The idea behind the AMPS project is to investigate the use of natural language processing and machine learning to predict the outcome of Maltese court cases, specifically those within the Small Claims Tribunal, which deals with claims of up to €5000. The objective is to provide a tool to help adjudicators decide cases more efficiently while respecting and integrating the ethical and safety concerns which inevitably arise. By aligning with best practices and guidelines, the project team intends not only to develop a tool for the courts but also to advance the use of the Maltese language in relation to Artificial Intelligence (AI).
AMPS started when Dr Ivan Mifsud, Dean of the Faculty of Laws at UM, was inspired to introduce the fast-evolving world of AI to the Maltese legal system. While already being applied to the legal landscape in other countries such as the US, AI in Malta presents both a unique opportunity and a challenge, particularly because Maltese is considered to be a low-resourced language when it comes to training AI models. Nonetheless, the team behind AMPS, which includes Dr Charlie Abela and Dr Joel Azzopardi from the Faculty of ICT, plan to innovate around these obstacles and develop a prototype that adjudicators in the Small Claims Tribunal can use.
‘My long-term vision, in simple terms, is to have a mobile app where the user can describe a problem and receive a solution,’ states Mifsud. ‘Is a case at the Small Claims Tribunal likely to be won or not? What are the comparable facts and reasoning behind other cases and judgements?’
Maltese Legal Context
The Small Claims Tribunal is not as small as it used to be, and this presents growing challenges. When set up in 1993, it handled cases up to 100 Maltese lira in value, for which one did not need a lawyer. Cases were not settled based on strict interpretations of the letter of the law but simply on principles of equity. However, successive Ministers for Justice increased the amount which the Tribunal handles, until eventually, it became crucial to involve lawyers, turning simple cases into multi-session affairs because of the higher stakes.
As the Tribunal has slowed, the backlog of cases has also grown, as it can take up to six sittings to settle a case. A tool to assist adjudicators in making a decision might therefore present a partial solution. Fortunately, the Tribunal offers researchers a host of publicly available data, which is relatively simple to build a dataset with, compared to cases in higher courts.
Challenges and Opportunities
Building the dataset is no easy feat, the researchers explain. It is not just about gathering the data but also translating it into something a machine learning program can use. Furthermore, aside from legal and ethical issues that must be tackled, there is also the language aspect. With Maltese being a low-resource language, most AI models available to the researchers have primarily been trained in English.
‘What we want to do is build a dataset of judgements from data that is freely available on the e-Courts website. Tools normally available for such projects include software to extract text from documents, such as the names of the plaintiff and defendant, the amount of money involved, and other important data points mentioned in the judgement. As the judgements are in Maltese, however, the tools available are limited,’ Abela elaborates. ‘We have thought of two ways to address this problem. We can either translate the judgments into English, and the second option is to build on the work other researchers have done, such as machine learning models already trained on Maltese.’
As the researchers chart the way forward, they indicate that translating the judgements poses the risk of a loss of nuance. An inaccurate translation might mean that some of the original meaning is lost, particularly because the accuracy of translation software itself depends on how a document is written in the context of available Maltese resources.
To settle on a way forward, the researchers are focusing on identifying the best practices and leading examples in the industry and their areas of research. Azzopardi states that the project must look at case studies around the world where other low-resource languages have found innovative solutions to the same challenges AMPS faces today. The team will be looking at how similar research in such countries collects data, the size of the data samples, and what kind of data is used to train the various potential models. Once they have tackled the language question, the researchers will set up appropriate AI models to provide predictions on the basis of the available data.
‘The model will be trained to identify the salient features of cases won and lost. In this way, we will build the training dataset. Cases can be segmented in various ways through rhetorical role labelling, an AI technique which already exists to automatically segment case documents. However, such models are typically trained on data from specific legal jurisdictions. We will have to fine-tune them to the Maltese context, as research has shown that they are not so generalisable across different jurisdictions,’ Azzopardi explains.
Ethical and Safety Implications
Addressing the ethical and legal concerns behind the use of such an app, Mifsud points out how such a tool might cause problems if used incorrectly. People already give a lot of weight to what they learn from social media, the radio, television, or from reading the news, in that they may think they are informed on legal issues but are more likely to make bad decisions in a legal context. It would be easy to misunderstand or misinterpret the results from such an app, which is ultimately a probabilistic language model. AI, after all, can be wrong and does not represent real thinking (yet), but rather mimics such processes. It is for this reason that the Ministry for Justice offered its support to the project on the condition that it is limited to developing an app for adjudicators to use as a decision support system, not turned into a product for the market.
The ethical dimension is not just a question of who uses such an app or how they use it but also extends to how it is programmed. Mifsud warns that software is already used in the US, which leads to profiling, where minorities are identified as more likely to be criminals, for example. There are inherent biases which must be tackled, and that is where the question of safety comes in.
‘Most cases taken to the Tribunal are won, but it would be inaccurate for the app to tell everyone that they are likely to win as a result. The reason most cases are won is that when they first go to a lawyer, the lawyer will advise them if they are likely to win or lose. Therefore, most cases are won because only such cases make it to the Tribunal, while others are more likely to be settled out of court. However, if not trained properly, the AI will simply equate a court case with victory. Furthermore, you can simply never tell which way a court case will go,’ Mifsud warns.
Abela underlines how central the safe use of AI is to the AMPS project. The researchers will be ensuring their work is aligned with the Council of Europe’s ‘European Ethical Charter for the use of Artificial Intelligence in judicial systems and their environment’, particularly through an evaluative checklist. Furthermore, it will be important to ensure a broad understanding of how the tool should be used, how it functions, and how to interpret its outputs, which should not be taken as objective truths.
Having started three months ago, the project, funded by the Malta Council for Science and Technology Research Excellence Programme, is to last for 18 months with no chance of renewal or extension. However, the researchers are already looking ahead and are envisioning how a successful project tackling the Small Claims Tribunal could lead to a similar approach for other courts. If the team cracks the Tribunal, they believe it would not be too difficult to tackle cases dealing with higher amounts, though the criminal courts present a separate set of challenges entirely and would potentially require a rethink.
Applying AI to the Maltese context presents challenges and opportunities alike, but with breakthroughs made by ethical researchers at UM, there is the promise of alignment with best practices and policy. With the project at a preliminary stage, the initiative is likely to make a substantial contribution not only to the Maltese legal context but also to the very use of the Maltese language in AI, building a stronger foundation for research to come.
Comments are closed for this article!