Abstract
In the present paper we describe our approach to semi-automatic text annotation based on clustering. Given a large collection of announcements of cultural events from several websites, we group them based on their content and infer respective semantic categories that can be used for annotation (e.g. lecture, sports, food, music).
We experiment with various models for vectorising the texts, including pretrained multilingual Sentence Transformers and multilingual ELMo models. The produced text embeddings are then clustered using K-means.
We evaluate our clustering results using a stratified sample of texts with pre-existing categories (collected from websites listing the events) as well as intrinsic evaluation measures. The rationale behind this work is to produce a single categorisation covering texts from various sources and in two languages - English and Russian.
The labelled collection of texts is intended for use in a Digital Humanities project aimed at describing cultural life in a selected location, for example, comparing types of events in Russian and British cities.