Unsupervised Controllable Text Formalization
Authors: Parag Jain, Abhijit Mishra, Amar Prakash Azad, Karthik Sankaranarayanan6554-6561
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on public datasets demonstrate the efficacy of our model towards: (a) transforming a given text to a more formal style, and (b) varying the amount of formalness in the output text based on the specified input control. |
| Researcher Affiliation | Industry | Parag Jain, Abhijit Mishra, Amar Prakash Azad, Karthik Sankaranarayanan IBM Research {pajain34,abhijimi,amarazad,kartsank}@in.ibm.com |
| Pseudocode | No | The paper describes the system architecture and training process in narrative text and with a diagram (Figure 2), but it does not include any formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code and datasets are released for academic use. [...] The source code and dataset are publicly available. [...] The source code and dataset are available at https://github.com/parajain/uctf. |
| Open Datasets | Yes | Our dataset contains unlabeled text which are simple and informal in nature. It comprises of 14432 informally written sentences collected from Enron Email Corpus4, Corpus of Late Modern English Prose5, non-spam emails from Spam Dataset6, and essays for kids7. [...] 4http://bailando.sims.berkeley.edu/enron_email.html 5http://ota.ox.ac.uk/desc/2077 6http://csmining.org/index.php/spam-assassin-datasets.html 7http://www.mykidsway.com/essays/ |
| Dataset Splits | Yes | The data is split into a train:valid:test split of 80% : 12% : 8%. |
| Hardware Specification | No | The paper describes the software setup and model parameters but does not specify any hardware details like GPU models, CPU types, or memory used for running the experiments. |
| Software Dependencies | No | The paper mentions several software tools and libraries used (e.g., 'Spacy', 'Ken LM', 'Marian toolkit') and provides references for some, but it does not specify exact version numbers for these software dependencies, which is required for reproducibility. |
| Experiment Setup | Yes | We used a bidirectional GRU based encoder with 2 layers; the forward and backward hidden representations from the encoder are concatenated. The embedding matrix with a word embedding size of 300 was shared between the encoder and decoder. For encoder, 2 layers of GRUs were used and the hidden dimension size was set to 250. The decoder had 2 layers with hidden dimension size set to 500 . The weights related to language model, sentence similarity and readability scorers, given in Equation 3 are empirically set to 0.6, 0.2 and 0.2 respectively. During loss computation (Equation 7) we tried different values of parameter λ [0.1, 0.3, 0.5] and settled with 0.1. The ζ1 and ζ2 in equation 4 were set to 1.05 and 1.1 respectively; these threshold values were incremented by 5% for each category to suit practical purposes. The temperature parameter τ was set to 0.001; we did not use annealing. For both encoder-decoder and control predictor, we employed the Adam optimizer with the learning rate set to 0.001. [...] For each input, we chose max K = 100 samples. Exploitation occurs for 20 epochs with early stopping enabled. |