Using Rules To Optimize Neural Network

In recent years,deep learning has developed rapidly and has made remarkable achievements in many areas.Practice shows that deep convolutional neural networks are very useful in many artificial intelligence applications,especially in the areas of recommendation systems,image recognition and segmentation,and speech recognition,and many applications have proven that these methods are very effective.On the visual side,deep convolutional neural networks have succeeded in identifying human faces,objects,and traffic signs in addition to powering the vision of robots and self-driving cars.At the same time,more and more people are trying to apply deep convolutional neural networks in the field of natural language processing.Sentiment analysis is a basic research in natural language processing.Sentiment analysis is a method that can determine whether a piece of text is positive or negative or neutral.It is also commonly referred to as opinion mining to obtain the speaker's opinion or attitude.A common use case of this technology is to find people's views on specific topics and provide public opinion mining for companies,governments,and other departments.Suppose you want to know that people on Weibo think of a star acting.You can use the message data on Weibo to answer this question.You can even find out why people think the star is good or bad by extracting exactly what keywords people like or do not like.For example,if "appearance" appears as a common topic for negative comments,then we immediately understand why the audience is not happy.Usually we can use the word vector to represent the words numerically to perform semantic emotion classification tasks.Word vectors are simpler and more natural than one-hot codes.Its greatest advantage comes from its sparseness,which is a sparse coding of words,a low dimensional vector space formed by an implicit layer through iterative learning.In this low-dimensional space,word vectors can get some important grammatical features.Synonyms and synonyms have a closer Euclidean distance.Convolutional neural networks use multiple convolution filter layers to extract higher abstract features.The convolutional neural network was first converted from computer vision technology.Under normal circumstances,we can consider the convolutional layer as a kind of filter.Convolutional neural networks can provide artificial intelligence solutions for a variety of tasks.Similar to image classification,convolutional neural networks also work well in the field of natural language processing.Simple and efficient,the convolutional neural network is the best choice for data modeling.Although the traditional neural network provides a powerful parameter learning mechanism for the training of big data models,it does not mean that the neural network handles any problems.The neural network model is still flawed.It is inflexible and unexplainable when it is used.It has always been a major problem that plagues neural networks and deep learning.Sometimes our model is even counter-intuitive.Conversely,these flaws also make it difficult to direct the model directly with human intent.For a long time,due to the nature of its black box,convolutional neural networks often require repeated duplication and trials.This is a painful process.When an error occurs,we cannot precisely locate the position and the problem of the parameter.Later in the article,we will focus on textual emotions as an example to discuss how to apply convolutional neural networks to areas where it is not good at.Combining deep learning and structured rules is very beneficial for interpreting neural networks.We also need to combine neural networks and structured logic rules to increase the flexibility of the model.We use the "distillation method" to transform structured information into parameters or weights of a neural network.Using this framework,we can strengthen neural networks simply by adding some simple structural rules.At the end of the article,we took text sentiment analysis as an example,and added some grammar rules to the model to successfully achieve the goal of optimizing the accuracy of the convolutional neural network and accelerating the training speed.In the practice of competition or industry,some intuitions and insights can be quickly integrated into the final results to improve model indicators.