Adversarial Learning for Chinese NER From Crowd Annotations
Authors: YaoSheng Yang, Meishan Zhang, Wenliang Chen, Wei Zhang, Haofen Wang, Min Zhang
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In our experiments, we create two data sets for Chinese NER tasks from two domains. The experimental results show that our system achieves better scores than strong baseline systems. |
| Researcher Affiliation | Collaboration | 1School of Computer Science and Technology, Soochow University, China 2Alibaba Group and 3Shenzhen Gowild Robotics Co. Ltd 4School of Computer Science and Technology, Heilongjiang University, China |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide a direct link or explicit statement about the availability of the source code for the methodology described in this paper. It only references third-party tools' source code or external resources for data. |
| Open Datasets | No | In our experiments, we create two data sets for Chinese NER tasks in the dialog and e-commerce domains. We hire undergraduate students to annotate the sentences... Finally, we have 16,948 sentences annotated by the students... Finally, we obtain 2,337 sentences for EC-MT and 2,300 for EC-UQ. The paper does not provide specific access information (link, DOI) for these newly created datasets. |
| Dataset Splits | Yes | Among them, we use 300 sentences as the development set and the remaining 700 as the test set. The rest sentences with only student annotations are used as the training set. (for DL-PS). Among them, we use 100 sentences as the development set and the remaining 300 as the test set. The rest sentences with only crowdsourcing annotations are used as the training set. (for EC-MT/UQ). |
| Hardware Specification | No | The paper does not provide any specific hardware details such as GPU or CPU models, or memory specifications used for running experiments. |
| Software Dependencies | No | The paper mentions tools like 'Crfsuite' and 'word2vec' but does not specify their version numbers or any other software dependencies with specific versions. |
| Experiment Setup | Yes | Concretely, we set the dimension size of the character embeddings by 100, the dimension size of the NE label embeddings by 50, and the dimension sizes of all the other hidden features by 200. We exploit online training with a mini-batch size 128 to learn model parameters. The max-epoch iteration is set by 200, and the best-epoch model is chosen according to the development performances. We use RMSprop (Tieleman and Hinton 2012) with a learning rate 10 3 to update model parameters, and use l2-regularization by a parameter 10 5. We adopt the dropout technique to avoid overfitting by a drop value of 0.2. |