What is Trampoline SONAR

Trampoline’s SONAR platform brings a fresh approach to information management, by harnessing the social behaviour that occurs within organizations. SONAR plugs into the corporate network and connects to existing systems, including email servers, contact databases and document stores. It analyses this data to map social networks, information flows, expertise and individuals’ interests throughout the enterprise. Finally, a truly useful application for social networking, or rather professional networking! Rather than the LinkedIn-like approach, where users need to keep their profiles updated, SONAR is an automatic knowledge mapping system.

SONAR allows an organization to leverage the embedded intelligence of the whole community, as well as information stored electronically. Unlike traditional knowledge management applications, SONAR takes account of the ‘soft’ or hidden knowledge assets contained within an organization – for example the expertise employees pick up during their tenure at a company, but which is never formally recorded. With SONAR, individuals get the information they need instantly, unrecognized expertise becomes visible, the enterprise increases the reuse and value of its knowledge assets, and the firm improves its agility and competitiveness.

At the core of SONAR’s technology is a combination of social network analysis and natural language processing, which unites information contained in multiple electronic forms and the social behaviour surrounding it.

SONAR works around two key concepts: Connections and Themes. Connections are the people who users communicate with. They might be project team mates, other employees or contacts in outside organizations. Themes are topics of interest. They might be projects, deals or areas of expertise – whatever people are communicating about.

About SONAR’s Theme Extraction Technology

Theme Extraction is one of the main services provided by SONAR's Intelligence Core module. It automatically identifies the expertise and interests of individuals and groups inside an enterprise, as well as external contacts in communication with the enterprise.

SONAR’s Theme Extraction method is designed to maximize the value of results and efficiency of deployment in enterprise environments. SONAR Theme Extraction functions without the need for lengthy consultancy to compile a taxonomy at the deployment stage, so the system begins to give value as soon as it is connected to an email server, instant messaging platform or document repository.

Inside Theme Extraction

SONAR uses multiple language models to enable comparisons between the results it extracts from text. Core measures focus on the pair of characteristics ‘phraseness’ and ‘informativeness’ – i.e. whether a string of words is a meaningful phrase rather than a chance juxtaposition, and whether a combination of words has significance within its context. The combination of phraseness and informativeness into a single score produces a measure of ‘keyphraseness’, the extent to which a phrase encapsulates a core idea of the text. SONAR scans pairs and triplets of words in their context, ranks them by keyphraseness and displays the top results to the user as themes.

Foreground and background content

The system is predicated on an understanding that the corpus in question is composed of ‘foreground’ and ‘background’ content. The foreground content is the text in focus at a particular moment, and the background is the wider group of data from which it is taken. For example, a single email can be seen as foreground or focus information, with the background being the user’s entire email archive; or the foreground could be the user’s entire email corpus in comparison to the background of the whole enterprise’s email content. Additionally, to expedite the highlighting of relevant information, function words with little meaning - such as ‘and’, ‘it’ or ‘she’ - are filtered out at the start of the process.

Language models

Theme extraction is facilitated by SONAR’s language models. In simple terms, the language models take a body of text and decompose it into blocks of words for analysis. The size of the blocks which a language model analyses give the model its name. A model which analyses individual words is 'order 1' (1-gram), a model which counts the frequency of pairs is 'order 2' (2-gram), and so on. In a 2-gram model, the probability of the occurrence of a word is dependent on the previous word.

Establishing keyphraseness

To establish keyphraseness, SONAR undertakes two comparisons – a foreground phraseness comparison and a foreground/background informativeness comparison – which are then combined to reveal a rating of keyphraseness. The foreground phraseness comparison compares the probability of a phrase appearing in a lesser-gram model, versus a larger-gram model within the foreground data only. This could, for example, be used to pull out core phrases from an email.

The informativeness comparison compares the probability of the occurrence of phrases discovered by foreground models to background gram models. Phrases which appear more frequently within the foreground will stand out as significant to the individual user (perhaps an area of their particular expertise) and phrases which appear more in the background will fade out as insignificant (such as general discussion of company expenses policy).

By combining these results, SONAR reveals the keyphraseness of elements of information, and with it the core themes of the corpus.

Conclusion: Trampoline's Business

I think Trampoline is one of the best applications for the enterprise that leverages web 2.0 principles, in its harnessing of social behaviors. It is a complex system and fairly unique at this point. You can learn more about Trampoline and SONAR at my site, where I have an interview with CEO Charles Armstrong. We discuss SONAR's first big customer (Raytheon), their integration requirements, and the competitive landscape, as well as funding and related topics. We also discuss the Extended Enterprise in quite a bit of detail.