项目作者： LaurentVeyssier
项目描述 ：
Use end-to-end memory networks architecture for Question & Answering NLP system
高级语言： Jupyter Notebook
项目地址: git://github.com/LaurentVeyssier/Question-Answering-with-End-to-End-Memory-Network.git
Question-Answering-with-End-to-End-Memory-Network
Use end-to-end memory networks architecture for Question & Answering NLP system
Project objective
This project uses a end-to-end memory network architecture to build a chatbot model able to answer simple questions on a text corpus (‘story’). Learning capabilities allow logical deduction on memorized corpus. The model is written in keras.
Dataset
The project uses the bAbI dataset from Facebook Research. The dataset is available here. bAbI dataset is composed of several sets to support 20 tasks for testing text understanding and reasoning as part of the bAbI project. The aim is that each task tests a unique aspect of text and reasoning, and hence test different capabilities of learning models. The datasets are in english.
- Each task tests a unique aspect of learning capabilities: Dialog in the restaurant domain, children’s book missing word test, Movie dialog, questions-detailed answers dataset, path or localization problems….
- For our task, there are 10,000 samples for training, and 1,000 for testing. A sample item in the set is composed of a story (several short sentences), a question and the answer to the question for training purpose. In our case, the answers are simply Yes / No answers. A sample of the dataset is show below.
Memory Networks Architecture
This deep learning neural network architecture was published in 2015 and you can refer to the original paper for its detailed description. The architecture shares some early principles with attention model.
The model takes two different inputs: A story (represented as a list of sentences all required to answer the question) and a question. The model must take the entire story context into consideration to answer the query. The use of end-to-end memory network becomes handy in this use-case.
The model performs calculation in order to combine these inputs and predict the answer. We can split the network into several functions:
- Input Encoder m: This section transforms all input sentences into vectors of given embedding size and length of sentence_max_length. size: batch x sentence_max_length x embedding_size
- Input Encoder c: This section transforms all input sentences into vectors of embedding size query_max_length and length of sentence_max_length. size: batch x sentence_max_length x query_max_length.
- Question Encoder u: This section vectorizes the input question with given embedding size and query_max_length. size: batch x query_max_length x embedding_size
Calculation steps:
- Calculation input weights p: dot product between m and u followed by a softmax activation function generating weights p (batch x sentence_max_length x query_max_length)
- Response vector O from the addition of p and input Encoder c => (batch x sentence_max_length x query_max_length)
- Concatenation of Response vector O with Question Encoder u resulting into answer object of shape (samples x query_max_length x [sentence_max_length + embedding_size])
- The answer object is then passed through an LSTM layer (dimension reduction) followed by a dense layer resulting into output vector of the size of the vocabulary (output shape = samples x vocabulary_size). Finally, a sigmoid generates a probability distribution over the vocabulary size. In this project, due to the training objectives, the probability arbitrates over 2 words of the vocabulary: ‘yes’ and ‘no’.
Memory Networks model representation:
All parameters (embeddings, weight matrix to determine predicted answer) are learned during training.
Model limitation: The whole vocabulary must be known during training phase. Only words which are part of the corpus (training and testing) can be used during inference.
Results
The model is trained very quickly over 120 epochs using RMSprop and lr = 0.01. Other hyperparameters: Embedding size of 128, batch size of 256. Accuracy on unseen test data reaches over 97%.
Excellent prediction on complex story.
Story:
