Knowledge and tools in building GRANDJUR 1.1

This paper will describe selected aspects of the first phase of GRANDJUR, an ongoing project to assist Ohio prosecutors in preparing cases and selecting charges for presentation to a grand jury. In this project we have tried to identify, categorize, and make available a range of types of information for the user to choose among, compare and contrast, and make better evaluations. In addition, we sought to allow prosecutors to obtain and use the information in making decisions without their abdicating control or responsibility. A few points regarding our experience comparing different implementation tools will also be discussed. For several reasons, we felt it was important to give much responsibility for decision-making to users. Many lawyers become impatient with programs that leave no flexibility for the lawyer to direct program flow. In particular, they seem to wish to manipulate depth of detail. Typically a lawyer, will skim familiar or less important areas and direct attention to more problematic areas. Because of the importance of the decisions being made, we also feel it would be unethical to relieve a prosecutor of much responsibility for deciding appropriate charges. Moreover, having the user provide an additional check on knowledge quality is important. Errors in coding seem inevitable. The knowledge base continues to grow and evolve; the law changes and we add more expertise over time. Requiring the user to participate actively rather than passively helps detect errors. Working with prosecutors reduces the likelihood of liability for us as system producers. Shifting responsibility for decisions by making an expert system user's role more active seem wise to help minimize potential liability.1 Therefore, we designed the program as an aid, not a replacement for user decision-making. At each step the user is provided with information and choices, but he or she, not the program makes the substantive decisions. For example, the program will provide a statutory definition of a crime, and give the user a choice of further types of information available. Using the system to analyze a particular case, the prosecutor may search deeper in the knowledge base, obtaining additional assistance in understanding and satisfying elements. The user, however, controls how much assistance to obtain, as well as making the substantive decision that the evidence in the case file sufficiently matches that suggested by the program to satisfy the crime. This reinforces the prosecutor's ethical responsibility to avoid bringing charges without probable cause,2 and, we hope, encourages the user to evaluate the assistance provided more closely. The portion of the program described here provides two approaches. Using an informational approach, the user can query the system for information about various crimes, their elements, types of proof, possible defense responses, etc.3 Diagram 1 provides a schematic representation of sample types of knowledge and their interrelationships. We tried to identify a variety of knowledge types that prosecutors with varying degrees of expertise could use, and to make those types of information available as the user wishes. Thus each crime is broken down into elements, and the user may search further depths of the knowledge base for additional help. Attached to the various nodes in the network are also comments, expert suggestions, citations, and other information the user might wish. The other phase, allows the user to identify proof and elements satisfied in the case being analyzed, and then compare that proof to crimes in the knowledge base. For example, under the general heading of homicide are such crimes as aggravated murder, aggravated felony murder, murder, and voluntary manslaughter. Associated with each are related crimes (lesser-included, lesser non-included, and superior offenses). Thus, the user examining aggravated murder may see there are two forms with different elements, and that murder and voluntary manslaughter are, under appropriate circumstances, considered lesser included offenses. The program's main analysis is of crime elements. For example, a primary factor differentiating homicides is the mens rea element. The Ohio murder statute prohibits “purposely” causing the death of another.5 “Purposely” remains elusive and vague, so the user may obtain additional information, including a statutory definition,6 and expert comment. The statute's language alone could mislead without additional explanation. It seems that the second clause of the definition - that of intending the conduct regardless of the result - probably does not apply in a murder case. Therefore additional expert comment clarifies this point. Statutes, and appellate opinions, tend to leave intent elements rather abstract. To some extent, this cannot be overcome, as the law itself remains unclear until such time as an authority clarifies it. Inconsistency, incompleteness, and lack of clarity, as attributes of the domain, are reflected in the knowledge base as well. We included conflicting interpretations where they existed, with information for the user to make his or her own evaluation and choice between them.7 Vagueness of the law, and in some cases the open texture, however, are constrained to some extent in daily practice by practical application in specific cases of the law to facts. Lawyers dealing with an abstract concept such as intent, may translate it into prototype fact situations with which they associate the concepts, or interpretations into facts and lay concepts that can be shown and explained to a jury. For prosecutors, stereotypical cases, based on general fact patterns that satisfy requisite elements can form a working definition of the law.8. Therefore, we also included, where applicable, reference to typical factual or evidentiary patterns, listing evidence types and samples, and standard jury instructions. For example, one expert suggests most homicides, at least among those that are solved, fall into one of perhaps three or four categories - domestic disputes, bar fights, or neighborhood disagreements are most common. Killings associated with another crime, such as a robbery or attempted robbery also account for a significant number of homicides. For proof, many prosecutors, (and, to some degree, the police who derive evidence the prosecutor uses) tend to rely on certain types of evidence which vary according to the prototype patterns. Our primary expert tells prosecutors in training sessions there are basically four ways to prove crime elements - a confession, accomplice or associate testimony, eyewitness testimony, or other circumstantial evidence. With the “purposeful” element in murder, as in many instances, the types of proof within these categories tend to fall into more narrow patterns. Accomplice or associate testimony frequently relates statements by the defendant regarding his plans or feelings toward the victim. Or, for another example, in the category of other circumstantial evidence to show intent, the program user may examine such types of evidence as motive, type of weapon or instrument used, manner of infliction of the wound, preparation or planning, etc. Another step deeper, the user may see, from typical cases, that use of a deadly weapon such as a gun alone may suffice to support an inference of intent, as may a coroner's testimony that the death was caused by an especially repeated forceful blows to the head. Recognizing that mere evidence of use of a dangerous weapon will suffice to support an inference of intent to kill gives that requirement quite a different meaning.9 Linking together elements, one sees that certain evidence may support more than one element, and that patterns of groups of evidence supporting the various elements tend to fall together, forming pattern cases.10 To satisfy user interest in implications of his or her choices, predictions about potential defense tactics are available. Expert comment, for example, suggests that reliance on evidence of a particularly violent or forceful injury to show intent probably increases the likelihood the defense will claim self defense. As with other types of information, the source is identified so the user may evaluate the credibility and utility of the information. We are experimenting with two implementations of the programs described here. We built the first prototype using a Prolog language, but have now begun to experiment with a hypertext tool. Both are high level, powerful languages; one can do nearly whatever one wishes in either. Each implementation, for example, allows backward chaining, and modification of the knowledge base during a run. Each language allows tracing a user's search path, or the solutions selected, etc. The primary distinctions reduce to convenience in representation and control. The hypertext tool lends itself well to handling long strings, and allows converting strings to lists for processing, and back. It features nice screen handling, and the highlighting of selected text. Easy windowing and menu production provide for quick development of friendly user interaction. While Prolog also has features that allow manipulation of text, its strength is more in handling symbols. In the hypertext medium we found it easy to represent statutory or case material in text form; with the Prolog form, it seemed easier to manipulate the concepts as symbols, then translate them into text for delivery to the screen. The representation in the hypertext medium seemed driven toward using the actual statutory language as a basis for representation, while in Prolog, a representation of crimes by elements, and a network of concepts seemed more natural. This Prolog representation required additional coding to make a pleasant presentation, but allowed a quick and easy development of a prototype without fancy screen handling. The use of Prolog predicates to identify types of knowledge, with an argument to identify the