A social affective text mining approach for detecting human emotions on specific topics in twitter data

Abstract

The increase in availability of public unstructured data has led to an increasing interest in analysing and understanding its contents. In this context, text mining techniques have been developed, most of which classify text with respect to its positive or negative polarity.More sophisticated emotion mining and social affective text mining techniques deal with the detection of specific emotions in text. Horizon scanning is a research field which may benefit a lot from text mining but little work has been done to support this idea. Its aim is to identify weak signals for emerging issues, which is traditionally done through producing a list of topics by manual scanning of text documents. According to current research in horizon scanning, it can be argued that an indicator for topic relevance is the occurrence of emotions in the written context of a specific topic. Based on this assumption, the aim of this work was to design an emotion mining approach for Twitter micro-blogging posts which supports the horizon scanning process. It considers Twitter-specific factors, such as the use of a limited character length and the use of social media language. The proposed approach was evaluated by measuring emotion mining accuracy as well as its applicability to horizon scanning. This was done by using three Twitter corpora: (1) an accuracy evaluation corpus, (2) a corpus containing Tweets from November 2013 to March 2014 with hashtags related to the Ukraine and (3) a corpus containing Tweets from 1 November 2013 posted from UK locations. Precision values of the proposed approach reached an average of 50% and two out of four identified horizon scanning criteria were compatible with the proposed emotion mining approach. These results show that the proposed novel approach is an appropriate tool for emotion mining and horizon scanning based on Twitter data. Future work may aim to increase emotion mining accuracy by performing text dependency parsing and considering factors such as negation and adjectives which are modified by adverbs. Furthermore, Twitter corpus limitations concerning specific locations and hashtags should be altered in order to examine more broadly under which circumstances a corpus may generate usable results for emerging issue identification.