The legal industry is characterized by the presence of dense and complex documents, which necessitate automatic processing methods to manage and analyse large volumes of data. Traditional methods for extracting legal information depend heavily on substantial quantities of annotated data during the training phase. However, a question arises on how to extract information effectively in contexts that do not favour the utilization of annotated data. This study investigates the application of Large Language Models (LLMs) as a transformative solution for the extraction of legal terms, presenting a novel approach to overcome the constraints associated with the need for extensive annotated datasets. Our research delved into methods such as prompt-engineering and fine-tuning to enhance their performance. We evaluated and compared, to a rule-based and BERT systems, the performance of four LLMs: GPT-4, Miqu-1-70b, Mixtral-8x7b, and Mistral-7b, within the scope of limited annotated data availability. We implemented and assessed our methodologies using Luxembourg’s traffic regulations as a case study. Our findings underscore the capacity of LLMs to successfully deal with legal terms extraction, emphasizing the benefits of one-shot and zero-shot learning capabilities in reducing reliance on annotated data by reaching 0.690 F1 Score. Moreover, our study sheds light on the optimal practices for employing LLMs in the processing of legal information, offering insights into the challenges and limitations, including issues related to terms boundary extraction.

The legal industry is characterized by the presence of large volumes and complex documents. Given the continuous evolution of these documents, there is a growing interest in automating the processing of legal texts to streamline compliance. One key step of this process is the extraction of legal entities. State-of-the-art methods for legal entity extraction, including rule-based systems, Bi-LSTM, and BERT, require substantial annotated data to be effective, a task that is time-intensive for domain experts. With the rise of Large Language Models (LLMs), research has increasingly focused on leveraging their capabilities and exploring zero-shot approaches. In this paper, we present a hybrid system that distils GPT-4 knowledge through rule-based methods into a CamemBERT model. This approach not only reduces the need for expert involvement compared to the standard CamemBERT system but also outperforms the GPT-4-only system, enhancing the F1 score for legal entities by 9–24% points.

Compliance with legal documents related to industrial maintenance is the company’s obligation to oversee, maintain, and repair its equipments. As legal documents endlessly evolve, companies are in favour of automatically processing these texts to facilitate the analysis and compliance. The automatic process involves first, in this pipeline, the extraction of legal entities. However, state-of-the-art, like BERT approaches, have so far required a large amount of data to be effective. Creating this training dataset however is a time-consuming task requiring input from domain experts. In this paper, we bootstrap the legal entity extraction by levering Large Language Models and a semantic model in order to reduce the involvement of the domain experts. We develop the industrial perspective by detailing the technical implementation choices. Consequently, we present our roadmap for an end-to-end pipeline designed expressly for the extraction of legal rules while limiting the involvement of experts.

In different domains, compliance with legal documents about industrial maintenance is crucial. Legal industrial maintenance is the legal commitment of a company to control, maintain and repair its equipments. With the evolution of legal texts, companies are increasingly adopting automatic processing of legal texts in order to extract their key elements and to support the task of analysis and compliance. To perform such a task of knowledge extraction, a number of state-of-the-art proposal relies on a semantic model. Based on existing models from both legislative and industrial maintenance domains, we propose a new semantic model for the legal industrial maintenance: SEMLEG (SEmantic Model for the LEGal maintenance). This model results from an analysis of documents extracted from the Légifrance French governmental website.

This paper describes our two deep learning systems that competed at SemEval-2022 Task 1 “CODWOE: Comparing Dictionaries and WOrd Embeddings”. We participated in the subtask for the reverse dictionary which consists in generating vectors from glosses. We use sequential models that integrate several neural networks, starting from Embeddings networks until the use of Dense networks, Bidirectional Long Short-Term Memory (BiLSTM) networks and LSTM networks. All glosses have been preprocessed in order to consider the best representation form of the meanings for all words that appears. We achieved very competitive results in reverse dictionary with a second position in English and French languages when using contextualized embeddings, and the same position for English, French and Spanish languages when using char embeddings.