Towards Credible Human Evaluation of Open-Domain Dialog Systems Using Interactive Setup

Authors: Sijia Liu, Patrick Lange, Behnam Hedayatnia, Alexandros Papangelis, Di Jin, Andrew Wirth, Yang Liu, Dilek Hakkani-Tur

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our comprehensive human evaluation results shed light on how to conduct credible human evaluations of open domain dialog systems using the interactive setup
Researcher Affiliation Industry Amazon Alexa AI {sijial, patlange, behnam, papangea, djinamzn, wirandre, yangliud, hakkanit}@amazon.com
Pseudocode No The paper does not contain any sections explicitly labeled 'Pseudocode' or 'Algorithm', nor are there any structured code-like blocks describing procedures.
Open Source Code No The paper mentions leveraging 'the Hugging Face s transformers library for all our models' and provides a link to its GitHub repository, but it does not state that the authors' own code for the presented methodology is open-source or publicly available.
Open Datasets Yes GPT2-XL/GPT2-M fine-tuned on Blended Skill Talk (BST) Dataset (Smith et al. 2020); GPT2-XL fine-tuned on Topical Chat (TCS) Dataset (Gopalakrishnan et al. 2019); GPT2-XL fine-tuned on Wizard-of-Wikipedia (Wo W) Dataset (Dinan et al. 2018).
Dataset Splits No The paper discusses the datasets that GPT2 models were fine-tuned on and statistical power for evaluations, but it does not provide explicit details about train/validation/test dataset splits used for the interactive evaluation data collected in their experiments.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU types, or cloud computing instance specifications used for running their experiments or fine-tuning the models.
Software Dependencies No The paper mentions using 'the Hugging Face s transformers library for all our models' but does not specify a version number for this library or any other software dependencies like Python or specific deep learning frameworks with their versions.
Experiment Setup No The paper describes the setup of the human interactive evaluation mechanisms (e.g., SOBA, SATA) and how models were fine-tuned (e.g., 'nucleus sampling'), but it does not provide concrete hyperparameter values such as learning rates, batch sizes, optimizer settings, or detailed training schedules that would constitute a reproducible experimental setup.