BERT, NLP, and the Future of Enterprise Search

For three decades, enterprise search has been fundamentally broken. Users type keywords into a box. The system pattern-matches across documents and returns results ranked by relevance scores that often miss the point entirely. A finance manager searching for "cash flow problems" gets documents containing those words, not documents actually about cash flow issues.

Then in February 2018, Allen Institute's ELMo paper showed what was possible when you use deep language understanding instead of keyword matching. I read it late one night, then read it again the next morning. The implications for enterprise software are enormous.

The traditional search problem isn't really a search problem. It's a language understanding problem. When you search, you're not asking the system to find words; you're asking it to understand what you mean. Most systems fail spectacularly at this because they don't understand language. They understand patterns.

ELMo is the first system that genuinely understands context. The same word means different things in different contexts—"bank" means something different in a financial document versus a geology report. ELMo's bidirectional learning captures that context. It reads the words before and after to understand what's actually happening.

Why This Matters for Enterprise

Imagine an invoice management system where instead of searching for "invoice number 45782," you could search for "all invoices to vendors we stopped using last year." Or a manufacturing system where instead of "lot number ABC-123," you could ask "what materials from this supplier are currently in inventory?" The system would understand the semantic meaning of your query, not just match keywords.

This is enormous because enterprise data is messy. Purchase orders are formatted differently across departments. Customer records contain errors and duplicate entries. Inventory descriptions are inconsistent. Systems that rely on precise matching fail constantly. Systems that understand language can navigate inconsistency and ambiguity the way humans do.

The capability gap between keyword search and semantic search is similar to the gap between a library card catalog and a librarian. One can find books if you know the exact title. The other understands what you're looking for and helps you find what you actually need.

We're starting to build search into some of our ERP systems by integrating these language models. Instead of users memorizing part numbers or invoice codes, they can query in natural language. "Show me the inventory items we've been ordering more of lately" becomes a conversation instead of a database query that requires technical knowledge.

The training data question is thorny for enterprises though. These language models are trained on general internet text. They understand common English, but they don't understand your industry jargon, your abbreviations, your specific data model. The next wave will be fine-tuning these models on enterprise data, which requires solving for privacy and security.

But the direction is clear. Enterprises are drowning in data they can't effectively search because search has been keyword-based for decades. Language models solve this. By the end of 2018, this shift from keyword matching to semantic understanding won't be cutting-edge anymore—it'll be table stakes.