A neural network that reads contract clauses and rates them Low, Medium or High risk, the way a legal reviewer would. It was trained on clauses from 1,300 reviewed contracts and tested on the clauses of this portfolio's 200 live contracts.
1D convolutional neural network (Keras)Compared with Naive Bayes, logistic regression, SVM, MLP, BiLSTMOcclusion explainabilitySynthetic clauses, patterned on Indian construction contracts
Pick a portfolio clause or paste your own. The network runs in your browser; nothing is sent anywhere.
Party names are normalised to "employer" and "contractor" before rating, exactly as in training.
How the models compare
Results on the 800 portfolio clauses, which no model saw during training or tuning. Macro-F1 weighs Low, Medium and High equally; High recall is the share of truly high-risk clauses caught.
Contract review worklist
Clause ratings rolled up to the 200 portfolio contracts. Select a contract to see its clauses.
How it works
6,539 historical clauses rated by six reviewers; 26 unreadable scans removed; party names normalised.
Trained on contracts signed 2019 to Sep-2024 and tested on the 200 portfolio contracts (Oct-2024 onwards). Tuning used 5-fold cross-validation grouped by contract.
The 1D-CNN learns 64-number word embeddings, then 96 filters scan three-word windows for risky phrases.
Explanation by occlusion: each word is blanked out in turn, and the drop in the probability of High shows how much the network relied on it.
Key terms (payment days, notice periods, retention, escalation) come from rule-based extraction.
Honest limits
Synthetic clauses are more regular than real contracts; accuracy must be re-measured on a client's own reviewed clauses.
Reviewer labels are noisy: cross-validated agreement ranges from 85% to 89% by reviewer, so part of the remaining error is human disagreement.
Words the network never saw in training are treated as unknown. Unusual drafting can be misread.
A screening aid for legal review, not legal advice. Every flagged clause needs a human decision.