Source-linked AI summary
How Does NLP Benefit Legal System: A Summary of Legal Artificial Intelligence
Haoxi Zhong, Chaojun Xiao, Cunchao Tu, Tianyang Zhang, Zhiyuan Liu, Maosong Sun
TL;DR
LegalAI addresses text-heavy legal work by applying NLP and AI to tasks such as judgment prediction, case matching, and legal question answering. The paper surveys symbolic and embedding-based methods, reviews applications and datasets, and conducts experiments showing both useful gains and substantial remaining gaps. It concludes that LegalAI should combine performance with interpretability while supporting, rather than replacing, legal professionals.
Problem
LegalAI seeks to apply AI, especially NLP, to legal tasks that involve extensive documents and require substantial professional expertise.
Method
The paper surveys LegalAI methods and applications from NLP and legal-professional perspectives, then evaluates representative models across multiple datasets.
Results
Existing models still perform poorly on legal question answering and remain far behind humans, while BERT is limited by legal documents reaching 56,694 tokens versus its 512-token maximum.
Takeaways & Limitations
LegalAI should combine embedding-based performance with symbol-based interpretability and serve as a supporting reference for legal professionals.
Takeaways & Limitations
Legal question answering remains constrained by inadequate multi-hop reasoning and legal-concept understanding.
Abstract
from arXiv · showhide
Legal Artificial Intelligence (LegalAI) focuses on applying the technology of artificial intelligence, especially natural language processing, to benefit tasks in the legal domain. In recent years, LegalAI has drawn increasing attention rapidly from both AI researchers and legal professionals, as LegalAI is beneficial to the legal system for liberating legal professionals from a maze of paperwork. Legal professionals often think about how to solve tasks from rule-based and symbol-based methods, while NLP researchers concentrate more on data-driven and embedding methods. In this paper, we introduce the history, the current state, and the future directions of research in LegalAI. We illustrate the tasks from the perspectives of legal professionals and NLP researchers and show several representative applications in LegalAI. We conduct experiments and provide an in-depth analysis of the advantages and disadvantages of existing works to explore possible future directions. You can find the implementation of our work from https://github.com/thunlp/CLAIM.
1 Introduction
LegalAI applies AI, especially NLP, to text-heavy legal tasks, aiming to reduce professionals’ workload and provide affordable legal assistance. The paper reviews methods, applications, experiments, and datasets while highlighting tensions between symbolic interpretability and neural performance.
- Most LegalAI tasks use NLP because legal resources primarily appear as judgments, contracts, and legal opinions.
- LegalAI can reduce the time legal professionals spend retrieving and understanding complex legal documents.
- Qualified LegalAI systems should support tedious legal work and provide reliable reference for people unfamiliar with law.
- The field has progressed from hand-crafted rules and features toward deep learning and benchmark datasets.
- Symbolic models offer interpretability but limited effectiveness, whereas stronger embedding methods may be difficult to interpret and risk ethical issues.
- The paper surveys LegalAI from NLP and legal-professional perspectives, details three applications, evaluates methods across datasets, and summarizes benchmarks.
2 Embedding-based Methods
Embedding-based methods represent legal text, facts, and knowledge in vector spaces for prediction, but legal vocabulary and domain knowledge remain difficult to model. The paper discusses domain-specific pretraining and knowledge integration as routes toward stronger LegalAI systems.
- Embedding-based methods represent legal facts and knowledge in embedding space and apply deep learning to LegalAI tasks.
- Character and word embeddings convert discrete text into continuous vector space, supporting downstream NLP tasks.
- Existing embeddings may not learn professional legal terminology directly from factual descriptions, motivating representations that combine grammar and legal knowledge.
- Legal knowledge modelling is important because many LegalAI results depend on legal rules and knowledge.
- Legal knowledge graphs face costly construction, country-specific meanings, and a mismatch between legal-concept explanations and ordinary entity relations.
- Directly applying general-domain pretrained language models to legal text can perform unsatisfactorily because terminology and knowledge differ.
- Legal-domain pretrained models provide stronger baselines, while future work should integrate knowledge into pretrained models to improve reasoning between legal concepts.
3 Symbol-based Methods
Symbol-based methods extract and use legal structures such as entities, relations, events, and elements to improve interpretability and support downstream reasoning. Experiments indicate that domain-specific pretraining helps element extraction, although existing methods remain insufficient for applications.
- Symbol-based methods use legal symbols and knowledge for structured prediction, providing interpretability while allowing deep learning for performance.
- Information Extraction: Legal information extraction applies ontologies, global consistency, rules, CRF, joint models, and neural methods to identify entities, relations, and events.
- Information Extraction: Extracted legal symbols should be evaluated for their benefits and interpretability, not only for extraction performance.
- Information Extraction: Relation extraction in inheritance disputes can explain judgment results by identifying parties’ relationships to the deceased.
- Information Extraction: Event timeline extraction supports criminal judgment prediction by representing what participants did and the order of events.
- Legal Element Extraction: Legal elements encode crucial crime-related facts, can provide intermediate supervision for judgment prediction, and make predictions more interpretable.
- Legal Element Extraction: Element extraction is formulated as multi-label classification across divorce, labor, and loan disputes and is evaluated with neural encoders including BERT.
- Legal Element Extraction: Domain-specific pretrained language models outperform general-domain models for element extraction, but existing methods remain insufficient for corresponding applications.
4 Applications of LegalAI
LegalAI applications span legal judgment prediction, similar case matching, and legal question answering, covering central functions of legal reasoning and consultation. Experiments show that current systems benefit from combining neural representations with legal structure, yet remain limited by long documents, infrequent labels, insufficient case similarity signals, and difficult legal reasoning.
- Applications: LegalAI applications include legal judgment prediction, similar case matching, and legal question answering.These tasks cover core judgment functions in Civil Law and Common Law systems and provide consultancy for people unfamiliar with law.
- Legal Judgment Prediction: C-LJP contains over 2.68 million Chinese legal documents and evaluates relevant articles, applicable charges, and penalty terms.The first two subtasks are multi-label classification, while penalty-term prediction is regression.
- Legal Judgment Prediction: Most models perform promisingly on high-frequency charges or articles, but low-frequency labels expose a gap between micro-F1 and macro-F1.BERT also performs unsatisfactorily because legal documents can be extremely long: the maximum length is 56,694 and 15% exceed 512 tokens.
- Legal Judgment Prediction: Combining symbolic task order with neural fact encoding improves judgment prediction, supporting the necessity of hybrid embedding-based and symbol-based methods.TopJudge formalizes topological order between LJP tasks and uses TextCNN to encode fact descriptions.
- Similar Case Matching: Neural models outperform TF-IDF in similar case matching, but performance remains insufficient because term-level and semantic-level similarity alone misses legally important case elements.Future directions include element-based representations and incorporating legal knowledge for performance and interpretability.
- Legal Question Answering: Existing models cannot answer legal questions well, leaving a large gap between model performance and humans in legal question answering.LQA remains difficult because questions vary in form and require legal multi-hop reasoning and legal-concept understanding.
5 Conclusion
The paper reviews LegalAI’s development and argues that future systems should combine embedding-based performance with symbol-based interpretability while addressing knowledge modelling, legal reasoning, and ethical risks. LegalAI should support rather than replace professionals, with model outputs treated as references and simple cases still reviewed.
- The paper surveys LegalAI tasks, applications, development status, and future directions from both NLP and legal-professional perspectives.
- Knowledge modelling, legal reasoning, and interpretability remain foundational challenges for reliable LegalAI.The paper states that substantial research progress is still needed.
- Future research should combine embedding-based and symbol-based methods to address LegalAI’s principal challenges.
- Direct deployment can raise ethical issues such as gender bias and racial discrimination, limiting confidence in model results.
- LegalAI should assist rather than replace professionals, using model outputs as references while requiring review even for simple cases.