Culturally Aware Natural Language Inference

Jing Huang, Diyi YangEMNLP 2023 · OpenReview

Abstract

Humans produce and consume language in a particular cultural context, which includes knowledge about specific norms and practices. A listener's awareness of the cultural context is critical for interpreting the speaker's meaning. A simple expression like I didn't leave a tip implies a strong sense of dissatisfaction when tipping is assumed to be the norm. As NLP systems reach users from different cultures, achieving culturally aware language understanding becomes increasingly important. However, current research has focused on building cultural knowledge bases without studying how such knowledge leads to contextualized interpretations of texts. In this work, we operationalize cultural variations in language understanding through a natural language inference (NLI) task that surfaces cultural variations as label disagreement between annotators from different cultural groups. We introduce the first Culturally Aware Natural Language Inference (CALI) dataset with 2.7K premise-hypothesis pairs annotated by two cultural groups located in the U.S. and India. With CALI, we categorize how cultural norms affect language understanding and present an evaluation framework to assess at which levels large language models are culturally aware. Our dataset is available at https://github.com/SALT-NLP/CulturallyAwareNLI.

Introduction

Most Natural Language Processing (NLP) systems operate under the assumption of a single "ground truth" for any given text. However, human language is deeply embedded in cultural contexts. A sentence like "I didn't leave a tip" implies deep dissatisfaction in the U.S., but remains a neutral statement in cultures where tipping is not customary. Current NLP models often struggle with these nuances, leading to poor performance for global users, the propagation of stereotypes, or "label bias" where one culture's perspective is treated as the universal standard.

An example of cultural variation in Natural Language Inference (NLI) regarding tipping norms.

This research moves beyond simply building cultural knowledge bases to asking: Can models actually use that knowledge to make context-specific inferences? To address this, the authors introduce CALI (Culturally Aware Natural Language Inference), a dataset designed to highlight how cultural norms—such as tipping or wedding customs—change the meaning of text.

The CALI Dataset

The researchers operationalized cultural awareness through a Natural Language Inference (NLI) task, determining if a hypothesis follows from a premise. The dataset construction involved several key steps:

  1. Script-Based Generation: Instead of simple "rules of thumb," the authors used "scripts"—causal chains of behavior (e.g., the steps of a wedding or a business negotiation).
  2. Model-in-the-Loop: ChatGPT was used to generate premises and hypotheses based on these scripts, followed by rigorous human editing to remove "cultural indicators" (like names or country mentions) that might bias the model.
  3. Cross-Cultural Annotation: Annotators from the U.S. and India labeled the dataset. Cultural variation was measured by the disagreement between these two groups.

The resulting CALI dataset contains 2,700 premise-hypothesis pairs centered around normative behaviors.

Evaluation Framework

The study proposed a diagnostic framework to assess models at three distinct levels:

  • Level 1 (Knowledge): Does the model know the norm?
  • Level 2 (Context): Can the model recognize the norm in a sentence?
  • Level 3 (Inference): Can the model adjust its conclusion based on the norm?

Key Findings

The findings reveal a striking "awareness gap" in Large Language Models (LLMs). While models like GPT-4 often possess factual knowledge of cultural norms, they frequently fail to recognize when those norms apply or how to adjust their reasoning accordingly.

The Knowledge-Application Gap

GPT-4 is remarkably proficient at Level 1 (Knowledge), achieving over 90% accuracy on cultural quizzes. However, it fails significantly at Level 3 (Inference). Even when the model is explicitly provided with the cultural norm, it maintains a 30% error rate in updating its inference. This suggests that reasoning is a significantly harder bottleneck to overcome than simple memory.

Disagreement as a Signal

In standard NLP, annotator disagreement is often discarded as noise or error. In CALI, disagreement is treated as a primary metric for identifying cultural "pivot points" in language. About 30% of the pairs in CALI resulted in different labels between U.S. and Indian annotators, proving that culture significantly impacts NLI.

Syntactic Bias and Fine-tuning

Models often rely on linguistic shortcuts, such as word overlap or sentence structure, rather than cultural reasoning. Furthermore, models fine-tuned on massive "culturally-agnostic" datasets (like MultiNLI) showed no specific cultural bias but also demonstrated zero cultural awareness—performing equally mediocre across all cultural contexts.

Conclusion and Future Work

While GPT-4 showed a 30% improvement in recognizing norms (Level 2) compared to GPT-3.5, the struggle to apply that knowledge persists. The study acknowledges limitations, such as using national locales (U.S. and India) as proxies for complex, diverse cultures. Future work should move toward more granular, multi-cultural frameworks and develop methods that allow LLMs to dynamically switch cultural "lenses" based on the user's context.

Paper Citation: Jing Huang and Diyi Yang.

  1. Culturally Aware Natural Language Inference. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7591–7609.

Summary updated

Abstract

Humans produce and consume language in a particular cultural context, which includes knowledge about specific norms and practices. A listener's awareness of the cultural context is critical for interpreting the speaker's meaning. A simple expression like *I didn't leave a tip* implies a strong sense of dissatisfaction when tipping is assumed to be the norm. As NLP systems reach users from different cultures, achieving culturally aware language understanding becomes increasingly important. However, current research has focused on building cultural knowledge bases without studying how such knowledge leads to contextualized interpretations of texts. In this work, we operationalize cultural variations in language understanding through a natural language inference (NLI) task that surfaces cultural variations as label disagreement between annotators from different cultural groups. We introduce the first Culturally Aware Natural Language Inference (CALI) dataset with 2.7K premise-hypothesis pairs annotated by two cultural groups located in the U.S. and India. With CALI, we categorize how cultural norms affect language understanding and present an evaluation framework to assess at which levels large language models are culturally aware. Our dataset is available at https://github.com/SALT-NLP/CulturallyAwareNLI.
BibTeX
@inproceedings{huang2023culturally,
  author = {Jing Huang and Diyi Yang},
  title = {{Culturally Aware Natural Language Inference}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2023},
  url = {https://openreview.net/forum?id=XEBHsJpFY9},
}
Figure