Latest-generation knowledge management solutions for the enterprise have been gaining traction of late, led by high-profile players such as Autonomy. Given that many of the well-known solutions for organizing information in corporate intranets promise near-total automation of the process, it’s interesting to hear how seriously many enterprise taxonomy software vendors are taking the human factor.
Peter Morville, a specialist in information architecture, recently commented: “We love information technology because it has made our jobs necessary by enabling the creation and connection of tremendous volumes of content, applications and processes. We hate IT because it constantly threatens to replace the need for us.” Although advocates of human intervention in the classification of online content, such as the founders of the Open Directory Project, are fond of slogans such as “humans do it better,” it seems as if some taxonomy software vendors are bent on removing humans from the process entirely. Somewhere in between, there exist hybrid solutions which allow either humans or software to tag and classify documents, depending on the company’s specific needs. (See, for example, “Portal Maximizer Helps Portals Soar“).
Quiver, a recent arrival on the scene, is a perfect example of the evolution of thinking in this field. At its inception, Quiver’s main intellectual property was algorithm-related. Avi Segal and Ofer Mendelevitch co-founded Quiver and continue to oversee the development of technologies that can classify information in a range of 70-80% accuracy. The “first pass” through the Quiver classifier is a “naive Bayes” analysis which creates “clusters,” and the second pass involves a “Kleinberg filter” to refine the results. (Still with us?) This continues to be a key part of Quiver’s offering, and the company has assembled an impressive advisory board of computer scientists to ensure that the technology stays current with the thinking in the field.
There are good reasons that companies need some automation in classifying information. While a global consulting firm such as Ernst and Young can hire seventy full time knowledge managers to oversee the task, many companies cannot justify such an outlay.
However, Quiver management now emphasizes the inadequacies of the very technologies which served as the foundation for its earlier beta product release. “If you want to get anything close to 100% accuracy in taxonomy,” argues Andrew Feit, Quiver’s Executive VP of Sales and Marketing, “you need a way to integrate live human input.” Even with a very simple taxonomy building tool such as Gossamer Threads, argues Feit, one can be relatively certain that there are “no problems with the information if someone you trust put it there.” But there can be scalability problems with lower-end, all-human-input solutions.
Quiver’s new product, the QKS Classifier (QKS= “Quiver Knowledge Suite”), is a hybrid system that allows document management on a large scale. For most applications, the automated filters are still used at the outset, and from there, depending on the set of “business rules” that have been set up to apply to documents of different types, employees may be involved in the taxonomy process.
In other words, Quiver’s most recent efforts have been more focused on solving specific pain points for those administering corporate intranets, specifically those relating to workflow. In fact, according to Feit, Quiver has placed the majority of its recent efforts into designing the workflow and human factors sector of its process. In working with development partners, Quiver learned that the most efficient solution for many corporations would be to involve human input heavily in some instances, and not at all in others. “For some kinds of documents,” points out Feit, “you really need 100% accuracy. But for other things, companies may not want to tie up staff time and decide that 70% or 80% accuracy is ‘good enough'”.
To perform the initial task of spidering all the content in the corporate intranet, Quiver chose Inktomi after experimenting with in-house technology and reviewing several potential partners. Inktomi will also help Quiver to market its QKS line of products, with the Classifier being the first one released. In general, the QKS is being developed as an API which can plug into various search engines and work well in various environments. Cutting edge features include “workflow memory” – in which documents that change slightly don’t need to be reconsidered by a live editor – and feedback loops and learning systems intended to allow human input to train the automated classifier to better perform its job.
Quiver’s hybrid emphasis seems to fit well with what some respected analysts are saying about knowledge management challenges. As part of an overview of various software offerings of relevance to information architecture in general, Peter Morville comments that buyers of “automated category generation” products (Semio Taxonomy, Autonomy Portal-in-a-Box) should “proceed with great caution,” whereas his group sees “great promise” in “software that leverages human-defined rules or pattern matching algorithms to automatically assign index terms to documents.” Products such as Autonomy Categorizer and Inxight Categorizer seek “to integrate human expertise in designing taxonomies with software that populates those taxonomies quickly, consistently, and inexpensively.” Yet even these have drawbacks that may not be overcome without more human input.
In short, products seeking to leverage corporate knowledge assets must come to terms with the fact that in a large corporate environment, ownership of the taxonomy is typically distributed and decentralized. It must be possible for the relevant employees to become as involved as necessary in managing the information.
Many of the companies now working on hybrid solutions, then, may still place too much emphasis on the automation aspect. Given the obvious human factors which go into knowledge management for companies dealing with the proper classification of and controlled access to complex, high-level information, Quiver’s decision to devote more resources to workflow and business rules issues – in short, the human factors – seems sound. QKS Classifier is currently considering applications for its beta program, and half a dozen high profile development partners, including a top three investment bank and a government systems integrator, are providing feedback on the product. The 1.0 version of the product is expected to launch August 8, 2001.