Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Unsupervised Translation of Emergent Communication
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Applies unsupervised neural machine translation (XLM-based, with back-translation and denoising) to translate emergent communication messages from referential games of varying complexity into English using only monolingual data and MSCOCO captions.
Problem
Emergent communication between AI agents is hard to interpret and relate to natural language, and prior translation approaches need parallel EC-NL data that is unavailable for AI-generated languages.
Approach
Treat EC messages from referential games of four complexity levels (Random, Category, Supercategory, Inter-category) as a monolingual corpus and translate them to English with UNMT, using MSCOCO captions as the English monolingual prior, so no parallel EC-NL pairs are needed; translatability is then compared across game complexities.
Key results
- Inter-category game EC achieved the highest UNMT BLEU (9.21) and ROUGE-L (0.370) among EC game types, versus 6.08 BLEU for Supercategory.
- Random-agent baseline EC translation scored near-zero BLEU (0.071), while trained EC translations scored 5.6-9.2 BLEU.
- Category game EN-to-EC translation reached BLEU 39.31, far above other complexities (14.55 to 26.77).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Inter-category rather than Supercategory EC for UNMT translation | UNMT EC→English · MSCOCO captions | BLEU: 6.08 -> 9.21 | Helped |
| Use Inter-category rather than Supercategory EC for UNMT translation | UNMT EC→English · MSCOCO captions | METEOR: 0.276 -> 0.310 | Helped |
| Use Inter-category rather than Supercategory EC for UNMT translation | UNMT EC→English · MSCOCO captions | ROUGE-L: 0.343 -> 0.370 | Helped |
| Use Inter-category rather than Random-agent messages | UNMT EC→English · MSCOCO captions | BLEU: 0.071 -> 9.21 | Helped |
| Use Category rather than Supercategory game complexity | Referential game agents · 10 candidates | accuracy: 77.35 { \(\pm\) 2.96} -> 70.51 { \(\pm\) 0.88} | Hurt |
| Use Category rather than Supercategory EC for translation | UNMT EC→English · MSCOCO captions | BLEU: 6.08 -> 7.41 | Helped |
| Use Category rather than Supercategory EC for translation | UNMT English→EC · MSCOCO captions | BLEU: 14.55 -> 39.31 | Helped |
| Use Supercategory rather than Category game complexity | Referential game agents · 10 candidates | vocabulary usage: 55.95 { \(\pm\) 2.04} -> 73.10 { \(\pm\) 9.77} | Helped |
Surprising findings
- Translation performance is not correlated across different EC instantiations (seeds or games) evaluated on the same test set, suggesting each run develops distinct protocols.
- Category game, with the lowest vocabulary usage and entropy, achieved higher translation scores than Supercategory, suggesting simpler protocols are easier to translate.
Limitations
- Number of distractors and communication channel settings (vocabulary size, message length) were fixed after preliminary experiments and not fully explored.
- Sharing a pre-trained ResNet visual module may limit generality of the emergent protocols.
- Translations sometimes hallucinate extra details, attributed to narrow caption-length distribution in training.
Experimental setup
Datasets: MSCOCO MSCOCO-CN emergent communication messages from referential games
Benchmarks: MSCOCO captions (max over 5 references) MSCOCO-CN calibration
Metrics: BLEU METEOR ROUGE-L BERTScore CLIP Score Jaro similarity TTR novelty game accuracy vocabulary usage
Baselines: random agent communication
Builds on: EGG framework XLM unsupervised neural machine translation infoNCE
Closest papers by meaning
- Emergent Translation in Multi-Agent Communication (ICLR 2018 · similarity 0.89)
- Expressivity of Emergent Languages is a Trade-off between Contextual Complexity and Unpredictability (ICLR 2022 · similarity 0.89)
- Lewis's Signaling Game as beta-VAE For Natural Word Lengths and Segments (ICLR 2024 · similarity 0.87)
- Semantics and Spatiality of Emergent Communication (NeurIPS 2024 · similarity 0.87)
- Speaking Your Language: Spatial Relationships in Interpretable Emergent Communication (NeurIPS 2024 · similarity 0.86)
- A compressive-expressive communication framework for compositional representations (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.