Recently, there have been opportunities and challenges for both legal practitioners and AI researchers due to the abundance of legal documents that are available in digital form. This development requires a lot of assistance and work to present this content in a helpful and distributed manner, which could have the potential to offer legal professionals many advantages. While Pretrained Language Models (PLMs) have demonstrated impressive performance in various Natural Language Processing (NLP) tasks, developing effective strategies for legal practice tasks remains challenging. This is mainly because legal documents often consist of lengthy and non-uniformly structured text, which sets them apart from general documents. Consequently, the PLMs may not fully capture the intricate legal details present in legal documents. This is because most language models struggle to grasp the extensive connections between various structures in the text. To address these challenges, we propose an approach that utilizes outputs from multiple intermediate layers of pre-trained transformer models while considering the entire legal document. Our primary objective is to generate more accurate and meaningful embedding representations, enabling the identification of semantic and syntactic relationships among sentences in the document. This approach also aims at preserving the contextual interactions with adjacent sentences, thereby enhancing the overall understanding of the document’s content. The performance of our proposed approach is showcased through experimental results conducted on publicly available legal datasets over different tasks.
Legal judgment documents contain complex legal terminology, long-range contextual dependencies, and multiple interrelated entity references, making automated legal document understanding a challenging Natural Language Processing (NLP) task. Existing studies primarily address Named Entity Recognition (NER), Coreference Resolution (CR), and Judicial Outcome Classification (JOC) independently, resulting in fragmented contextual understanding and limited semantic interaction among complementary legal NLP tasks. To address this limitation, this study proposes the Legal Unified Modular Pipeline (LUMP), a sequential transformer-based framework that integrates XLNet for extracting domain-specific legal entities, DistilBERT for resolving pronoun–antecedent relationships to maintain entity consistency, and LegalBERT for classifying judicial outcomes using contextually enriched legal representations. Owing to the absence of a unified benchmark dataset, each module was independently optimized using the most appropriate benchmark dataset and subsequently integrated during inference through sequential contextual enrichment. Experimental evaluation demonstrated that the NER, CR, and JOC modules achieved an entity-level F1-score of 93.94
Recent advances in AI make it conceivable to delegate legal decision-making to machines, or to enhance human adjudication through AI assistance. Using classic normative conflicts — the trolley problem and comparable moral dilemmas — as a proof of concept, we examine the alignment between AI legal reasoning and human judgment. In our baseline experiment, we find a pronounced mismatch between decisions made by GPT and those of human subjects. This misalignment raises substantive concerns for AI-powered legal decision-aids. We investigate whether explicit normative guidance can address this misalignment, with mixed results. GPT-3.5 is susceptible to such intervention, but frequently refuses to decide when faced with a moral dilemma. GPT-4 is outright utilitarian, and essentially ignores the instruction to decide on deontological grounds. GPT-o3-mini faithfully implements this instruction, but is unwilling to balance deontological and utilitarian concerns if instructed to do so. We replicate the experiment with four LLMs from different providers. Claude-sonnet-4.6 comes closest to human respondents. Gemini-2.5-flash-lite is most sensitive to normative instructions. Llama-4-scout and Mistral-nemo have a strong utilitarian bias, and do not strongly respond to normative interventions. At least for the time being, explicit normative instructions are not fully able to realign AI advice with the normative convictions of the population, or the legislator deciding on its behalf.
As artificial intelligence (AI) systems move from isolated tools to continuous collaborative partners, a critical question arises: when AI and humans work together on a task, who is accountable for the output? We discuss how accountability shifts from the human to the AI over the transition of six states: Autonomous Human, AI-Assisted Human, High-Synergy AI-Assisted Human, Human–AI Hybrid, Supervised AI, and Autonomous AI. This taxonomy is built on two formal measures - the Human Causal Primacy Ratio measures what fraction of an output is causally attributable to the human and serves as a proxy for human accountability, and the Synergy Coefficient ( ) measures the performance gain the human–AI pair achieves over either party working alone. Agency Erosion reveals how the two measures are causally related: organizations pursue higher , and the cost of that pursuit is lower HCPR which is the transfer of accountability from human to AI. We use the Agency Erosion Index to track how human causal contribution changes over time. Two case studies – integrated software development environments and autonomous vehicles – are presented to demonstrate how agency erosion occurs when human-AI pairs evolve from Autonomous Human to Autonomous AI.
Court proceedings typically involve comprehensive case narratives that encompass facts, evidence, witness testimony, and arguments for and against the convicted individual. While previous research has worked towards identifying these high-level rhetorical classes from case documents, there has been limited exploration into identifying more fine-grained events related to a case, e.g., whether a testimony was challenged, whether a crime involved political faction, etc. From an NLP perspective, this is a multi-label text classification problem that is more challenging for legal documents due to the scarcity of annotated data. To address that issue, we propose an approach that combines Large Language Models (LLMs) with human expertise to create an efficient way for making the ground truth annotation. We address this task with an in-context learning (ICL) based approach, the advantage being that, unlike a supervised model, it does not require a large quantity of training data. We augment a standard ICL approach in two ways: first, instead of finding localized examples by their similarities with a test instance, we show that an ontology of class descriptions to construct the set of localized examples yields better results; second, we show that working with a set of examples that are diverse in content and also in class labels lead to more effective predictions. We make both the datasets and code publicly available ( https://anonymous.4open.science/r/LLM_LawAI-9456/ ).
Soil and water conservation (SWC) laws and regulations are hierarchically complex and frequently revised, and grassroots compliance urgently needs efficient, intelligent knowledge tools. Locally deployable small language models (SLMs) avoid the computational cost and data-leakage risks of large models, but their capability boundaries in the SWC legal domain and the pathways for enhancing them remain unclear. Using 71 national and local SWC regulations and the newly promulgated Ecological and Environmental Code of the People’s Republic of China as source materials, we built an evaluation dataset spanning four task types—theoretical knowledge, operational specifications, law-enforcement practice, and comprehensive analysis—across three regulatory levels, and evaluated five SLMs (Qwen3-14B, Qwen3-30B, Phi-4, Gemma3-12B, and Gemma4-31B) under local closed-book conditions. Guided by Cognitive Load Theory, we designed a structured prompt (role setting, task identification, reasoning guidance, and output constraints), ablated each component against an equal-length placebo, and compared reasoning guidance with direct provision of statutory text; beyond accuracy and cost, we added process-level metrics including citation hit rate, rule-selection quality, and suspected hallucination. Baseline accuracy ranged from 68.9
The growing adoption of artificial intelligence in judicial applications exposes a critical limitation of neural network-based legal AI systems: their decision-making processes are inherently opaque, which undermines the reliability of their outcomes, especially in cases involving personal liberty. Existing interpretability methods primarily rely on instance-based analysis and local input-output attributions; however, they are unable to capture the models’ global decision-making mechanisms, which greatly limits their applicability in legal contexts where both interpretability and predictive accuracy are essential. To address this challenge, this work proposes a logic-based interpretability and analysis framework that integrates the neural network prediction process with formal logical representations to construct transparent decision traces, capturing the relationships between legal elements and model outputs. Based on this representation, we introduce a method for evaluating the logical relationship between the extracted model logic and statutory legal elements, define the concept of conditional satisfiability in the legal domain, and apply influence-based metrics to quantify the importance of legal elements across different cases. Our analysis of the influence rankings reveals significant inconsistencies between the internal decision logic of legal AI systems and actual statutory logic, confirming that even highly accurate models may fail to align with legal decision-making principles. Our framework offers a globally interpretable analysis methodology for inspecting neural legal AI models, thereby supporting risk identification and trustworthiness assessment in legal AI applications, providing a foundation for trustworthy and interpretable AI in law.
Legislative text translation requires unusually high fidelity because terminological substitutions, modal shifts, and discourse-level reformulations may alter rights, duties, or institutional meaning. While Large Language Models (LLMs) and multi-agent systems have advanced general-purpose translation, their probabilistic generation processes remain weakly aligned with the constrained interpretive conditions of statutory language. Existing agentic frameworks, which primarily rely on horizontal collaboration among peer agents, often improve fluency without providing sufficient vertical control over the layered structure of legal language. This paper proposes HL-Trans, a Hierarchical Control Framework for legislative translation that operationalizes fidelity through three progressively constrained layers: terminology anchoring, syntactic shaping, and discourse polishing. Rather than treating Skopos theory as a license for free functional adaptation, we reinterpret the legislative translation brief as a constrained functional objective: the target text should preserve legal concepts, deontic force, and institutional genre while remaining subject to expert human review. Experiments on a Chinese-English-Japanese legislative corpus show that HL-Trans improves terminology consistency, deontic modality preservation, and discourse conformity over strong LLM and multi-agent baselines. The analysis further identifies a “Gain-Interference-Rescue” pattern: terminology constraints first improve conceptual fidelity; syntactic constraints may temporarily reduce surface fluency while protecting legal force; and discourse-level polishing can restore readability without relaxing lower-level constraints. The results suggest that hierarchical control is a promising design principle for trustworthy legal translation, while also underscoring the continuing need for transparent resources, expert validation, and accountable human oversight.
Understanding case law evolution contributes to the comprehension of legal principles, the identification of trends in judicial reasoning, and testing consistency between judgments. Automation of the analysis has the potential to make it easier to keep up with relevant case law. While legal network analysis typically focuses on identifying precedents, it has largely overlooked tracking historical developments in judicial reasoning. This study applies Main Path Analysis (MPA) to legal citation networks to uncover how key decisions have shaped the law over time. This study illustrates MPA with consumer protection law cases decided by the Court of Justice of the European Union (CJEU). MPA was applied to a citation network of 446 consumer rights cases from 1980 to Oct 2023. The method involves two steps: (1) weighting citations to measure knowledge flow and (2) extracting significant paths that trace the evolution of judicial reasoning. The Top-3 key-route main paths from five time slices were merged into an overlay network and analysed for alignment with judicial interpretations and evolving consumer protection themes. The topics on the paths are compared to those prevalent in the literature on consumer protection. The analysis uncovered key paths in consumer protection jurisprudence shaped by pivotal cases. Organised around five themes—unfair terms, equivalent effect, comparative advertising, free movement of goods, and product liability— paths align closely with the CJEU’s interpretative practices. While MPA highlights significant pathways, it has caveats and challenges, particularly in achieving comprehensive and representative results and in large networks. Future improvements could include integrating semantic adjustments and key case identification to enhance both aspects, as well as developing complementary methods to identify derivative or branching paths in domain development.
In many jurisdictions, academia is at the service of legal practice. Law professors write commentaries that summarize the state of the art of doctrine, chiefly of jurisprudence. In the spirit of a proof of concept, using the guarantee of freedom of assembly in the European Convention on Human Rights, we show that this task can be completely outsourced to large language models. Using standard NLP metrics and an LLM as a judge approach, we develop an evaluation pipeline that works without costly human annotation. The commentaries fully written by GPT 4o, Gemini 2.5 flash or Kimi K2 Instruct are on par with their best human written competitor, the Guide provided by the Court itself.
Legal principles stated in judicial decisions are a key reference for judges and legal professionals. However, typically, decisions are long and complex documents. The automatic extraction of legal principles therein may thus be beneficial to many. Unfortunately, this is an under-investigated task for which few resources and implementations exist, especially with regard to EU law. To promote research in this domain, we develop a novel corpus of 64 judicial decisions on Fiscal State Aid by the CJEU (the Court of Justice of the European Union), in which legal principles are identified and tagged. With this new resource, we conduct extensive experimentation to identify the best combinations of representation methods, classifiers and data augmentation methods for the task at hand, obtaining an F1 score of 0.72. Additionally, we consider the collections of legal principles obtained by running the extractive task and collating the output paragraphs together. We show, through the use of lexical metrics and human evaluation, that the quality of generated collections is comparable to that of expert-made ones.
The rapid development of large language models (LLMs) such as ChatGPT has catalyzed significant advancements in legal AI systems, exemplified by models like LawGPT. However, most existing legal LLMs primarily focus on tasks such as legal knowledge Q A, case retrieval, and statute search, lacking the capacity to model the procedural complexity, adversarial dynamics, and multi-role interactions inherent in real courtroom settings, failing to address the critical demands of practical legal education. In legal pedagogy, simulated courtroom serves as an indispensable tool for cultivating high-caliber legal professionals. Recognizing this issue, we propose CourtSim-LLM—a specialized LLM designed for Chinese mock court education, leveraging in-context learning to simulate comprehensive courtroom proceedings. Our methodology involves two key innovations: (1) the construction of the CourtSim-Law Dataset, a multi-turn dialogue corpus capturing authentic courtroom dynamics, especially adversarial courtroom debates and syllogistic legal reasoning, and (2) role-specific Low-Rank Adaptation (LoRA) fine-tuning of a Chinese open-source legal LLM to simulate plaintiff and defendant interactions. We employed a hybrid evaluation method comprising 11 indicators and organized 15 senior legal experts to conduct a double-blind evaluation to comprehensively validate the performance of CourtSim-LLM in generating simulated courtroom dialogues. Experimental results show that this model significantly outperforms other models in key dimensions such as "text quality," "procedural integrity," "adversarial argumentation ability," and "judgment logic," scoring approximately 1.5 times that of the baseline model and nearly twice that of other comparative models. This not only demonstrates the model's superior performance in procedural rigor and adversarial argumentation but also promotes the application of AI in simulating multi-role, multi-task court proceedings, effectively bridging the gap between legal theory and practice.
Legal dispute resolution hinges on the weighing of evidence and arguments across parties, yet human biases and bounded rationality can influence outcomes. Simulating this process with large language models can clarify both the capabilities and the anthropomorphic behavior. Most existing studies focus on algorithmic support for adjudication, such as online courts, risk assessment, and sentencing tools. At the same time, the role of AI agents in mediation and conciliatory negotiation remains largely untested. To address this gap, the study designs six scenario types and evaluates them on a multi-agent platform. It addresses two questions: (i) To what extent can LLM agents produce realistic simulations from a juridical perspective? We find that LLMs can reproduce diverse dispute scenarios, but performance varies by role; in particular, judge agents sometimes commit serious legal errors when interpreting clauses and may infer property rights rather than apply the correct rules. (ii) Do sensitive factors and intrinsic properties affect legal dispute resolution? Sensitivity analyses across model-level configurations, framework design choices, and procedural settings confirm the robustness of our findings; we further observe irrational patterns yielding mechanical or unexpected decisions and measurable effects from intrinsic properties. These findings show strengths in fact-heavy money bargaining and mixed remedies, while highlighting limits where careful discretion and normative justification are required. Code and scenarios are publicly available for replication ( https://github.com/TOM-ZHOUch/Legal_Dispute_Resolution ).
The paper explores how the construction of Bayesian networks for legal evidence can be assisted by scenario models, especially how spatio-temporal constraints can be transferred from scenario models to Bayesian models. The paper introduces two new methods for constructing scenario-based Bayesian networks: non-Boolean time-nodes with event-states and non-Boolean event-nodes with time-states.
Legal Case Retrieval not only involves the identification and selection of similar cases, but also encompasses an in-depth comparison and abstraction of their holdings, reasoning structures, and underlying legal principles, serving as a critical pillar of judicial digital transformation and the development of smart courts. With the rapid advancement of neural networks, deep learning, and pretrained language models, retrieval paradigms have gradually evolved from traditional keyword matching toward intelligent approaches such as semantics-based dense retrieval. Owing to the high level of domain specificity and structural complexity of legal documents, case-based retrieval has long remained a focal research topic in this field. To leverage the strengths of traditional methods in capturing global semantics while fully exploiting the contextual modeling capabilities of pretrained language models, this paper proposes ALTER, A Lightweight Topic-awarE Representation legal case retrieval system. Specifically, ALTER incorporates topic modeling to enhance the global semantic representations learned by pretrained language models, and introduces a Co-Attention layer to explicitly facilitate bidirectional interactions between the topic encoder and the text encoder. In addition, a multi-label classification task is employed to guide the text encoder to implicitly learn topical information, resulting in topic-aware text representations. During the online retrieval stage, only the text encoder is used to construct the index, enabling efficient and lightweight inference while achieving joint matching between text and topics. Experimental results on the Chinese LCR dataset LeCaRD and the English LCR dataset COLIEE23-24 demonstrate that ALTER consistently achieves state-of-the-art performance across multiple evaluation metrics.
Legal argumentation requires sophisticated reasoning, fact analysis, and a subtle appreciation of statutory and case law. The project provides an AI-based framework that uses large language models (LLMs) to aid legal practitioners by creating context-specific counter-arguments, mimicking courtroom-style argumentation, and case outcome simulation. The model is incorporated with a refined GPT-3.5 Turbo model that was trained on expert-curated triplets of case summaries, legal arguments, and counter-arguments from actual court transcripts and legal discussions. To improve factual grounding and minimize hallucinations, the model is supplemented with retrieval-based legal context from statutory law corpora. A new assessment metric—designed with feedback from working attorneys—was proposed to measure the quality of arguments on grounds of persuasiveness, legal appropriateness, and completeness. Quantitative metrics such as cosine similarity and tailored scoring measures demonstrated that AI-generated counter-arguments attained an average similarity of 0.598 with human-written counterparts and were 6.91
Transformer-based language models have sparked a revolutionary change in Legal NLP, endowing lawyers with unparalleled tools to effectively navigate, understand, and draft large volumes of text. However, the dearth of large-scale datasets from authoritative sources hampers further progress. The available resources are primarily single-task, English-only, and written in layman’s terms. To bridge this gap, we introduce Comma, a multi-task and multi-lingual archive of 14K verdicts drawn from the Constitutional Court of the Italian Republic, grounded in a non-common law system. Documents in Comma diverge from ordinary legal manuscripts as they address fundamental principles and rights, involve technical jargon, exhibit an articulated structure, are diachronic, have extended length, and demand more significant expertise and interpretation. By embracing 4 widespread languages, Comma tackles a panoply of necessity-driven tasks: multi-granular abstractive summarization, decision generation, article retrieval, and ruling classification. We systematically benchmark a catalog of language models in both few-shot and full settings, uncovering substantial headroom for improvement. We contribute to the new era of Legal NLP systems by openly releasing Comma and best-performing models (https://github.com/disi-unibo-nlp/comma).
Effective legal case retrieval is critical for fair and just adjudication of similar cases. Unfortunately, current methods fall short of understanding legal cases at a semantic level. To address the issue, this paper presents a BERT-based two-stage ranking model for Chinese legal case retrieval. The first stage uses the well-known BM25 ranking function to quickly retrieve top n case candidates from the candidate pool. In the second stage, firstly the fine-tuned BERT is used for semantic analysis of the retrieved case candidates (i.e., obtained their embeddings) and then a ranking mechanism is used to re-rank the retrieved cases. We fine-tune the BERT model by integrating the point-wise method with auxiliary learning to better understand the deep semantics of a legal judgement text, and we further train the re-ranking neural network based on the fine-tuned BERT model in conjunction with the pair-wise method. Our extensive experiments reveal that the proposed method surpasses state-of-the-art methods on the Legal Case Retrieval Dataset (LeCaRD) by improving the ranking quality of retrieved legal cases, leading to more relevant cases appearing higher in the ranked results.
Indonesian court decisions contain rich legal knowledge about how judges interpret statutes, assess evidence, and determine sentences, yet much of this remains hidden in semi-structured, inconsistent documents. Rule-based systems fail on lengthy narratives, while machine learning models overfit to noise, and large language models risk factual errors. This study introduces a hybrid framework that combines rule-based extraction for structured sections with a BERT-based pipeline for narrative text. This framework was developed using a dataset of 9,109 narcotics and corruption cases; from this corpus, seven legal experts manually annotated 200 decisions. A relevance filter cleaned the text before entity extraction using fine-tuned LegalBERT. The entity extraction pipeline achieved an average F1-score of 84.39
Legal Multi-Choice Question Answering (LMCQA) presents unique challenges due to the complexity and subtlety of legal reasoning. This paper presents DeAtt-LMCQA, a novel model that integrates the DeBERTa language model with attention mechanisms to mitigate classification bias and enhance performance in legal multiple-choice tasks. The system first uses BM25 to retrieve relevant legal articles for each question-option pair, grounding answer evaluation in authoritative legal content. It then encodes these triples via DeBERTa and employs a two-stage attention mechanism for semantic fusion and matching across articles, questions, and answer options. A binary classification strategy is adopted to assess each option independently, improving interpretability and robustness. Extensive experiments on the Judicial Examination Challenge - Question Answering (JEC-QA) dataset from China’s judicial examination demonstrate that DeAtt-LMCQA outperforms multiple strong baselines and achieves third place in the Challenge of AI in Law (CAIL) 2022 competition. Furthermore, it shows competitive advantages over several contemporary large language models in both accuracy and interpretability, highlighting its practical value in domain-specific legal question answering. Our code and model checkpoint are available at https://github.com/gbchen99/chatGPT_result/tree/main/source_code .
