Neuramanteau

Accompanying code for the paper - Neuramanteau: A Neural Network Ensemble Model for Lexical Blends.

Try out the demo here

Requirements

• Python 3.5+
• Tensorflow 1.1+
• Numpy 1.11+

Usage

1. Train the model
   • python blender.py train --data wiktionary-blends-shuffled.csv --num-epochs 16
2. Evaluate the model
   a. Validation set: python blender.py eval --data wiktionary-blends-shuffled.csv --partition-name validation
   b. Test set: python blender.py eval --data wiktionary-blends-shuffled.csv --partition-name test
   c. External dataset: python blender.py eval --data <dataset csv file> --no-partitions
3. Sample outputs:
   • python blender.py sample --data wiktionary-blends-shuffled.csv --source-words work alcoholic

By default the model uses 6 experts. To change the number of experts use the flag --num-experts 15 (for 15 experts). You must pass the same values during sampling and evaluation.

To use dominances add the flag --add-dominances during training. During sampling set dominances via --dominances 1 0 (1st word dominant).

Dataset

The blend dataset we provide here was scraped from Wiktionary on March 2017. It consists of 3250 blends (unfiltered). Each line of the dataset contains the blends in the following format:

target blend, first source word, second source word

Three CSV files are provided:

• wiktionary-blends-shuffled.csv : The entire blend dataset
• wiktionary-blends-shuffled-cs.csv: Contains subset of the blend dataset which have overlapping characters (common string)
• wiktionary-blends-shuffled-nocs.csv: Contains subset of the blend dataset which have no overlapping characters

DISCLAIMER: The above dataset is provided as is and we are not liable for its contents