Relation-Constrained Decoding for Text Generation

Authors: Xiang Chen, Zhixian Yang, Xiaojun Wan

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To evaluate our method, we first construct an RCD benchmark based on dependency relations from treebanks with annotated dependencies. Experimental results demonstrate that our approach can achieve better preservation of the input dependency relations compared to previous methods. To further illustrate the effectiveness of RESEAL, we apply our method to three downstream tasks: sentence summarization, fact-based text editing, and data-to-text generation. We observe an improvement in generation quality.
Researcher Affiliation Academia Xiang Chen , Zhixian Yang , Xiaojun Wan Wangxuan Institute of Computer Technology, Peking University Center for Data Science, Peking University The MOE Key Laboratory of Computational Linguistics, Peking University caspar@pku.edu.cn, yangzhixian@stu.pku.edu.cn, wanxiaojun@pku.edu.cn
Pseudocode Yes Algorithm 1 RESEAL (overview) ... Algorithm 2 Probability Surgery and RG-Top-K
Open Source Code Yes The source code is available at https://github.com/Caspar Swift/RESEAL.
Open Datasets Yes We then construct the dataset for dependency placement task from the English-EWT [39] corpus... Dataset We conduct experiments on English Gigaword dataset [36]... We adopt the Web Edit dataset provided by Iso et al. [17]... We adopt Web NLG dataset [11]... We use the data provided by Ribeiro et al. [35]
Dataset Splits Yes We then construct the dataset for dependency placement task from the English-EWT [39] corpus4, which contains 16,621 sentences with dependency annotations and standard train/dev/test set split. ... We use the validation and test set provided by Zhou et al. [49] with 8,000 and 2,000 sentence pairs, respectively. ... Web Edit dataset ... which contains 181K/23K/29K instances as train/valid/test set. ... Web NLG dataset ... contains 18,102/872/1,862 instances as train/valid/test set.
Hardware Specification Yes Our models are trained on NVIDIA V100 GPU for 30 epochs with batch size 32. We use Adam optimizer [19] with learning rate 3e-5.
Software Dependencies No The paper mentions software tools like Stanza and spaCy, and models like BART, BERT, GPT-2, and T5, but does not provide specific version numbers for these software components or libraries, which are necessary for reproducible dependency description.
Experiment Setup Yes Our models are trained on NVIDIA V100 GPU for 30 epochs with batch size 32. We use Adam optimizer [19] with learning rate 3e-5. ... During decoding, we use standard beam search with beam size k = 20. ... The decay factor λ introduced in Section 3.1 is another important hyperparameter of RESEAL. ... The detailed experimental settings can be found in Appendix C.4.