New Method Reveals How Language Models Build Meaning from Definitions

A team of researchers has developed a new framework for understanding how language models translate written definitions into abstract concepts—bridging classical linguistic theory with modern neural network analysis.

The Gap in Current Understanding#

Mechanistic interpretability research has made progress mapping how transformer models work for factual retrieval tasks, pinpointing which components handle specific functions. Yet most of this insight comes from narrow experimental settings. The open question remains: how do these same architectural pieces handle the harder task of conceptual interpretation—taking a definition and extracting its underlying meaning across multiple dimensions (like distinguishing the core subject from its distinguishing features).

DSRA: A New Interpretability Toolkit#

The researchers introduce DSRA (Definitional Semantic Role Analysis), which combines causal tracing techniques with semantic role annotations rooted in Argument Structure Theory, a linguistic framework for understanding how word meanings relate to their structural roles. Applied to the reverse dictionary task—where a model must infer a word from its definition—DSRA identifies which internal activations map to which semantic components of a definition (e.g., the genus or differentiating qualities).

What the Evidence Shows#

Testing on GPT-J-6B (English) and BERTIN GPT-J-6B (Spanish) revealed consistent patterns:

• Early MLP layers bind content-rich tokens (those carrying meaning) to high-specificity semantic categories
• Middle-to-upper MHA layers integrate information across the sequence, with strongest concentration at the final token position
• Upper hidden states accumulate and consolidate meaning before the model's final output

The alignment between restored neural states and linguistic semantic roles held steady across both languages, suggesting the mechanism is not an artifact of English but a genuine structural feature.

Why It Matters#

These findings extend interpretability beyond simple token localization, offering a richer vocabulary for understanding compositional meaning-making in neural networks. By grounding mechanistic analysis in linguistic theory, the work provides a bridge between how language scholars think about meaning and how machine learning researchers study it.