Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ContractNLI: A Dataset for Document-level Natural Language Inference for Contracts

Yuta Koreeda, Christopher D. Manning

TL;DR

ContractNLI tackles document-level natural language inference for contracts by introducing a large NDA-based dataset (607 contracts) and a span-aware Transformer model (Span NLI BERT) that jointly handles evidence identification and NLI. The approach uses dynamic context segmentation and [SPAN] tokens to manage long documents and fixed hypotheses, enabling concise, self-contained evidence spans. Empirical results show substantial gains over baselines, with larger models and domain-specific pretraining yielding the best performance, though challenges such as negation by exception and discontinuous spans limit progress. The work provides a practical, extensible framework for contract review automation and highlights directions for extending to other contract types and refining hypothesis design.

Abstract

Reviewing contracts is a time-consuming procedure that incurs large expenses to companies and social inequality to those who cannot afford it. In this work, we propose "document-level natural language inference (NLI) for contracts", a novel, real-world application of NLI that addresses such problems. In this task, a system is given a set of hypotheses (such as "Some obligations of Agreement may survive termination.") and a contract, and it is asked to classify whether each hypothesis is "entailed by", "contradicting to" or "not mentioned by" (neutral to) the contract as well as identifying "evidence" for the decision as spans in the contract. We annotated and release the largest corpus to date consisting of 607 annotated contracts. We then show that existing models fail badly on our task and introduce a strong baseline, which (1) models evidence identification as multi-label classification over spans instead of trying to predict start and end tokens, and (2) employs more sophisticated context segmentation for dealing with long documents. We also show that linguistic characteristics of contracts, such as negations by exceptions, are contributing to the difficulty of this task and that there is much room for improvement.

ContractNLI: A Dataset for Document-level Natural Language Inference for Contracts

TL;DR

ContractNLI tackles document-level natural language inference for contracts by introducing a large NDA-based dataset (607 contracts) and a span-aware Transformer model (Span NLI BERT) that jointly handles evidence identification and NLI. The approach uses dynamic context segmentation and [SPAN] tokens to manage long documents and fixed hypotheses, enabling concise, self-contained evidence spans. Empirical results show substantial gains over baselines, with larger models and domain-specific pretraining yielding the best performance, though challenges such as negation by exception and discontinuous spans limit progress. The work provides a practical, extensible framework for contract review automation and highlights directions for extending to other contract types and refining hypothesis design.

Abstract

Reviewing contracts is a time-consuming procedure that incurs large expenses to companies and social inequality to those who cannot afford it. In this work, we propose "document-level natural language inference (NLI) for contracts", a novel, real-world application of NLI that addresses such problems. In this task, a system is given a set of hypotheses (such as "Some obligations of Agreement may survive termination.") and a contract, and it is asked to classify whether each hypothesis is "entailed by", "contradicting to" or "not mentioned by" (neutral to) the contract as well as identifying "evidence" for the decision as spans in the contract. We annotated and release the largest corpus to date consisting of 607 annotated contracts. We then show that existing models fail badly on our task and introduce a strong baseline, which (1) models evidence identification as multi-label classification over spans instead of trying to predict start and end tokens, and (2) employs more sophisticated context segmentation for dealing with long documents. We also show that linguistic characteristics of contracts, such as negations by exceptions, are contributing to the difficulty of this task and that there is much room for improvement.

Paper Structure

This paper contains 26 sections, 5 equations, 5 figures, 11 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. ContractNLI Dataset
  3. Task Formulation
  4. Data Collection
  5. Data Statistics
  6. Span NLI BERT for ContractNLI
  7. Experiments
  8. Baselines
  9. Experiment Settings
  10. Results
  11. Discussion
  12. Controlled Experiments
  13. Is Hypothesis Information Useful?
  14. Can Better Evidence Identification Lead to Better NLI?
  15. Challenges of ContractNLI
  16. ...and 11 more sections

Figures (5)

  • Figure 1: An overview of document-level NLI for contracts. Given a contract, a system must classify whether each hypothesis is entailed by, contradicting to or not mentioned by the contract and identify evidence for the decision as spans in the contract.
  • Figure 2: The NLI label distribution. Hypothesis names are used only for a human readability purpose. A full list of hypotheses can be found in \ref{['tab:hypotheses']}.
  • Figure 3: Number of evidence spans in each document.
  • Figure 4: Model architecture of proposed Span NLI BERT
  • Figure 5: Question answering with evidence annotation interface