A Connectionist Model to Justify the Reasoning of the Judge

One of the main obstacles to the use of Artificial Neural Network (ANN) in the legal domain comes from their inability to justify their reasoning. Justification indeed is crucial for the judge because it assures him that the reasoning carried out by a legal machine is legally founded. We propose in this paper a method able to overcome this constraint by developing an algorithm of justification applied to connectionist prototypes (Multilayer Perceptron) implemented at the Court of Appeal of Versailles. We will first describe the algorithm. We will then discuss the two main advantages offered by the ANN with regard to rule based systems. A first advantage consists of their suitability for some types of reasoning not based on explicit rules, which are specially numerous in the discretionary field of the judge. Another advantage can be emphasised as a result of our experiment: these models can be used for improving the self justification process of a decision maker (making it more precise) and even for predicting (or suggesting) new lines of reasoning based on implicit knowledge. Some examples extracted from a knowledge base on the contract of employment (clause of non-competition) will illustrate this point. 1 Artificial Neural Networks and Law (ANN): State of Works The models of artificial intelligence generally have as their objective to mimic human reasoning. But they need first to know and to formalize the rules which are the base of the decision-making process, and ways of to using them. When it is about the law, we generally think that these rules are given by the legislator. We can’t, however, give to the whole law the shape of rules1. If we try, we notice quickly that the basic elements of the law (its ontology) can be vague or indefinite and that in reality, law delegates to the judge the task of adding the implicit rules of interpretation of these elements to allow him to legally resolve real cases; it is what is called the discretionary power of the judge. An ANN is particularly adapted to such vague, fuzzy and noisy knowledge. Numerous tests were made to use connectionist models in the treatment of the law ([4, 2]). Their aptness to handle the most open textured parts of the reasoning was recognized ([1]) in spite of the criticisms which are made of them ([5]). Their main defect lies in their difficulty to express the rules which underlie their results. In certain domains (treatment of the signal, treatment of the speech, recognition of faces), the experimenter does not need to know which pixel or which neuron tipped over the decision of the network. On the other hand to know which criterion or sub-criterion took decision is indispensable for the magistrate and for the affected citizens. 1For example, principles require a different representation. Filipe Borges, Raoul Borges and Danièle Bourcier, ‘A Connectionist Model to Justify the Reasoning of the Judge’ in T.J.M. Bench-Capon, A. Daskalopulu and R.G.F. Winkels (eds.), Legal Knowledge and Information Systems. Jurix 2002: The Fifteenth Annual Conference. Amsterdam: IOS Press, 2002, pp. 113-122. 114 Filipe Borges, Raoul Borges and Danièle Bourcier Our hypothesis consists in showing that an algorithm can justify the ”decision” of a neural network. The first stage is still to teach a network the knowledge of the judges, who sometimes are unaware certain important elements they use. The network can then be used by the decision-maker to auto-justify or to criticize its own criteria. 2 Theory of a Connectionist Justification 2.1 Obstacle: The Dispersal of the Information in the Structure The reading of a ruling shows that the judge justifies himself generally by quoting the elements of facts or law which directed his decision. This clause is not applicable because the contract of employment foresaw the payment of a financial counterpart and because this one was not paid ; that dismissal is without real and serious cause because the letter of dismissal does not express real and precise motives; this privilege of jurisdiction is pushed aside because an agreement foresees rules of particular competence, and so on. This way, a satisfactory justification emanating from a connectionist model would consist in stating which elements of facts or law (criteria) were the most important in the decision-making process. At present, this model can at the most (at least for the perceptron) list the inputs and show that their combination had a particular effect on the output. It is not a justification but simply a description. What is easy for an expert system (to justify the reasoning) is hard for an ANN, because of the way the knowledge is structured. The problem is that the ANN scatters the information through its structure and the weights of its connections. This makes the procedure of recovery of relevant information particularly complex. 2.2 Hypothesis: to Track Down the Signal in the Structure The action of the perceptron consists in converting the value of its inputs to a signal which propagates along its connections. This signal undergoes several mathematical transformations and is output by the last neuron as a value which is interpreted by the experimenter. In fact, the main difficulty consists in understanding the way the global signal propagates and to discover retroactively how every neuron acted on this signal. In this way, we hope to go back to the inputs of the network and to determine the influence of each of them. To ask what role is played by every entry on the output signal of a perceptron is like asking what role is played by every tributary in the final flow of a river. Different sources (the inputs) feed various tributaries (the connections of the network) which eventually end up on the same rivers (connections of the hidden neurons), and throw themselves into the sea by the way of the same river mouth (the output of the network). If we can measure the flow of water at the mouth as well as the flow of the various rivers we can calculate the contribution of every river to the final flow, and interpret it. A weak flow can be explained by the fact that such rivers have an extremely weak flow themselves. On the contrary, one or several rivers with an enormous flow can explain why we observe a very important flow of water at the mouth. The connectionist model used here presents a similarity with this image, at least for its connections. But it presents a supplementary complication : the existence of neurons. These neurons can convert the signal so that if we take the previous image we can obtain tributaries which, besides crossing each others, see their flow increasing and decreasing several times during the same journey, and, moreover, it happens regularly that their course is reversed. All A Connectionist Model to Justify the Reasoning of the Judge 115 these complications justify the use of data processing ’to track down’ and to measure all the transformations of the signal as it propagates within the neural network. The hypothesis is that in the case of a binary output (ranging from 0 to 1 and interpreted), certain neurons have, by the way of their connections, a tendency to decrease the signal while the others tend to increase it. By adding the influence of every connection relative to every input and by comparing them with the others we can determine the sense of influence of every input (the global sense of the source) and the relative importance of their influence on the final result. We also put forward the hypothesis that the most influential criteria within a neural network correspond to the most determining elements within a decision-making process, and identifying them is enough to justify the reasoning. 3 Protocol of Experiment 3.1 Development of an Algorithm The method developed is valid for a multilayer perceptron with continuous inputs and outputs (ranging from 0 to 1), with one hidden layer, and one final neuron. The neural network used in this experiment has 14 entries, 1 hidden layer, 3 hidden neurons, and 1 neuron on the output layer. The values of the entries and of the output are continuous (from 0 to 1) but were used in a boolean way (0 or 1). – We find the weights ‘Wnm‘ of every connection – We calculate the signal ‘S’ sent by every entry to every hidden neuron, from the input value ‘X’ (which can be 0 or 1) and the weight which connects it to the various hidden neurons :