Solving Sequential Text Classification as Board-Game Playing
Authors: Chen Qian, Fuli Feng, Lijie Wen, Zhenpeng Chen, Li Lin, Yanan Zheng, Tat-Seng Chua8640-8648
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive results on three representative datasets show that the proposed approach outperforms the state-of-the-art methods with statistical significance. We conduct extensive experiments to answer the following three research questions: RQ1 Does our proposed approach, Gu Go, outperform the currently state-of-the-art STC solutions? RQ2 How do the different labeling orders and the speedup strategies affect performance? RQ3 What are the main differences between jump labeling and successive labeling apart from the labeling order? |
| Researcher Affiliation | Academia | Chen Qian Tsinghua University qc16@mails.tsinghua.edu.cn Fuli Feng National University of Singapore fulifeng93@gmail.com Tsinghua University wenlj@tsinghua.edu.cn Zhenpeng Chen Peking University czp@pku.edu.cn Li Lin Tsinghua University veralin1994@gmail.com Yanan Zheng Tsinghua University zhengyanan932@gmail.com Tat-Seng Chua National University of Singapore chuats@comp.nus.edu.sg |
| Pseudocode | No | The paper describes the steps of Monte Carlo Tree Search (MCTS) conceptually and with equations, and includes an architecture diagram (Figure 4), but it does not provide a formal pseudocode block or algorithm listing. |
| Open Source Code | Yes | The code is publicly available at https://github.com/qianc62/Gu Go. |
| Open Datasets | Yes | MAM (Qian et al. 2019) contains manuals from a wikibased site4 to teach people to fix various devices such as phones, cameras, cars, etc. (4https://www.ifixit.com). COR (Feng, Zhuo, and Kambhampati 2018) is a collection of user-generated recipes with textual descriptions of cooking procedures from a food-focused social network5. (5https://www.recipe.com). We Bis (Chen et al. 2019) contains various consumer reviews6... (6https://webis.de). |
| Dataset Splits | Yes | For comparison, all datasets are divided into train/dev/test sets using an 8:1:1 ratio. |
| Hardware Specification | Yes | All of our experiments are run on a machine equipped with an Intel Core i7 processor, 32 GB of RAM, and an NVIDIA Ge Force-GTX-1080-Ti GPU. |
| Software Dependencies | Yes | We implement Gu Go via Python 3.7.3 and Pytorch 1.0.1. |
| Experiment Setup | Yes | We use BERT (Devlin et al. 2019) as the language model and perform average pooling to obtain the fragment embeddings with dimension of 768. For the encoding layer, we also set its output dimension as 768. In the game-playing layer, we set a hard threshold (α=0.05) in chi-squared testing to select emission and transition features for game bonus evaluation (Equation 2). In order to obtain the prior probability of each fragment (Equation 3), we use a one-hidden-layer MLP (the size of hidden layer is 200) with Re LU as an activation function and the Adam optimizer (Kingma and Ba 2015) with learning rate 10 4. The training process includes two main steps: 1) We pretrain the encoding layer and the MLP module with at most 5,000 epochs, a mini-batch size of 32 and the cross entropy as the loss function. |