A fine-grained recognition of Named Entities in ELTeC collection using cascades - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2022

A fine-grained recognition of Named Entities in ELTeC collection using cascades

Résumé

In the scope of the COST action “Distant Reading for European Literary History” (Schöch et al. 2021; Patras et al. 2021) the working group 2 (WG2) responsible for methods and tools suggested a set of seven named entity (NE) categories to be used for annotating novels (the so-called “level-2” text version). Tags to be used for this set are: PERS, LOC, ORG, WORK, EVENT, ROLE, DEMO (Frontini et. al 2020; Šandrih Todorović et al. 2021). The level-2 version of Serbian novels was produced using this set of categories and tags (Krstev et al. 2019). For Serbian and French the fine-grained named entity recognition systems were developed based on exhaustive lexicons of corresponding languages and rules implemented in the form of cascades of finite-state automata (Maurel and Friburger 2014; Krstev et al. 2014). These systems were developed using the open-source corpus processing suite Unitex/GramLab and its module CasSys. Both systems recognize and tag a rich set of NE categories and subcategories and allow entity embedding; moreover, the French system recognizes NEs that correspond to TEI guidelines, chapter 13 (TEI P5). An example that illustrates this in French is (Marquis de la Lande factories): usines de la Lande Similarly, in Serbian (Queen Elizabeth of Hungary): kraljice Ugarske Elizabete Moreover, both systems recognize beside broad categories suggested by WG2 the other categories such as temporal or measurement expressions. In both Serbian and French systems, the recognition module is separated from the annotation module, which enables production of output as needed. In this paper we will illustrate this on a few Serbian and French novels from ELTeC corpus chosen to match in respect to corpus balance criteria, namely author’s gender, novel’s size, year of first publication. The novels will be annotated with the simplified tags needed for level-2 text format, and with more elaborate TEI compliant tags that reflect all nuances of recognized NEs. Two output formats for Serbian and French novels will be uploaded into TXM corpus processing systems which will enable both quantitative and qualitative analysis (Krstev et al., 2019). Besides statistical analysis of annotated NER, we will perform contrastive analysis of Serbian and French NEs and for both languages between fine-grained and simplified versions of annotation. The qualitative analysis will reveal interesting examples of annotation, open issues and hard cases. Textometrie analysis in TXM will be illustrated for both fine-grained and simplified versions of annotated samples. Finally, we will go back to the research questions that were posed by Action’s working group 3 (literary theory and history) when the Action started. Namely the first idea and wish of the WG3 was to produce fine grained annotations that will allow, for instance, distinction between cities and villages, different person’s roles (professions, family relations, etc.), person’s gender, types of locations (continent, country, region, city, village, mountain, waterbody, astronym), etc. After the analysis of availability of NER tools, the fine-grained approach was substituted with a much simpler schema. With this research we would like to reopen these questions and establish whether it is possible to meet the need for more detailed literary analysis based on Named Entities.
Fichier non déposé

Dates et versions

hal-03615219 , version 1 (21-03-2022)

Identifiants

  • HAL Id : hal-03615219 , version 1

Citer

Cvetana Krstev, Denis Maurel, Ranka Stanković. A fine-grained recognition of Named Entities in ELTeC collection using cascades. Final Action Event of COST Action Distant Reading for European Literary History, Christof Schöch, Apr 2022, Krakow, Poland. ⟨hal-03615219⟩
59 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More