Heavy-tailed Representations, Text Polarity Classification & Data Augmentation
Authors: Hamid Jalalzai, Pierre Colombo, Chloé Clavel, Eric Gaussier, Giovanna Varni, Emmanuel Vignon, Anne Sabourin
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on synthetic and real text data show the relevance of the proposed framework and confirm that this method generates meaningful sentences with controllable attributes, e.g. positive or negative sentiments. |
| Researcher Affiliation | Collaboration | Hamid Jalalzai LTCI, T el ecom Paris Institut Polytechnique de Paris hamid.jalalzai@telecom-paris.fr Pierre Colombo IBM France LTCI, T el ecom Paris Institut Polytechnique de Paris pierre.colombo@telecom-paris.fr Chlo e Clavel LTCI, T el ecom Paris Institut Polytechnique de Paris chloe.clavel@telecom-paris.fr Eric Gaussier Univ. Grenoble Alpes, CNRS, Grenoble INP, LIG eric.gaussier@imag.fr Giovanna Varni LTCI, T el ecom Paris Institut Polytechnique de Paris giovanna.varni@telecom-paris.fr Emmanuel Vignon IBM France emmanuel.vignon@fr.ibm.com Anne Sabourin LTCI, T el ecom Paris Institut Polytechnique de Paris anne.sabourin@telecom-paris.fr |
| Pseudocode | Yes | A detailed description of the training step of GENELIEX is provided in Algorithm 2 in Appendix A.3, see also Appendix A.2 for an illustrative diagram. |
| Open Source Code | No | The paper does not provide any explicit statement about releasing source code for the described methodology or a link to a code repository. |
| Open Datasets | Yes | In our experiments we rely on two large datasets from Amazon (231k reviews) [41] and from Yelp (1,450k reviews) [58, 36]. |
| Dataset Splits | No | The paper states that 'LHTR is trained on 2250 examples and a testing set of size 750' for a toy example, and mentions 'The extreme region contains 6.9k samples for Amazon and 6.1k samples for Yelp, both corresponding roughly to 25% of the whole test set size.' and 'For experiment C4 the test set contains 104 sequences.' However, it does not provide explicit training, validation, and test dataset splits (e.g., percentages or exact counts for all splits) for the main Amazon and Yelp experiments. It does not mention a 'validation' set explicitly. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used to run the experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper mentions software tools like BERT, fastText, and Transformer Decoder, but it does not specify any version numbers for these or other software dependencies, which are necessary for reproducibility. |
| Experiment Setup | Yes | The proportion of extreme samples in the training step of LHTR is chosen as κ = 1/4. The regularly varying target distribution is chosen as a multivariate logistic distribution with parameter δ = 0.9. For each train set s sequence considered as extreme, 10 new sequences are generated using each algorithm. GENELIEX is evaluated with scaling factor λ ranging from 1 to 1.5. Appendix B.6 describes the MLP architectures, and Appendix B.7 mentions 'Table 7 for hyperparameters'. |