Machine reading comprehension, MRC for short, is one of the recent research hotspots in the field of natural language processing, and it is also a long-term goal of artificial intelligence in the process of processing and understanding human language. Thanks to the development of deep learning technology and large-scale labeled data sets, end-to-end neural networks have made great progress in solving reading comprehension tasks.
Human beings can easily read and understand their mother tongue, while machines are difficult to read and understand natural language. In order for the machine to read and understand the natural language, it is necessary to convert the natural language into a numerical form that it can be used to read, store and calculate. After several natural languages are converted into numerical values, the machine determines the relationship between them through a series of operations on these values, and then determines the position of an individual in the whole (complete set) according to the relationship between individuals in a complete set.
Machine reading comprehension is a technology that uses algorithms to make computers understand the semantics of articles and answer relevant questions. Because articles and questions are in the form of natural language, machine reading comprehension belongs to the category of natural language processing, and it is also one of the latest and hottest topics. In recent years, with the rapid development of machine learning, especially deep learning, the research of machine reading comprehension has made great progress and emerged in practical application.
With the development of machine reading comprehension technology, the task of reading comprehension is also upgrading. From the early “cloze form” to “single document reading comprehension” based on Wikipedia, such as the task of taking squiad designed by Stanford University as the data set; It is further upgraded to “multi document reading and understanding” based on Web data. The typical representative of this form is the task with Microsoft ms-marco and Baidu dureader as data sets.
At present, researchers have designed a variety of models for different reading comprehension tasks and achieved preliminary results. However, in the multi document reading comprehension task, because there are many documents related to the question, there are more ambiguities, which may eventually lead to the wrong answer in the reading comprehension model. Facing these problems, human thinking mode is usually: find multiple candidate answers first, select the final answer by comparing the contents of multiple candidate answers, so as to find the answer with the highest accuracy.
Most of the early reading comprehension models were based on retrieval technology, that is, search in the article according to the questions and find the relevant sentences as the answers. However, information retrieval mainly depends on keyword matching, and in many cases, the answers found by simply relying on the text matching of questions and article fragments are not related to the questions. With the development of deep learning, machine reading comprehension has entered the era of neural network. The progress of related technology has greatly improved the efficiency and quality of the model, and continuously improved the accuracy of machine reading and understanding model.
Although the machine reading comprehension model based on deep learning has different structures, it has gradually formed a stable framework after years of practice and exploration. The input of machine reading comprehension model is articles and questions. Therefore, first of all, these two parts should be digitally coded to become an information unit that can be processed by computer. In the process of coding, the model needs to retain the semantics of the original statements in the article. We call the coding module in the model the coding layer.
In the coding layer, due to the correlation between articles and problems, the model needs to establish the relationship between articles and problems. This can be solved by the attention mechanism in natural language processing. In this process, the reading comprehension model combines the semantics of articles and questions to further deepen the model’s understanding of them. We call this module the interaction layer.
Through the interaction layer, the model establishes the semantic relationship between the article and the question, and can predict the answer to the question. The module that completes the prediction function is called the output layer; Since there are many types of answers to machine reading comprehension tasks, the specific form of the output layer needs to be associated with the answer type of the task. This can be solved by natural language processing technology.
Natural language processing is an important technical cornerstone to realize the vision of machine and human-computer interaction. Machine reading comprehension can be regarded as one of the jewels in the field of natural language processing. Machine reading comprehension will make knowledge acquisition unrestricted by the human brain; But for the ultimate goal of “understanding and thinking” of machine reading comprehension, it is only the beginning of the long march.
Experts believe that end-to-end deep neural network can better find some potential features in natural language processing, so as to improve the accuracy of machine reading comprehension. The deeper induction and summary of natural language, knowledge citation, reasoning attribution, knowledge map and transfer learning will be the future development direction of machine reading comprehension.
As an important branch of artificial intelligence technology, machine reading comprehension will be more and more applied to various industries. As Professor Zhou Haizhong, an internationally renowned scholar, once predicted: “with the progress of science and technology, the era of artificial intelligence is coming; At that time, artificial intelligence technology will be widely used in various disciplines and will produce unexpected results. “