Original URL: http://www.theregister.co.uk/2006/08/08/ibm_avatar/
IBM gets handle on unstructured data
Almaden incarnates search and BI
It is perhaps easy to assume that the notion of BI (business intelligence) for the masses - or 'DIYBI', as espoused here, is most likely to involve a sawn-off version of an existing BI tool - probably a mature one where the development costs have already been recovered.
In practice, this is somewhat less than likely, if only because most corporate BI tools are focused on working with structured data which is more found more extensively in large enterprises. It is already well understood that the majority of data in play in any business is unstructured and therefore not immediately well-suited to BI manipulations - and this is even more likely in the smaller businesses for which DIYBI might be attractive.
Developing technologies that can not only work with unstructured data but actually extract information from it of real value for the user is an important step along the way to DIYBI for the masses, but it also involves technology that goes well beyond what might be called a typical BI tool of today.
At one level, the basics of the 'DIY' toolset already exist in the form of the search engines such as Google, Yahoo! and MSN among many others. But search is a very minor part of BI, and in any environment where there is an embarrassment of unstructured data riches, can by itself be more of a hindrance than help.
The key here, according to Nelson Mattos, vice president of information & interaction research at IBM's Almaden Research Laboratories in California, is the ability to provide semantic analysis on both structured and unstructured data within a common environment.
"Users want to find whatever it is they want, regardless of where it is stored," he said. "They also want to continue to working with the tools they have and know, such as spreadsheets and Powerpoint presentations."
To that end, user interfaces are of equal importance to them, which means that search engine technologies have now become the UI of choice, says Mattos. "Everyone is already familiar with the search engine model - everyone can type a few words and get a result - and I want to use that paradigm in the context of a business intelligence environment."
IBM's research work is not intended to service the DIYBI market but it fits in with the notion of BI for the masses, for it is designed to support users running tasks associated with their jobs rather than be a tool for BI specialists. This is the target for Project Avatar, currently under development at Almaden.
Its object is to provide the tools that allow to users extract insight out of both structured and unstructured data. "This is a very broad area that requires management of structured and unstructured data and the use of traditional search technologies," Mattos said, "though that is not sufficient, because if you look at data across the enterprise there are huge amounts, so unless there is some intelligence to help find the insights we just overwhelm users with the amount of information."
Mattos sees search engines and BI coming together as the world of unstructured data is reeled in to the business need for intelligence. “At the moment 80-85 per cent of the data stored on computers is unstructured,” he said, “so it would be good to have a common framework that will allow users to analyse a record or historical data to identify problems, issues or trends – analysing structured and unstructured data together.”
The paradigms used to interface with that combined framework may look like a search engine but they will be different. The text related search technologies have been, until recently, solely built on keywords not semantics. There have been some sophisticated algorithms developed that can look at keywords in a context, recognising company names, technical components and the like. But they were developed in a proprietary fashion and that, Mattos suggests, is why they have never taken off.
“A few years ago IBM got together with a few agencies from the US Government, plus several research institutions, where it was decided to develop a standard framework which would allow all the proprietary algorithms to be plugged in so that others could take advantage of them,” he said. “It was announced last year that we were going to give the implementation of this framework (called Unstructured Information Management Architecture - UIMA) to the open source community. It can now be used to use search arguments to identify documents where a connection has been inferred, even though none of the keywords of the search argument are in them.”
Project Avatar puts a layer of intelligence on top of the primitive interface layer. This performs semantic analyses of search requests in order to surface more comprehensive results. To demonstrate, Mattos used the simple example of searching for the words 'John' and 'Phone'.
"This will normally lead to all documents containing those words. But Avatar will infer the likelihood that the required answer is actually John's phone number. The system will then search for a document containing John's phone number, even if it does not contain those two words together. I may even use some description that I know I can associate with 'John' and the system will find a string that it associates with 'phone number'."
According to Mattos, the objective here is to use intelligence in the search interface to allow users to start looking for concepts rather than keywords. From here it is then possible to start extracting the concepts out of documents, which in turn will allow users to start generating information even before anyone has asked for specific facts. An example would be examining records from a call centre and being able to determine the percentage of calls complaining about quality and/or new maintenance contract terms, and from that rapidly pinpoint areas in customer and product support that need to be addressed to improve the customers’ experience.
This is, arguably, the essence of BI, surfacing possible answers to questions that can impact the ongoing performance of a business, before the business user has even formulated them.
One of the most important types of unstructured data is the voice, and Almaden has already developed speech recognition that can record the voice of the caller and do speech to text conversion in any of the major languages. "We can even do language to language translation, so could take worldwide records, translate them all into English and do analysis on the results. This information can then be incorporated with typed records.
An interesting side issue here is the added ability to analyse customers' spoken interactions with company staff to assess factors such as customer satisfaction. "This is done using sentiment analysis, which can only be obtained from the voice, whether someone was angry, upset or whatever, not the text,” Mattos said. "That can't be done in real time, however."
From a users' point of view, Project Avatar will make it possible to have a single, customer-defined `company-standard’ UI to make inquiries of both structured and unstructured data, where users will be able to search for concepts.
As this is now out in the open source community, it is likely that DIYBI tools will start appearing that are a mixture of search engine and BI tool. It is unlikely to come from IBM of course, though the company has delivered a search engine for enterprises called Omnifind which is built on the technology. ®