natural-language-processing

My works for "Natural Language Processing" Coursera course.

Week1

Predict tag of Stackoverflow with linear model. (multi-tag) Used GridSearch CV for optimisation on tfidf model.

F1-score weighted :0.654025

Code is outdated see nlp-2

Week2

Named entity recognition with LSTMs. No tweaks and optimisation.

Train set quality:
precision: 76.68%; recall: 77.86%; F1: 77.26

Validation set quality:
precision: 47.95%; recall: 32.59%; F1: 38.80

Test set quality:
precision: 51.64%; recall: 36.42%; F1: 42.72

Week3

Find duplicate questions by their embedding in word2vec. Google embeddings : DCG@ 1: 0.320 | Hits@ 1: 0.320
DCG@ 5: 0.387 | Hits@ 5: 0.448
DCG@ 10: 0.404 | Hits@ 10: 0.500
DCG@ 100: 0.438 | Hits@ 100: 0.669
DCG@ 500: 0.460 | Hits@ 500: 0.844
DCG@1000: 0.477 | Hits@1000: 1.000

Starspace embeddings made with training set : DCG@ 1: 0.516 | Hits@ 1: 0.516
DCG@ 5: 0.614 | Hits@ 5: 0.699
DCG@ 10: 0.633 | Hits@ 10: 0.757
DCG@ 100: 0.664 | Hits@ 100: 0.906
DCG@ 500: 0.674 | Hits@ 500: 0.981
DCG@1000: 0.676 | Hits@1000: 1.000

Week4

Seq2Seq model, encoder-decoder to learn to addition and substraction. Model could be used for other tasks.

Dataset size =200000 (2 * originally given size) No tweaks.

Epoch: 10, MAE: 8.472910, Invalid numbers: 10155/200,000

Week5

TODO