Naver AI Hackathon 2018

Naver AI Hackathon 2018 task was to predict movie ratings and the similarities between two questions.
I participated as a team Deeppangyo.

Summary of approach

• Movie rating prediction: Ranked 9th (over 200 teams, MSE: 2.86229)
• Two question similarity prediction: Ranked 8th (1st round) / 11th (2nd round) (ACC: 0.960798/0.960328)

Features

Movie rating prediction

• Fully dockerized environments
• Bidirectional LSTM (Phoneme + Char + Word)
• L2 regularization
• Word CNN
• Self Attention (but not used)

Two question similarity prediction

• Siamese network using Manhattan distance
• Contrastive loss
• Convolution + Bidirectional GRU
• Self Attention
• GAP/GMP
• Word CNN (but not used)

Usage

Movie rating prediction on NSML

Login with nsml, and run commands as follows:

$ cd movie-review
$ nsml run -d movie_final -e main.py -a "--epochs 10 --batch 1000 --strmaxlen 117 --embedding 64 --dropout 0.3"

Two question similarity prediction on NSML

Login with nsml, and run commands as follows:

$ cd kin
$ nsml run -d kin_final -e main.py -a "--epochs 200 --batch 2000 --strmaxlen 250 --embedding 128 --valrate 0.3"

Requirements

• python 3
• keras
• torch
• numpy
• matplotlib
• tensorflow
• pandas
• sklearn