Deep-learning-accentuated-natural-language-processing

This project is focused on exploring different methods to classify our text information.

Doc2vec (treating each sentence as a document) visual on tensorboard , added colors for our five different classes

Word2vec visual on tensorboard

Motivation

Vast amount of textual information available on the web, we try to classify the the sentences specific to food from a raw html as belonging to one of the five different classes Preparation, Ingredients, Nutrition, Recipe, Title.

Approach

• Set up different architectures and then monitor their metrics to identify the best ones
• Build an end to end system on google cloud
• Use GPUs to run models over several epochs
• Build a front end flask-app and run the model on unseen data
• Print results on webage, host the app on the host IP Address

Deep-Glove-embeddings-conv-autoencoders.ipynb

• This notebook contains the model with the best results
• Shows implementing glove on using deep learning
• Also tries an unsupervised approach using Autoencoders
• Refer notebook for code and explanations

Glove-cnn-lstm-gpu.ipynb

• The best model ran on GPUs , the notebook is modified run locally over 1 epoch
• On the cloud the model was run over 60 epochs , screenshots of which are attached as png files
• Color coded confusion matrix and also a function to identify the misclassified sentences
• Model checkpointing to identify the weights of the model with highest accuracy
• keras callbacks to get plots of accuracy and histograms , screenshots are attached
• Refer notebook for code and explanation

Padded-CNN-LSTM.ipynb

• Running CNN - LSTM network on the padded sentences/sequences
• Runnnig LSTM on the padded sentences/sequences
• Using the keras default embedding layer
• Refer notebook for code and explanation

Word2vec-Doc2vec-FFN-Tensorboard.ipynb

• Word2vec , doc2vec models using gensim , fed into tensors for a tensorboard visualisation
• The doc2vec vectors are fed into a deep feed forward neural network
• Keras preprocessing is carried out as always
• Refer notebook for code and explanation

Flask-spyder.py

• The front end for our model
• Generic to allrecipes.com , any recipe url from the website can be passed
• the page is then scrapped using beautiful soup to obtain the textual information
• The sentences are then passed into our established model
• Predictions are made for these sentences(unseen data)
• The final webpage displays the title , image and the predictions for that particular url recipe page
• Refer python script for code and explanation

Results

• The results produced by the glove-cnn-keras model were very encouraging. An accuracy of close to 97 percent was obtained. Pre-trained wiki glove embeddings played an important role in getting up the accuracy. The fucntionality in keras to specify your own embedding layer was helpful.

Future Work

• Extend this problem to a multi-class , multi-label problem. Set-up the front-end on the Google Cloud to extend the pipeline that has already been established.