AttnConvnet at SemEval-2018 Task 1 : Attention-based Convolutional Neural Networks for Multi-label Emotion Classification

Tensorflow implementation of AttnConvnet

Warning!! Dirty version source code

The source code will be beautified and restructured ASAP

1. Model

   • Embedding

     • Pre-trained Glove embeddings
     • random initialized embeddings

   • Multi-head Dot-product Attention

     • using partial code from google's Transformer v1.0.8

   • 1-layer Convolutional Neural Network

2. Dataset

   • SemEval-2018 Task 1 : Affect in Tweets

     • Subtask 5 : Detecting Emotions (multi-label classification)
       • train-eng
       • dev-eng
       • test-eng

   • Lexicon

     • Word-Emotion and Word-Sentiment Association lexicons
       • NRC Word-Emotion Association Lexicon

Requirements

• python 2.7
• numpy
• pandas
• Tensorflow 1.4
• nltk

Usage

1. Download data to data/(In the root directory)

mkdir data
cd data
mkdir processed
wget http://saifmohammad.com/WebDocs/AIT-2018/AIT2018-DATA/E-c/English/2018-E-c-En-train.zip
wget http://saifmohammad.com/WebDocs/AIT-2018/AIT2018-DATA/E-c/English/2018-E-c-En-dev.zip
wget http://saifmohammad.com/WebDocs/AIT-2018/AIT2018-DATA/AIT2018-TEST-DATA/semeval2018englishtestfiles/2018-E-c-En-test.zip

# unzip
unzip 2018-E-c-En-train.zip
unzip 2018-E-c-En-dev.zip
unzip 2018-E-c-En-test.zip

2. Emoji-to-meaning preprocessing(To be updated)

3. Get pre-trained embedding(Optional)

wget http://nlp.stanford.edu/data/glove.840B.300d.zip
unzip glove.840B.300d.zip

4. Process data

cd ..
python process_data.py
python process_embedding.py # (optional) get pre-trained embedding, it will take a couple of minutes

5. Get & process nrc Lexicon file(Optional)
   • Put the file NRC-Emotion-Lexicon-Wordlevel-v0.92.txt inside data/

python process_nrc.py # it will take a couple of minutes

6. Train & Test model

# Train
bash run.sh [parameter set] train
# Example : bash run.sh basic_params train

# Text
bash run.sh [parameter set] pred
# Example : bash run.sh basic_params pred

Use pre-trained embedding

Open params.py and change embedding = None to the path of pre-trained embedding file

Example : embedding = 'data/processed/glove_embedding.npy'

Use lexicon

Open params.py and change lexicon_effect = None to other values

• AC + nrc1 in paper
  • lexicon_effect = 'nrc1'
• AC + nrc2 in paper
  • lexicon_effect = 0.4