About Emotion Identification in Visual Sentiment Analysis
Olga Kanishcheva, Galia Angelova · Recent Advances in Natural Language Processing · 2015
In this paper we present an approach for anal- ysis of sentiments and emotions in image tag- ging using SentiWordNet as an external lin- guistic resource of emotional words. Our aim is to design and implement algorithms that as- sess the emotions and polarity given a set of image tags. The approach is not limited to ob- ject analysis only (considering informational keywords) but deals with the involvement tags and employs some techniques used for senti- ment analysis in social networks. We consider the issue of tag sense disambiguation when image keywords are mapped to SentiWordNet. The Lesk algorithm helps to identify correctly the meaning of about 50% of the ambiguous single keywords of 200 images. The total number of tags we process is about 10,000. Calculating a sentiment score for each im- age, the system classifies images into three classes (positive, negative, neutral). These classes are compared to emotional assessments done (i) by humans and (ii) by training of a SVM classifier that provides the baseline of 69.7% precision, 29.9% recall and 41.8% F- measure. Our approach works with 63.53% precision, 58.7% recall and 61.02% F- measure. The experiments are performed using the annotations of the industrial auto-tagging platform Imagga that identifies automatically image objects with high precision.