Abstract
Artificial agents that learn to communicate in order to accomplish a given task acquire communication protocols that are typically opaque to a human. A large body of work has attempted to evaluate the emergent communication via various evaluation measures, with compositionality featuring as a prominent desired trait. However, current evaluation procedures do not directly expose the compositionality of the emergent communication. We propose a procedure to assess the compositionality of emergent communication by finding the best-match between emerged words and natural language concepts. The best-match algorithm provides both a global score and a translation-map from emergent words to natural language concepts. To the best of our knowledge, it is the first time that such direct and interpretable mapping between emergent words and human concepts is provided.
Introduction
In the field of Emergent Communication (EC), agents are trained to collaborate by developing their own signaling protocols. While these protocols often lead to task success, they are typically unintelligible to humans, appearing as sequences of discrete tokens with no obvious semantic grounding. A primary goal is to ensure these languages are compositional—able to combine simple concepts (e.g., "red", "triangle") into complex ones ("red triangle").
Existing metrics like Topographic Similarity (TopSim) act as "black boxes." They measure structural correlations but fail to explain what the agents are communicating. This lack of interpretability makes it difficult to diagnose agent failures or verify concept acquisition. To bridge this gap, researchers from Technion have introduced Concept-Best-Matching (CBM).
Methodology: Concept-Best-Matching (CBM)
CBM is a framework that leverages bipartite graph matching to map emergent AI "words" directly to human-understandable concepts. Unlike previous metrics that provide a single abstract score, CBM offers a human-interpretable translation map, revealing exactly which AI tokens represent specific attributes like color or shape.
The CBM procedure involves four primary steps:
- Data Collection: Observe the messages (EC words) an agent sends when presented with specific target objects defined by natural language (NL) concepts (e.g., "Blue," "Triangle").
- Graph Construction: Build a weighted bipartite graph. On one side are the "emergent words" the AI used; on the other are the "ground-truth concepts" provided by an Oracle.
- Weighting: The edges between a word and a concept are weighted by how often they co-occur.
- Optimal Matching: The algorithm uses the classical Hungarian algorithm to find the optimal one-to-one mapping that maximizes the total weight.
The resulting CBM score (from 0 to 1) represents how well the AI's vocabulary aligns with human concepts.
Experimental Results
The researchers evaluated CBM using two standard benchmarks: SHAPE (17 concepts) and THING (50 concepts). They compared two types of communication channels: Gumbel-Softmax (GS) and Quantized (QT) communication.
Interpretability and Translation
CBM's primary advantage is the generation of a human-interpretable translation map. As shown in Figure 2, the algorithm successfully matches specific emergent tokens to concepts like colors, shapes, and symbols, while also identifying "paraphrase" words that weren't part of the primary mapping.
Identifying Failure Modes
Beyond scoring, CBM allows researchers to categorize why a language might lack compositionality:
- Ambiguity: One AI word is being used for multiple distinct concepts.
- Paraphrasing: The AI has multiple words for the same single concept.
- Unmatched Concepts: The AI simply lacks a word for a specific attribute, leaving a "hole" in its vocabulary.
Our analysis revealed that Quantized (QT) channels tend to produce more interpretable, one-to-one mappings than GS channels. In the THING dataset, GS communication often resulted in high rates of unmatched concepts (up to 0.55), suggesting agents were ignoring significant portions of the conceptual space.
Conclusion
Concept-Best-Matching provides a significant step toward transparent AI communication. By moving from abstract similarity scores to concrete translation maps, researchers can better understand the internal logic of emergent protocols. While currently requiring labeled data and focusing on conjunctive expressions, future work will explore scaling CBM to larger vocabularies and more complex logical structures.
Summary updated