Assessing named entities in the ELTeC-por collection

Abstract

This paper reports on the NER annotation of the ELTeC-por collection, a collection of hundred Portuguese novels published between 1840 and 1920, compiled in the scope of the COST action "Distant reading for European literary history".

In addition to discussing its compilation, the choices taken and what remains to be done, we provide an initial characterization of the novels according to size, subgenre, publication place, author gender and which edition was used.

Then we present PALAVRAS-NER, the NER system which we use to annotate the collection, explaining the way it works.

We then focus on a subcollection of eight novels fully human revised, which we use to both evaluate the performance of the automatic system, and to characterize the population of the full collection. These novels can be further subdivided according to two different features: historical versus contemporary novels, on the one hand, and original vs. modernized ortography, on the other. Also some works are canonical while others are not.

In addition to the quantitative analysis of the annotation results and process, we present some qualitative description of the human revision as well.

We offer a detailed analysis of some categories, demonstrating how the most mentioned places, professions and demonyms can be good indicators for distant reading.

We end the paper comparing briefly with other work using named entities for literary texts and suggesting future work.