SACR: A Drag-and-Drop Based Tool for Coreference Annotation

SACR: A Drag-and-Drop Based Tool for Coreference Annotation Bruno Oberle LiLPa (EA 1339), Université de Strasbourg 14 rue Descartes, F-67084 Strasbourg Cedex, France oberleb@unistra.fr Abstract This paper introduces SACR, an easy-to-use coreference chain annotation tool, which is used to annotate large corpora for Natural Language Processing (NLP) applications. Coreference annotation is usually considered as costly both in terms of time and human resources. So, in order to find the easiest annotation strategy, we will first of all compare several annotation schemes implemented in existing tools. Since interface ergonomics is also an important part of our research, we then focus on identifying the most helpful features to reduce the strain for annotators. In the next section of the paper, we present SACR in details. This tool has been developped specifically for coreference annotation, and its intuitive user interface has been designed to facilitate and speed up the annotation process, making SACR equally suited for students, occasional and non-technical users. In order to create coreference chains, elements are selected by clicking on the corresponding tokens. Coreference relations are then created by drag-and-dropping expressions one over the other. Finally, color frames around marked expressions help the user to visualize both marked expressions and their relations. SACR is open source, distributed under the terms of the Mozilla Public License, version 2.0, and freely available online. Keywords: coreference annotation, annotation tool, coreference chain, interface ergonomics 1. Introduction and Context The development of statistical methods in automatic lan- guage processing leads to an increased need for annotated resources. In particular, training algorithms that aim to de- tect coreference1 requires manually annotated texts. Be- cause such a task is costly in terms of time and human re- sources, there are few annotated corpora with coreference relations. In French, there is only one such a corpus large enough to be used with supervised methods: the ANCOR corpus (Muzerelle et al., 2014), with 488,000 lexical units, 116,000 referring expressions (only nouns and pronouns) and 51,300 relations. But it contains only spoken French. The “Democrat” project2 aims, among others, at providing a large corpus annotated with coreference chains. It is ex- pected to have one million words and 100,000 elements of coreference chains. Several annotation strategies have been discussed (Landragin et al., 2017) in order to balance sci- entific needs with annotation speed. The main purpose of this paper is to introduce SACR3, a tool specifically designed to facilitate and speed up anno- tation of coreference chains. We first compare coreference annotation strategies used by different tools, both in terms of annotation scheme (section 2) and interface usability 1There is coreference when two linguistic expressions refer to the same referent, that is, the same entity. For example, in My cat is drinking milk. It is really thirsty, both my cat and it refer to the same entity (the cat I own) and thus are coreferential. All expres- sions that refer to the same entity are said to be in a coreference chain. 2DEscription et MOdélisation des Chaînes de Référence: out- …

Content truncated to 500 words