compliant (TEI consortium 2016) reference annotations to the case of syntactic dependencies. Our preliminary experiments with using automatic syntactic analysis for mention detection, described in (Grobol, Tellier, et al. 2017) give us confidence that in the absence of gold syn- tactic analysis, automatic parsing is a valuable asset for devel- oping an end-to-end coreference detection system. Moreover, unrelated works on ANCOR, such as Temporal@ODIL (An- toine et al. 2017), are already using semi-automatic parsing as a support for further manual annotations, which suggests that standardized syntactic annotations would be useful for further uses of ANCOR beyond coreference. 2. Context 2.1. Related works As mentioned earlier, there already exist several coreference treebanks, most notably the AnCoRa corpus (Taulé, Martí, and Recasens 2008) (for Spanish and Catalan), the Prague Dependency Treebank (Nedoluzhko et al. 2016) (for Czech) and the omnipresent OntoNotes (Pradhan, Hovy, et al. 2007). To this day, the latter is still the largest coreference corpus, with nearly 3M words in three languages – Arabic, English and Chinese – and the evaluation standard for coreference detection system since its use for the CoNLL-2011 (Pradhan, Ramshaw, et al. 2011) and CoNLL-2012 (Pradhan, Mos- chitti, et al. 2012) coreference detection shared tasks. Also relevant for our case is the NXT-Switchboard corpus (Cal- houn et al. 2010), a coreference treebank of oral English transcriptions. For some other languages, only coreference corpora with rich morphological or shallow syntactic annotations exist, such …

Content truncated to 500 words