Appropriate Answer Prediction

(科技大擂台與AI對話)

Feature

A CNTK (Microsoft deep learning toolkit) implementation of CS565600 competition
We use LSTM + attention to do this task
For more model information, please refer to the report
If you meet any problem in this repo, feel free to contact zlsh80826@gmail.com

Requirements

Here are some required libraries for training

General

python3
cuda-9.0 (CNTK required)
openmpi-1.10 (CNTK required)
gcc >= 6 (CNTK required)

Python

Please refer requirements.txt

Usage

We recommand you to run all the scripts in script directory

cd AppropriateResponsePrediction/script

Each script contain helper, you can check it for customed settings.

python <some script>.py --help

Preprocess

This script will convert the text format program to processed npy format.

You can specify --threads to indicate how many threads you want to use.

python preprocessing.py

Train Fasttext

This script will train the Traditional Chinese Embedding with processed data.

python train_fasttext.py

Generate The Training Data

Default settings will generate 4 million training data, which will consume about 8 GB disk space.

python gen_training.py

Convert tsv to ctf

CNTK support large training file, but we need to convert it to ctf format.

python tsv2ctf.py

Train

Default settings will run 300 epochs and save the checkpoint of each epoch.

python train.py

Inference

Inference script will read the checkpoint and do the inference. So you can inference while training, 4 - 10 epochs result is good enough in my experimence.

python inference.py

Performance

Based on the Kaggle Leaderboard, our implementation is second prize (the first two are fake).

	Public score	Private score
Single Model	73.6	71.6
Ensemble Model	76.4	72

zlsh80826/AppropriateResponsePrediction