Smart Contract Analysis with NLP

At Siemens' scale contracts arrive constantly, and every one passes through the legal team before signature. The team corrects the same clauses for the same reasons over and over: liability caps, payment terms, penalties, jurisdiction. That knowledge exists in the organisation - but only in a few experienced heads and in old files.

The result is two risks, both bad: slowness and inconsistency. Reading a new contract end to end takes hours, and whether a clause gets caught depends on who happened to review it that day.

The challenge was to move that accumulated judgement out of individuals and into a system.

I built a contract review system that learns from the organisation's own contract history. The idea is simple: instead of hand-coding the policy, look at what the legal team actually changed in the past.

1. I extracted a memory from past revisions. Five real contract sets were read in both their original draft and revised form (Word documents with tracked changes). The two versions were compared sentence by sentence to identify exactly what changed, building a corpus of "original clause → corrected clause" pairs.
2. Matching by meaning, not by words. Each clause was converted into a vector using a sentence-embedding model, so the system also catches a clause written in different words that carries the same risk - in legal text, the same provision is rarely written the same way twice.
3. New contracts are read against that memory. When a document is uploaded it is split into sentences, each compared against the revision corpus, and close matches are flagged - alongside the edit previously made to that clause, offered as a suggestion.
4. An interface built for the reviewer. A Streamlit app with document upload, flagged clauses in a sortable table, and a diff view of the proposed change. Reviewers focus on what the system surfaces instead of reading the whole contract.
5. A classical comparison line. To show the semantic approach was genuinely needed, results were compared against classical word-count classifiers (logistic regression, SVM).

What emerges is not a rule list but the organisation's contract memory: decisions the team made over years become automatically reusable on new contracts. The judgement stays with the lawyer - the system only says "we've looked at this before, and here's how."

Stack: Python, python-docx, NLTK sentence tokenisation, difflib, sentence-transformers (MiniLM) embeddings, cosine similarity, scikit-learn, Streamlit.