Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

KPT: Keyword-Guided Pre-training for Grounded Dialog Generation

Authors: Qi Zhu, Fei Mi, Zheng Zhang, Yasheng Wang, Yitong Li, Xin Jiang, Qun Liu, Xiaoyan Zhu, Minlie Huang

AAAI 2023 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments on various few-shot knowledge-grounded generation tasks, including grounding on dialog acts, knowledge graphs, persona descriptions, and Wikipedia passages. Our comprehensive experiments and analyses demonstrate that KPT consistently outperforms state-of-the-art methods on these tasks with diverse grounding knowledge.
Researcher Affiliation Collaboration 1Co AI Group, DCST, IAI, BNRIST, Tsinghua University 2Huawei Noah s Ark Lab
Pseudocode Yes Algorithm 1: Prepare keyword-guided pre-training data
Open Source Code No The paper mentions using 'Conv Lab-3 (Zhu et al. 2022) for dataset loading and model training' which is a third-party toolkit, but does not provide a link or statement for the authors' own implementation code.
Open Datasets Yes As shown in Table 1, our pre-training datasets include Daily Dialog (Li et al. 2017), Schema-Guided Dialog (Rastogi et al. 2020), Taskmaster-1/2/3 (Byrne et al. 2019, 2021), Meta LWOZ (Li et al. 2020), DSTC8-Reddit (Lee et al. 2019), and Wiki Dialog (Dai et al. 2022), covering chit-chats, goal-oriented dialogs, and information seeking dialogs.
Dataset Splits Yes We randomly split the data into training (70%), validation (15%), and test set (15%). We fine-tune the models until the validation loss does not decrease for 5 consecutive epochs. Models with the lowest validation losses during training are selected as the final models.
Hardware Specification Yes We set the batch size per GPU to 64 and use 8/2 Tesla V100 32G GPUs for pre-training/fine-tuning.
Software Dependencies No The paper mentions software components like 'T5 (Raffel et al. 2020)', 'GPT-2 Large (Radford et al. 2019)', 'Dialo GPT Large (762M)', and 'Conv Lab-3 (Zhu et al. 2022)'. While specific models/toolkits are named, explicit version numbers for T5, GPT-2, or PyTorch/CUDA are not provided.
Experiment Setup Yes We consider two sizes of model: 60M T5-small and 220M T5-base. For both RG and KPT, we pre-train the models for 1 epoch. During pre-training, we set the keyword ratio α to 0.3... We use Adafactor optimizer with a constant learning rate 1e-3 for both pre-training and fine-tuning. We set the batch size per GPU to 64...