Fast Text Based Classification of News Snippets for Telecom Assurance

Abstract

The quality of Telecom companies’ mobile service can be seriously compromised by the occurrence of different types of events, whether they are expected or not. The goal of this work is to automatically identify online news that report such events. Three possible topics are searched for: “fire”; “meteorologic” and “public gatherings”. Remaining news’ topics should be ignored. Each category was specifically chosen by its relevance towards the most known network providers’ problems. The data is highly unbalanced. We compare different lightweight models for text classification using information collected from several Portuguese online newspapers: Support-Vector Machines (SVM), Fuzzy Fingerprints and K-Nearest Neighbours (KNN). More complex deep models, such as Bert or RoBerta, are dismissed due to the requirement of a fast response. The proposed models predict the categories based entirely on the title and the short news’ snippets that are freely available. Preliminary results indicate F1-scores above 0.78 for each of the three topics.