Context-Dependent Sentiment Analysis in User-Generated Videos

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 873–883 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1081 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 873–883 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1081 Context-Dependent Sentiment Analysis in User-Generated Videos Soujanya Poria Temasek Laboratories NTU, Singapore sporia@ntu.edu.sg Erik Cambria School of Computer Science and Engineering, NTU, Singapore cambria@ntu.edu.sg Devamanyu Hazarika Computer Science and Engineering, NITW, India devamanyu@sentic.net Navonil Mazumder Centro de Investigacin en Computacin, IPN, Mexico navonil@sentic.net Amir Zadeh Language Technologies Institute, CMU, USA abagherz@cs.cmu.edu Louis-Philippe Morency Language Technologies Institute, CMU, USA morency@cs.cmu.edu Abstract Multimodal sentiment analysis is a de- veloping area of research, which involves the identification of sentiments in videos. Current research considers utterances as independent entities, i.e., ignores the inter- dependencies and relations among the ut- terances of a video. In this paper, we pro- pose a LSTM-based model that enables utterances to capture contextual informa- tion from their surroundings in the same video, thus aiding the classification pro- cess. Our method shows 5-10% perfor- mance improvement over the state of the art and high robustness to generalizability. 1 Introduction Sentiment analysis is a ‘suitcase’ research prob- lem that requires tackling many NLP sub-tasks, e.g., aspect extraction (Poria et al., 2016a), named entity recognition (Ma et al., 2016), concept ex- traction (Rajagopal et al., 2013), sarcasm detec- tion (Poria et al., 2016b), personality recognition (Majumder et al., 2017), and more. Sentiment analysis can be performed at differ- ent granularity levels, e.g., subjectivity detection simply classifies data as either subjective (opin- ionated) or objective (neutral), while polarity de- tection focuses on determining whether subjec- tive data indicate positive or negative sentiment. Emotion recognition further breaks down the in- ferred polarity into a set of emotions conveyed by the subjective data, e.g., positive sentiment can be caused by joy or anticipation, while negative sen- timent can be caused by fear or disgust. Even though the primary focus of this paper is to classify sentiment in videos, we also show the performance of the proposed method for the finer- grained task of emotion recognition. Emotion recognition and sentiment analysis have become a new trend in social media, help- ing users and companies to automatically extract the opinions expressed in user-generated content, especially videos. Thanks to the high availability of computers and smartphones, and the rapid rise of social media, consumers tend to record their re- views and opinions about products or films and upload them on social media platforms, such as YouTube and Facebook. Such videos often con- tain comparisons, which can aid prospective buy- ers make an informed decision. The primary advantage of analyzing videos over text is the surplus of behavioral cues present in vo- cal and visual modalities. The vocal modulations and facial expressions in the visual data, along with textual data, provide important cues to bet- ter identify affective states of the opinion holder. …

Content truncated to 500 words