Early Text Classification Using Multi-Resolution Concept Representations

Proceedings of NAACL-HLT 2018, pages 1216–1225 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Early Text Classification using Multi-Resolution Concept Representations A. Pastor L´opez-Monroy⋆, Fabio A. Gonz´alez†, Manuel Montes-y-G´omez‡§, Hugo Jair Escalante‡ and Thamar Solorio⋆ ⋆ Department of Computer Science, University of Houston,Texas USA † Systems and Computer Engineering Department, Universidad Nacional de Colombia ‡ Instituto Nacional de Astrof´ısica, ´Optica y Electr´onica, Puebla M´exico § PRHLT Research Center, Universitat Polit`ecnica de Val`encia, Spain alopezmonroy@uh.edu, fagonzalezo@unal.edu.co, {mmontesg, hugojair}@ccc.inaoep.mx, solorio@cs.uh.edu Abstract This paper proposes a novel document repre- sentation, called Multi-Resolution Represen- tation (MulR), to improve the early detection of risks in social media sources. The goal is to effectively identify the potential risk us- ing as little evidence as possible and with as much anticipation as possible. MulR al- lows us to generate multiple “views” of the text. These views capture different semantic meanings for words and documents at differ- ent levels of granularity, which is very use- ful in early scenarios to model the variable amounts of evidence. The experimental eval- uation shows that MulR using low resolution is better suited for modeling short documents (very early stages), whereas large documents (medium/late stages) are better modeled with higher resolutions. We evaluate the proposed ideas in two different tasks where anticipation is critical: sexual predator detection and de- pression detection. The experimental evalua- tion for these early tasks revealed that the pro- posed approach outperforms previous method- ologies by a considerable margin. 1 Introduction Everyday there is a huge amount of people in- teracting in many social media sites. Unfortu- nately this immense cyber-world has been misused by cyber-criminals, who hide in the depths of the web. For this reason, the social media informa- tion has been increasingly studied in the context of applications related to security, forensics and e-commerce. Recently the early prediction sce- narios have attracted the attention of the scientific community (Losada et al., 2017), which aims to prevent major threats in a number of practical situ- ations by analyzing the text as evidence (e.g., sex- ual harassment, cyberbullying, etc). In Natural Language Processing this emerging field is called early text classification and the goal is to identify risky-target categories by using as few text as possible and with as much anticipa- tion as possible. In real scenarios the amount of evidence available from users under analysis is continuously growing. Consider for instance chat rooms, or posts and comments in social net- works, these text sources comprise cumulative evi- dence for early prediction that can be used to better capture the phenomenon under study (Escalante et al., 2017; Losada et al., 2017). This scenario has challenging particularities. For example, in early stages where 10% or 20% of the informa- tion is available it is necessary to model very short length documents, which tend to produce sparse and low discriminative representations. On the other hand late stages require to exploit as much evidence as possible to make accurate predictions. This dynamism between the document …

Content truncated to 500 words