README 2.78 KB
# DEFT 2017 - Sentiment Analysis

- Authors: Mickael Rouvier and Pierre-Michel Bousquet
- Version: 1.0
- Date: 26/06/17

These scripts provide the LIA system that I used for the DEFT 2017 - Sentiment Analysis. The LIA system is a multi-view ensemble of Convolutional Neural Networks (CNN). Four different word embeddins are used to initialize the input of CNN : lexical embedding, sentiment embedding (multi-task learning), sentiment embedding (distant learning) and sentiment embedding (negative sampling). The system is a fusion at the score level of the different CNNs variants.


You can reproduce my results or freely adapt my code for your experiments.


Warning, before to run the system execute the makefile:

```shell
make
```

This executable split the training corpus (K-Fold) and tokenize the tweets:

```shell
sh run_corpus.sh
```

This executable train the different word embeddings:
```shell
sh run_word2vec.sh
```

This executable learn the models:

```shell
sh run_cnn.sh
```

This executable run the model on dev and test:

```shell
sh run_extract_dev.sh
sh run_extract_test.sh
```

At this point you can score the CNNs:
```shell
ruby bin/scoring.rb data/task1_test.tokenize  results_test/cnn_task1_0_distant_size100_123.txt
ruby bin/scoring.rb data/task2_test.tokenize  results_test/cnn_task2_0_distant_size100_123.txt
ruby bin/scoring.rb data/task3_test.tokenize  results_test/cnn_task3_0_distant_size100_123.txt
```


This executable run the fusion system:
```shell
sh run_fusion.sh
```


Finally, you can score the full-system:
```shell
ruby bin/scoring.rb data/task1_test.tokenize  output/equipe-8_tache1_run3.csv
ruby bin/scoring.rb data/task2_test.tokenize  output/equipe-8_tache2_run1.csv
ruby bin/scoring.rb data/task3_test.tokenize  output/equipe-8_tache3_run3.csv
```



# Results


## Baseline

We reproduce the sentiment analysis system of Kim (based on Word embeddings and CNN):


| Corpus       | Baseline      | 
| ------------ |:-------------:|
| Task1        |  59.55        |
| Task2        |  77.18        |
| Task3        |  57.59        |




## DEFT 2017

These results are those SENSEI-LIF system presented in SemEval 2016 Sentiment Analysis:

| Corpus      | Task1         | Task2         | Task3         |
| ----------- |:-------------:|:-------------:|:-------------:|
| Run1        |  60.23        |  78.31        |  57.83        |
| Run2        |  63.44        |  77.39        |  58.49        |
| Run3        |  65.00        |  77.43        |  59.39        |


# Citing

The system is described in this paper:

    @inproceedings{rouvier2017,
      author    = {Mickael Rouvier and Pierre-Michel Bousquet},
      title     = {LIA @ DEFT’2017 : Multi-view Ensemble of Convolutional Neural Network},
      booktitle = {DEFT 2107},
      year      = {2017},
      address   = {Orleans, France}
    }