Text-summarization Model

As a beginner of artificial intelligence, I explored NLP field in 2020 winter holidy. This is the presentation of my project, which is used to automatically generate a text summary . Since I am new to text-summarization, there may be some mistakes in the project, so I appreciate any kind of advice and feedback. The project is based on one of the popular neural network---- Seq2Seq Model .

(1)Introdoction:

In deep learning, the seq2eq model is actually a encoder-decoder model , which is used to transform a sequence to another by the recurrent neural network (RNN). Most public projects on Github are usually applied to the language translation models, however, seq2seq model is able to carry great value to text summarization. Based on the paper Sequence to Sequence Learning with Neural Networks and some current translation models, the model is developed.
The model can be divided into three parts: Encoder , Decoder , Attention Mechanism

• Encoer : Comprised by a series of RNN (in my model , classical RNN model is replaced by GRU, which is an advanced RNN model), the main work of the encoder part is to encode the input source into a sigle vector (the vector is called the context vector in the paper). Firstly, the input sentence would be embedded and then input into the encoder layers. In the stage of the encoder layers, for each step t, the input to the encoder RNN is both the embedding of the current word e(x_t), as well as the hidden state from the previous time-step h(t-1)_ , and the encoder RNN outputs a new hidden state h_t. It can be presented as:

• Decoder : Comprised by a series of RNN (in my model , classical RNN model is replaced by GRU, which is an advanced RNN model), the main work of the decoder part is to decode the context vector into a target sequence (the summarization of the input). At each time-step, the input to the decoder RNN is the embedding of current word, d(y_t), as well as the hidden state from the previous time-step, s(t-1) , where the initial decoder hidden state, s0, is the context vector, s0 = z = h_T, i.e. the initial decoder hidden state is the final encoder hidden state. Thus, similar to the encoder, we can represent the decoder as:

• Attention : Attention mechanism is a method to improve the performance of classical encoder-decoder framework, which usually has trouble with gradient disappearence and gradient blast. Instead of only getting information from the context vector, the attention mechanism allow the decoder part use all of token information in the final hidden layer by the diffrent weights. If you want view the detail of this technique , please take a look at the code of Attention class in model2.trainclass.py or a famous paper Neural Machine Translation by Jointly Learning to Align and Translate .
  We can visualize the method as the picture below:

• Note : Apart from the parts above , there are also some important and common motheds to process data , such as using torchtext to automatically generate dataset we need , and loading english dictionaty from spacy library to embede the words of the input. If you are confused about the codes , please read the explanation around the codes.

(2)Model Information :

Prerequisites:

• Python == 3.6
• numpy == 1.18.0
• pandas == 0.24.2
• torch == 1.0.0
• torchtext == 0.4.0
• spacy == 2.1.8
• GPU (optional)

Warning: I suggest you to create a new virtual environment by anaconda for this project . Since if your packages can not match with each other , some strange and tough problems will appears. (I was stucked here for a time 😭 .)

Data:

• My Data in kaggle

How to run:

The project can be run on both local pycharm envirenment and online notebook (Pycharm is recommended). Here is some main steps:

• use dataPreparention.py file to load and clean data
• change the information in Model2_Setting.py file (especially the path of files and the parameters of neural network)
• run Model2_GetModel.py file to make sure your network works
• run Model2_Train.py to train the model
• run Model2_prediction.py

My Result:

My result

(3)My Process:

The basic information of my model:

The process of training is below:

Usr the result we get to make some prediction :

Actually, I have to admit that my model does not perform well, and it can be seen from the prediction below . I have spent a lot of time to improve it , such as changing the parameters (learning rate, dropout rate , epoch times, clip value so on ), and checking whether my neural network is right or not . However, I fail. I observed that no matter what parameters the model has, the value of loss funtion always can not decrease after the first some epochs (about 10-20 epochs).

Since I have upload all of my codes and data of the project , it is not difficult for deep-leaners to repeat the work . So if you can make an improvement about the model , you are welcome to pull your codes to the repository. And I will also study and research in machine learning in my spare time in university , so I will also make an effort to imporve the model . Updating soon ...