Unsupervised Learning of Deterministic Dialogue Structure with Edge-Enhanced Graph Auto-Encoder
Authors: Yajing Sun, Yong Shan, Chengguang Tang, Yue Hu, Yinpei Dai, Jing Yu, Jian Sun, Fei Huang, Luo Si13869-13877
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on several public datasets demonstrate that the novel model outperforms several alternatives in aggregating utterances with similar semantics. The effectiveness of the learned dialogue structured is also verified by more than 5% joint accuracy improvement in the downstream task of low resource dialogue state tracking. |
| Researcher Affiliation | Collaboration | Yajing Sun1,2, Yong Shan3, Chengguang Tang4*, Yue Hu1,2*, Yinpei Dai4, Jing Yu1,2, Jian Sun4, Fei Huang4, Luo Si4 1Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China 2School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China 3Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China 4Alibaba Group, Beijing, China |
| Pseudocode | Yes | Algorithm 1: Conversational Graph Initialization |
| Open Source Code | No | The paper does not provide explicit statements or links to the authors' own open-source code for the methodology described. It only links to third-party libraries (PyTorch, HuggingFace Transformers). |
| Open Datasets | Yes | Dataset We conduct the main experiments on the three public dialogue corpus: DSTC2 (Henderson, Thomson, and Williams 2014), Cam Rest676 (Rojas-Barahona et al. 2017) and SGD (Rastogi et al. 2019). Cam Rest676 contains a total of 676 dialogues in this dataset about finding restaurants in Cambridge, UK. ... For further verifying the effectiveness of our method, We choose the low-resource dialogue state tracking as our downstream dialogue task, and conduct experiments on WOZ 2.0 dataset (Wen et al. 2017). |
| Dataset Splits | Yes | We split the single-domain dataset by domain and select home, bus and flight for training. The statistical result is shown in Table 1. Note that the nodes and edges of our conversational graph are predefined based on Table 1, which are equal to the dialogue acts and states annotations in the dataset. For further verifying the effectiveness of our method, We choose the low-resource dialogue state tracking as our downstream dialogue task, and conduct experiments on WOZ 2.0 dataset (Wen et al. 2017). WOZ 2.0 dataset is a single restaurant reservation domain, in which belief trackers estimate three slots (area, food, and price range). Specifically, we sample 10% data from the original training set to construct the low-resource training set. The validation set and the test set are not changed. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, memory) used for running the experiments. |
| Software Dependencies | No | Our model is implemented with Py Torch 1 (Paszke et al. 2019). We employ the pre-trained BERT model in the SUM module that has 12 layers of 784 hidden units and 12 self-attention heads 2. It is published as bert-base-uncased model in a Py Torch version of BERT: https://github.com/huggingface/pytorchtransformers. |
| Experiment Setup | Yes | The learning rate is 1e-5. In the SIM module, The number of layers and hidden size in EGAE are set to 2 and 768 for the graph convolution encoder, respectively. Adam optimizer (Kingma and Ba 2015) is employed for optimization with learning rate and warmup proportion set to 1e-3 and 0.1. The batch size is set to 64. |