Sentiment analysis, stance detection, and intent detection on social media texts are all significant research problems with several application opportunities. In this chapter, the authors explore the possible contribution of sentiment and intent information to machine learning-based stance detection on tweets. They first annotate a Turkish tweet dataset with sentiment and proprietary intent labels, where the dataset was already annotated with stance labels. Next, they perform stance detection experiments on the dataset using sentiment and intent labels as additional features. The experiments with SVM classifiers show that using sentiment and intent labels as additional features improves stance detection performance considerably. The final form of the dataset is made publicly available for research purposes. The findings reveal the contribution of sentiment and intent information to the solution of stance detection task on the Turkish tweet dataset employed. Yet, further studies on other datasets are needed to confirm that our findings are generalizable to other languages and on other topics.
TopIntroduction
Social media analysis has emerged as one of the most popular research areas, especially due the vast amount of social media posts produced each and every day. Significant subproblems of social media analysis include sentiment analysis, stance detection, and intent detection on social media posts, and these three research topics constitute the main focus of our chapter.
In this book chapter, we investigate the possible contribution of sentiment and intent information to stance detection in Turkish tweets. To achieve this objective, first, we extend an existing tweet corpus annotated with stance information, by annotating it further with sentiment and convenient intent labels. Next, we conduct our stance detection experiments on the ultimate dataset by using a machine learning algorithm (SVM) together with sentiment and intent information as features. By conducting these experiments, we aim to unravel the possible effects (improving or impeding effects) of sentiment analysis and intent detection to stance detection.
Turkish is a morphologically-rich language however it is also a low-resource language with respect to different natural language processing (NLP) tasks. Although there exists previous work on sentiment analysis in Turkish such as (Balahur et al., 2014; Çoban et al., 2021), to the best of our knowledge, there is limited work on stance detection in Turkish (Küçük, 2017; Küçük & Can, 2019), and no study on intent detection in Turkish social media content.
This book chapter includes descriptive information about the tweet dataset which has sentiment, stance, and intent annotations, all at once. The dataset which is further annotated within the context of the current book chapter is made publicly available for research purposes at https://github.com/dkucuk/Turkish-Tweet-Dataset-Stance-Sentiment-Intent. We describe the settings and results of the related experiments and discuss the corresponding findings.
We anticipate that these findings will not be limited to Turkish content only, and instead will be generalizable to other languages as well. Our book chapter is concluded together with further research tasks based on the experiments described, regarding these three significant research topics of social media analysis.
The contributions of our book chapter can be summarized as follows:
- •
To the best of our knowledge, we present the first tweet dataset jointly annotated with sentiment, stance, and intent information. We present the details of the annotation process and the annotation classes employed for each task. This annotated dataset is also made publicly available for research purposes.
- •
We also present the settings and results of our SVM-based stance detection experiments, where sentiment and intent annotations are used as the features for SVMs in addition to other features based on unigrams, hashtags, and named entities. Thereby, we reveal the effects of sentiment and intent information to the task of stance detection.
Sentiment analysis, stance detection, and intent detection have a variety of significant application areas. These areas include recommender systems, personalized advertising, market analysis, information retrieval, and predictions for elections, among others. Hence, the findings of our study which aims to determine the effects of the sentiment analysis and intent detection on stance detection, can be used practically in the aforementioned application areas. Therefore, our study will make both theoretical and practical contributions to the related literature. The findings of the current study can readily be used by future work on determining the interrelationships between these three significant subproblems of social media analysis and NLP.