Search

Understanding and explaining legal systems is very challenging due to their complex structure, specialized terminology, and multiple interpretations. Legal AI models are currently undergoing drastic advancements due to the development of Large Language Models (LLMs) that have achieved state-of-the-art performance on a wide range of tasks and are currently undergoing very rapid iterations. As an emerging field, the application of LLMs in the legal field is still in its early stages, with multiple challenges that need to be addressed. Our objective is to provide a comprehensive survey of legal LLMs, not only reviewing the models themselves but also analyzing their applications within the legal systems in different geographies. The paper begins by providing a high-level overview of AI technologies in the legal field and showcasing recent research advancements in LLMs, followed by practical implementations of legal LLMs. Two databases (i.e., SCOPUS and Web of Science) were considered alongside additional related studies that met our selection criteria. We used the PRISMA for Scoping Reviews (PRISMA-ScR) guidelines as the methodology to extract relevant studies and report our findings. The paper discusses and analyses the limitations and challenges faced by legal LLMs, including issues related to data, algorithms, and judicial practices. Moreover, we examine the extent to which such systems can be effectively deployed. The paper summarizes recommendations and future directions to address challenges, aiming to help stakeholders overcome limitations and integrate legal LLMs into the judicial system.

The research presented in the paper aims at increasing the capacity to identify security weaknesses in programming languages that are less supported by specialized security analysis tools, based on the knowledge gathered from securing the popular ones, for which security experts, scanners, and labeled datasets are, in general, available. This goal is vital in reducing the overall exposure of software applications. We propose a solution to expand the capabilities of security gaps detection to downstream languages, influenced by their more popular “ancestors” from the programming languages’ evolutionary tree, using language keyword tokenization and clustering based on word embedding techniques. We show that after training a machine learning algorithm on C, C++, and Java applications developed by a community of programmers with similar behavior of writing code, we can detect, with acceptable accuracy, similar vulnerabilities in C# source code written by the same community. To achieve this, we propose a core cross-language representation of source code, optimized for security weaknesses classifiers, named CLaSCoRe. Using this method, we can achieve zero-shot vulnerability detection—in our case, without using any training data with C# source code.

Intent classification is a central component of a Natural Language Understanding (NLU) pipeline for conversational agents. The quality of such a component depends on the quality of the training data, however, for many conversational scenarios, the data might be scarce; in these scenarios, data augmentation techniques are used. Having general data augmentation methods that can generalize to many datasets is highly desirable. The work presented in this paper is centered around two main components. First, we explore the influence of various feature vectors on the task of intent classification using RASA’s text classification capabilities. The second part of this work consists of a generic method for efficiently augmenting textual corpora using large datasets of unlabeled data. The proposed method is able to efficiently mine for examples similar to the ones that are already present in standard, natural language corpora. The experimental results show that using our corpus augmentation methods enables an increase in text classification accuracy in few-shot settings. Particularly, the gains in accuracy raise up to 16% when the number of labeled examples is very low (e.g., two examples). We believe that our method is important for any Natural Language Processing (NLP) or NLU task in which labeled training data are scarce or expensive to obtain. Lastly, we give some insights into future work, which aims at combining our proposed method with a semi-supervised learning approach.

Software developers represent the bastion of application security against the overwhelming cyber-attacks which target all organizations and affect their resilience. As security weaknesses which may be introduced during the process of code writing are complex and matching different and variate skills, most applications are launched intrinsically vulnerable. We have advanced our research for a security scanner able to use automated learning techniques based on machine learning algorithms to recognize patterns of security weaknesses in source code. To make the scanner independent on the programming language, the source code is converted to a vectorial representation using natural language processing methods, which are able to retain semantical traits of the original code and at the same time to reduce the dependency on the lexical structure of the program. The security flaws detection performance is in the ranges accepted by software security professionals (recall > 0.94) even when vulnerable samples are very low represented in the dataset (e.g., less than 4% vulnerable code for a specific CWE in the dataset). No significant change or adaptation is needed to change the source code language under scrutiny. We apply this approach on detecting Common Weaknesses Enumeration (CWE) vulnerabilities in datasets provided by NIST (Test suites–NIST Software Assurance Reference Dataset).

This paper proposes a supervised machine learning system to detect fake news in online sources published in Romanian. Additionally, this work presents a comparison of the obtained results by using recurrent neural networks based on long short-term memory and gated recurrent unit cells, a convolutional neural network, and a Bidirectional Encoder Representations from Transformers (BERT) model, namely RoBERT, a pre-trained Romanian BERT model. The deep learning architectures are compared with the results achieved by two classical classification algorithms: Naïve Bayes and Support Vector Machine. The proposed approach is based on a Romanian news corpus containing 25,841 true news items and 13,064 fake news items. The best result is over 98.20%, achieved by the convolutional neural network, which outperforms the standard classification methods and the BERT models. Moreover, based on irony detection and sentiment analysis systems, additional details are revealed about the irony phenomenon and sentiment analysis field which are used to tackle fake news challenges.

Computer-Supported Collaborative Learning tools are exhibiting an increased popularity in education, as they allow multiple participants to easily communicate, share knowledge, solve problems collaboratively, or seek advice. Nevertheless, multi-participant conversation logs are often hard to follow by teachers due to the mixture of multiple and many times concurrent discussion threads, with different interaction patterns between participants. Automated guidance can be provided with the help of Natural Language Processing techniques that target the identification of topic mixtures and of semantic links between utterances in order to adequately observe the debate and continuation of ideas. This paper introduces a method for discovering such semantic links embedded within chat conversations using string kernels, word embeddings, and neural networks. Our approach was validated on two datasets and obtained state-of-the-art results on both. Trained on a relatively small set of conversations, our models relying on string kernels are very effective for detecting such semantic links with a matching accuracy larger than 50% and represent a better alternative to complex deep neural networks, frequently employed in various Natural Language Processing tasks where large datasets are available.