Improved Text Matching by Enhancing Mutual Information
Authors: Yang Liu, Wenge Rong, Zhang Xiong
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments are conducted on Zhihu, Wiki QA and Sem Eval datasets and the results demonstrate that the rewriting module improves ranking performance consistently. |
| Researcher Affiliation | Academia | 1State Key Laboratory of Software Development Environment, Beihang University, China 2School of Computer Science and Engineering, Beihang University, China {liuyang1123, w.rong, xiongz}@buaa.edu.cn |
| Pseudocode | No | The paper describes algorithmic steps in narrative form and with equations, but it does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper states 'We make Zhihu Rank public available5 for research community' with a goo.gl link to a dataset, but it does not provide any specific link or explicit statement about the open-source code for their methodology. |
| Open Datasets | Yes | We conduct AS experiments on Sem Eval 2016 task 3 subtask A2 dataset and the Wiki QA3(Yang, Yih, and Meek 2015) dataset. We make Zhihu Rank public available5 for research community. We select 20,000 useable questions and their answers from the data warehouse of Zhihu to train and test our model. |
| Dataset Splits | No | The paper mentions 'learning rate will be de decayed if results have no improvement in valid data', implying the use of a validation set. However, it does not provide specific percentages or sample counts for the training, validation, and test splits needed for reproduction. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware (e.g., GPU/CPU models, memory, or cloud instances) used to run the experiments. |
| Software Dependencies | No | The paper mentions software components like 'Adam optimizer' and 'Glove embedding' and model architectures like 'bidirectional GRU' and 'CNN model', but it does not specify version numbers for any software libraries, frameworks, or programming languages used. |
| Experiment Setup | Yes | We use Adam optimizer with learning rate 0.001 to update parameters, learning rate will be de decayed if results have no improvement in valid data. The batch size is 20 and word embeddings are initialized using 100 dimensions glove embedding. For the generating module, hidden state size of the decoder RNN is 256. The discriminator is a regular CNN model with convolutional kernel heights being 2 and 3, the width equals word embedding size. For the fusion matching module, both Bi GRU1 and Bi GRU2 have 512 hidden units and initial states are zero. The batch size is 20. L2 regularization is used to prevent model from overfitting. The feature extractor CNN consists of convolution and pooling layers and a dense layer, with 3 3 kernel size and 2 2 pooling window size. Dropout is employed to the hidden outputs with probability 40% during training phase. |