Introduction
In psychology, word association tasks—such as providing the first word that comes to mind when prompted with "Apple"—serve as a primary tool for mapping the structure of human semantic knowledge. As Large Language Models (LLMs) become increasingly integrated into human-centric tasks, researchers have sought to perform similar "brain surgery" on these models to determine if their internal conceptual maps align with human intuition.
Traditionally, this has been measured using word embeddings, where the cosine similarity between mathematical vectors represents the strength of an association. However, this approach suffers from two fundamental flaws. First, cosine similarity is inherently symmetric ($A \approx B$ is identical to $B \approx A$), whereas human association is often a one-way street (e.g., "North Korea" strongly triggers "China," but the reverse is much weaker). Second, converting contextualized LLM embeddings into static vectors often requires external templates or corpora, which can introduce noise and biases not present in the model itself.
The WordTies Framework
To address these limitations, the authors introduce WordTies, a novel probing algorithm that moves away from embedding space and into sampling space. Instead of measuring distances between vectors, WordTies forces the model to generate text under specific constraints.
Methodology
- Constrained Sampling: For Masked Language Models (MLMs) like BERT, the algorithm uses a modified Gibbs Sampling approach. It fixes a "cue" word within a sequence of [MASK] tokens and allows the model to fill in the surrounding context.
- Associativity Score: By generating thousands of these constrained sentences, the system calculates a conditional probability: Given that the cue word is present, how likely is the model to generate the response word?
- Reasoning Paths: To interpret these associations, the authors link word pairs to external knowledge graphs: WordNet for lexical relations (synonyms/hypernyms) and ASCENT++ for commonsense properties (e.g., "apples" are "sour").
Key Results and Performance
The researchers evaluated WordTies against the Small World of Words (SWOW) dataset, the gold standard for human word associations, using BERT, RoBERTa, and DistilBERT.
WordTies significantly outperformed traditional embedding-based baselines (such as Vocab and C2S) across all metrics. On the BERT model, WordTies achieved a Precision@1 of 0.368, compared to the best baseline of 0.247. Furthermore, the Spearman’s correlation ($\rho$) with human association strengths was more than double that of the baselines, indicating that WordTies captures the nuances of human thought far more faithfully.
The Asymmetry Advantage
A standout feature of WordTies is its ability to capture the directionality of associations. The method achieved over 97% precision in identifying which word in a pair was the stronger "trigger." This capability is mathematically impossible for symmetric distance metrics like cosine similarity, proving that sampling-based probing is a superior method for understanding the "flow" of information in an LLM.
Insights into Model Knowledge
The study revealed several intriguing patterns in how LLMs organize information:
- Commonsense vs. Lexical: LLMs appear slightly more proficient at associating words based on shared properties (commonsense knowledge) than on purely dictionary-based relations like synonyms.
- The "Hop" Penalty: There is a strong negative correlation between the length of a reasoning path in a knowledge graph and the model's association strength. If two concepts are separated by multiple "hops" in a graph, the LM is significantly less likely to associate them.
- Sampling as Faithful Probing: The results suggest that sampling discrete words is a more "faithful" way to probe an LM's knowledge because it examines the model's actual output behavior rather than its internal geometric representations.
Limitations and Future Directions
While effective, WordTies is computationally intensive, requiring approximately 12 seconds per cue to generate enough samples for a stable score. Additionally, while the method works seamlessly for MLMs, applying it to Causal LMs (like GPT-4) requires more complex sampling techniques like Langevin Dynamics.
Beyond measuring general knowledge, the authors suggest that WordTies could become a powerful tool for detecting social biases (gender, race) in models with higher precision and fewer confounding variables than current embedding-based methods.
Summary updated