Towards Author-informed NLP: Mind the Social Bias

Inbar Pendzel, Einat MinkovEMNLP 2025

TL;DR

This paper introduces a novel approach to author-informed Natural Language Processing (NLP) by representing users in a "social embedding space" learned from large-scale Twitter network data. This method significantly improves performance on tasks like stance detection and toxicity detection by incorporating user context, but also highlights the critical issue of social biases being inadvertently modeled and reinforced.

What Problem Does This Paper Address?

Traditional NLP tasks often struggle with understanding the nuances of human communication, especially when opinions are conveyed implicitly, sarcastically, or with limited context. For instance, the phrase "I so love Biden" could be genuine or sarcastic. To address this, researchers have recognized the need to incorporate the author's social and cultural context. While previous work has used user information, it often relied on transductive learning tailored to specific tasks or datasets. This paper aims to create generalizable, inductive user embeddings that capture broader social semantics, enabling better author-informed NLP across various tasks and datasets.

Methods

The core of this work is the creation of "social user embeddings." Instead of relying on text alone, the authors leverage pre-trained entity embeddings from a large-scale Twitter network. These embeddings, learned from popular accounts that users co-follow, capture "social semantics." Individual users are then projected into this social embedding space by averaging the embeddings of the accounts they follow. This inductive approach allows for the creation of user representations that are not tied to a specific task or dataset.

For author-informed NLP tasks, like stance detection, the authors combine these social user embeddings with semantic embeddings derived from transformer-based text encoders (like BERT). This combined representation is then fed into a dedicated classifier. They also explore using a user's historical tweets as an additional source of author context.

Datasets

The paper utilizes several publicly available datasets for stance detection:

  • SemEval: Targets include "Hillary Clinton" and "Feminism."
  • COVID-19: Focuses on opinions about COVID-19 vaccinations.
  • P-Stance: Covers political stances on "Biden" and "Sanders."

Additionally, they introduce two user-level datasets:

  • Yoga: Users identified as yoga fans based on hashtag usage.
  • Vegetarianism: Users identified by self-descriptions containing "vegetarian" or "vegan."

For these user-level datasets, random users are used as counter-examples.

Key Insights and Ideas

  • Social Semantics in Network Structure: The structure of social networks, specifically who users follow, encodes rich "social semantics" that can be learned and leveraged for NLP tasks.
  • Inductive User Embeddings: Learning user representations from pre-trained entity embeddings allows for transfer learning and generalization across different tasks and datasets, overcoming limitations of transductive approaches.
  • Social Biases are Inherent: A significant finding is that user stances are correlated with socio-demographic traits encoded in their embeddings. This means author-informed NLP approaches can inadvertently model and reinforce existing social biases.
  • Generalizability Across Topics: Author-informed models, particularly those using social embeddings, show improved generalization for cross-topic stance prediction, suggesting that social context helps bridge topical gaps.

Key Results

  • Stance Detection Improvement: Author-informed stance prediction consistently outperformed text-only baselines, showing substantial gains (4-21%) across multiple datasets.
  • User Embeddings Outperform Sparse Representations: Dense social user embeddings generally performed better than sparse, discrete representations (e.g., one-hot vectors of followed accounts) for stance detection, especially in cross-topic scenarios.
  • Toxicity Detection Benefits: The approach also showed improvements in author-informed toxicity detection, demonstrating its broad applicability.
  • Socio-Demographic Correlations: Detailed analysis revealed distinct socio-demographic profiles for users supporting or opposing different targets, highlighting the presence of social biases in stance-labeled datasets. For example, supporters of certain political figures showed higher proportions of Democrats and women.

Limitations and Future Work

  • Twitter Dependency: The research is primarily based on Twitter data, and its applicability to other social networks might require diverse network data indicative of personal interests and beliefs.
  • Data Volatility: Twitter data is subject to account deletions and post removals, posing challenges for reproducibility and requiring robust data recovery methods.
  • Bias Mitigation: A major limitation is the inadvertent modeling and reinforcement of social biases. Future work needs to explore methods for mitigating these risks.
  • Integration with Large Language Models (LLMs): An open question is how to effectively integrate social author context into state-of-the-art LLMs, potentially through multi-modal architectures.
  • Ethical Considerations: The authors emphasize the need for transparency with users about the risks of social bias in NLP and the importance of obtaining consent. They also plan to make their datasets and user representations available to researchers with IRB authorization.

Summary updated

BibTeX
@inproceedings{pendzel2025towards,
  author = {Inbar Pendzel and Einat Minkov},
  title = {{Towards Author-informed NLP: Mind the Social Bias}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure