A Graph Fusion Approach for Cross-Lingual Machine Reading Comprehension
Authors: Zenan Xu, Linjun Shou, Jian Pei, Ming Gong, Qinliang Su, Xiaojun Quan, Daxin Jiang
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experiments on two benchmark datasets for cross-lingual MRC show that our approach outperforms all strong baselines, which verifies the effectiveness of syntax information for cross-lingual MRC. |
| Researcher Affiliation | Collaboration | Zenan Xu1 , Linjun Shou2, Jian Pei3, Ming Gong2, Qinliang Su1,4 , Xiaojun Quan1, and Daxin Jiang2 1School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China 2Microsoft Search Technology Center Asia (STCA), Beijing, China 3School of Computing Science, Simon Fraser University 4Guangdong Key Laboratory of Big Data Analysis and Processing, Guangzhou, China |
| Pseudocode | No | The paper describes algorithms but does not provide structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper states: 'We implement on top of Hugging Face s Transformers (Wolf et al. 2019) and report results on both base and large models, i.e., GFMRCbase and GFMRClarge. We initialize our base model by the pre-trained XLM-Rbase model released by Hugging Face5, which contains 12 layers...' and provides a footnote link: '5https://huggingface.co/xlm-roberta-base'. This refers to a third-party framework/model, not the specific code for their proposed GFMRC approach. |
| Open Datasets | Yes | Datasets: MLQA (Lewis et al. 2020) and Ty Di QA-Gold P dataset (Clark et al. 2020) are two recent public benchmark datasets for cross-lingual machine reading comprehension... Following FILTER (Fang et al. 2021), we use SQu AD v1.1 (Rajpurkar et al. 2016) English training data as additional data during the fine-tuning stage. |
| Dataset Splits | No | The paper mentions a 'development set' for Ty Di QA-Gold P and 'test set' for MLQA and Ty Di QA-Gold P, but it does not provide specific details on the train/validation/test splits (e.g., exact percentages or sample counts for all splits across all datasets used, especially the combined translated training data), which would be needed to fully reproduce the data partitioning. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper mentions using 'Hugging Face s Transformers (Wolf et al. 2019)' and 'Stanza toolkit3 (Qi et al. 2020)' but does not specify their version numbers or other software dependencies with specific versions. |
| Experiment Setup | Yes | We set the number of intermediate layers, i.e., the Syntax-Enhanced and Alignment-Aware Transformer layers, to 10 in the base model and to 22 in the large model. The first bottom Transformer layer is used for encoding the raw input sentences and the top layer converts the joint representation of the sentences in the source and target languages back to individual language spaces. |