Conversational Context for Social Media Stance Detection

19 concepts14 papers1 proposal

Traditional stance detection—the task of determining whether a user is "for" or "against" a specific target—has long treated social media posts as isolated snippets of text. However, a growing body of research argues that this "bag-of-posts" approach is fundamentally flawed. In the rapid-fire environment of platforms like X (formerly Twitter) or Reddit, users rarely restate the topic of conversation. Instead, they rely on the preceding dialogue to provide the necessary context for their opinions.

The Ambiguity of Isolation

The core problem is that social media language is highly elliptical. As Niu et al. observe, a simple comment like "I agree" or "That’s not true" is impossible to classify in isolation; its stance depends entirely on the claim it is responding to. When human annotators are forced to look at posts without their conversational history, they fail to reach a consensus in nearly half of all cases (Li et al.).

To resolve this, researchers have shifted toward modeling the conversational hierarchy. Rather than viewing a tweet as a standalone unit, Lukasik et al. and Derczynski et al. propose treating discourse as a tree structure. In this view, the meaning of a "child" post is intrinsically linked to its "parent." By incorporating these extra-linguistic constraints—such as how users interact in threads and their ideological consistency—Hasan et al. demonstrated that models could more accurately identify rebuttals and opposing stances that would otherwise look like generic disagreement.

Modeling the Thread as a Sequence

If a conversation is a tree, any individual post is the endpoint of a specific path or "branch." Several influential models leverage this sequential nature:

  • Branch-LSTMs: Kochkina et al. and Procter et al. pioneered the use of sequential classifiers to process conversation branches. By treating the thread as a linear sequence of interactions, these models can track how a rumor is supported or denied as the discussion evolves.
  • Branch-BERT: Building on this, Li et al. introduced architectures that feed the entire ancestral lineage of a post into a transformer. This allows the model to resolve ambiguities by "seeing" the full context that the user was reacting to.
  • Masked Hierarchical Transformers: To handle multi-party conversations where many branches interleave, Zhu et al. developed a masking mechanism. This ensures that a post only "attends" to its direct ancestors, preventing the model from being distracted by irrelevant side-discussions happening in the same thread.

Beyond Textual Context

The need for context extends beyond simple text. In multimodal settings, where users reply to images with text or vice versa, Niu et al. found that incorporating prior turns in a dialogue is essential for Large Language Models (LLMs) to accurately capture public opinion.

Furthermore, stance is not always a static label. Marreddy et al. highlight that in long conversations, user opinions and "dogmatism" (intellectual rigidity) can shift. Understanding these dynamics requires looking at the social graph of the interaction. Su et al. utilized Graph Attention Networks (GATs) to model these relationships, treating the conversation as a web of agreements and disagreements rather than just a sequence of words.

The Frontier: Zero-Shot and Multicultural Stance

As the field matures, the focus has moved toward more challenging scenarios. Zheng et al. introduced the Stanceosaurus dataset to emphasize that conversational context is vital for interpreting reactions to misinformation across different languages and cultures. Similarly, Ding et al. have pushed into zero-shot conversational stance detection, where models must capture speaker interaction dynamics to identify opinions on targets they have never seen before.

Across all these facets, the conclusion is consistent: to understand what a person thinks, you must first understand who they are talking to and what was said before they spoke.

Key Foundational Work

Stance Classification in Rumours as a Sequential Task Exploiting the Tree Structure of Social Media Conversations (2016) — contributed 1 concept elements to this direction.
Extra-Linguistic Constraints on Stance Recognition in Ideological Debates (2013) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_conversational_context_dependency_contex_3a994d1d,
  title = {conversational context dependency (context-aware stance detection)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/conversational-context-dependency-context-aware-stance-detection/rdr_3a994d1d4cda4b1ea9ac19e9ddaf5fd5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_conversational_context_dependency_contex_3a994d1d_v1,
  title = {conversational context dependency (context-aware stance detection)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/conversational-context-dependency-context-aware-stance-detection/rdr_3a994d1d4cda4b1ea9ac19e9ddaf5fd5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake