How JP Morgan Chase Slashed Legal Costs
- JP Morgan Chase wanted to automate its processes for reviewing commercial-loan agreements.
- Feature Selection is a useful machine learning concept, helping place a text into suitable categories.
- The bank used Feature Selection to tell apart different types of contracts.
- As a result, it saved over 360,000 hours of menial legal work.
Contract reviews were becoming a problem
JP Morgan Chase set out to improve the process of reviewing commercial-loan agreements.
Having employees menially review thousands of contracts annually was straining the bank’s human resources and limiting potential growth. The bank also sought the added benefit of decreasing loan-servicing mistakes stemming from human error.
Concept: Feature Selection
Imagine wanting to explain how to tell apart oranges from apples to someone who’s never seen fruit.
Colour alone is not enough (oranges can be green), nor is size (apples can be big).
Both colour and size are features, and to successfully tell apart apples from oranges, you’ll need more than just these two. Ideally, you’d like the number of features to be low.
In machine learning, Feature Selection is the process of finding the best features for helping a machine separate one group from another. These can often be very different to what humans would find helpful.
JP Morgan’s researchers have identified 150 different features used to tell apart different types of contracts. As the example here shows, sometimes the text can offer clues for useful features:
1. What’s the text before the first full stop?
2. Does the clause feature crucial words, like ‘amount?
3. Are there numbers? Is there a currency symbol?
Feature selection speeds up contract reviews
The bank implemented a program called COIN, which stands for Contract Intelligence.
The software identifies and categorizes repeated clauses. It does so by classifying clauses according to about 150 different “attributes”. (Also known as features, see NLP Concept above.) COIN analyses contract documents to find words or phrases relevant to these attributes.
Based on these, the system extracts from the contract the relevant sections warranting human review. If the system fails to analyze a contract, it directs it to human reviewers, for them to manually search the document.
The bank reported that the solution saves lawyers and loan officers work 360,000 hours annually.