Hello everyone! First of all, I would like to thank the Institute for Human Rights Law and the Law School of Huazhong University of Science and Technology for their invitation. I’m honored to have this opportunity to speak here. The topic I’ll talk about today is “Artificial Intelligence (AI) Helps Create Unified Standards for Judicial Trials”.

 

We have been studying judicial smart technologies for several years. We have learned from big data that the court is increasingly understaffed. In Zhejiang, a civil judge hears more than 300 cases a year and the number is still increasing. However, most of these cases are relatively simple and there is relatively clear trial logic for the judicial trial of them. So, this is where AI comes in. Based on this fact, we have created a big picture of judicial products. Now let’s have a look at the entire architecture from bottom to top. The bottom layer, which is something like the infrastructure part of the cloud, ensures data security and computing efficiency. Next comes the data layer, which is our judicial data. These data include not only laws and regulations, judgment documents, and electronic files, but also the judicial knowledge label system and the logic graph of judicial trials. These data with logic and knowledge labels are the professional knowledge of a judge. The combination of AI and industry knowledge allows AI to better empower the industry. Further up, we provide multi-modal algorithm capabilities including natural language understanding, image/visual algorithms, and speech recognition algorithms. These atomic algorithm capabilities, combined with our self-learning platform, can quickly form the basic engines for this industry, that is, the six basic engines we list here. Among them are the law and regulation engine, the judicial document analysis engine, the judicial document retrieval engine, the judicial human-computer dialogue engine, the judicial logic reasoning engine, and the judicial knowledge graph engine. These 6 engines well support the upper-layer application scenarios. Taking the judicial document analysis engine as an example, it can well support the processing of various documents, and the extraction of elements and events in judicial scenarios. In a general judicial scenario, we offer a variety of atomic capabilities, such as those to convict and sentence, report similar cases, generate dispute focuses, and so on. At the application layer, we’ve created the first-ever smart trial throughout the process. Now, it works for multiple causes of action. Our goal is to create unified standards for trials to improve trial efficiency and quality.

 

Next, I will mainly share details about our work on smart financial trials. As you can see in this figure, the whole trial process includes consultation, case filing, trial, and judgment. In each one of the above aspects, we provide smart capabilities. In terms of consultation, we provide self-service Q&A robots to improve user-guiding efficiency. In terms of case filing, the automatic verification of user information for completeness and relevance improves the efficiency of case filing. In terms of trial and judgment, I’d like to share some details about the specific work we’ve done in important smart links.

 

Specifically, automatic questioning includes flow-based questioning and fact-based questioning. Flow-based questioning is relatively easier to be understood and examples include the automatic questioning of the discipline and process of the entire trial. For fact-based questioning, the system will automatically generate some scripts available for the judge based on the real-time scripts of both the plaintiff and the defendant and then push them to the judge. Once chosen by the judge, the scripts will be automatically played using voice-based technologies. Of course, the judge may use his/her speech rather than choosing the algorithm-recommended ones, if the recommended results of the algorithm are not what the judge wants to say. Those scripts chosen by the judge and his/her scripts will be automatically recorded by our system in real-time to facilitate the iterative optimization of our next model. At present, in terms of fact-based questioning, the probability that a judge chooses our recommended speeches is about 85%. By using our recommended scripts, what is said by a judge is more unified and standard.

 

The real-time push of legal provisions. Based on the dialogue information of the trial, if the legal provision recommendation conditions are met, the system will push relevant legal provisions to the judge and litigants in real-time. At present, it works for many causes of action more than 90% of the time.

 

Automatic generation of dispute focuses. During the trial process, the information volume of the trial record is huge. For example, a trial record of 10,000-30,000 characters will be generated for a trial of 1 hour. We can not only better generate dispute focuses, but also highlight the sentences about them and finally generate a summary of the trial. As we found out through experiments, the accuracy of dispute focus generation is more than 90% and more than 80% of the trial summaries are simplified.

 

Automatic classification of evidence. In financial lending cases, there are as many as 45 types of evidence including IOUs, invoices, bank account statements, payment vouchers, transfer records, and chat records. If these different types of evidence are not well classified, it will take a lot of time for the judge to review them. The evidence submitted by litigants can be accurately classified by using technologies such as text recognition, image recognition, and natural language understanding. Now, our classification accuracy is more than 90%.

 

Automatic generation of the evidence chain. To allow the judge to review the evidence sequentially involved in the case, facts mentioned in the trial are identified, summed up, and extracted by using algorithm models and based on the real-time transcription of the trial. Then the facts are associated with relevant evidence. By taking advantage of visual interaction with the judge, as the trial progresses, the evidence chain generation helps the judge clarify his/her thinking, which helps sort out complex cases.

 

Risk point prediction. By using AI reasoning technologies and trial knowledge graphs created by legal experts, information that may have a decisive effect on the trial results and may be misjudged by algorithms is predicted to help the judge with his/her decision-making. The accuracy of risk point prediction is more than 90%.

 

The reasoning of judgment results. Based on the prediction results of algorithms and with the judge’s correction of the algorithm results, the prediction of judgment results is finally made. Now, the prediction accuracy is 99%. The generation of judgment documents. Our ultimate purpose is to generate judgment documents. Based on the judge’s trial logic graph and the real-time semantic understanding of the plaintiff, the defendant, the judge, evidence materials, and so on. during the trial process, the results are presented to the judge on time. The judge may intervene and correct the results at any time, so our judgment documents are also generated in real-time. After the judge modifies the information, the algorithm will timely update and re-generate the information based on the judge’s modification.

 

Smart trial. As we know, in a traditional trial with processes like the above ones, the judge or his/her assistants need to proofread the information and judge the results of each stage. However, because different judges may have different levels of professionalism and experience, their judgments of a case may vary greatly. With smart technologies and assistance throughout the whole process, judges’ scripts, trial logic, and the writing style of judgment documents become more unified and a new judge may better learn from experienced ones to improve his/her trial efficiency and quality. Today, smart trial throughout the process has been widely used in the entire Zhejiang Province and the tried cases are numbered in thousands. We’ve found out from the results that the improvement in trial efficiency is very obvious. In nearly 90% of the cases, the results can be pronounced at court and the trial time is reduced from the previous 40 days to 40-50 minutes. What’s more, AI-assisted trials haven’t seen their second trials up to date, which means the trial quality is greatly improved and the trial standards are more unified. In addition, we are also cooperating with the Higher People’s Court of Zhejiang to explore more trial models such as asynchronous trials and multi-case consecutive trials, which allow the cases to be tried by using litigants’fragmented time anytime and anywhere.

 

To sum up, we’ve explored the in-depth application of AI technologies in the entire trial process with financial loan cases and created the first-ever smart trial model characterized by “human-machine integration, smart trial, and rapid judgment”. By centering on building knowledge graphs and model algorithms, we’ve developed an AI trial system that simulates legal thinking, used AI to assist judges in trying cases, and improved trial efficiency and accuracy.

 

This brings me to the end of my speech today. Thank you for your attention.