Grammar-Based Grounded Lexicon Learning

Authors: Jiayuan Mao, Freda Shi, Jiajun Wu, Roger Levy, Josh Tenenbaum

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate G2L2 on two domains: visual reasoning and language-driven navigation. Results show that G2L2 can generalize from small amounts of data to novel compositions of words. and We evaluate G2L2 on two domains: visual reasoning in CLEVR [21] and language-driven navigation in SCAN [25]. Beyond the grounding accuracy, we also evaluate the compositional generalizability and data efficiency, comparing G2L2 with end-to-end neural models and modular neural networks.
Researcher Affiliation Academia Jiayuan Mao MIT Haoyue Shi TTIC Jiajun Wu Stanford University Roger P. Levy MIT Joshua B. Tenenbaum MIT
Pseudocode Yes Algorithm 1 The CKY-E2 algorithm.
Open Source Code Yes Project page: http://g2l2.csail.mit.edu. and Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes]
Open Datasets Yes We evaluate G2L2 on two domains: visual reasoning in CLEVR [21] and language-driven navigation in SCAN [25].
Dataset Splits Yes Since CLEVR does not provide test set annotations, for all models, we held out 10% of the training data for model development and test them on the CLEVR validation split.
Hardware Specification No The main paper does not contain specific hardware details for running experiments. The checklist states, “Details can be found in the supplementary material.”
Software Dependencies No The main paper does not provide specific ancillary software details with version numbers. The checklist indicates that more detailed information might be in the supplementary material.
Experiment Setup Yes We train different models with either 10% or 100% of the training data and evaluate them on the validation set. and Instead of using manually defined heuristics for curriculum learning or self-paced learning as in previous works [28, 26], we employ a curriculum learning setup that is simply based on sentence length: we gradually add longer sentences into the training set. and We tuned the hidden size (i.e., the dimension of intermediate token representations) within {100, 200, 400}, as well as the number of layers (for both the encoder and the decoder) from {2, 4, 8}.