Tractable Control for Autoregressive Language Generation

Authors: Honghua Zhang, Meihua Dang, Nanyun Peng, Guy Van Den Broeck

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we demonstrate the effectiveness of Ge La To on challenging benchmarks for constrained generation: Common Gen (Lin et al., 2020), Yelp!Review (Cho et al., 2019) and News (Zhang et al., 2020); in particular, we focus on Common Gen for detailed analysis. For both unsupervised and supervised settings, Ge La To achieves state-of-the-art performance in terms of various automatic evaluation metrics including BLEU score while guaranteeing 100% constraint satisfaction. Main evaluation results are presented in Table 1.
Researcher Affiliation Academia Honghua Zhang * 1 Meihua Dang * 1 Nanyun Peng 1 Guy Van den Broeck 1 1Department of Computer Science, University of California, Los Angeles, USA.
Pseudocode Yes Algorithm 1 Constrained Sampling with Ge La To
Open Source Code Yes In this section, we demonstrate the effectiveness of Ge La To2 https://github.com/UCLA-StarAI/GeLaTo on challenging benchmarks for constrained generation
Open Datasets Yes Common Gen (Lin et al., 2020) is a benchmark for constrained generation with lexical constraints... We also evaluate Ge La To on the Yelp!Review (Cho et al., 2019) and the News (Zhang et al., 2020) datasets.
Dataset Splits Yes For hyper-parameter tuning, we conduct cross-validation on a small subset of the training set and report evaluation results for both validation (dev) and test set.
Hardware Specification Yes all methods are evaluated on a single NVIDIA A100 GPU with 40 GB memory
Software Dependencies No The paper mentions the use of 'Juice.jl framework (Dang et al., 2021)' and 'LemmInflect3', but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes Unsupervised Setting: we fine-tune the model for 1 epoch with learning rate = 1e-6. Supervised Setting: for 3 epochs with learning rate = 1e-6. Training HMMs: we train HMMs with the expectation-maximization (EM) algorithm for 40 epochs, and we resample 0.2 million examples for each epoch. Decoding: We adopt beam search to greedily search for x1:n that maximizes p(x1:n | α); we experiment with different beam sizes: 16, 32, 64 and 128.