https://github.com/daemon/musent
Extremely lightweight sentiment classifier
https://github.com/daemon/musent
Last synced: about 1 year ago
JSON representation
Extremely lightweight sentiment classifier
- Host: GitHub
- URL: https://github.com/daemon/musent
- Owner: daemon
- License: mit
- Created: 2018-01-27T06:16:54.000Z (over 8 years ago)
- Default Branch: master
- Last Pushed: 2018-01-30T21:45:47.000Z (over 8 years ago)
- Last Synced: 2025-02-04T13:43:47.564Z (over 1 year ago)
- Language: Python
- Size: 10.7 KB
- Stars: 0
- Watchers: 1
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# musent
Used for several NLP tasks, musent is a collection of small-footprint models. For example, at inference time for the SST binary sentiment classification task, the micro-model performs no more than a few thousand multiplies per word vector to achieve comparable accuracy, with respect to previous state-of-the-art. Even more drastically, the nano-model uses only a few dozen multiplies per **sentence** to achieve interesting results.
## Installation
1. Run `./getData.sh`
2. Run `python microsem` to train a model (SST binary by default)
## Results
| Model | SST-fine | SST-bin | # mult/word |
| -------------- |---------------|------------|--------------|
| CNN-static | 45.5 | 86.8 | ~90k |
| CNN-multi | 47.4 | 88.1 | ~180k |
| CNN-nonstatic | 48.0 | 87.2 | ~90k |
| musent-micro | 47.7 | 85.9 | ~1.6k |
| musent-nano | 40.6 | 83.4 | < 100 |
WIP