Response Generation by Context-Aware Prototype Editing
Authors: Yu Wu, Furu Wei, Shaohan Huang, Yunli Wang, Zhoujun Li, Ming Zhou7281-7288
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiment results on a large scale dataset demonstrate that our new paradigm significantly increases the relevance, diversity and originality of generation results, compared to traditional generative models. Furthermore, our model outperforms retrieval-based methods in terms of relevance and originality. |
| Researcher Affiliation | Collaboration | Yu Wu, Furu Wei, Shaohan Huang, Yunli Wang, Zhoujun Li, Ming Zhou State Key Lab of Software Development Environment, Beihang University, Beijing, China Microsoft Research, Beijing, China {wuyu,wangyunli,lizj}@buaa.edu.cn {fuwei, shaohanh, mingzhou}@microsoft.com |
| Pseudocode | No | The paper includes equations and an architectural diagram (Figure 1) but does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | To facilitate further research, related resources of the paper can be found at https://github.com/Mark Wu NLP/Response Edit. |
| Open Datasets | No | We collected over 20 million human-human contextresponse pairs (context only contains 1 turn) from Douban Group 4. After removing duplicated pairs and utterance longer than 30 words, we split 19,623,374 pairs for training, 10,000 pairs for validation and 10,000 pairs for testing. Although the data was collected from a public source (Douban Group), the paper does not provide a direct link, DOI, or citation to *their processed version* of the dataset for public access. |
| Dataset Splits | Yes | After removing duplicated pairs and utterance longer than 30 words, we split 19,623,374 pairs for training, 10,000 pairs for validation and 10,000 pairs for testing. |
| Hardware Specification | No | The paper does not specify any particular hardware used for the experiments, such as specific GPU or CPU models, or memory details. |
| Software Dependencies | No | We implement our model by Py Torch 3. ... We use a Pytorch implementation, Open NMT (Klein et al. 2017) in the experiment. The paper mentions PyTorch and OpenNMT but does not provide specific version numbers for these software dependencies. |
| Experiment Setup | Yes | We employ the Adam algorithm (Kingma and Ba 2015) to optimize the objective function with a batch size of 128. We set the initial learning rate as 0.001 and reduce it by half if perplexity on validation begins to increase. We will stop training if the perplexity on validation keeps increasing in two successive epochs. ... In practice, the word embedding size and editor vector size are 512, and both of the encoder and decoder are a 1-layer GRU whose hidden vector size is 1024. Message and response vocabulary size are 30000, and words not covered by the vocabulary are represented by a placeholder $UNK$. ... All generative models use beam search to yield responses, where the beam size is 20 except S2SA-MMI. |