Reproduce “Grammar-Based Grounded Lexicon Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiayuan Mao, Haoyue Shi, Jiajun Wu, Roger P. Levy, Joshua B. Tenenbaum
The G2L2 framework from MIT and Stanford introduces a neuro-symbolic approach to grounded language learning, leveraging a lexicalist grammar to achieve robust compositional generalization and data efficiency. It demonstrates superior performance on challenging "out-of-sample" compositional tests on the CLEVR dataset and attains perfect 100% accuracy on all SCAN generalization splits, often with only 10% of the training data.
Key insights
- Incorporating lexicalist linguistic theories, which emphasize rich lexicon entries and simple combinatory rules, provides a powerful and computationally tractable foundation for achieving robust compositional generalization in AI.
- Disentangling concept learning (via neural embeddings) from symbolic reasoning (via program execution) within a neuro-symbolic architecture is critical for robust generalization to unseen compositions.
- The "expectation invariance" of neuro-symbolic program execution can be leveraged to dramatically reduce the computational cost of compositional parsing and learning, addressing the combinatorial explosion challenge.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.