KID-Review: Knowledge-Guided Scientific Review Generation with Oracle Pre-training

Authors: Weizhe Yuan, Pengfei Liu11639-11647

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally, we perform a comprehensive evaluation (human and automatic) from different perspectives. Empirical results have shown the effectiveness of different types of knowledge as well as oracle pre-training.
Researcher Affiliation Academia Weizhe Yuan, Pengfei Liu Language Technologies Institute, Carnegie Mellon University, Pittsburgh, PA, USA {weizhey, pliu3}@cs.cmu.edu
Pseudocode No The paper describes the model architecture and processes using text and diagrams (Fig. 2, 3, 4) and mathematical equations, but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes We make all code, relevant datasets available: https://github.com/yyy-Apple/KIDReview as well as the KIDREVIEW system: http://nlpeer.reviews.
Open Datasets Yes We use ASAP-Review dataset introduced by Yuan, Liu, and Neubig (2021) for our experiment. It consists of ICLR papers from 2017-2020 and Neur IPS papers from 2016-2019, together with their aligned reviews.
Dataset Splits Yes To make a fair comparison, we use the same training, validation, and test split as them. The basic statistics of this dataset are shown in Tab. 1. Figure 5: Train Validation Test 8,000 6,993 Unique papers Aligned papers
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running experiments.
Software Dependencies No The paper mentions using BART (Lewis et al. 2020) and a specific BART checkpoint ('bart-large-cnn') but does not provide version numbers for BART or any other key software dependencies.
Experiment Setup Yes We set the embedding size to be 128 when learning citation embeddings. We use two GAT layers for the concept graph, each with 4 attention heads, and we set the hidden size to be 200. To get the initial concept graph embeddings, we set l = Nenc/2, where Nenc denotes the total number of layers in BART encoder. For each BART decoder layer, we add another cross-attention module to attend to entity node representations on top of the regular cross attention module. We use beam search decoding during generation and adopt the same parameters following Yuan, Liu, and Neubig (2021) for all systems.