Fast and Constrained Absent Keyphrase Generation by Prompt-Based Learning
Authors: Huanqin Wu, Baijiaxin Ma, Wei Liu, Tao Chen, Dan Nie11495-11503
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on keyphrase generation benchmarks have demonstrated the effectiveness of our approach. In addition, we evaluate the performance of constrained absent keyphrases generation from an information retrieval perspective. |
| Researcher Affiliation | Collaboration | 1 Tencent AI Platform Department, China 2 Peking University {huanqinwu, thinkweeliu, vitochen, kathynie}@tencent.com, mabaijiaxin@stu.pku.edu.cn |
| Pseudocode | No | The paper describes the approach steps and shows architectural diagrams (Figures 2, 3, 4) but does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions publicly available implementations for baseline models and datasets, e.g., 'We use the publicly implementation available in https://github.com/thinkwee/Uni Keyphrase.' and 'We follow the official Git Hub repository to prepare datasets which are available on https://github.com/memray/Open NMT-kpgrelease.' However, it does not provide a specific link or explicit statement about the availability of the source code for their own proposed methodology. |
| Open Datasets | Yes | We follow the setup widely used in the keyphrase prediction task, which is training the model on the KP20K (Meng et al. 2017) dataset, and giving an evaluation on three more benchmark datasets: INSPEC (Hulth 2003), NUS (Nguyen and Kan 2007) and SEMEVAL (Kim et al. 2010). We use the data from KP20K validation set as validation data and apply them to identify optimal checkpoints for testing. We follow the pre-process, post-process setting of Meng et al. (2017, 2019); Yuan et al. (2020)1. In addition, we apply ACM-CR (Boudin and Gallina 2021)2 to build retrieval tasks for evaluating the consistency between generative absent keyphrase and document. 1We follow the official Git Hub repository to prepare datasets which are available on https://github.com/memray/Open NMT-kpgrelease. 2https://github.com/boudinfl/redefining-absentkeyphrases/blob/main/data/acm-cr/acm-cr.v1.tar.gz |
| Dataset Splits | Yes | We use the data from KP20K validation set as validation data and apply them to identify optimal checkpoints for testing. |
| Hardware Specification | Yes | It takes about 50 minutes per epoch to train the model on 4 Nvidia Tesla V100 GPU cards with mixed-precision training. For a fair comparison, we use the same device (NVIDIA V100) to evaluate the inference time on the KP20K test set. |
| Software Dependencies | No | The paper states 'our model is implemented using Py Torch' but does not specify a version number for PyTorch or any other software libraries or dependencies. It also mentions using a 'pre-trained prefix LM' but without a version. |
| Experiment Setup | Yes | The weight of positive label wp in PKE is set to 5.0. The weight of positive label wkw in KWE is set to 10.0. The learning rate is 1e-5 and the proportion of warmup steps is 0.1. We set the batch size to 200 and the maximum length to 384. The number of [MASK] tokens in each K-Prompt is 4 (two [MASK] tokens on each side of the keyword) and the number of [MASK] tokens in NK-Prompt is 8. We set the K of the top keywords as 6. |