NLP
Natural language processing (NLP) is the field of getting software to read, interpret and produce human language — classifying documents, extracting entities and figures, answering questions, translating, summarising, detecting sentiment, and conversing. It moved from hand-written rules through statistical methods to transformer models that learned language from text at scale, and the large language models absorbed most of its tasks into one model. For a fintech it is document processing, support automation and compliance screening.
- Also known as
- Natural language processing, Natural-language processing, Text analytics, Language AI
- Used at
- document classification and extraction (contracts, statements, invoices, KYC documents), customer-service automation and routing, complaint detection and conduct surveillance, sanctions and adverse-media screening (matching names and reading news), voice-of-customer analysis, chat and voice assistants, regulatory-change monitoring, and, in trading, news and filings analysis
- Binding
- data protection on the text processed (customer communications are personal data); conduct rules on automated communications with customers; surveillance obligations (trade and communications surveillance under market-abuse rules) that NLP helps meet; the EU AI Act's transparency duties for AI interactions
- Typical tasks
- tokenisation, part-of-speech and parsing (now mostly implicit), named-entity recognition (people, companies, amounts, dates), classification, sentiment, question answering, summarisation, translation, information extraction into structured schemas, and generation; the modern stack does most of them by prompting or fine-tuning a transformer
- Typical history
- ELIZA and rule systems; statistical machine translation and hidden Markov models; word embeddings (word2vec, 2013); sequence models (LSTMs); the transformer (2017); BERT for understanding and GPT for generation (2018); and the large language models that made a general model competitive with task-specific ones on most benchmarks (2020 onward)
- Parties
- the research community, the labs whose models now define the field, the vendors of document and communications AI, and the fintech teams applying it — mostly through rented models and fine-tuned open ones
How NLP works in practice
A lender receives 30,000 customer emails a month. An NLP pipeline classifies each by intent (payment query, complaint, hardship, fraud report, general), extracts entities (account references, amounts, dates), detects vulnerability indicators the conduct rules require it to act on, routes complaints to the complaints team within the regulatory clock, drafts replies for the routine ones from templates and the customer's record, and flags the hardship cases for a human with the relevant history summarised. Five years ago this was a set of separate models — a classifier, an entity extractor, a sentiment model — each trained on labelled data; today it is one language model prompted for each task, evaluated on the same labelled sets, with a fine-tuned small model for the high-volume classification. The result is the same discipline with less engineering: an evaluation set per task, monitoring, and a human in the loop for anything that matters.
The regulated uses have their own shape. Communications surveillance reads employee messages for market-abuse indicators, where the false-positive rate decides whether the compliance team can keep up. Adverse-media screening reads news in many languages to find whether a customer is the person in the article — entity resolution under uncertainty, where a wrong match is a wrongly declined customer. Complaint detection must find complaints the customer did not label as such, because the regulator's definition is broad. Each is NLP applied with a threshold set by the cost of the two kinds of error, and the language models raised the accuracy while leaving the threshold-setting to people. For a buyer, a fintech's NLP capability is its labelled data, its evaluation sets and its pipeline discipline; the models are rented or open, and the value is in what the company knows about its own language.
For a fintech, NLP is the technology behind everything it reads and writes at scale, delivered now by general models and governed by the same evaluation discipline as any model; for an acquirer, the labelled sets and the pipelines are the asset.
Example
A payments company's compliance team screens 12,000 merchant applications a month against adverse media. The old pipeline — keyword search on names, a human reading every hit — produces 4,000 hits a month, 96% false positives, and a review team of eight. The new one: a language model reads each article and answers whether it concerns the applicant entity (matching on name, location, industry and dates from the application), what the allegation is, and how severe; a fine-tuned small model pre-filters obvious non-matches; every decision is logged with the article and the reasoning. Hits requiring review fall to 600 a month at 40% false positives; the team falls to three; true adverse findings rise from 110 to 150 a month because the model reads articles the keyword search missed. The evaluation set — 800 hand-reviewed cases — is rerun monthly, and the regulator's review of the screening process reads its results.
Common mistakes
- Building task-specific models where one prompted or fine-tuned model would do. - No labelled evaluation set, so accuracy is a feeling. - Thresholds set by convenience rather than by the cost of each error. - Automated customer communications without a human path for what matters. - Reading NLP vendors' accuracy claims without a test on the company's own text.