judicial/trial engine with knowledge graph
Judicial/trial engine is used to make a judgement during law/legal case. Instead of replacing people judger,
it is working together with people judger and make judger's job more efficient, less time consuming and easier.
Judicial/trail engine achieve this purpose by incorporating knowledge from team of experienced judgers who have many years of judicial practical experiences.
These people judgers work together and transform the knowledge of how they make a judgement for law cases into a graph.
for simplification, we call this graph as judicial knowledge graph.
Once judicial knowledge graph is formed, our job is to transform it into a runnable module in AI system.
The key idea of judicial engine is start reasoning from claim, using judicial knowledge graph, associate evidences to nodes of knowledge graph,
then we can make a conclusion that we support a claim or not.
Another way of saying, if there are three claims in a law case form who file a lawsuit(plaintiff). for each claim, we will ask plaintiff to provide evidences,
and these evidences will be used to prove something on the knowledge graph, then once meet all the requirement for a claim, we can support the claim.
2.Main Objects and Relationship
For a [legal case], it has serveral [claim]s, interested parties should provide with [evidence]s to support their [claim]s (or called appeals),
AI judger will reasoning among [logical tree], which is a form of [knowledge graph], based on proof of evidences, and then reach a [conclusion].
Each type of case has its own logical tree. the main branches of logical tree are claims.
Legal case combine all object together(claim, evidence and logical tree).
3.Type of Claims and Logical Operators
There are many type of law cases, e.g. traffic accident, private lending, marriage and so on. For each type of law case,
people judgers will abstract key claims use their knowledge and experience.
Below we will list some key claims for private lending which account for one of highest percentage among all civil cases.
For each claim, there will be a logical tree associated with. by default, the root of this logical tree is the claim.
Key claims of private lending:
claim of principal: demand to pay back the principal
claim of interest: demand to pay interest of principal
claim for loss of contract breach: cause some loss by a party who violate a contract
claim of mutual debt of husband and wife: may apply to a person in a marriage whose spouse involve in liability
claim of guarantee liability: someone or organization take guarantee responsibility for lending.
Logical operators are used to reasoning among sub nodes or parents nodes in knowledge graph.
Two main types of Logical Operator:
AND: true if meet all requirements
OR: true if meet one requirement among all requirement
There are other type of operators except these main types, including NOT, XOR(exclusive OR), we will omit for now.
command to run:
It will print some logs, including the claim, evidences supported this claim, conclusion.
If the claim is support, it will gives the reason(path) to support it; if it is overturn, the reason why it is not support.
Python 2.7 or Python 3.x
We use three steps to reasoning on the knowledge graph. basically, it is a form of bottom-up solution.
For your better understanding, we will describe main steps first, then in the following section #7 Case Study, we will give you two examples.
Step one, compute path starting from node (node_start)of knowledge graph where evidence point to, to a claim(or appeal, as node_end) which is a sub root
node in knowledge graph.
you can do this simply by getting parent node of node_start, then get parent node of this parent node...util it reach root node.
Step two, for each path, try to set value of all nodes along the path, util you are not able to.
Essentially, if a node is proofed by a evidence, you can set it to True(Activate); and if its parent node's(node_p) operator is OR,
it means node_p is true if one condition of sub nodes is met, you can set node_p's value to True(Activate).
However if its node_p's operator is AND, you need check and see whether all sub nodes's value is True.
Step three, start from the claim, you can check the value of the sub nodes of this claim,
and get a conclusion to support the claim or not.
For more detail, you can check function of logic_operation from LogTreeOperationObject.py
We will give two examples, one a claim is supported, another a claim is overthrown.
Case Study 1: a claim with evidences that AI judger support
the claim is [appeal of principal 1], it has two evidences approval of following nodes: [written agreement 4] and [receipt of cash(bill) 11], the question is should a AI judger support this claim or not?
Step 1, that's compute path from evidence to claim. take first evidence [written agreement 4], it parent node is [loan agreement 2],
[loan agreement 2]'s parent node is [appeal of principal1]. so the path is: [written agreement 4, loan agreement 2, appeal of principal 1], or [4, 2, 1] for short. for evidence of [receipt of cash(bill)], you can also get its path: [receipt of cash(bill) 11, cash(bill) delivery 8, delivery of money 3, appeal of principal 1], or [11, 8, 3, 1] for short. these numbers are IDs representing nodes in knowledge graph,
Step2, try to set value of nodes along the path.
1) take the first path: [written agreement 4, loan agreement 2, appeal of principal 1], or [4, 2, 1]. as there is an evidence point to [written agreement 4], we set this node 4 to be True. what's the value of [loan agreement 2]? [loan agreement 2]'s operator is [or], means it will be True if one of its sub nodes is True. as we can see from knowledge graph that [written agreement 4] is its sub node, and its value is True, so value of [loan agreement 2] will be True. for [appeal of principal 1], as its operator is [and], it need both of its sub nodes are True, but we only know [loan agreement 2] is True, so we will not able to set its value. so far, we know :[written agreement 4:True, loan agreement 2:True, appeal of principal 1:False]. we will set value in a field named as [result] in each node. for any node we can't set to True, its value will be False( as default value). 2) take the second path [receipt of cash(bill) 11, cash(bill) delivery 8, delivery of money 3, appeal of principal 1], and use same logic, we can get the computation result:[ recepit of cash(bill):True, cash(bill) delivery 8:True, delivery of money 3: True, appeal of principal 1:False].
Step 3, check sub nodes of the claim, and compute the value of claim.
the claim is [appeal of principal 1], its operator is [and], it has two sub nodes:[loan agreement 2],[delivery of moeny 3]. In order to support the claim of [appeal of principal 1], both of [loan agreement 2] and [delivery of money 3] should to be True. From step 2, we already know that both of these two nodes are True, so we can get final result of claim[ appeal of principal 1] is True. Thus AI judger can reach a conclusion to suppose this claim.
Input： Claim(or appeal): appeal of principal 1 Evidence list: written agreement 4,receipt of cash(bill) 11 Output： Conclusion:support(yes) Path to support: written agreement4-->loan agreement2-->appeal of principal1 receipt of cash(bill)11-->cash(bill) delivery8-->delivery of money3-->appeal of principal1
Case Study 2: a claim with evidences that AI judger overthrow
the claim is [appeal of principal 1], and it has two evidences approval of following nodes:
[expression of borrowing money 9] and [signature of interested parites 10], the question is should a AI judger support this claim or not?
Input： Claim(or appeal): appeal of principal 1 Evidence list: expression of borrowing money 9,signature of interested parites 10 Output： Conclusion:overthrow(no) Path why not support: delivery of money3-->appeal of principal1
We leave it you to do exercise to reasoning for this case 2, and get conclusion.
7.What you didn't see
So far you can already know how AI judger works with knowledge graph, and the main key ideas of how its reasoning. However, what you see here is only
the prototype of this AI judger. In real practices, things become more complex, and more business demand are involved. Including, but not limit to:
1) thousands of nodes in a knowledge graph. it will become a very big logical tree. 2) serveral claims are included in a law case, instead of one claim. although we will still look after one claim at each time, but you need to associate evidences with specific claim by yourself. 3) not only the plaintiff(the person who fill a lawsuit) will start some original claims( as forward direction), but also defendant(the person who was accused) will start some claim from opposite direction to overthrow original claim. in this case, even claim(s) is hold to be True in the beginning, but it still possible to rejected it by AI judger. additional logic will be need to reach an accurate and comprehensive conclusion. 4) person judger may interactive with AI judger by providing some feedback, such as declares some of evidences not to be trusted. then we have to adopt the feedback and decide the priority of the feedback.
Up to now, we've covered all necessary components of AI judger with knowledge graph. We not only describe key ideas
of how we reasoning along the logical tree in judicial engine , but also give two case study. we hope you enjoy it.
For any idea or feedback, you can contact with me: firstname.lastname@example.org