The SIMATIC Knowledge Manager

In this paper, we present details about the SIMATIC Knowledge Manager (SKM), a Textual CBR system that uses existing documents, such as FAQs and other useroriented documentation, and finds the most relevant documents for a given problem description. The major difference of the SKM compared to standard Information Retrieval tools and WWW search engines is that knowledge about the application domain can be brought into play when assessing the relevance of documents. Thus, not only the names of products, devices, and software components can be represented but also their relationships, such as dependencies between a series of products. Furthermore, the structure of the domain can be taken into account thus allowing a clustering of products into categories that express common properties. The SKM employs a case-based approach in that it considers the existing documents as cases and a user’s request as a query in the sense of the CBR paradigm. Also, by relying on these documents, a separate case authoring process is avoided which would require a substantial amount of both initial knowledge engineering when setting up the system as well as maintenance while the system is running. Problem Description Siemens is selling a wide range of automation systems within its SIMATIC program world-wide. Subsidiaries of Siemens as well as other companies are engaged in repairing and maintaining this equipment. To support technicians when trying to solve problems at the customer’s side, Siemens operates a hotline for second level customer support which answers telephone calls. This hotline serves about 65,000 customers world-wide and 85 employees manage approximately 13,000 calls per month. The hotline has to struggle with two major problems: ̄ Firstly, there is a huge demand for information from the clients’ side (we will refer to both external technicians as well as Siemens internal staff searching for information as clients or users). Consequently, the hotline staff is always busy and sometimes requests from clients are queued and can only be answered some hours later. ̄ Secondly, the hotline is contacted again and again because of the same problem due to different clients facing the same difficulties when maintaining SIMATIC components. As the hotline staff itself consists of 60 people, such situations are rarely recognized and, hence, reuse of problem solving knowledge hardly ever occurs. To overcome these problems, Siemens decided to utilize the increasing popularity of the World Wide Web and to provide information, such as updates of drivers or news about the latest products, via WWW pages. An immediate consequence of that decision was that some kind of system would be required allowing users to search the document collection. This was recognized as a crucial requirement for the success of that strategy because the primary question is whether or not a particular information is given in a set of documents but whether or not users would be able to find it. Due to the expected growth of the collection, a simple folder--oriented categorization very soon would have required a tremendous amount of maintenance and, at the same time, would have limited the benefit for users. Siemens very soon realized that a standard Information Retrieval (IR) approach [16] would not be appropriate despite numerous tools being available. This has the following reasons: ̄ The documents frequently refer to names of products, devices, hardware and software components. These sometimes have a kind of code, such as CP .I473 MAP, and sometimes consist of a group of words, such as USER TECHNOLOGY MODULE. Also, for a single product several names may exist, such as the commonly used name, the correct product identifier, and a code similar to the above. Representing such names in IR tools would be hard if not impossible. ̄ Products cannot be considered in isolation. Rather, relationships exist among the various products and components which should be taken into account when searching for relevant documents. For example, the above mentioned component CP 14 73 NAP appears to be highly similar to another component named CP .1430. Also, some products may be highly similar because they belong to the same series whereas another 40 From: AAAI Technical Report WS-99-10. Compilation copyright © 1999, AAAI (www.aaai.org). All rights reserved. group of products shows completely different properties. A wide range of such relationships exist and demand for means for explicitly representing this type of knowledge. ̄ SIMATIC is not a single range of products but rather consists of more than a dozen different programs. Some of the products can be used for many programs whereas others are highly specific for a single one. Again, this is a specific type of knowledge that has to be somehow represented in a search engine. ̄ Although the documents primarily consist of textual descriptions, more structured elements, such as feature values, are widely used, too. Obviously, all the above remarks indicate that a knowledge-based approach is required which utilizes the various pieces of knowledge about the domain in order to implement an assessment of documents beyond plain keyword matching. To identify appropriate technologies and tools, Siemens in Autumn 1997 started a 3--month trial in which several tools have been tested with respect to their applicability to the task, the expected costs, and the required maintenance while running the system. During this period, it also became obvious that the initial idea of letting the hotline staff use the tool would not be feasible. The reason for this is that the employees running the hotline are highly skilled and, hence, would make use of such a system only in rare circumstances if particularly difficult problems have to be solved. Of course, building a system for these situations is also highly difficult. Instead, Siemens decided to design the system for use by the clients with the objective of achieving a call avoidance at the hotline due to clients solving their problems at least partially in a self-service manner. Application Description