Multiple-Attribute Text Rewriting
Authors: Guillaume Lample, Sandeep Subramanian, Eric Smith, Ludovic Denoyer, Marc'Aurelio Ranzato, Y-Lan Boureau
ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments demonstrate that the fully entangled model produces better generations, even when tested on new and more challenging benchmarks comprising reviews with multiple sentences and multiple attributes. |
| Researcher Affiliation | Collaboration | 1Facebook AI Research, 2MILA, Universit e de Montr eal 3Sorbonne Universit es, UPMC Univ Paris 06 |
| Pseudocode | No | The paper describes its approach and model architecture but does not include a dedicated pseudocode or algorithm block. |
| Open Source Code | No | The source code and benchmarks will be made available to the research community after the reviewing process. |
| Open Datasets | Yes | We use data from publicly available Yelp restaurant and Amazon product reviews following previous work in the area (Shen et al., 2017; Li et al., 2018) and build on them in three ways to make the task more challenging and realistic. (...) Yelp Reviews This dataset consists of restaurant and business reviews provided by the Yelp Dataset Challenge2. (...) Amazon Reviews The amazon product review dataset (He & Mc Auley, 2016) |
| Dataset Splits | No | The few models that met the specified threshold on the validation set were evaluated by humans on the same validation set and the best model was selected to be run on the test set. The paper mentions a validation set but does not provide specific details on its size or split percentage for reproducibility. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory used for running the experiments. |
| Software Dependencies | No | We used the Adam optimizer (Kingma & Ba, 2014) with a learning rate of 10 4, β1 = 0.5, and a batch size of 32. (...) Fluency is measured by the perplexity assigned to generated text sequences by a pre-trained Kneser Ney smooth 5-gram language model using Ken LM (Heafield, 2011). (...) We normalize, lowercase and tokenize reviews using the moses (Koehn et al., 2007) tokenizer. The paper mentions tools and their citations but lacks specific version numbers for software dependencies. |
| Experiment Setup | Yes | We used the Adam optimizer (Kingma & Ba, 2014) with a learning rate of 10 4, β1 = 0.5, and a batch size of 32. (...) The hyper-parameters of our model are: λAE and λBT trading off the denoising auto-encoder term versus the back-translation term (...), the temperature T (...) and the pooling window size w. |