Stance Detection and Open Research Avenues

Stance detection is a research problem focusing on people’s positions towards specific targets, in natural language texts Küçük and Can (2020, 2021, 2022). It can be considered as a subproblem of affective computing, along with related research topics such as sentiment analysis. Stance classification, stance analysis, and stance extraction are also used to refer to the problem of stance detection in the related literature.

One of the important milestones of stance detection research is the stance detection shared task on English tweets performed in 2016 Mohammad et al. (2016a). Within the course of this competition, a stance-annotated tweet dataset is created and publicly shared Mohammad et al. (2016b). This shared task is followed by other similar competitions on texts in Chinese Xu et al. (2016), Spanish and Catalan Taulé et al. (2017), Italian Cignarella et al. (2020), and Basque Agerri et al. (2021).

Significant subproblems of stance detection and closely-related problems have been previously discussed in details in Küçük and Can (2020, 2021, 2022). Based on the related figure in Küçük and Can (2020), a revised schematic representation demonstrating stance detection, its subproblems, and related problems is given in Figure 1. The newly-added problems are contextual stance detection Cignarella et al. (2020); AlDayel and Magdy (2021), intent detection Küçük (2021a), and stance quantification Küçük (2021b), which are shown in blue in the figure.

Contextual stance detection does not only use the input text, but makes use of contextual information (social media user profiles, interactions between posts, etc.) as well during the stance detection procedure. Hence, contextual stance detection is a subproblem of generic stance detection. Intent detection (together with slot filling), on the other hand, is a research problem of dialogue systems and natural language understanding, where the goals of the users are extracted from their utterances Niu et al. (2019). Stance quantification aims the determine the percentages of the textual items belonging to distinct stance classes, instead of labeling each item with its stance label Küçük (2021b). We aim to cover all of the research problems in Figure 1 during our tutorial.

Most of the recent work on stance detection employs deep learning approaches Zhao and Yang (2020); Schiller et al. (2021). Yet, ensemble methods are also commonly observed in stance detection research Chen et al. (2021).

It has been previously reported that stance detection studies were published on several languages including English, Chinese, Spanish, Catalan, Italian, Japanese, Turkish, Czech, Russian, and Arabic Küçük and Can (2020). Recent work also reveal that there are related studies performed on Basque Agerri et al. (2021), German Mascarell et al. (2021), Portuguese Won and Fernandes (2022), and Persian Nasiri and Analoui (2022), too.

This stance detection tutorial will consist of two main parts where the first part will be devoted to presentation of basic concepts, related research problems, stance detection competitions, machine learning and deep learning based approaches, and stance detection resources like the datasets, with particular emphasis on publicly shared datasets. The second part of the tutorial will mostly cover open research topics related to stance detection. Basically the following open research topics will be considered in the second part:

1. (1)

   Stance Detection in Data Streams: Data streams constitute a significant research area Bonab and Can (2016); Gözüaçık and Can (2021); Gulcan and Can (2022) and stance detection can also be applied to large volumes of particularly social media posts provided as streams. Hence, stance detection in data streams is one of the fruitful open research areas of stance detection Bechini et al. (2021).

2. (2)

   Finer Grained Stance Detection: Most of the work on stance detection performs three-way or two-way classification using the stance classes of favor, against, none, and neutral. However, considering related work on sentiment analysis, corresponding polarity classes occasionally include classes that posses additional semantic information such as strongly positive, strongly negative, weakly positive etc. In order to extract richer stance-related information from the underlying texts, finer-grained stance detection can be employed similarly: by extending the basic stance label set with new labels such as strongly favor, weakly favor, strongly against, etc.

3. (3)

   Stance Detection on Legal Texts and Other Text Genres: Recently, most of the stance detection research has been conducted on social media texts, particularly on tweets. But stance detection can also be performed on legal texts and on other text genres such as generic news articles Mascarell et al. (2021) which may lead to fruitful results for the corresponding domains.

4. (4)

   Cross-lingual and Multilingual Stance Detection: In cross-lingual stance detection, stance-annotated dataset in one language is used to improve stance detection in another language Mohtarami et al. (2019); Hardalov et al. (2022) where the latter language is usually a low-resource one. On the other hand, in multilingual stance detection Lai et al. (2020), a stance detection approach is usually applied to datasets in different languages to test the portability of the employed approach. Both cross-lingual and multilingual stance detection are part of important open research topics of stance detection.

Second part of the tutorial will also cover important application areas of stance detection such as topics in information retrieval Al-Ghadir et al. (2021), fake news detection Shu et al. (2019), and rumour detection and resolution Yang et al. (2022); Huang et al. (2022).

The main two benefits of this tutorial are listed below:

• •

  The tutorial will firstly cover the basic concepts, approaches, and resources for stance detection. While covering approaches and resources such as datasets, more emphasis will be given to recent research on stance detection.

• •

  Secondly, significant open research topics related to stance detection will be outlined, along with the main application areas of stance detection. Although these two aspects can be thought as intersecting, here we will put more emphasis on open research avenues of stance detection.

Target audience of this tutorial include researchers and practitioners interested in stance detection, social media analysis, affective computing, natural language processing, and information retrieval. We believe that the tutorial will be beneficial to those researchers/practitioners who have prior information about stance detection as well as to those who do not have, since the tutorial will cover basics, current state-of-the-art, and significant future research opportunities.

There are no prerequisites for this tutorial.

Below provided is the outline of our half-day tutorial proposal on stance detection and open research problems related to stance detection.

1. (1)

   Tutorial Part I: Basics, Competitions, Approaches, and Datasets

   1. (a)

      Basic concepts of stance detection

      1. (i)

         Stance detection: problem definition

      2. (ii)

         Subproblems of stance detection

      3. (iii)

         Problems related to stance detection

   2. (b)

      Stance detection competitions (shared tasks)

      1. (i)

         Shared task on English tweets (2016)

      2. (ii)

         Shared task on Chinese microblogs (2016)

      3. (iii)

         Shared task on Spanish and Catalan tweets (2017)

      4. (iv)

         Shared task on Italian tweets (2020)

      5. (v)

         Shared task on Basque tweets (2021)

   3. (c)

      Approaches to stance detection

   4. (d)

      Stance detection datasets

2. (2)

   Tutorial Part II: Open Research Avenues and Application Areas

   1. (a)

      Open research avenues

      1. (i)

         Stance detection in data streams

      2. (ii)

         Finer grained stance detection

      3. (iii)

         Stance detection on legal documents and on other text genres

      4. (iv)

         Cross-lingual and multilingual stance detection

   2. (b)

      Application areas

      1. (i)

         Information retrieval

      2. (ii)

         Rumour classification

      3. (iii)

         Fake news detection