Team members :

• Melinda Femminis
• Estelle Valerie Tsague Mbialeu
• Catherine Pedroni

Thsi project is part of a kaggle competition. The main goal of this competition is to be able to predict the difficulty level of a french text according to the The Common European Framework of Reference for Languages that describes 6 level of language : A1, A2, B1, B2, C1, C2.

For this project we have two dataset. The first dataset called training_data has three features : id, sentence, difficulty. The second dataset called unlabelled_test_data has only two features : id and sentence. By training several models we are going to try and predict the difficulty level for the unlabelled_test_data with the best accuracy score possible.

We can clearly see on this table that the best model is the logistic regression. An interesting see to observe is that for both the regression and the decision tree, all score (Precision, Recall, F1, Accuracy) have the same value. This means that we have as many false positives as false negatives. For our knn model, we can observe a clear difference in the scores' values. The difference between precision and recall will probably mean that the confusion matrix won't display a clear diagonal as opposed to the regression matrix or decision tree matrix. Overall, this results could still improve but are however above the baseline of 0.169

After Loading the training_data and doing some Exploratory Data Analysis, we did several steps to build our classifier :

• split the dataset in 2 sets; 80% for training and 20 % for test set
• create the metrics function that prints a classification report (accuracy, precision, recall, f1 score)
• create TfIdf vectorisers with different (sublinear_tf= True, tokenizer=word_tokenize, ngram_range= (1,1))
• test multiple models such as Logistic regression , K-Nearest Neighbors, Decision Tree and Random Forest to find the one that has the best accuracy. After this step, we tested our models on the unlabeled data to submit it to the Kaggle competition.

What we did to improve the accuracy :

• Preprocessing: remove punctuation, remove stopword, remove X most/least frequent words, stemming
• Dimentionality reduction: reduce the number of feature using the TruncatedSVD
• Try other models: multinomail naive Bayes, Support Vector Machine, Neural Networks with Keras

Unortunatelly, all these methods and models didn't improve the accuracy.