Meta search engine

[0001] This application is a conversion of copending provisional application 60/062,958, filed Oct. 10, 1997.

[0002] A number of useful and popular search engines attempt to maintain full text indexes of the World Wide Web. For example, search engines are available from AltaVista, Excite, HotBot, Infoseek, Lycos and Northern Light. However, searching the Web can still be a slow and tedious process. Limitations of the search services have led to the introduction of meta search engines. A meta search engine searches the Web by making requests to multiple search engines such as AltaVista or Infoseek. The primary advantages of current meta search engines are the ability to combine the results of multiple search engines and the ability to provide a consistent user interface for searching these engines. Experimental results show that the major search engines index a relatively small amount of the Web and that combining the results of multiple engines can therefore return many documents that would otherwise not be found.

[0003] A number of meta search engines are currently available. Some of the most popular ones are MetaCrawler, Inference Find, SavvySearch, Fusion, ProFusion, Highway 61, Mamma, Quarterdeck WebCompass, Symantec Internet FastFind, and ForeFront WebSeeker.

[0004] The principle motivation behind the basic text meta search capabilities of the meta search engine of this invention was the poor precision, limited coverage, limited availability, limited user interfaces, and out of date databases of the major Web search engines. More specifically, the diverse nature of the Web and the focus of the Web search engines on handling relatively simple queries very quickly leads to search results often having poor precision. Additionally, the practice of “search engine spamming” has become popular, whereby users add possibly unrelated keywords to their pages in order to alter the ranking of their pages. The relevance of a particular hit is often obvious only after waiting for the page to load and finding the query term(s) in the page.

[0005] Experience with using different search engines suggests that the coverage of the individual engines was relatively low, i.e. searching with a second engine would often return several documents which were not returned by the first engine. It has been suggested that AltaVista limits the number of pages indexed per domain, and that each search engine has a different strategy for selecting pages to index. Experimental results confirm that the coverage of any one search engine is very limited.

[0006] In addition, due to search engine and/or network difficulties, the engine which responds the quickest varies over time. It is possible to add a number of features which enhance usability of the search engines. Centralized search engine databases are always out of date. There is a time lag between the time when new information is made available and the time that it is indexed.

[0007] An object of this invention is to improve meta search engines.

[0008] Another object of the present invention is to provide a meta search engine that analyzes each document and displays local context around the query terms.

[0009] A further object of this invention is to provide a search method that improves on the efficiency of existing search methods.

[0010] A further object of this invention is to provide a meta search engine that is capable of displaying the context of the query terms, advanced duplicate detection, progressive display of results, highlighting query terms in the pages when viewed, insertion of quick jump links for finding the query terms in large pages, dramatically improved precision for certain queries by using specific expressive forms, improved relevancy ranking, improved clustering, and image search.

[0011] These and other objectives are attained with a computer implemented meta search engine and search method. In accordance with this method, a query is forwarded to a number of third party search engines, and the responses from the third party search engines are parsed in order to extract information regarding the documents matching the query. The full text of the documents matching the query are downloaded, and the query terms in the documents are located. The text surrounding the query terms are extracted, and that text is displayed.

[0012] The engine downloads the actual pages corresponding to the hits and searches them for the query terms. The engine then provides the context in which the query terms appear rather than a summary of the page (none of the available search engines or meta search services currently provide this option). This typically provides a much better indication of the relevance of a page than the summaries or abstracts used by other search engines, and it often helps to avoid looking at a page only to find that it does not contain the required information. The context can be particularly helpful whenever a search includes terms which may occur in a different context to that required. The amount of context is specified by the user in terms of the number of characters either side of the query terms. Most non-alphanumeric characters are filtered from the context in order to produce more readable and informative results.

[0013] Results are returned progressively after each individual page is downloaded and analyzed, rather than after all pages are downloaded. The first result is typically displayed faster than the average time for a search engine to respond. When multiple pages provide the information required, the architecture of the meta engine can be helpful because the fastest sites are the first ones to be analyzed and displayed.

[0014] When viewing the full pages corresponding to the hits, these pages are filtered to highlight the query terms and links are inserted at the top of the page which jump to the first occurrence of each query term. Links at each occurrence of the query terms jump to the next occurrence of the respective term. Query term highlighting helps to identify the query terms and page relevance quickly. The links help to find the query terms quickly in large documents.

[0015] Pages which are no longer available can be identified. These pages are listed at the end of the response. Some other meta search services also provide “dead link” detection, however the feature is usually turned off by default and no results are returned until all pages are checked. For the meta search engine of this invention however, the feature is intrinsic to the architecture of the engine which is able to produce results both incrementally and quickly.

[0016] Pages which no longer contain the search terms or that do not properly match the query can be identified. These pages are listed after pages which properly match the query. This can be very important—different engines use different relevance techniques, and if just one engine returns poor relevance results, this can lead to poor results from standard meta search techniques.

[0017] The tedious process of requesting additional hits can be avoided. The meta search engine understands how to extract the URL for requesting the next page of hits from the individual search engine responses. More advanced detection of duplicate pages is done. Pages are considered duplicates if the relevant context strings are identical. This allows the detection of a duplicate if the page has a different header or footer.

[0018] U.S. Pat. No. 5,659,732 (Kirsch) presents a technique for relevance ranking with meta search techniques wherein the underlying search engines are modified to return extra information such as the number of occurrences of each search term in the documents and the number of occurrences in the entire database. Such a technique is not required for the meta search engine of this invention because the actual pages are downloaded and analyzed. It is therefore possible to apply a uniform ranking measure to documents returned by different engines. Currently, the engine displays pages in descending order of the number of query terms present in the document (if none of the first few pages contain all of the query terms, then the engine initially displays results which contain the maximum number of query terms found in a page so far). After all pages have been downloaded, the engine then relists the pages according to a simple relevance measure.

[0019] This measure currently considers the number of query terms present in the document, the proximity between query terms, and term frequency (the usual inverse document frequency may also be useful (Salton, G. (1989), Automatic text processing: the transformation, analysis and retrieval of information by computer, Addison-Wesley.) $R=c_1{N}_p+\left(c_2-\frac{\sum _{i=1}^{N_p-1}\sum _{j=i+1}^{N_p}\min \left(d\left(i,j\right),c_2\right)}{\sum _{k=1}^{N_p-1}\left(N_p-k\right)}\right)/\frac{c_2}{c_1}+\frac{N_t}{c_3}$
where N_p is the number of query terms that are present in the document (each term is counted only once), N_t is the total number of query terms in the document, d(i, j) is the minimum distance between the ith and the jth of the query terms which are present in the document (currently in terms of the number of characters), c₁ is a constant which controls the overall magnitude of the relevance measure R, c₂ is a constant specifying the maximum distance between query terms which is considered useful, and c₃ is a constant specifying the importance of term frequency (currently c₁=100, C₂=5000, and c₃=10C₁) This measure is used for pages containing more than one of the query terms; when only one query term is found the term's distance from the start of the page is used.

[0020] This ranking criterion is particularly useful with Web searches. A query for multiple terms on the Web often returns documents which contain all terms, but the terms are far apart in the document and may be in unrelated sections of the page, e.g. in separate Usenet messages archived on a single Web page, or in separate bookmarks on a page containing a list of bookmarks.

[0021] The engine does not use the lowest common denominator in terms of the search syntax. The engine supports all common search formats, including boolean syntax. Queries are dynamically modified in order to match each individual query syntax. The engine is capable of tracking the results of queries, automatically informing users when new documents are found which match a given query. The engine is capable of tracking the text of a given page, automatically informing the user when the text changes and which lines have changed. The engine includes an advanced clustering technique which improves over the clustering done in existing search engines. A specific expressive forms search technique can dramatically improve precision for certain queries. A new query expansion technique can automatically perform intelligent query expansion.

[0022] Additional features which could easily be added to the meta search engine of this invention include: Improved relevance measures, Alternative ordering methods, e.g. by site, Field searching e.g. page title, Usenet message subject, hyperlink text, Rules and/or learning methods for routing queries to specific search engines, Word sense disambiguation, and Relevance feedback.

[0023] Further benefits and advantages of the invention will become apparent from a consideration of the following detailed description, given with reference to the accompanying drawings, which specify and show preferred embodiments of the invention.

[0057] One of the fundamental features of the meta search engine of this invention is that it analyzes each document and displays local context around the query terms. The benefit of displaying the local context, rather than an abstract or summary of the document, is that the user may be able to more readily determine if the document answers his or her specific query. In essence, this technique admits that the computer may not be able to accurately determine the relevance of a particular document, and in lieu of this ability, formats the information in the best way for the user to quickly determine relevance. A user can therefore find documents of high relevance by quickly scanning the local context of the query terms. This technique is simple, but can be very effective, especially in the case of Web search where the database is very large, diverse, and poorly organized.

[0058] The idea of querying and collating results from multiple databases is not new. Companies like PLS, Lexis-Nexis, and verity have long since created systems which integrate the results of multiple heterogeneous databases. Many other Web meta search services exist such as the popular and useful MetaCrawler service. Services similar to MetaCrawler include SavvySearch, Inference Find, Fusion, ProFusion, Highway 61, Mamma, Quarterdeck WebCompass, Metabot, Symantec Internet FastFind, and WebSeeker.

[0059] FIG. 1 shows the home page of the meta search engine of this invention. The bar 12 at the top contains links for the options page, the help page, and the submission of suggestions and problems. Queries are entered into the “Find:” box 14. The selection of which search engines to use for the query is made by clicking on the appropriate selection on the following line. The options are currently:

[0060] 1. Web—standard Web search engines: (a) AltaVista, (b) Excite, (c) Infoseek, (d) HotBot, (e) Lycos, (f) Northern Light, (g) WebCrawler, and (h) Yahoo.

[0061] 2. Usenet Databases—indexes of Usenet newsgroups: (a) AltaVista, (b) DejaNews, (c) Reference.com.

[0062] 3. Press—indexes of press articles and news wires: (a) Infoseek NewsWire, Industry, and Premier sources—c/o Infoseek—Reuters, PR NewsWire etc., and (b) NewsTracker—c/o Excite—online newspapers and magazines.

[0063] 4. Images—image indexes: (a) Corel—corel image database, (b) HotBot—HotBot images, (c) Lycos—Lycos images, (d) WebSeer—WebSeer images, (e) Yahoo—Yahoo images, and (f) AltaVista—AltaVista images.

[0064] 5. Journals—academic journals: (a) Science.

[0065] 6. Tech—technical news: (a) TechWeb and (b) ZDNet.

[0066] 7. All—all of the above.

[0067] The constraints menu 16 follows which contains options for constraining the results to specific domains, specific page ages, and specific image types. The main options menu 20 follows which contains options for selecting the maximum number of results, the amount of context to display around the query terms (in characters), and whether or not to activate clustering or tracking.

[0068] The options link on the top bar allows setting a number of other options, as shown at 22 in FIG. 2. These options are: 1. The timeout (per individual page download), 2. Whether or not to filter the pages when viewed, 3. Whether or not to filter images from the pages when viewed, 4. Whether each search displays results in a new window or not, and 5. Whether or not to perform image classification (for manual classification of images). Additionally, the options page shows at 24 and 26 which queries and URLs are being tracked for changes, and allows entering a new URL to track.

[0069] FIGS. 3 to 8 show a sample response of the meta search engine of this invention for the query nec and “digital watermark”. FIG. 3 shows the top portion of the response from the search. The search form can be seen at the top, followed by a tip 30 which may be query sensitive. Results which contain all of the query terms are then displayed as they are retrieved and analyzed (as mentioned before, if none of the first few pages contain all of the query terms then the engine initially displays results which contain the maximum number of query terms found in a page so far). The bars 32 to the left of the document titles indicate how close the query terms are in the documents—longer bars indicate that the query terms are closer together. The engine which found the document, the age of the document, the size of the document, and the URL follow the document title.

[0070] After the pages have been retrieved, the engine then displays the top 20 pages ranked using term proximity information (FIG. 4). In descending order, and referring to FIGS. 5 to 8, the engine then displays those pages which contain fewer query terms, those pages which contain none of the query terms, those pages which contain duplicate context strings, and those pages which could not be downloaded. Links to the search engine pages which were used are then provided, followed by terms which may be useful for query expansion. With reference to FIG. 8, the engine then displays a summary box with information on the number of documents found from each individual engine, the number retrieved and processed, and the number of duplicates.

[0071] FIG. 9 shows a sample of how the individual pages are processed when viewed. The links 40 at the top jump to the first occurrence of the query terms in the document, and indicate the number of occurrences. The [Track Page] link activates tracking for this page—the user will be informed when and how the document changes.

[0072] The engine comprises two main logical parts: the meta search code and a parallel page retrieval daemon. Pseudocode for (a simplified version of) the search code is as follows:

• [0073] Process the request to check syntax and create.. ..regular expressions which are used to match query.. ..terms
• [0074] Send requests (modified appropriately) to all.. ..relevant search engines
• [0075] Loop for each page retrieved until maximum number.. ..of results or all pages retrieved
• [0076] If page is from a search engine
  • [0077] Parse search engine response extracting hits.. ..and any link for the next set of results
  • [0078] Send requests for all of the hits
  • [0079] Send requests for the next set of results.. ..if applicable
• [0080] Else
  • [0081] Check page for query terms and create.. ..context strings if found
  • [0082] Print page information and context strings if all.. ..query terms are found and duplicate context.. ..strings have not been encountered before
  • [0083] Endif
• [0084] End loop
• [0085] Re-rank pages using proximity and term frequency.. ..information
• [0086] Print page information and context strings for pages.. ..which contained some but not all query terms
• [0087] Print page information for pages which contained no.. ..query terms
• [0088] Print page information and context strings for pages.. ..which contain duplicate context strings
• [0089] Print page information for pages which could not be.. ..downloaded
• [0090] Print summary statistics.

[0091] FIG. 10 shows a simplified control flow diagram 50 of the meta search engine. The page retrieval engine is relatively simple but does incorporate features such as queuing requests and balancing the load from multiple search processes, and delaying requests to the same site to prevent overloading a site. The page retrieval engine comprises a dispatch daemon and a number of client retrieval processes. Pseudocode for (a simplified version of) the dispatch daemon is as follows:

• [0092] Start clients
• [0093] Loop
  • [0094] Check for timeout of active clients
    • [0095] Send any queued requests if possible, balancing.. ..load for requests from multiple search.. ..processes
    • [0096] If there is a message from a client
      • [0097] If message is “replace me” replace the.. ..client with a new process
      • [0098] If message is “done” update client.. ..information
      • [0099] If message is “status” return status
      • [0100] If message is “get” then
        • [0101] If all clients are busy or a request.. ..has been made to this site.. ..within the last x seconds then.. ..queue the request
      • [0102]  Otherwise send request to a client
      • [0103]  Endif
    • [0104] Endif
• [0105] End loop

[0106] The client processes simply retrieve the relevant pages, handling errors and timeouts, and return the pages directly to the appropriate search process.

[0107] The algorithm used for image meta search in the meta search engine of this invention is as follows:

• [0108] Process the request to check syntax and create.. ..regular expressions which are used to match.. ..query terms
• [0109] Send requests (modified appropriately) to all.. ..relevant image search engines
• [0110] Loop for each page retrieved until maximum number of.. ..images or all pages retrieved
  • [0111] If page is from a search engine
    • [0112] Parse search engine response extracting.. ..hits and any link for the next set.. ..of results
    • [0113] Send requests for all of the hits
    • [0114] Send request for the next set of results.. ..if applicable
      • [0115] Else if page is an image
        • [0116] Add image to the display queue
      • [0117] Else
        • [0118] Analyze query term locations in the page.. ..and predict which (if any) of the.. ..images on the page corresponds to.. ..the query—send a request to.. ..download this image
      • [0119] Endif
      • [0120] If n images are in the display queue
        • [0121] Create a single image montage of the.. ..images in the queue
        • [0122] Display the montage as a clickable image.. ..where each portion of the image.. ..corresponding to the original.. ..individual images shows a detail.. ..page for the original image
      • [0123] Endif
• [0124] End loop
• [0125] If any images are in the display queue
  • [0126] Create a single image montage of the images in.. ..the queue
  • [0127] Display the montage as a clickable image where.. ..each portion of the image corresponding.. ..to the original individual images shows.. ..a detail page for the original image
• [0128] Endif
• [0129] Print summary statistics

[0130] FIG. 11 shows a simplified control flow diagram 60 for the image meta search algorithm.

[0131] Image Classification

[0132] The Web image search engine WebSeer attempts to classify images as photographs or graphics. WebSeer extracts a number of features from the images and uses decision trees for classification. We have implemented a similar image classification system. However, we use a different feature set and use a neural network for classification. FIGS. 12 and 13 show the response of the meta search engine of this invention to the image query koala, with the images filtered for photos. FIG. 14 shows the response when filtering for graphics.

[0133] Document Clustering

[0134] Document clustering methods typically produce non-overlapping clusters. For example, Hierarchical Agglomerative Clustering (HAC) algorithms, which are the most commonly used algorithms for document clustering (Willet, P. (1988), “Recent trends in hierarchical document clustering: a critical review”, Information Processing and Management 24, 577–597), start with each document in a cluster and iteratively merge clusters until a halting criterion is met. HAC algorithms employ similarity functions (between documents and between sets of documents).

[0135] A document clustering algorithm is disclosed herein which is based on the identification of co-occurring phrases and conjunctions of phrases. The algorithm is fundamentally different to commonly used methods in that the clusters may be overlapping, and are intended to identify common items or themes.

[0136] The World Wide Web (the Web) is large, contains a lot of redundancy, and a relatively low signal to noise ratio. These factors make finding information on the Web difficult. The clustering algorithm presented here is designed as an aid to information discovery, i.e. out of the many hits returned for a given query, what topics are covered? This allows a user to refine their query in order to investigate one of the subtopics.

[0137] The clustering algorithm is as follows:

• [0138] Retrieve pages corresponding to the query
• [0139] For each page
  • [0140] For n=1 to MaximumPhraseLength
    • [0141] For each set of successive n words
      • [0142] If this combination of words has not.. ..already appeared in this.. ..document then add the set to a.. ..hash table for this document.. ..and a hash table for all.. ..documents
    • [0143] End for
  • [0144] End for
• [0145] End for
• [0146] For n=MaximumPhraseLength to 1
  • [0147] Find the most common phrases of length n, to a.. ..maximum of MaxN phrases, which occurred.. ..more than MinN times
  • [0148] Add these phrases to the set of clusters
• [0149] End for
• [0150] Find the most common combinations of two clusters.. ..from the previous step, to a maximum of MaxC.. ..Combinations, for which the combination.. ..occurred in individual documents at least.. ..MinC times
• [0151] Delete clusters which are identified by phrases.. ..which are subsets of a phrase identifying.. ..another cluster
• [0152] Merge clusters which contain identical documents
• [0153] Display each cluster along with context from a set.. ..of pages for both the query terms.. ..and the cluster terms.

[0154] FIG. 15 shows the clusters 70 produced by this algorithm for the query “joydeep ghosh”, and FIG. 16 shows the first two cluster summaries 72 and 74 for these clusters. FIGS. 17 and 18 show the clusters 76 and 80 produced by HuskySearch for the same query. FIG. 19 shows the clusters 82 produced by AltaVista. FIGS. 20 and 21 show the clusters 84 and 86 produced by the meta search engine of this invention for another two queries: “neural network” and typing and injury.

[0155] Query Expansion

[0156] One method of performing query expansion is to augment the query with morphological variants of query terms. Word stemming (Porter, M. F. (1980), “an algorithm for suffix stripping”, Program 14, 130–137.) can be used in order to treat morphological variants of a word as identical words. Web search engines typically do not perform word stemming, despite the fact that it would reduce the resources required to index the Web. One reason for the lack of word stemming by Web search engines is that stemming can reduce precision. Stemming considers all morphological variants. Query expansion using all morphological variants often results in reduced precision for Web search because the morphological variants often refer to a different concept. Reduced precision using word stemming is typically more problematic on the Web as compared to traditional information retrieval test collections, because the Web database is larger and more diverse. A query expansion algorithm is disclosed herein which is based on the use of only a subset of morphological variants. Specifically, the algorithm uses the subset of morphological variants which occur on a certain percentage of the Web pages matching the original query. Currently, the query terms are stemmed with the Porter stemmer (Porter, M. F. (1980), “An algorithm for suffix stripping”, Program 14, 130–137.) and the retrieved pages can be searched for morphological variants of the query terms. Variants which occur on greater than 1% of the pages are displayed to the user for possible inclusion in a subsequent query. No quantitive evaluation of this technique has been performed, however observation indicates that useful terms are suggested. As an example, for the query nec and “digital watermark”, the following terms are suggested for query expansion: digitally, watermarking, watermarks, watermarked.

[0157] Currently the technique does not automatically expand a query when first entered, because the query expansion terms are not known until the query is complete. However the technique can be made automatic by maintaining a database of expansion terms for each query term. The first query containing a term can add the co-occurring morphological variants to the database, and subsequent queries can use these terms, and update the database if required.

[0158] Specific Expressive Forms

[0159] Accurate information retrieval is difficult due to the possibility of information represented in many ways—requiring an optimal retrieval system to incorporate semantics and understand natural language. Research in information retrieval often considers techniques aimed at improving recall, e.g. word stemming and query expansion. As mentioned earlier, it is possible for these techniques to decrease precision, especially in a database as diverse as the Web.

[0160] The World Wide Web contains a lot of redundancy. Information is often contained multiple times and expressed in different forms across the Web. In the limit where all information is expressed in all possible ways, high precision information retrieval would be simple and would not require semantic knowledge—one would only need to search for one particular way of expressing the information. While such a goal will never be reached for all information, experiments indicate that the Web is already sufficient for an approach based on this idea to be effective for certain retrieval tasks.

[0161] The method of this invention is to transform queries in the form of a question, into specific forms for expressing the answer. For example, the query “What does NASDAQ stand for?” is transformed into the query “NASDAQ stands for” “NASDAQ is an abbreviation” “NASDAQ means”. Clearly the information may be contained in a different form to these three possibilities, however if the information does exist in one of these forms, then there is a high likelihood that finding these phrases will provide the answer to the query. The technique thus trades recall for precision. The meta search engine of this invention currently uses the specific expressive forms (SEF) technique for the following queries (square brackets indicate alternatives and parentheses indicate optional terms or alternatives):

• • [0162] What [is|are]x?
  • [0163] What [causes|creates|produces] x?
  • [0164] What do you think [about|of|regarding] x?
  • [0165] What does x [stand for|mean]?
  • [0166] Where is x?
  • [0167] Who is x?
  • [0168] [Why|how] [is|are](a|the) x y?
  • [0169] Why do x?
  • [0170] When is x?
  • [0171] When do x?
  • [0172] How [do |can] i x?
  • [0173] How (can)[a|the] x y?
  • [0174] How does [a|the] x y?

[0175] As an example of the transformations, “What does x [stand for|mean]?” is converted to “x stands for” “x is an abbreviation” “x means”, and “What [causes|creates|produces] x?” is transformed to “x is caused” “x is created” “causes x” “produces x” “makes x” “creates x”.

[0176] Different search engines use different stop words and relevance measures, and this tends to result in some engines returning many pages not containing the SEFs. The offending phrases are therefore filtered out from the queries for the relevant engines.

[0177] FIG. 22 shows at 90 the response of the meta search engine of this invention for the query “What does NASDAQ stand for?” The answer to the query is contained in the local context displayed for about 5 out of the first 6 pages. FIG. 23 shows at 92 the response of Infoseek to the same query. The answer to the query is not displayed in the page summaries, and which, if any, of the pages contains the answer is not clear. FIGS. 24 and 25 show, at 94 and 96, the meta search engine of this invention and Infoseek responses to the query “How is a rainbow created?” Again, the answer is contained in the local context shown by the meta search engine of this invention but it is not clear which, if any, of the pages listed by Infoseek contain the answer to the question. FIG. 26 shows at 100 a third example of the response from the meta search engine of the invention for the query “What is a mealy machine?”

[0178] It is reasonable to expect that the amount of easily accessible information will increase over time, and therefore that the viability of the specific expressive forms technique will improve over time. An extension of the above-discussed procedures is to define an order over the various SEFs, e.g. “x stands for” may be more likely to find the answer to “What does x stand for” than the phrase “x means”. If none of the SEFs are found then the engine could fall back to a standard query.

[0179] Search tips may be provided by the meta engine. These tips may include, for example, the following:

• • [0180] Use quotes for phrases, e.g. “nec research”.
  • [0181] You can hide the various options above to save screen space by clicking on the “hide” links.
  • [0182] Window option: clicking on a hit brings up the page in the same window for multiple searches or a new window for each new search.
  • [0183] Filter option: filters pages when viewed to highlight query terms. Faster due to local caching of the page.
  • [0184] The letter(s) after the page titles identify the search engine which provided the result (e.g. A==AltaVista).
  • [0185] The second field after the page titles is the time since the page was last updated (e.g. 5 m=5 months, 1y=1 year).
  • [0186] The third field after the page titles is the size of the page.
  • [0187] The context option selects the number of characters to display either side of the query terms.
  • [0188] The timeout option is the maximum time to download each individual page.
  • [0189] Searching in “Press” is useful for higher precision with current news topics.
  • [0190] Image option: remove images from the pages when viewed (for faster viewing).
  • [0191] When viewing a filtered page, clicking on a query term jumps to the next occurrence of that term. Clicking on the last occurrence of a term jumps back to the first occurrence.
  • [0192] You can use “-term” to exclude a term.
  • [0193] You can search for links to a specific page, e.g. link:www.neci.nj.nec.com/homepages/giles. Self links are excluded.
  • [0194] When in doubt use lower case.
  • [0195] This meta engine makes more than three times as many documents available as a single search engine. Constraining your search can help, e.g. if you want to know what NASDAQ stands for, searching for “NASDAQ stands for” rather than “NASDAQ” can find your answer faster although the information may also be expressed in alternative ways.
  • [0196] Clicking on the search engine links in the “Searching for:” line will show the search engine response to the current query.
  • [0197] You can search for images by selecting the “images” button, e.g. “red rose”.
  • [0198] The bar to the left of the titles is longer when the query terms are closer together in the document.
  • [0199] The query term links in the “Searching for” line lead to the Webster dictionary definitions.
  • [0200] If you select Tracking: Yes, then your query will be tracked and new hits will be displayed on your customized home page similar to the “recent articles about NEC Research”.
  • [0201] Select Cluster: Yes to cluster the documents and identify common themes.
  • [0202] You can filter images using a neural network prediction of whether each image is a photo or a graphic using the Images: option.
  • [0203] A listing of pages ranked by term proximity is shown after all of the documents have been retrieved.

[0204] Tracking Queries and URLs

[0205] Services such as the Informant (The Informant, 1997) track the response of Web search engines to queries, and inform users when new documents are found. The meta search engine of this invention supports this function. Tracking is initiated for a query by selecting the Track option when performing the query. A daemon then repeats the query periodically, storing new documents along with the time they were found. New documents are presented to the user on the home page of the search engine, as shown at 102 in FIG. 27. The engine does not currently inform users if the documents matching queries have changed, although this could be added.

[0206] The meta search engine of this invention also supports tracking URLs. Tracking is initiated by clicking the [Track page] link when viewing one of the pages from the search engine results. Alternatively, tracking may be initiated for an arbitrary URL using the options page. A daemon identifies updates to the pages being tracked, and shows a list of modified pages to the user on the home page, as in FIG. 27. The [Page] link displays the page being tracked and inserts a header at the top showing which lines have been added or modified since the last time the user viewed the page (e.g. see FIG. 28).

[0207] Estimating the Coverage of Search Engines and the Size of the Web

[0208] As the World Wide Web continues to expand, it is becoming an increasingly important resource for scientists. Immediate access to all scientific literature has long been a dream of scientists, and the Web search engines have made a large and growing body of scientific literature and other information resources easily accessible. The major Web search engines are commonly believed to index a large proportion of the Web. Important questions which impact the choice of search methodology include: What fraction of the Web do the search engines index? Which search engine is the most comprehensive? How up to date are the search engine databases?

[0209] A number of search engine comparisons are available. Typically, these involve running a set of queries on a number of search engines and reporting the number of results returned by each engine. Results of these comparisons are of limited value because search engines can return documents which do not contain the query terms. This may be due to (a) the information retrieval technology used by the engine, e.g. Excite uses “concept-based clustering” and Infoseek uses morphology—these engines can return documents with related words, (b) documents may no longer exist—an engine which never deletes invalid documents would be at an advantage, and (c) documents may still exist but may have changed and no longer contain the query terms.

[0210] Selberg and Etzioni (Selberg, E. and Etzioni, O. (1995), Multi-service search and comparison using the MetaCrawler, in “Proceedings of the 1995 World Wide Web Conference”.) have presented results based on the usage logs of the MetaCrawler meta search service (due to substantial changes in the search engine services and the Web, it is expected that their results would be significantly different if repeated now). These results considered the following engines: Lycos, WebCrawler, InfoSeek, Galaxy, Open Text, and Yahoo. Selberg and Etzioni's results are informative but limited for several reasons.

[0211] First, they present the “market share” of each engine which is the percentage of documents that users follow that originated from each of the search engines. These results are limited for a number of reasons, including (a) relevance is difficult to determine without viewing the pages, and (b) presentation order affects user relevance judgements (Eisenberg, M. and Barry, C. (1986), Order effects: A preliminary study of the possible influence of presentation order on user judgements of document relevance, in “Proceedings of the 49th Annual Meeting of the American Society for Information Science”, Vol. 23, pp. 80–86).

[0212] The results considered by Selberg and Etzioni are also limited because they present results on the percentage of unique references returned and the coverage of each engine. Their results suggest that each engine covers only a fraction of the Web, however their results are limited because (a) as above, engines can return documents which do not contain the query terms—engines which return documents with related words or invalid documents can result in significantly different results, and (b) search engines return documents in different orders, meaning that all documents need to be retrieved for a valid comparison, e.g. two search engines may index exactly the same set of documents yet return a different set as the first x.

[0213] In addition, Selberg and Etzioni find that the percentage of invalid links was 15%. They do not break this down by search engine. Selberg and Etzioni do point out limitations in their study (which is just a small part of a larger paper on the very successful MetaCrawler service).

[0214] AltaVista and Infoseek have recently confirmed that they do not provide comprehensive coverage on the Web (Brake, D. (1997), “Lost in cyberspace”, New Scientist 154(2088), 12–13.) Discussed below are estimates on how much they do cover.

[0215] We have produced statistics on the coverage of the major Web search engines, the size of the Web, and the recency of the search engine databases. Only the 6 current major full-text search engines are considered herein (in alphabetical order): AltaVista, Excite, HotBot, Infoseek, Lycos, and Northern Light. A common perception is that these engines index roughly the same documents, and that they index a relatively large portion of the Web.

[0216] We first compare the number of documents returned when using different combinations of 1 to 6 search engines. Our overall methodology is to retrieve the list of matching documents from all engines and then retrieve all of the documents for analysis. Two important constraints were used.

[0217] The first constraint was that the entire list of documents matching the query must have been retrieved for all of the search engines in order for a query to be included in the study. This constraint is important because the order in which the engines rank documents varies between engines. Consider a query which resulted in greater than 1,000 documents from each engine. If we only compared the first 200 documents from each engine we may find many unique URLs. However, we would not be able to determine if the engines were indexing unique URLs, or if they were indexing the same URLs but returning a different subset as the first 200 documents.

[0218] The second constraint was that for all of the documents that each engine lists as matching the query, we attempted to download the full text of the corresponding URL. Only documents which could be downloaded and which actually contain the query terms are counted. This is important because (a) some engines can return documents which they believe are relevant but do not contain the query terms (e.g. Excite uses “concept-based clustering” and may consider related words, and Infoseek uses morphology), and (b) each search engine contains a number of invalid links, and the percentage of invalid links varies between the search engines (engines which do not delete invalid links would be at an advantage).

[0219] Other details important to the analysis are:

[0220] 1. Duplicates are removed when considering the total number of documents returned by one engine or by a combination of engines, including detection of identical pages with different URLs. URLs are normalized by a)removing any “index.html” suffix or trailing “/”, b) removing a port 80 designation (the default), c) removing the first segment of the domain name for URLs with a directory depth greater than 1(in order to account for machine aliases), and d) unescaping any “escaped” characters. (e.g. %7E in a URL is equivalent to the tilde character).

[0221] 2. We consider only lowercase queries because different engines treat capitalized queries differently (e.g. AltaVista returns only capitalized results for capitalized queries).

[0222] 3. We used an individual page timeout of 60 seconds. Pages which timed out were not included in the analysis.

[0223] 4. We use a fixed maximum of 700 documents per query (from all engines combined after the removal of duplicates)—queries returning more documents were not included. The search engines typically impose a maximum number of documents which can be retrieved (current limits are AltaVista 200, Infoseek 500, HotBot 1,000, Excite 1,000, Lycos 1,000, and Northern Light>10,000) and we checked to ensure that these limits were not exceeded (using this constraint no query returned more than the maximum from each engine, notably no query returned more than 200 documents from AltaVista).

[0224] 5. We only counted documents which contained the exact query terms, i.e. the word “crystals” in a document would not match a query term of “crystal”—the non-plural form of the word would have to exist in the document in order for the document to be counted as matching the query. This is necessary because different engines use different morphology rules.

[0225] 6. HotBot and Altavista can identify alternate pages with the same information on the Web. These alternate pages are included in the statistics (as they are for the engines which do not identify alternate pages with the same data).

[0226] 7. The “special collection” (premier documents not part of the publicly indexable Web) of Northern Light was not used.

[0227] Over a period of time, we have collected 500 queries which satisfy the constraints. For the results presented herein, we performed the 500 queries during the period Aug. 23, 1997 to Aug. 24, 1997. We manually checked that all results were retrieved and parsed correctly from each engine before and after the tests because the engines periodically change their formats for listing documents and/or requesting the next page of documents (we also use automatic methods designed to detect temporary failures and changes in the search engine response formats).

[0228] FIG. 29 shows the fraction of the total number of documents from the 6 engines which were retrieved by each individual engine. Table 1 below shows these results along with the 95% confidence interval. HotBot is the most comprehensive in this comparison. These results are specific to the particular queries performed and the state of the engine databases at the time they were performed. Also, the results may be partly due to different indexing rather than different databases sizes—different engines may not index identical words for the same documents (for example, the engines typically impose a maximum file size and effectively truncate oversized documents).

[0229]

TABLE 1
Search	Hot-	Ex-	Northern	Alta-
Engine	Bot	cite	Light	Vista	Infoseek	Lycos
Coverage	39.2%	31.1%	30.4%	29.2%	17.9%	12.2%
WRT 6
Engines
95%	+/	+/	+/	+/	+/	+/
confidence	−1.4%	−1.2%	−1.3%	−1.2%	−1.1%	−1.1%
interval

[0230] FIG. 30 shows the average fraction of documents retrieved by 1 to 6 search engines normalized by the number retrieved from all six engines. For 1 to 5 engines, the average is over all combinations of the engines, which is averaged for each query and then averaged over queries. Using the assumption that the coverage increases logarithmically with the number of search engines, and that, in the limit, an infinite number of search engines would cover the entire Web, f(x)=b(1-1/exp(ax)), where a and b are constants and x is the number of search engines, was fit to the data (using Levenberg-Marquardt minimization (Fletcher, R. (1987), Practical Methods of Optimization, Second Edition, John Wiley & Sons) with the default parameters in the program gnuplot) and plotted on FIG. 30. This is equivalent to the assumption that each engine covers a certain fixed percentage of the Web, and each engine's sample of the Web is drawn independently from all Web pages (c_i=c_i-1+c₁(1-c_i-1), i=2 . . . n where c₁ is the coverage of i engines and c₁ is the coverage of one engine).

[0231] There are a number of important biases which should be considered. Search engines typically do not consider indexing documents which are hidden behind search forms, and documents where the engines are excluded by the robots exclusion standard, or by authentication requirements. Therefore, we expect the true size of the Web to be much larger than estimated here. However search engines are unlikely to start indexing these documents, and it is therefore of interest to estimate the size of the Web that they do consider indexing (hereafter referred to as the “indexable Web”), and the relative comprehensiveness of the engines.

[0232] The logarithmic extrapolation above is not accurate for determining the size of the indexable Web because (a) the amount of the Web indexed by each engine varies significantly between the engines, and (b) the search engines do not sample the Web independently. All of the 6 search engines offer a registration function for users to register their pages. It is reasonable to assume that many users will register their pages at several of the engines. Therefore the pages indexed by each engine will be partially dependent. A second source of dependence between the sampling performed by each engine comes from the fact that search engines are typically biased towards indexing pages which are linked to in other pages, i.e. more popular pages.

[0233] Consider the overlap between engines a and b in FIG. 31. Assuming that each engine samples the Web independently, the quantity n_o/n_b where n_o is the number of documents returned by both engines and n_b is the number of documents returned by engine b, is an estimate of the fraction of the indexable Web, P_a, covered by engine a. Using the coverage of 6 engines as a reference point we can write P′_a=na/n₆, where n_a is the number of documents returned by engine a and n₆ is the number of unique documents returned by the combination of 6 engines. Thus, P′_a is the coverage of engine a with respect to the coverage of the 6 engines, we can write c=P′_a/P_a=n_an_b/n₆n_o. We use this equation to estimate the size of the Web in relation to the amount, of the Web covered by the 6 engines considered here. Because the size of the engines varies significantly, we consider estimating the value of c using combinations of two engines, from the smallest two to the largest two. We limit this analysis to the 245 queries returning ≧50 documents (to avoid difficulty when n₀=0). Table 2 shows the results. Values of c smaller than 1 suggest that the size of the indexable Web is smaller than the number of documents retrieved from all 6 engines. It is reasonable to expect that larger engines will have lower dependence because a) they can index more pages other than the pages which users register, and b) they can index more of the less popular pages on the Web. Indeed, there is a clear trend where the estimated value of c increases with the larger engines.

[0234]

TABLE 2
			AltaVista
	Lycos		&	Northern
Search	&	Infoseek &	Northern	Light &	Excite
Engines	Infoseek	AltaVista	Light	Excite	& HotBot
Engine	Smallest		→		Largest
Sizes
Estimated	0.6	0.9	0.9	1.9	2.2
c
95%	+/−0.04	+/−0.06	+/−0.04	+/−0.12	+/−0.17
confidence
interval

[0235] Using c=2.2, from the comparison with the largest two engines, we can estimate the fraction of the indexable Web which the engines cover: HotBot 17.8%, Excite 14.1%, Northern Light 13.8%, AltaVista 13.3%, Infoseek 8.1%, Lycos 5.5%. These results are shown at 120 in FIG. 32. The percentage of the indexable Web indexed by the major search engines is much lower than is commonly believed. We note that (a) it is reasonable to expect that the true value of c is actually larger than 2.2 due to the dependence which remains between the two largest engines, and (b) different results may be found for queries from a different class of users.

[0236] HotBot reportedly contains 54 million pages, putting our estimate on a lower bound for the size of the indexable Web at approximately 300 million pages. Currently available estimates of the size of the Web vary significantly. The Internet Archive uses an estimate of 80 million pages (excluding images, sounds, etc.) (Cunningham, M. (1997), ‘Brewster's millions’, http://www.irish-times.com/irish-times/paper/1997/0127/cmp1.html.) Forrester Research estimates that there are more than 75 million pages (Guglielmo, C. (1997), ‘Mr.Kurnit's neighborhood’, Upside September.) AltaVista now estimates that there the Web contains 100 to 150 million pages (Brake, D. (1997), ‘Lost in cyberspace’, New Scientist 154(2088), 12–13).

[0237] A simple analysis of page retrieval times leads to some interesting conclusions. Table 3 below shows the median time for each of the six major search engines to respond, along with the median time for the first of the six engines to respond when queries are made simultaneously to all engines (as happens in the meta engine).

[0238]

	TABLE 3
		Median Time for response
	Search Engine	(seconds)
	AltaVista	0.9
	Infoseek	1.3
	HotBot	2.6
	Excite	5.2
	Lycos	2.8
	Northern Light	7.5
	All engines	2.7
	First of 6 engines	0.8
	First result from the meta	1.3
	search engine of this
	invention

[0239] Histograms of the response times for these engines and the first of 6 engines are shown in FIGS. 33 and 34, and the median times are shown in FIG. 35. FIG. 36 shows the median time for the first of n engines to respond. These results are from September 1997, and we note that the relative speed of the search engines varies over time.

[0240] Looking now at the time to download arbitrary Web pages, FIG. 37 shows a histogram of the response time. FIG. 38 shows the median time for the first of n engines to respond. We can estimate the time for the meta engine to display the first result, which we create by sampling from the distributions for the first of 6 search engines (the meta engine actually uses more than 6 search engines but we concentrate on the major Web engines here), and the first of 10 Web pages (the actual number depends on the number returned by the first engine to respond), adding these together, and averaging over 10,000 trials.

[0241] FIG. 39 shows a histogram of this distribution. The median of the distribution is 1.3 seconds (compared to 2.7 seconds for the median response time of a search engine even without downloading any actual pages). For comparison, the average time MetaCrawler takes to return results is 25.7 seconds (without page verification) or 139.3 seconds (with page verification) (Selberg, E. and Etziono, O. (1995), Multi-service search and comparison using the MetaCrawler, in ‘Proceedings of the 1995 World Wide Web Conference’)(the underlying search engines and/or the Web appear to be significantly faster than they were when Selberg and Etzioni performed their experiment).

[0242] Therefore, on average we find that the parallel architecture of the meta engine of this invention allows it to find, download and analyze the first page faster than the standard search engines can produce a result although the standard engines do not download and analyze the pages. Note that the results in this section are specific to the particular queries performed (speed as a function of the query is different for each engine) and the network conditions under which they were performed. These factors may bias the results towards certain engines. The non-stationarity of Web access times is not considered here, e.g. the speed of the engines varies significantly over time (short term variations may be due to network or machine problems and user load, long term variations may be due to modifications in the search engine software, the search engine hardware resources, or relevant network connections).

[0243] The meta search engine of this invention demonstrates that real-time analysis of documents returned from Web search engines is feasible. In fact, calling the Web search engines and downloading Web pages in parallel allows the meta search engine of this invention to, on average, display the first result quicker than using a standard search engine.

[0244] User feedback indicates that the display of real-time local context around query terms, and the highlighting of query terms in the documents when viewed, significantly improves the efficiency of searching the Web.

[0245] Our experiments indicate that an upper bound on the coverage of the major search engines varies from 6% (Lycos) to 18% (HotBot) of the indexable Web. Combining the results of six engines returns more than 3.5 times as many documents when compared to using only one engine. By analyzing the overlap between search engines, we estimate that an approximate lower bound on the size of the indexable Web is 300 million pages. The percentage of invalid links returned by the major engines varies from 3% to 7%. Our results provide an indication of the relative coverage of the major Web search engines, and confirm that, as indicated by Selberg and Etzioni, the coverage of any one search engine is significantly limited.

[0246] While it is apparent that the invention herein disclosed is well calculated to fulfill the objects previously stated, it will be appreciated that numerous modifications and embodiments may be devised by those skilled in the art, and it is intended that the appended claims cover all such modifications and embodiments as fall within the true spirit and scope of the present invention.