[author]Wei Bin, ZHENG

[content]

Artificial Intelligence Assisting to Finding Facts of Criminal Cases

WEI Bin, ZHENG Zhifeng

Institute of Law for Artificial Intelligence, Southwest University of Political Science and Law, Chongqing 410020, China

Abstract：The emerging artificial intelligence (AI) is being applied to assist in finding facts of criminal cases for forensic investigation, initiating the forensic practice of AI into the expert’s system of criminal cases. From the early classical logic to the non-classical one, the foundation has been being enriched for AI’s logic to dig out the facts of cases. The ever-evolving Bayesian model is bringing the fact-finding of cases into quantitative determination from qualitative deduction. While computational argumentation model helps to clarify the structure of evidential argument for the facts of criminal cases, the revolution of big data, algorithms and block-chain correlation pulls the evidence approaching to the case facts into a more accurate, finer and more scientific course. AI is continuously overcoming its technical defects, getting closer to the thinking and umpirage by judges and juries to find the criminal facts, yet its role must be oriented at the assistant status for the judicial judgment to improve the accuracy of finding facts from cases and reduce the occurrence of wrong cases.

Key words: facts of criminal cases; artificial intelligence (AI); Bayesian model; computational argumentation model

The determination of criminal case facts is a difficult problem in criminal justice theory and practice. Unclear or even wrong determination of case facts is the main cause of criminal miscarriages of justice. In recent years, the cases of Huugjiltu, Nian Bin, Zhang's uncle and nephew, and Zhao Zuohai in my country were all caused by wrong determination of case facts. The determination of case facts is not only a theoretical problem in forensic science, but also a practical problem in the field of public security technology.

Traditional determination of criminal case facts mostly relies on deduction, induction, and abductive reasoning, but these methods alone are not enough in difficult criminal cases. In the actual identification process, due to the complexity of evidence and other elements in major criminal cases, the elements of the crime and important circumstances are often difficult to clarify, and it is still difficult to reconstruct the truth of the case facts. Therefore, it is urgent to open up new identification methods. With the breakthrough of artificial intelligence technology, its application in criminal case analysis has gradually become an important research direction in the field of forensic science in countries around the world. The research on case fact determination based on artificial intelligence technology has important theoretical and application value.

The criminal case expert system opened the prelude to the application of artificial intelligence in the determination of criminal case facts. With the continuous development of artificial intelligence theory and technology, the artificial intelligence method of criminal case fact determination has been continuously optimized, and the evidence pointing to the facts of criminal cases has become more accurate and clear. From the early classical logic to the non-classical logic, the artificial intelligence logical basis for determining case facts has been continuously enriched; the evolution of the Bayesian model has enabled the determination of case facts to move from qualitative research to quantitative research, which provides a precise reference for judges or jurors; the computable argumentation model helps to clarify the structure of evidence argumentation in criminal case facts, making the analysis and evaluation of evidence argumentation possible.

In recent years, the changes in big data, algorithms and blockchain technology have made breakthrough progress in the determination of case facts, and artificial intelligence-assisted systems that integrate multiple technologies have begun to appear, which marks that the application of artificial intelligence methods has begun to move towards the intelligent direction of truly meeting the needs of legal practice.

1Application of traditional computer software in criminal case analysis

1.1Evidence management system

Early artificial intelligence technology mainly focused on data storage, organization and analysis. Criminal evidence management system is one of the earliest computer software for criminal case analysis.

The "General Evidence Management System (GEMS)" developed by Eden Technology Company of Australia can be used to identify, store, mark, retrieve and display evidence and information of criminal cases (especially fraud and corruption crimes). It also has functions such as electronic imaging, optical character retrieval and evidence vector, which can help investigators improve the accuracy of case fact investigation.

Later, the evidence management system gradually developed the function of case fact determination. The evidence management system developed by the Institute of Evidence Science of China University of Political Science and Law consists of client computers, application servers and databases. The software function modules included in the application server are: case information input module, case information query module, statistics module and judgment generation module. The entire management system can initially achieve: input basic case information - select case nature - automatically generate judgment.

In this process, the facts of the case must be determined before generating the judgment. However, this type of management system is not specifically used to determine the facts of the case, and does not clearly show the logical relationship between evidence and case fact claims.

1.2Criminal Case Expert Assistance System

With the rise of artificial intelligence expert systems in the 1980s, criminal case expert assistance systems also came into being. Professor Zhao Tingguang of my country presided over the development of a practical criminal law expert system in 1993, which includes a consulting and retrieval system, an auxiliary qualitative system, and an auxiliary sentencing system. The consulting and retrieval system can retrieve the current criminal law and judicial interpretations, the auxiliary qualitative system can identify the form of crime, the number of crimes, and whether it is a joint crime, and the most core auxiliary sentencing system can give sentencing results for a single crime and multiple crimes.

The sentencing of this system is a fuzzy decision. First, the purpose of quantitative analysis is achieved through two steps: the first step is to correctly identify the weight "level" of each sentencing circumstance; the second step is to correctly evaluate the degree of leniency or severity of the sentencing circumstance. Finally, through these two quantitative analyses, the score of the circumstance is determined = weight level × severity level.

Foreign legal expert systems are divided into rule-based and case-based expert systems, and continental law and Anglo-American law countries have their own emphases. The HYPO system developed by Ashley is the first legal expert system based on legal precedents. It uses a fixed set of case facts as premises to construct a tripartite argument. If the current case is similar enough to the precedent in the database, then the principle of precedent will support the same conclusion. Although Ashley only discussed the application of HYPO in trade secret law, case-based expert systems have been used to assist in criminal case analysis.

Legal expert systems have the characteristics of strong applicability, strong reliability, and the ability to explain the deduction process. However, the law is an open system, it is not closed and unchanging. Legal concepts and legal reasoning are both revocable, and legal expert systems have weak autonomous updating capabilities and poor effectiveness, making it difficult to deal with emerging legal issues. In addition, legal expert systems also face the problem of "knowledge reception bottleneck".

2Artificial Intelligence Methods Based on Bayesian Model

2.1Application of Bayesian Model to Identify Case Facts

At the beginning of this century, driven by the cross-development of artificial intelligence and forensic science, Bayesian model began to be applied to the identification of case facts. Fink lestein and Fairley published "A Bayesian Method for Identifying Evidence" in the Harvard Law Review, which was the earliest literature to apply Bayesian theory to the study of evidence probability.

British scientist Fenton published an article in Nature magazine to demonstrate the scientific nature of Bayesian model application in the legal field. He pointed out that people's beliefs in identifying case facts will change with the continuous evidence given in court. Changes in beliefs about a piece of evidence will affect the beliefs about the case facts it supports or opposes, and such changes are exactly what Bayesian model is good at. The application of Bayesian model has attracted the attention of courts in European and American countries. For example, the British Court of Appeal ruled that the field of DNA evidence and "fields with a solid statistical foundation" allow the use of Bayesian model to evaluate evidence.

The principle of artificial intelligence based on the Bayesian model conforms to the law of people's understanding of case facts. This is because most of the time it is difficult for people to restore the absolute truth, but rather the probability of probability.

Obviously, the evidence in the fact finding of criminal cases is not expressed with absolute certainty, but a probabilistic and non-absolute expression. Anderson et al. believe that the conclusion based on evidence must be probabilistic in nature, because: 1) evidence is always incomplete; 2) evidence is generally non-conclusive; 3) the evidence people have is usually vague; 4) the body of evidence is often inconsistent, and one piece of evidence may support one proposition, while another piece of evidence supports other propositions; 5) the evidence comes from people's imperfect credibility level. The probability of evidence easily makes the fact finding of the case uncertain, and the Bayesian model provides a scientific tool for the judge's precise discretion.

2.2Application of the likelihood ratio method

The Bayesian model has evolved into a verification method based on the likelihood ratio, which has been applied in the fields of physical and chemical testing of physical evidence. Scholars such as Zadora and Martyna proposed a method based on the likelihood ratio method to evaluate the results of physical and chemical testing of physical evidence. They believe that when there are two opposing hypotheses of case facts, it is necessary to use the likelihood ratio method to evaluate evidence. The specific method applicable is to use the empirical cross entropy method to illustrate how to verify the likelihood ratio method.

Assume that the prosecution's case fact claim is Hp, and the defense's opposing fact claim is Hd. The conditional probabilities P(E|Hp) and P(E|Hd) represent the likelihood of Hp and Hd, respectively. They refer to the probability of evidence E occurring when Hp and Hd occur. The likelihood ratio LR refers to the ratio of P(E|Hp) to P(E|Hd). Generally speaking, if P(E|Hp)>P(E|Hd), then the prosecution's claim Hp is supported by evidence E, and vice versa. According to Bayes' theorem, the likelihood ratio can also be expressed as the following formula:

The likelihood ratio can reflect the degree to which the evidence supports the factual claim (as shown in Table 1), which provides an intuitive reference for judges or jurors to determine the facts of the case.

2.3Application of Bayesian Network

Bayesian network is widely used in deep learning algorithms. It is a probabilistic graphical model. Its essence is uncertain reasoning. It has been widely used in many fields such as prediction, anomaly detection, diagnosis and treatment, and automated testing. Bayesian network is built on a directed chain consisting of points and links. This directed chain cannot form a closed loop, so the Bayesian model is an acyclic graph. Bayesian network has been used to determine the facts of the case. The network structure can indicate the correlation or independence between two variables in the case. The graphical structure usually includes one or more hypothesis nodes (such as the suspect is guilty), evidence nodes (such as the discovery of fingerprints that match the suspect's fingerprints), and some intermediate nodes (such as the suspect at the crime scene).

When there is a certain probabilistic correlation between two variables, the two nodes are connected by arrows in the graphical structure. This arrow-connected relationship is a causal correlation. In the Bayesian network, each node has two values, true and false. Each value has a conditional probability table, which contains all the prior probabilities that can be obtained. The combination of these conditional probability tables constitutes a joint probability distribution.

When the prior probability of a node changes, the posterior probability of all nodes pointed to by the node also changes accordingly. The characteristic of the Bayesian network model is that it can intuitively show the causal relationship between evidence and factual claims in the facts of the case, accurately evaluate the probability of the occurrence of events, and serve as a scientific reference for determining whether the facts of the case are established.

However, the Bayesian network still faces three difficulties in subjective probability theory:

First, when multiple independent evidences simultaneously support a fact to be proved, according to the multiplication formula of subjective probability calculation, the probability of the conclusion they jointly support will be lower than the probability supported by any independent evidence, which is obviously contrary to intuition;

Second, subjective probability theory cannot deal with the impact of opposing evidence or counter-argument on the probability of the target evidence. Since the strength of evidence will be affected by counter-evidence, if a piece of evidence can win in the competition with counter-evidence, then the evidence can be accepted. In the process of determining the facts of the case, the trial party needs to judge how one piece of evidence affects other evidence, which is something that the subjective probability theory cannot describe;

Finally, people's credibility of a piece of evidence or argument does not necessarily require the additivity principle to be satisfied. This is because in the process of determining the facts of the case, even if the probability of a proposition Q is determined to be P, it cannot be believed that the probability of non-Q is 1-P, because there may still be situations where it is not certain.

This shows that the subjective probability theory based on the additivity principle cannot well describe this feature of evidence evaluation. These defects mean that the Bayesian model is an imperfect artificial intelligence method, which needs to cooperate with other methods to make up for these defects.

3Artificial Intelligence Methods Based on Non-Classical Logic

3.1From Classical Logic to Non-Classical Logic

The development of modern logic after Gödel has provided new tools for the identification of case facts. In the judicial decision stage, legal reasoning makes clear syllogistic reasoning a necessary condition for legal justification, and classical logic plays an important role in testing the validity of legal reasoning.

The application of classical logic can process the internal structure of more complex legal language through predicate expressions, which is difficult for traditional non-classical logic to do, but the monotonicity of classical logic makes it unable to handle the non-monotonic properties of legal reasoning. The essence of non-monotonic reasoning is that the expansion of the premise set can lead to changes in the conclusion, and this expansion is allowed by the open structure of the law. In the process of determining the facts of the case, with the continuous revision of legal rules and the addition of new evidence, the legal facts of the case will also change, thereby changing the original conclusion or even refutation.

In order to characterize the non-monotonic properties of case fact determination, logicians have developed non-classical logic branches represented by default logic and defeasible logic. For example, reason based logic developed by Hage is a non-monotonic logic that can characterize "reasoning based on default" and is suitable for expressing the non-monotonic properties of legal reasoning.

Applied to case fact determination, reasons can represent different types of legal facts. Facts generated by reasons are called conclusions or factual claims, and legal rules are modeled as rules in reason logic. Reason logic proves case facts through qualitative comparative reasoning. However, traditional non-classical logic cannot express people's changes in the credibility or acceptability of a certain evidence or case fact claim in legal arguments and dialogues, which makes traditional non-classical logic not completely applicable to the determination of case facts.

3.2Application of computable argumentation models

In order to overcome the shortcomings of traditional non-classical logic and Bayesian models, applications based on computable argumentation models have begun to emerge in forensic science. Computable argumentation model is a new branch of non-classical logic that has rapidly emerged in the field of artificial intelligence in recent years. It can overcome the problem that multiple pieces of evidence support a conclusion, which reduces the credibility of the conclusion. It can also well express the weakening or defeating effect of counter-evidence on the target evidence, and it does not have to comply with the additivity principle.

In addition, its characteristic is that it can avoid the "knowledge reception bottleneck" problem commonly faced by artificial intelligence, because it does not need to receive knowledge applied to solve problems, but directly connects with a relatively complete source of background legal knowledge archives. It makes the problem intuitive by constructing arguments, helping system users to clarify and solve problems. It also supports revocable arguments in case fact determination and court dialogue under different opinions.

The core issue of case fact determination research is how to analyze and evaluate the arguments that evidence supports case fact claims, including whether the fact claims are proven to meet the standard of excluding reasonable doubt (BRD).

This requires solving two problems: first, the analysis of case facts needs to clarify and clarify the way and structure of evidence supporting fact claims (conclusions); second, the evaluation of case facts needs to evaluate the degree of support of evidence for case fact claims, etc.

The European natural science project "Argumentation Service Platform with Integrated Components", led by French computer scientist Amgoud, is a knowledge-based service platform that has been applied to the study of fact finding in criminal cases. Its core part is a structured argumentation framework. The characteristics of this framework are that it can recursively define argumentation concepts, thereby expressing the internal structure of arguments, clarify the attack types and defeat relationships between arguments, and rely on the preferences of premises and inference rules to compare and evaluate arguments.

Prakken improved this framework and obtained a new ASPCI+. This improved model added questioning evidence as a new attack type on the basis of the original one, and applied it to study a real case (Popov v. Hayashi).

3.3Visualization software for computable argumentation models

For the methods of computable argumentation models, artificial intelligence scientists not only pay attention to its theoretical model, but also pay more attention to the visualization software generated based on the theoretical model, which has promoted the development of application systems for computable argumentation models. This type of application system can intuitively express the characteristics of abstract arguments. It is a type of auxiliary argumentation software generated by visual programming technology, also known as a structured argumentation system. For example, the Aver system developed by Braak is a software specifically used to identify the facts of criminal cases. The algorithm model of the software combines the argument and story models to obtain a composite model, in which the argument model is based on the evidence base in the case, while the story model is expressed by the abductive network of events.

The Aver system translates this composite model into an ontology that can express and define evidence arguments and stories. This ontology distinguishes different types of relationships, among which the explanatory relationship contains abductive defeasible inferences for connecting events to stories, and the indicative relationship contains evidential defeasible rules for connecting evidence to events. The Aver system uses this ontology to create a visualization platform for the composite model, and then uses a web interface to help system users analyze and compare arguments and stories in the case to achieve the purpose of identifying the facts of the case.

4Technological changes in big data, algorithms and blockchain

Currently, the changes in big data, algorithms and blockchain technology have made the analysis, storage and evaluation of evidence more intelligent, greatly affecting the progress of research on the identification of criminal cases. Big data far exceeds traditional databases in terms of data acquisition, storage, management and application analysis. It uses distributed mining and storage of data through cloud computing, and usually works together with image and voice recognition algorithms. At present, big data has greatly promoted the transformation of my country's criminal case investigation model. The "Skynet System" used by my country's public security organs uses a video image recognition system to identify people's facial information, and then matches and verifies it with the information in the database to lock in criminal suspects.

In addition, the public security organs also use DNA technology to screen criminal suspects through paternal kinship, which requires comparison with the data in the established big database, and then gradually narrows the scope of screening and improves the efficiency of locking in criminal suspects. The "Baiyin Murder Case" cracked by the public security organs is a typical case. The role of blockchain technology is to preserve evidence for the identification of case facts. It is different from ordinary electronic evidence storage. Instead, it uses distributed evidence storage methods to ensure the integrity and authenticity of evidence through keys, while ensuring that the evidence will not be tampered with. Blockchain technology makes evidence extraction efficient and orderly. Evidence can be identified and extracted through encryption algorithms and other technologies. Evidence can form an evidence chain through pre-designed smart contracts, thereby achieving standardization of evidence storage and extraction.

With the cross-application of big data, the Internet and deep learning algorithms, an artificial intelligence-assisted system integrating multiple technologies has been born. In 2017, iFLYTEK cooperated with the Shanghai Court, the Procuratorate and the Public Security Bureau to develop a set of "Criminal Case Intelligent Assistance System" (also known as Project 206). This system is different from the expert system based on the knowledge base. It uses machine learning algorithms and is continuously trained with the support of a large database, so as to continuously update and improve.

The large database of Project 206 includes case libraries, document libraries, legal and regulatory and judicial interpretation libraries, case handling business document libraries, evidence standard libraries and electronic file libraries. The number of cases in the case library determines whether similar cases can be effectively pushed.

The functions included in Project 206 include evidence standard guidance, evidence rule guidance, arrest condition guidance, single evidence verification, social danger assessment, illegal speech evidence exclusion, similar case push, sentencing reference and document generation. The system can also display the case entry and system function usage in real time.

The feature of Project 206 is that it can use deep neural network model algorithms and image recognition technology to learn the file materials in the database, and preliminarily realize the intelligent recognition, positioning and information extraction of various evidence such as printed text, handwritten text, signatures, handprints and tables. As the system is widely promoted and applied, judges will continue to generate data for learning during use, and there will be more and more materials for machine learning, which will make the system more mature and applicable.

Undoubtedly, Project 206 is a concentrated embodiment of my country's artificial intelligence technology in the judicial field in recent years. It guides the entry of evidence through evidence guidance, applies appropriate evidence standards, and helps to find defects in the chain of evidence that proves the facts of the case, thereby improving the efficiency and accuracy of case fact determination and reducing the occurrence of wrong cases.

5Dilemma and future of the application of artificial intelligence methods

The application of artificial intelligence methods to the determination of case facts faces many difficulties:

First, how to convert legal language into logical language and evidence into logical propositions is still a bottleneck that is difficult to break through in natural language translation. How to clarify the causal relationship between these logical propositions also requires more research.

Second, the quantification of the proof standard is still a difficult problem. Although Bayesian theory provides a quantitative solution, the results of determining case facts from a quantitative perspective are still controversial. It is necessary to find a balance between quantitative and qualitative research.

Third, it is still difficult to find a suitable solution for the evaluation of the credibility of evidence. This involves the evaluation of the objective authenticity of evidence, and the evaluation of the degree of support for factual claims by evidence also faces this problem.

Fourth, in my country, a complete chain of evidence is an additional standard outside the criminal proof standard. How to define a complete chain of evidence from the perspective of artificial intelligence is an important open question.

Fifth, the legitimacy of blockchain evidence storage and the qualifications of third-party evidence storage platforms are questioned, and whether the stored evidence has judicial effect is still undecided.

Sixth, the relevant constraints and characteristics of case fact determination in criminal litigation practice in continental legal system countries, Anglo-American legal system countries and my country are different. How to make artificial intelligence methods truly applicable to the theory and practice of criminal case fact determination in my country is a more urgent issue.

Although these difficulties are difficult to solve in a short period of time, the application of artificial intelligence in criminal case fact determination will still usher in more development opportunities in the future. It can be foreseen that more artificial intelligence methods will move from theory to application in the future, intelligent application software will be implemented and promoted, and a number of typical cases assisted by artificial intelligence systems will also be produced.

In the future, the construction of "smart courts" in my country should shift from the intelligent construction of court procedures to the intelligent construction of case fact determination. Relevant research and application should also be closely combined with my country's criminal justice reality, focusing on satisfying the "three characteristics" of evidence, that is, the objectivity and authenticity of evidence, the relevance of evidence to the facts of the case, and the legitimacy of evidence acquisition. Further, we should explore the free conviction model of judges or jurors in determining case facts, reduce the mistakes and errors that may occur in subjective judgment, and on this basis, gradually respond to the many difficulties faced by the application of artificial intelligence methods.

It should also be noted that in the future, artificial intelligence methods need to focus on cross-cutting research or be closely integrated with other theories. For example, the integration of Bayesian networks and computable argumentation models has become a new trend in current artificial intelligence and legal research, and the combination of artificial intelligence methods and narrative theory is also an important research direction.

No matter how far artificial intelligence develops, it should be clear that the application of artificial intelligence cannot restore the facts of criminal cases 100%, and it cannot replace the determination of judges and jurors. As an intelligent auxiliary tool, its role is to provide scientific decision-making suggestions for judges and jurors, help improve the efficiency and accuracy of case fact determination, thereby reducing the arbitrariness of judicial decisions and avoiding the occurrence of false and wrongful convictions.