FrNewsLink : a corpus linking TV Broadcast News Segments and Press Articles

FrNewsLink : a corpus linking TV Broadcast News Segments and Press Articles Nathalie Camelin1, G´eraldine Damnati2, Abdessalam Bouchekif1,3, Anais Landeau1, Delphine Charlet2, Yannick Est`eve1 (1) LIUM, Maine University, France – {firstname.surname}@univ-lemans.fr (2) Orange Labs, Lannion, France – {firstname.surname}@orange.com (3) LSE, EPITA, France – {firstname.surname}@epita.fr Abstract In this article, we describe FrNewsLink, a corpus allowing to address several applicative tasks that we make publicly available. It gathers several resources from TV Broadcast News (TVBN) shows and press articles such as automatic transcription of TVBN shows, text extracted from on-line press articles, manual annotations for topic segmentation of TVBN shows and linking information between topic segments and press articles. The FrNewsLink corpus is based on 112 (TVBN) shows recorded during two periods in 2014 and 2015. Concomitantly, a set of 24,7k press articles has been gathered. Beyond topic segmentation, this corpus allows to study semantic similarity and multimedia News linking. Keywords: content linking, semantic textual similarity, topic segmentation 1. Introduction Several corpora are dedicated to evaluation in the domain of topic segmention, semantic similarity or resource link- ing (textual or multi-modal). An overview of such corpora is proposed in Section 2. contextualizing our proposed cor- pus. We describe in this article the corpus FrNewsLink that can be downloaded on the LIUM website1 in its first ver- sion. During two different periods of time: 7 consecutive days in February 2014 and 2 consecutive days in January 2015, several TVBN shows from 8 different French chan- nels have been recorded and web press articles appearing on the Google News homepage have been downloaded. As a consequence, FrNewsLink resources (whose elements are described in Section 3. and detailed statistics are given in Section 5.) deal with various fields such as politics, sports, cinema, etc. and some events that are very dynamic during a given day. Then, a manual annotation process (described in Section 4.) has been performed in order to obtain: 1- topic segmentation annotations and 2- linking annotations between topic segments and press articles. Therefore, this corpus is very useful for many tasks such as topic seg- mentation, topic titling, video linking, semantic similarity, events follow-up and topic modeling (grouping documents addressing similar topics). Section 6. presents several tasks that can be addressed with our corpus, along with evalua- tions of theses tasks performed on it. 2. Related work 2.1. Corpora for topic segmentation The C99 corpus designed by (Choi, 2000) has been widely used for evaluating topic segmentation of written text. It is an artificial corpus composed of 7000 segments randomly selected from articles of the Brown corpus. These segments are grouped into 700 documents where each document is the concatenation of 10 segments. A segment is composed of the first n sentences of the original article. The ICSI corpus (Shriberg et al., 2004) (Shriberg et al., 2000) contains 75 documents transcribed automatically 1https://lium.univ-lemans.fr/frnewslink/ from meeting records (approximately one hour each). For each conversation turn the speaker, start time, end time, and word content are marked. This corpus has been exploited in several …

Content truncated to 500 words