W. Bruce Croft
Center for Intelligent Information Retrieval
Computer Science Department
University of Massachusetts, Amherst
Amherst, MA 01003-4610
[email protected]
D-Lib Magazine, November 1995
With the enormous increase in recent years in the number of text databases available on-line, and the consequent need for better techniques to access this information, there has been a strong resurgence of interest in the research done in the area of information retrieval (IR). For many years, IR research was done by a small community that had little impact on industry. Most applications of text retrieval focused on bibliographic databases, and the large information services such as DIALOG or WESTLAW were based on standard Boolean logic approaches to text matching and paid little attention to the results of research on topics such as retrieval models, query processing, term weighting and relevance feedback.
Today, however, the situation is considerably different. Retrieval techniques based on IR research have found their way into major information services (for example, West Publishing's WIN system, Individual's clipping service) and the World Wide Web (for example, InfoSeek and Lycos). Many of the features once considered too esoteric for the typical user, such as "natural language" queries, ranked retrieval results, term weighting, "query-by-example", and query formulation assistance, have become common and, indeed, necessary in most IR products (for example, PLS, Verity and Fulcrum).
Given the speed with which industry has adopted the results of IR research from the 1970s and 1980s, the IR community is faced with identifying major new directions. The emergence of new applications such as "digital libraries" is both an opportunity and a challenge. These applications provide unique opportunities as testbeds for evaluating and stimulating research, but the challenge for IR researchers is to define and pursue research programs that maintain their relevance in a rapidly changing environment. One problem is that the priorities that IR researchers place on research issues are not necessarily the same as those of companies and government agencies that use and sell IR systems. Understanding those priorities and the operational experience behind them will be part of the process of deciding which issues are of fundamental importance and which are more transient.
In this paper, I summarize the experience of the National Science Foundation (NSF) Center for Intelligent Information Retrieval (CIIR) in the area of industrial and government research priorities. The Center has more than 40 members from the computer and information industries, applications areas such as medicine and environmental technology, and a variety of government agencies. These members participate in a variety of research and technology transfer projects, and the CIIR supports a number of prototype and operational retrieval applications, such as THOMAS and American Memory at the Library of Congress, InfoSeek, STAT/USA at the Department of Commerce, the Lotus Help Desk, and the U.S. Business Advisor.
The following list describes ten of the most important issues we have encountered during our interactions with CIIR members (apologies to David Letterman). They are listed in approximate reverse order of importance, based purely on my own assessment.
Relevance feedback is a process where users identify relevant documents in an initial list of retrieved documents, and the system then creates a new query based on those sample relevant documents. Algorithms for automatic relevance feedback have been studied in IR for more than thirty years, and the research community considers them to be thoroughly tested and effective. Companies and government agencies that use IR systems also view relevance feedback as a desirable feature, but there are some practical difficulties that have delayed the general adoption of this technique.
Most of the relevance feedback experiments reported in the IR literature were based on small test collections of abstract-length documents. The central problems in relevance feedback are selecting "features" (words, phrases) from relevant documents and calculating weights for these features in the context of a new query. These problems are substantially more difficult in environments with large databases of full-text documents. In addition, people searching databases in real applications often use relevance feedback in different ways than anticipated by IR researchers. Feedback techniques were developed to improve an initial query and assumed that a few relevant documents (all those in the top ten, for example) would be provided. In many real interactions, however, users specify only a single relevant document. Sometimes that relevant document may not even be strongly related to the initial query, and the user is, in effect, browsing using feedback.
These factors mean that traditional feedback techniques can be unpredictable in operational settings. Research aimed at correcting this problem is underway and more operational systems using relevance feedback can be expected in the near future. Relevance feedback techniques are also an important part of building profiles in a routing system (issue 6), with the main difference being the number of example relevant documents available.
Information extraction techniques, primarily developed in the context of the Advanced Research Projects Agency (ARPA) Message Understanding Conferences (MUCs), are designed to identify database entities, attributes and relationships in full text. For example, for people interested in new joint ventures, an information extraction system could identify the names of the companies involved, the new company, the products, and the location, all from articles coming over a news feed. Companies and government agencies have considerable interest in these techniques, and see them as contributing significant "added-value" to the text databases they and others generate. Potential users also see these techniques as tools to help with data analysis, browsing, and mining using text databases. The current state of information extraction tools is such that it requires a considerable investment to build a new extraction application, and certain types of information are very difficult to identify. Research in this area is focused on reducing the effort required for new applications.
Extraction of simple categories of information is, on the other hand, practical and can be an important part of a text-based information system. Examples of this type of information include company and other organization names, peoples' names, locations, and dates.
Multimedia indexing and retrieval refers to techniques being developed to access image, video and sound databases without text descriptions. The perceived value of multimedia information systems is very high and, consequently, industry has a considerable interest in the development of these techniques. General solutions to multimedia indexing are very difficult and, where they currently exist, tend to be of limited utility. An example of this is indexing images by their color distribution. This technique can be effectively used in some applications, such as retrieving pictures of fabric in specified color shades, but in many other applications simply cannot be used. Some progress has been made in multimedia indexing for specific applications (for example, retrieval of photographs of faces), and in processing language-related multimedia. Examples of language-related multimedia include text in images, scanned document images, and speech. Given the number of industrial and academic research groups working in this area, steady improvement of the techniques available can be expected.
The development of effective retrieval techniques has been the core of IR research for more than 30 years. A number of measures of effectiveness have been proposed, but the most frequently mentioned are recall and precision. Finding text that satisfies a user's information need is not simple, and considerable progress has been made in developing ranking techniques that are significantly more effective than Boolean logic.
Contrary to some researchers' opinions, companies that sell and use IR systems are interested in effectiveness. Having a more effective retrieval engine is a major selling point. It is not, however, the primary focus of their concerns and I have indicated this by the quite low ranking of this issue in the top 10. With regard to effectiveness, companies are particularly interested in techniques that produce significant improvements (rather than a few percent average precision) and that avoid occasional major mistakes. A system that performs well on most queries but makes it difficult for users to recover from bad mistakes, or even understand why they were made, is likely to be considered unacceptable. These occasional mistakes have very little impact on the average recall/precision measures used in standard IR tests, but considerable impact on end users. An example of a technique that produces reliable (although small) improvements in effectiveness, is generally well-regarded by users, but is one of the main sources of occasional bad mistakes is stemming. Solutions include building better stemmers and doing stemming as part of query processing rather than indexing.
Information routing, filtering and clipping are all synonyms used to describe the process of identifying relevant documents in streams of information such as news feeds. Instead of comparing a single query to large numbers of archived documents, as is the case for IR, large number of archived profiles are compared to individual documents. Documents that match are sent to the users associated with the profile. A profile is a representation of a long-term information need and is usually more complex than a session-based query.
Companies and government agencies often indicate that routing is the main function required for a text-based system, with IR being a backup, secondary function. Given that, we can expect to see a significant increase in the use of routing systems.
Both efficiency and effectiveness are important for routing, and both have to be addressed in different ways than IR systems. Efficiency is important in order to deal with high-volume document streams (e.g. 100 MB/hour) and large numbers of profiles (tens of thousands). New indexing and memory-based architectures are being developed for these systems. In terms of effectiveness, the basic algorithms are very similar to retrieval but instead of producing a ranking, cutoffs must be used to separate relevant from non-relevant documents. Learning techniques are being studied both as a means of determining these cutoffs and automatically building profiles based on user feedback.
Effective interfaces for text-based information systems are a high priority for users of these systems. The interface is a major part of how a system is evaluated, and as the retrieval and routing algorithms become more complex to improve recall and precision, more stress is placed on the design of interfaces that make the system easy to use and understandable. Interfaces must support a range of functions including query formulation, presentation of retrieved information, feedback, and browsing. The challenge is present this sophisticated functionality in a conceptually simple way. Despite the importance of this issue, there has been relatively little relevant research done by either the IR or human-computer interface (HCI) communities. This, however, is changing and much more work on interfaces for information systems and information visualization will be appearing.
One of the major causes of failures in IR systems is vocabulary mismatch. This means that the information need is often described using different words than are found in relevant documents. Techniques that address this problem by automatic expansion of the query are often regarded as a form of "magic" by users and are viewed as highly desirable. Vocabulary expansion can result from transforming the document and query representations, as with Latent Semantic Indexing, or it can be done as a form of automatic thesaurus built by corpus analysis. Further research in this area will make these techniques more reliable and efficient.
One of the most frequently mentioned, and most highly rated, issues is efficiency. Many different aspects of a system can have an impact on efficiency, and metrics such as query response time and indexing speed are major concerns of virtually every company involved with text-based systems. In the past, efficiency was a secondary issue in much of the IR literature. This has changed with the increased accessibility of large, full-text databases and new algorithms designed to increase indexing and query speed are published regularly. There has also been substantial research on text compression algorithms for decreasing storage overheads and I/O times. New applications of text-based systems that involve real-time and multi-user constraints will also require more work on concurrency control, update, and recovery strategies appropriate for these applications. Web-based systems, such as Infoseek, have particularly strict efficiency requirements since they must deal with hundreds of! thousands of queries per day.
The other aspect of indexing that is considered very important is the capability of handling a wide variety of document formats. This includes both standards such as SGML, HTML, Acrobat, and WordPerfect to name a few, as well as the myriad formats used in text-based applications such as medicine, law, commerce, etc.
With the advent of the World-Wide Web and the huge increase in the use of the Internet, there has been a corresponding increase in demand for text retrieval systems that can work in distributed, wide-area network environments. This demand also comes from groupware applications such as Lotus Notes, which facilitate the rapid creation of databases distributed throughout an organization. One problem of this type is addressed by Web search engines, such as Infoseek and Lycos, which index Web pages and provide access to them. The more general problems are locating the best databases to search in a distributed environment that may contain hundreds or even thousands of databases, and merging the results that come back from the distributed search. The results must be merged in order to produce the overall ranking of retrieved items, instead of a collection of individual rankings. The difficulty of doing this effectively comes from the fact that the individual rankings may be incompatible in the sense that the numbers used to produce these rankings may not be directly comparable (they may even come from different IR systems). Research addressing these issues has begun to appear in the major conferences.
The work in this area, and that done in routing and relevance feedback, is closely related to the work on multi-agent systems that has received a lot of attention.
The most important problem from the point of view of companies using and selling text-based systems is integration with other systems. A text retrieval system is a tool that can be used to solve part of an organization's information management problems. It is not often, however, the complete solution. Typically, a complete solution requires other text-based tools such as routing and extraction, tools for handling multimedia and scanned documents such as OCR, a database management system for structured data, and workflow or other groupware systems for managing documents and their use in the organization. Currently, these systems must be integrated using customized software, and even then the integration is often very superficial. More work on standardized architectures and common platforms is needed. Examples of some of the standards efforts include the Z39.50 search protocol, the SQL-Multimedia proposals, and the ARPA TIPSTER architecture for integrating retrieval, routing, and extraction systems.
One of the most important aspects of developing common platforms is the integration of database management and IR systems. An effective integration of these systems, together with multimedia capabilities, would provide an information system that could be used to manage many of the applications that currently exist. Although partial solutions to this integration do exist, they do not address one of the fundamental issues, which is that database systems all retrieve using Boolean logic, whereas both text and multimedia require techniques involving uncertainty and ranking for effective retrieval. True integration of text and multimedia is likely to require significant changes in the standard database techniques for indexing and query optimization, and may even require new query languages. Research on these issues is underway and more effort and support for this work is likely.
This list of ten issues is not meant to be comprehensive. There are other research areas that are of significant interest to companies, such as multilingual IR, data mining in text databases, and text categorization (attaching categories to text from predefined set, such as news categories or diagnosis codes). Although the ranking of these issues may not coincide with the priorities that IR researchers (including myself) may put on their own research programs, there is almost total agreement that these issues all pose significant research challenges. Some of these challenges will be solved in the short-term, but others will be the basis of longer research projects. The opportunities and experience provided by the explosion of operational text-based systems will be invaluable for IR research.
TheCIIR Bibliography describes papers that have more details on the related research being done at the CIIR, as well as pointers to other peoples' work.
hdl://cnri.dlib/november95-croft