The media, software and data firm Bloomberg L.P. is well-known both for its computer terminals, which provide real-time financial data to traders, and for its news services. The website bloomberg.com maintains publicly accessible profiles on around 77,000 public and 1.8 million private companies. These one-page snapshots summarise the business activities of a firm, provide contact details, list key executives and include links to press releases and recent media articles. Bloomberg also has profiles on business people, linked to the company profiles.
Recently, we fed the Bloomberg public company data to our API, serving results in a stripped-down user interface:
Only the text contained under the heading “profile” is used, stating the main activities of the company. Clicking on a recommendation thus returns companies with similar activities. One thing to note: Lateral serves results based purely on what’s written in the company profile texts, and pays no mind to the “obviousness” of the suggestion (the brand recognition of the company to humans, for example).
This demo shows the Lateral content recommender API in action, in an example of interest to the finance sector. The use case we had in mind was that of an investor or analyst building an overview of the publicly-listed players in a particular market. The traditional research workflow in such instances might involve repeated Google keyword searches and link-following. Our approach is simpler: just find an appropriate chunk of seed text to get started — an outline of the activities of a solar cell manufacturer, for instance — ask the Lateral API for thematically similar documents, and you’re away.
Of course, the API doesn’t just work for business-related documents: it can recommend from any collection of documents you care to feed it. Indeed, the core algorithm behind the Lateral recommender API has learned to model themes autonomously, just by “reading”, during a training phase in which it distilled thematic patterns from tens of millions of pages of English text. The resulting model can read any text, and give sensible recommendations from any corpus.