Contract risk analysis

Client: Major bank

Goal: Increase the precision of risk assessment for lines of credits in order to reduce surcharges of risk-weighted assets

Solution:

  • Added a new dimension of text-based data to the existing number-based model
  • Automatically classified key aspects in contracts, e.g. 100 of the most common negative covenants
  • Correlated with past contract performance
  • Created an improved risk assessment model

Classification of messages

Client: Major bank

Goal: Automatically categorize messages into classes

Solution:

  • Generated flexible classifier fingerprints (with keywords, bags of words or sample texts)
  • Implemented an intelligent routing system
  • Achieved 50% higher recall with same precision compared to production system
  • Enabled immediate evaluation of changes to the system through real-time processing

Understand your customer

Client: Major bank

Goal: Extract topics from different data sources (e.g. emails, social media) and determine customers’ intents

Solution:

  • Extracted topics from text by filtering via meaning
  • Determined intent based on topics and sub-topics
  • Routed to correct department to take action
  • Permits statistical analysis of customer feedback

Voice of the customers

Client: Large media company

Goal: Monitor in real time what customers are saying

Solution:

  • Converted the Twitter firehose into a stream of semantic fingerprints
  • Created one filter per user (made possible by low processing requirements)
  • Compared the stream of fingerprints to the pre-defined filter fingerprints
  • Generated a real-time content sub-stream for each user

Clustering

Client: Car industry

Goal: Improvement of data analytics models with information from free text sources

Solution:

  • Intelligent system for clustering of unstructured free text and word group extractions
  • Analyzes very specific, car-related vocabulary
  • Makes fine-grained distinctions between each car-related topic within the vocabulary
  • Detects similarities between car-related topics that are connected on the technical level
  • Works even with very short and frequently misspelled texts

Document search

Client: Financial services company

Goal: Find documents based on meaning

Solution:

  • Created custom semantic space (Retina) for finance
  • Applied Retina semantic search
  • Query fingerprint is compared to fingerprints index in the Finance Retina
  • Results are ranked according to their semantic similarity with the query
  • Delivered fewer false positives and better recall

Topic detection

Client: Large media company

Goal: Identify topics over time

Solution:

  • Extracted minute-by-minute topics from closed captions of 2-hour TV shows
  • Topics were extracted based on the meaning of what is said - not keywords
  • Correlated the topics with the real-time viewership
  • As a result, topics that lead to a drop in viewership can be immediately replaced

View screenshot