Introduction
Measuring social bias in artificial intelligence is a critical field of study, yet traditional methods frequently encounter a significant "blind spot": in-domain stereotypes. Conventional techniques typically rely on word embeddings (such as Word2Vec or BERT) to measure the proximity between a social group identifier (e.g., "women") and specific descriptive traits (e.g., "nurturing").
However, the inherent "identity" of the word often dominates the embedding space. For instance, when analyzing how a group is framed in a scientific versus an artistic context, the pre-trained associations of the identity token itself can drown out the subtle nuances of the specific sentence. Even "neutral word" approaches, which replace group names with generic terms like "person," often fail because these neutral terms carry their own pre-existing semantic baggage. To address this, researchers have developed ADAPTIVE AXES, a pipeline designed to shift the focus from word-level identity to context-level framing.
The ADAPTIVE AXES Pipeline
The authors propose a three-stage pipeline that utilizes Large Language Models (LLMs) and state-of-the-art text embeddings to identify social stereotypes within specific domains.
1. Context Masking
Instead of embedding the social group name directly, the pipeline masks it (e.g., "The [MASK] military..."). The surrounding context is then embedded using high-performance models like UAE-large-v1. This ensures the analysis captures how the group is being discussed, rather than relying on the group's name itself.
2. Semantic Axis Enhancement
To measure framing, the system projects these embeddings onto "semantic axes" (e.g., Beautiful vs. Ugly). The pipeline refines these axes through two methods:
- Pruning: Existing axes from dictionaries like WordNet are trimmed using LLMs or clustering algorithms to remove noisy or irrelevant words, making the "poles" of the axis more distinct.
- Augmentation: The system uses GPT-4 to generate entirely new, domain-specific axes that are absent from standard dictionaries, such as political transparency vs. political opaqueness.
3. Projection and Scoring
The masked context embeddings are projected onto these refined axes to calculate a "stereotype score," providing a quantitative measure of the language's directional lean.
Datasets and Methodology
The research utilized the NOW Corpus (News on the Web), a massive dataset containing US news articles from 2010 to 2023, for real-world case studies. Seed semantic axes were sourced from WordNet and ConceptNet. To validate the pipeline, the authors recruited participants via Prolific to rank the quality of the stereotypes identified by ADAPTIVE AXES against those found by traditional methods.
Key Results and Insights
The ADAPTIVE AXES pipeline demonstrated significant improvements over baseline methods in several key areas:
- Higher Consistency: The pipeline's semantic axes achieved a higher internal consistency score ($C = 0.153$) compared to traditional GloVe ($0.101$) or BERT-based methods.
- Superior Human Alignment: In human evaluations, the pipeline was ranked as the most accurate at capturing in-domain framing, outperforming token-based embedding methods with a score of 1.675 (where lower is better, compared to 1.925 for baselines).
- Real-world Tracking: The model successfully tracked the 2018 US-China trade war. It detected a sharp, measurable spike in news framing of China toward "trade barriers" and "planned economy" precisely when US sanctions were announced.
Discussion and Future Work
The core innovation of this work is the realization that well-encoded contexts are sufficient for understanding stereotypes, potentially reducing the reliance on biased social group tokens found in pre-training. Furthermore, by using LLMs as "axis architects," researchers can brainstorm nuanced dimensions for specific fields—such as globalization vs. nationalism in economics—that generic good vs. bad axes cannot capture.
However, a primary limitation remains: the distinction between co-occurrence and stereotype. The model may associate a group with a trait simply because it is a factual geographic or demographic truth (e.g., associating Canada with "North") rather than a social bias. Future work will focus on refining axis selection to isolate traits specifically linked to social biases and developing better filtering techniques to separate factual context from stereotype-laden discourse.
Summary updated