US 20060161532 A1
This disclosure relates to performing a query for a search term of a database containing a plurality of structured documents. Those structured documents that do not include the search term are ferreted or filtered out during an initial search. Matched structured documents which are those structured documents that do contain the search term are evaluated by ranking the individual elements based on how well each individual element matches the search term, and indicating to the user the ranking of the individual elements wherein the individual elements can be accessed by the user.
1. A method, comprising:
processing a query for a search term of a database containing a plurality of structured documents;
ferreting out those structured documents that do not include the search term;
evaluating elements of matched structured documents which are those structured documents that do contain the search term by:
ranking the individual elements based on how well each individual element matches the search term, and
indicating to the user the ranking of the individual elements wherein the individual elements can be accessed by the user.
2. The method of
3. The method of
4. The method of
5. The method of
6. The method of
7. The method of
8. The method of
wherein the ferreting out those structured documents includes determining all elements containing at least one query term;
wherein the ranking the individual elements includes ranking the individual elements according to a frequency with which the query term appears in the individual elements; and
returning the individual elements in a descending order of the frequency with which the query term appears in the individual elements.
9. The method of
10. A computer readable medium having computer executable instructions that when executed by a general process computer is capable of performing a method, the method comprising:
processing a query for a search term of a database containing a plurality of structured documents;
filtering out those structured documents that do not include the search term; and
evaluating matched structured documents which are those structured documents that do contain the search term by ranking the individual elements based on how well each individual element matches the search term.
11. The computer readable medium which performs the method of
12. The computer readable medium which performs the method of
13. The computer readable medium which performs the method of
14. The computer readable medium which performs the method of
15. The computer readable medium which performs the method of
16. The computer readable medium which performs the method of
17. The computer readable medium which performs the method of
18. A method for retrieving a portion of at least one document from a plurality of documents, comprising:
processing a query for a search term on a plurality of the documents;
setting a threshold of a weight of each element to be retrieved;
ranking the individual elements of a document that exceeds the threshold; and
indicating to a user the ranking of the individual elements.
19. The method of
20. The method of
a numerical indication of the ranking; and
a number of graphic indicators to represent the ranking.
This application is a continuation application of U.S. patent application Ser. No. 10/337,138 entitled “RETRIEVAL OF STRUCTURED DOCUMENTS,” filed on Jan. 6, 2003.
This invention relates to retrieval of documents, and more particularly, to retrieval of structured documents.
With the increased usage of computers, networks, the Internet, etc., documents are often searched for certain terms. For example, an individual or student searching for a particular topic may search for an electronic document (hereinafter document) using a specific search engine on a networked computer, a stand-alone computer, or over the Internet for that term. The search engine will return a list of documents that contain the term using computer-based document retrieval technology. Often the documents retrieved for a query are ranked according to how well each particular document matches the queried term(s). The user often has to consider the entire document to determine where a particular search term exists.
Often, memory locations in computers store certain documents in a hierarchical structure. Certain structured computer languages, such as eXtensible Markup Language (XML) that rely on hierarchical structures use tags, or similar devices, that structurally organize data into particular sections or elements. In a retrieved structured document, the user is often not aware where the particular search term exists in each document to find context for the term in the document. This additional user time and effort may be considerable for extended queries.
Many document retrieval systems consider documents as relatively small sized discrete retrieval units that can be queried and returned, such that the documents cannot be further sub-divided. Often a retrieved document is too large for a user to analyze in a meaningful manner. Thus, users often have to carefully review entire retrieved documents in digital library computer applications to determine locations of relevant terms and/or the context of the relevant terms.
Passage retrieval is in principle similar to document retrieval, but involves the additional preliminary stage of extracting passages from documents. One aspect of passage retrieval returns briefer answers to the user. To accomplish this, a document can be decomposed into fixed-length or pre-defined portions using, e.g., the term frequency inverse document frequency (TFIDF) algorithm or a variance of this algorithms, to build index at passage or paragraph level. However, this indexing method on which many document retrieval systems rely does not maintain semantic relationships among the elements in documents. In addition, this indexing mechanism may result in many discrete retrieved elements in a form that requires considerable computer work to present meaningful text to users.
This disclosure relates to a scalable method and associated system for retrieving structured documents.
In one aspect, a process and associated system comprises performing a query for a search term of a database containing a plurality of structured documents. Those structured documents that do not include the search term are ferreted or filtered out during an initial search. Matched structured documents, which are those structured documents that do contain the search term, are evaluated by ranking the individual elements based on how well each individual element matches the search term. This ranking evaluation is indicated to the user in such a manner that at least certain individual elements can be accessed by the user.
In another aspect, a process involves querying a document having a plurality of elements for a search term. The relevancies of different ones of the elements are weighted based on the search term. Different ones of the elements are ranked in response to the weighted different ones of the elements. Additionally, a path structure of the plurality of elements within the document are displayed wherein the path structure indicates the relevance of different ones of the elements within the document. The elements of the structured documents are scaled based on the structure of the document.
The same numbers reference like features and components throughout the drawings.
Document retrieval systems rely on search engines to locate, and likely display, relevant electronic documents (hereinafter “documents”) based on one or more given search terms or words provided by a user. Document retrieval systems are used more (and are required to perform more challenging operations) in the era of the Internet. Different embodiments of a structured document retrieval system 100, different embodiments are illustrated in
Structured documents are particularly suitable for meaningful information retrieval since the structure (e.g., hierarchy) of the document can provide information about the data stored within each section. One aspect of this disclosure describes how a search engine can search and retrieve portions of a document (e.g., portions being differentiated by tags or some other similar mechanism) that are relevant to the search. The structured document retrieval system can return structured documents having a suitable granularity. The granularity of a document can indicate how much a document can be sub-divided. Granularity represents a measure of the divisibility of documents into distinct elements. Though generally structured documents have a high degree of granularity while non-structured documents lack considerable granularity, there are exceptions to this rule.
Another aspect of this disclosure describes how searching and using one path among a plurality of paths in a plurality of structured documents utilizes an indexing scheme. The structured document retrieval system can display the relevant portions of the structured document either by themselves or in some association to other portions of the document.
In different embodiments, the database portion 106 stores structured documents containing data stored in XML format, in HyperText Markup Language (HTML) format, in SGML format, or in other structured format. HTML is the markup language presently predominantly used over the Internet. HTML documents typically have their own structure. HTML is most commonly used to define color, font, etc. that is displayed over computer displays. It is more difficult to use HTML in a structured query in many embodiments of the structured document retrieval system 100 than, for example, XML and SGML because HTML has been traditionally applied with more concern to display style, not the content style. XML and SGML, by comparison, are user defined and are primarily concerned in documents with the data content. HTML, XML, or SGML can be applied to different embodiments of the structured document retrieval system though XML and SGML are particularly suited. There are varieties and modifications of markup languages that are within the intended scope of the present disclosure.
The structured document retrieval system 100 relies on the content and the context of the document in retrieving and displaying the documents. The structured document retrieval system therefore matches the content (e.g., language) of structured documents. The structured document also indicates to the user the context of the retrieved portions of the document. An approximate match can be used for information retrieval. After searching each structure portion, it is determined where the match occurs in each structured portion.
To explain how the structured document retrieval system retrieves structured documents, certain aspects of structured documents are described.
Whether a particular portion of the document (i.e., text, images, table, etc.) is characterized as a content document element 302, a content section element 304, or a content paragraph element 306 depends upon the tag associated with each particular element. The use of tags in XML, SGML, HTML, and other mark-up languages is well known and will not be further detailed herein except by noting that XML and SGML both utilize user defined tags. As such, the particular terminology of the content document element 302, the content section elements 304, and the content paragraph elements 306 within a particular document can be dictated by a user or a programmer. It is intended that the content document element 302, the content section elements 304, and the content paragraph elements 306 could be described with different names or terminology while remaining within the intended scope of the present disclosure (as is also true for the elements described relative to
The structured document retrieval system 100 the content, the context, and the hierarchical structure of documents.
Mapping the content structure 300 of documents to the hierarchical index 300 (as shown respectively in
Relevant document elements with arbitrary granularity can be obtained using a scalable element retrieve strategy by employing normal keyword-based queries. The structured document retrieval system 100 includes a hierarchical indexing mechanism and an algorithm. Index terms propagate across the tree structure of XML documents according to terms' frequency and distribution in elements at different levels. As such, each document index element 312, section index element 314, and paragraph index element 316 as shown in
The hierarchical index 400 shown in
The structured document retrieval system 100 overcomes, to some extent, the difficulties that may exist with text length normalization and threshold setting in structured document retrieval. An entire document is often too large for a user to effectively analyze or determine the context of certain terms. It may be beneficial for a user, particularly those that are analyzing longer documents, to analyze only a part of a document. Additionally, it may be important to indicate the context (i.e., relation) between the displayed portion of the document and other portions of the document.
XML and SGML (and also HTML) represent computer languages that provide structured documents in a configuration that are particularly directed to hierarchical indexing and structured document retrieval. Typical XML or SGML documents include a set of logical elements (such as title, sections, subsections, paragraphs, etc). XML and SGML documents are typically structured based on the hierarchical relation, and as such are highly applicable to hierarchical indexing and the structured document retrieval system 100. XML and SGML provide a standard and effective way to explicitly represent the structure of documents using user-defined tags.
In one aspect of the disclosure, document retrieval does not provide effective information retrieval unless the user can retrieve arbitrary levels of elements in documents. Retrieving a set of extremely long document in response to a query is often of little use to the user unless the user is providing with some mechanism to derive the context of the search terms (or index terms) within the document. Obtaining arbitrary levels of elements in documents allows the structured document retrieval system 100 to adapt for querying documents having different lengths and structures. Thus, the retrieval system can return only the relevant elements to the user as determined by the user.
One aspect of the structured document retrieval system 100 is to utilize retrieval techniques (e.g., using indexing techniques) that retrieve more directed data. The structured document retrieval system 100 uses keyword-based queries (such as commonly used in document retrieval to retrieve relevant document elements) that are modified by having arbitrary granularity. One aspect of this disclosure includes two portions—a novel hierarchical indexing mechanism and an algorithm to select and display suitable document elements based on their hierarchical index 400 as shown in
Certain prior approaches only assign index terms to the leaf elements or fixed-length passages. Elements at different levels are assigned equal and proper index terms. The structured document retrieval system 100 uses the hierarchical indexing mechanism to assigns index terms that can precisely describe the inherent concept for each element while filtering out terms with too broad or too narrow semantics. Consider that a sample document about “China” contains a section “History”, which in turn also contains subsections such as “Tang dynasty”, “Ming dynasty” and “Qing dynasty”, etc. Selecting suitable index terms for the section “History” is challenging. “China” appears to be a relatively broad term that will provide too many results. “Tang”, “Ming”, and “Qing” by comparison appear to be narrow terms that may miss relevant documents. Terms such as “history” and “dynasty” appear to be good candidate terms.
The hierarchical indexing mechanism described herein automatically selects proper index terms for each element following a bottom-up propagation fashion. In one embodiment, the indexing method of the structured document retrieval system 100 ensures that an index term that appears in an element does not appear in any children elements of this element. The indexing method saves a considerable amount of storage space and retrieval time. Moreover, this hierarchical indexing mechanism avoids the potential problem of results aggregation since the retrieval process can obtain desirable elements directly due to the thorough distribution of index terms across the whole document tree. This savings in time over a large document search and retrieval process will lead to a considerable savings in time.
The hierarchical indexing mechanism continues to decision 478 in which it is determined whether there are any additional elements to consider in the document. Note that 472, 474, and 476 create a loop that considers each element in the hierarchical index 400 for index terms. In one embodiment, the user can limit certain sections of the hierarchical index 400 from consideration either by a prompt provided by the computer during operation or by a preliminary user limiting selection of the areas that the user wishes to search.
Following the looping through the 472, 474, and 476 portion of the hierarchical indexing mechanism 400, the hierarchical indexing mechanism continues to 478. In 478, for each index term in the hierarchical index, the appropriate terms are propagated from the child elements to the parent element in the hierarchical index 400. Such appropriate propagation of terms from the child elements to the parent elements can be equated to the “filtering process” described relative to as illustrated in
For example, when a user submits the query “history of China”, the title of a retrieved document is “China”, and a section is “history”. The most suitable path is through the heading “China” is to the section “History” since the “History” section is most relevant to the heading “China”. As such, the sections “China” and “History” form the index path from the query “history of China”; a user querying history of China returns this path.
Another part of the method uses a scalable algorithm to select suitable document elements to meet the user's information need. The term “scalable” as applied to the structured document retrieval system 100 indicates that the algorithm can retrieve any granularity of relevant document elements according to the extent of the concept in the user's query. Since the hierarchical index 400 as shown in
In one aspect, the document retrieval system 100 can be scalable. Certain embodiments of the structured document retrieval system allow a user to retrieve documents and display the sections documents in a wide variety of configurations and/or granularities. For example, a user can perform a broad retrieval to access the entire document. As such, using the China example, the user can access the “China” index term yielding the entire hierarchical index 400 of the retrieved document. Additionally, the user can perform a medium retrieval process to retrieve the section index elements 314 as shown in
A path ranking algorithm as described below allows retrieval and display of documents in a meaningful manner. Using specific indices to retrieve data with structure information has applications in many areas (e.g., database, IR, and XML). The structured document retrieval system 100 includes a hierarchical index 400 that maps directly to the structure of the document. The hierarchical indexing mechanism determines which retrieves document elements with various granularities, while imposing no limits on users' query language. For instance, a structured document retrieval system 100 can be applied to extremely long documents as well as relatively brief documents. For each element or component in the document tree there is a corresponding index element or component in the index tree. When a query comes in, the search engine uses the index tree to locate the most relevant component path or paths (that is most relevant to user's search). This component path(s) is through one or more of the section elements 304 (from the root path or document element 302), to the relevant individual paragraph element 306 as shown in
For each document, a hierarchical index 400 having the same structure as the document is established. The structure of the hierarchical index 400 therefore mirrors the content structure 300 of the original document. Index terms are distributed in all elements according to how generally they represent a concept in the document. To effectively assign an index term to an element, the index term should characterize the concept of the element while differentiating the element from other elements. Thus, to select index terms for an element, the term frequency and distribution in the element should be considered. Also the weight of the index term is compared with other terms in this element. The top ranked index terms are selected as index terms for this element. If a term appears frequently and is distributed evenly in the text of an element, and also has a rank beyond a threshold value, then the index term should be considered as an index term for this element.
The structured document retrieval system 100 takes advantage of the hierarchical structure of certain structured documents. Hierarchical documents (which represent one type of structured documents) include associations between parent elements and children elements. The distribution of a term in a parent element is quantifiable by considering the term's appearances in the children elements of this parent element. In one embodiment, this distribution quantification applies only to children elements that are directly descended from the parent element because the topic of each element should be supported by its directly descendant elements. If a term is distributed evenly in the children element of a section or document, this term would be a good candidate index term for that section or document. Information theory defines entropy as a special concept for capturing information content (or information uncertainty). Entropy is a useful criterion in one embodiment to measure the distribution of a term in an element.
Paragraph elements 306 as shown in
Weight (ti, Pj) stands for the weight of the term ti in the paragraph Pj. “tf (ti, Pj)” is the term frequency of ti in this paragraph. N denotes the number of documents in the corpus and ni represents the number of documents containing the term ti.
For elements at other levels that are not atomic elements (e.g., the section element 304 and the document element 302 ), the term frequency and term distribution are combined to provide the weight for a term. Thus the weight of the term ti in an arbitrary composite element Ej can be defined as the following:
“I (tii,Ej)” is the entropy measure of the term ti in element Ej and is defined as:
where subk stands for the kth directly descendant element of Ej and N(sub) the number of such descendant elements.
In calculation 3, the term
plays an important role in the indexing mechanism. Term frequency varies greatly in different elements due to the great variance of text length of them. Entropy measure may encounter the same length normalization problem as other document or passage retrieval methods face. The entropy measure allows generating the entropy when hypothesizing that all appearances of this term in an element are exactly equal in each children element of this element, which is the theoretic maximum value for this term's entropy. The proportion of this theoretic maximum value is the distribution measure of a term. Later experiments proved that this calculation could almost tradeoff the side effects of various term frequencies in elements at different levels.
Term weights are further normalized in order to compare weights in different elements. Term weights obtained by calculations 1 and 2 are divided by the maximum weight of all terms in the same element. Thus all terms' weights would fall into the range between 0 and 1.
As mentioned above, a term or element whose weight exceeds a certain threshold should propagate to the upper level since it may stand for a more general concept. The threshold can be dynamically adjusted according to statistics of terms' weights in a specific element. Different embodiments of this dynamic threshold setting can either be performed by the user or operator of the computer environment as illustrated in
1. For each paragraph element 306, calculate the terms' weight according to calculation 1.
2. For any element Ej at the one upper level (i.e., the section elements 304 or the document element(s) 302) following a bottom-up fashion, calculate term weights using calculation 2. If Weight (ti,Ej)≧average(Ej)+std_dev(Ej), the term ti is selected as an index term of the element Ej and all sub-elements of Ej would eliminate ti from their index term list. Here average (Ej) denotes the arithmetic average of all terms' weights in the element Ej, and std_dev(Ej) the standard deviation of these weights.
3. Repeat 2 until the root element, i.e., the document element 302, is reached.
This indexing solution utilizes the internal structure of documents thoroughly. Since all terms are compared with each other in the same level and a theoretic maximum value as normalization factor is put in the calculation, the effect of variant lengths of text in different elements is minimized. In addition, the threshold of average value plus standard deviation provides a dynamic local optimum value for determining index terms for a specific element. In addition, an index term of an element is not necessarily required to appear in all sub-elements of this element. Thus more representative index terms other than just a few words in titles are detected.
One embodiment of a structured document retrieval process 400 is illustrated in
The structured document retrieval method starts with 402 in which the user submits a query, using index words, to an initial search engine in the structured document retrieval system 100 to search multiple structured documents. The initial search engine responds to the original query by filtering out those entire documents that contain one or more of the search terms from those entire documents that do not contain one of more of the search terms.
In certain embodiments of the structured document retrieval system, the query submitted by the user in 402 may be the final input from the user before the returned documents are displayed. The method continues to 404 in which the search engine performs the search on all of the structured documents for the index terms. The search engine used in 404 can be configured to search the entire document together, or search all of the individual elements for the index terms. During 404, however, all documents are entirely searched for the given search term(s). This initial search can be performed by any one of a wide variety of search engines, many of which are commercially available and in common usage. As such, 402 and 404 act in a similar manner as most search engines used today by ferreting out those documents that do not include any search terms included in the query by the user.
The structured document retrieval process 400 continues to decision 406 in which it is determined whether the particular structured document contains the index term(s). If the answer to decision 406 is no, then the structured document retrieval process 400 continues to 408 in which those documents that do not contain any of the search terms are ferreted (i.e., filtered) out. Those documents that 408 applies to do not have to be further considered by the structured document retrieval process 400 (i.e. the structured document retrieval process 400 is terminated) since they do not contain any of the relevant index terms used in the search, and therefore these documents cannot be of interest to the user.
If the answer to decision 406 is yes, then the structured document retrieval process 400 continues to 410 in which the user and/or the structured document retrieval system 100 selects those structured documents to search further. 410 is optional, and the process 400 can instead continue directly to 412. The structured document retrieval system 100 continues to 412 in which the structured document retrieval system 100 automatically establishes the hierarchical index 400 having the same structure as the structured document using mapping. An example of this hierarchical mapping process is described relative to
The structured document retrieval process 400 continues to 414 in which the index terms in each element of the structured document are weighted. This weighting determines the relevancy of each element in the document. The weighting in 414, which considers each element distinctly, differs from 404, which searches each document in its entirety. Further details about 414 are described below. The structured document retrieval process 400 continues to 416 in the XML structure, in which it searches for highest level and match, and then searches the next lower level for any matched terms.
The structured document retrieval process 400 continues to 418 in which those documents having the closest match are displayed. In one embodiment of the disclosure, the structured document retrieval system 100 displays an element, as well as the context of the element. For example, as illustrated in
Term weighting is now described relative to the structured documents shown in
In this section, an embodiment of the scalable retrieval algorithm to select suitable document elements is described which displays to the user based on the hierarchical index 400. Since the hierarchical index 400 provides an effective mechanism to retrieve document elements with any granularity directly, the main task in the retrieval phase is online searching and ranking candidate elements. Certain embodiments of the structured document retrieval system 100 present an interface that displays each retrieved element with their structural context to facilitate users to browse the retrieval results in a convenient way. The displayed user interface plays an important role to the user since in many embodiments it displays the rank (relevance) of each element based on the index term input by the user.
For each document, the structured document retrieval system 100 uses a path ranking algorithm to calculate relevance values of all candidate elements to a query. A path for an element is considered to be all elements between the root (the document element) and this element (including the root element and this element) in the document tree. If one element is the ancestor of another element, its path is the subset of the path of the latter. According to one embodiment of the hierarchical indexing mechanism, an element does not share any index terms with its descendants. As such, an element is completely described by its path. Therefore, the element ranking problem can be transformed to a path ranking problem, that is, the goal is to find those element paths with high relevance values to the query.
Considering the China query described above, for the returned document “China”, index terms at document level may include “China”, “Chinese” while index terms for the section “History” may be “history”, “dynasty”, etc. To the query “history of China”, the element path of the section “History” contains all query terms while the element path of the document “China” contains only one query term. Thus the section “History” is a better answer to the query.
An element's rank is important in determining the relevance of the element to the query. A term's weight for an element path is defined as its weight in an element along the path whose index term list contains this term. The rank for a path according to a given query is defined as:
1n N/ni is the inverse document frequency (IDF) value of query term ti, which represents the query term's weight, and Q stands for the number of query terms in a query.
Given a coming query, traditional document retrieval techniques can be used to get a list of relevant documents first in order to narrow down the search space. Then when the user selects one of the relevant documents, the system searches for all candidate elements of this document and then ranks their paths according to the query.
The overall process to rank an element is described as below. All elements containing at least one query term are found. Paths are obtained for all candidate elements, and query terms' weights are assigned for elements to paths respectively. The weighted paths are ranked according to calculation 4. The elements corresponding to all, or a select few, of the ranked paths are returned (and typically displayed) in a descending order.
A long-standing problem in structured document retrieval is how to return proper elements which can best satisfy (e.g., are relevant to) the user's query needs. The structured document retrieval system 100 uses the average of all retrieved elements' rank as a dynamic threshold. Elements having a rank beyond the threshold are returned as results. Later experiments show that accurate element retrieval can be attained based on this dynamic threshold.
All returned elements are displayed by the structured document retrieval system 100 with their structural context, which can indicate the position of a returned element in the document tree as well as its surrounding elements as shown in
This section of the disclosure evaluates the performance of the document retrieval method 400 as described with the structured document retrieval system 100. The effects of the threshold settings are also considered. Experiments were conducted on Encarta® Multimedia Encyclopedia corpus, which as of the date of writing this disclosure, contains over 40,000 well structured XML documents. The query set includes 10 queries as listed in Table 1, which appear to be similar queries such as may be input by a typical user of the structured document retrieval system 100. Each query can best be answered by displaying to a user only a portion (i.e. an element) of the relevant documents.
Preliminary experimental results show that not only does the structured document retrieval system 100 perform significantly better than the compared method, but also it overcomes, to some extent, the difficulties of text length normalization and elements selection threshold, which are two long standing problem in structured document retrieval and passage retrieval.
To illustrate the benefits of the present document retrieval system, a passage retrieval system is implemented using the TFIDF Para, which is compared with the structured document retrieval system 100. In one embodiment as per the TFIDF Para system, only paragraphs are exploited as passages while other structural information is ignored. A term's weight in a paragraph can be defined by the conventional TFIDF measure without normalization, which provides similar results as calculation 1.
Previous work on passage retrieval or structured document retrieval focused their evaluations mainly on the impact of passage level evidence on retrieving whole documents. None of these conducted special experiments dedicated to evaluate the effectiveness of element retrieval. As a portion of the development of the structured document retrieval system 100, as described above, a series of experiments were run in order to test the ability of the scalable retrieval method to find elements of proper granularity for user queries.
Suppose a user who studies history of military operations intends to find out “what military aircrafts were used in Desert Storm”. An initial search may yield two articles Military Aircrafts and Gulf War as highly-ranking results both containing a portion with relevant content. The user has to scan each document (often quite long) to look for relevant content from the entire returned document.
Relevance judgments are made by human assessors. For each query, the relevant elements in documents are judged and selected by the assessors. The measurements used here are similar to those employed in document retrieval. Recall is the fraction of relevant elements, which have been retrieved and precision as the fraction of the retrieved elements which are judged relevant. Since the returned elements should be considered as having a relatively small number of relevant elements, various precision values are not computed in different recall levels. Instead, the overall recall and precision level are calculated and employed to contribute the major portion of a combined F-Value as described in calculation 5.
When judging what fraction of retrieved elements should be returned to users as answer, both fixed threshold values from 0.1 to 0.9 and dynamic threshold values are used. The dynamic threshold is determined by the average of the rank values of all retrieved elements (for one selected document) (Avg) and the standard deviation of these values (Std_Dev). Thus “Avg” and “Avg+Std_Dev” are exploited as dynamic thresholds in the experiment. The F-Value obtained by these two methods with various thresholds are compared and illustrated in Table 2.
The F-values on Table 1 shows improvement of the structured document retrieval process 400 compared to prior methods. F-value is a measure of matching that ranges from 0 to 1, where 1 represents a perfect match and 0 represents no match. Table 2 shows that the scalable retrieval system having various threshold settings displays a significant improvement in retrieval performance over another retrieval method that involves applying TFIDF measure to the paragraph level directly.
Additionally, users in general are more comfortable in dealing with the briefer and more directed results presented by the scalable retrieval method (according to the feedback of assessors). This additional user-comfort is largely due to the hierarchical context structural presentation of the structured document retrieval system 100 in which the user can view the organization of the entire document, the context of a retrieved element relative to the entire element, and the contents of the particular element (or a portion of the element).
The literal context of the element is preserved in results since the scalable retrieval method can return elements in various granularities which may be paragraphs, sections or whole documents according to the specification of queries while previous structured document or passage retrieval methods return only fixed level passages. For example, the query “Qing Dynasty” can be supported best by a section with the title “Manchu Qing Dynasty” under the section of “History” in the document “China”. Dozens of paragraphs are contained in this section. In the structured document retrieval system, this section is presented in the first place for this document to the query. However, separate paragraphs with paragraph level ranks are ordered in the TFIDF Para system. Users cannot make clear the correlations or context between these paragraphs ordered only by their individual evidence to this query. Therefore, the hierarchical indexing plus scalable retrieval achieves a relatively high retrieval performance and additionally provides a user-friendly results presentation since it exploits the original structural information of documents thoroughly.
Threshold setting is important for structured document retrieval to obtain desirable elements. In certain prior document retrieval systems, the threshold is fixed at 0.2. Experiments related to the structured document retrieval system 100 indicate that using a single threshold cannot make the system always perform best for all queries since documents that will be queried vary greatly in structure and length. Therefore, the structured document retrieval system 100 utilizes dynamic threshold instead of unchanged (i.e., static) threshold.
Consider how various thresholds affect the retrieval performance of the two methods.
Second, consider the effects of dynamic threshold.
Encarta online and Encarta CD versions can both be modified by integrating the scalable structured document retrieval system 100 to add scalable retrieval function thereto. For any database products that contain structured documents (e.g., using XML, SGML, certain applications of HTML, etc.), this technique can be used to improve search and retrieval performance.
The computer environment 500 illustrated in
The computer environment 500 includes a general-purpose computing device in the form of a computer 502. The computer 502 can be, for example, one or more of a stand alone computer, a networked computer, a mainframe computer, a PDA, a telephone, a microcomputer or microprocessor, or any other computer device that uses a processor in combination with a memory. The components of the computer 502 can include, but are not limited to, one or more processors or processing units 504 (optionally including a cryptographic processor or co-processor), a system memory 506, and a system bus 508 that couples various system components including the processor 504 and the system memory 506.
The system bus 508 represents one or more of any of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, and a processor or local bus using any of a variety of bus architectures. By way of example, such architectures can include an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MCA) bus, an Enhanced ISA (EISA) bus, a Video Electronics Standards Association (VESA) local bus, and a Peripheral Component Interconnects (PCI) bus also known as a Mezzanine bus.
The computer 502 typically includes a variety of computer readable media. Such media can be any available media that is accessible by the computer 502 and includes both volatile and non-volatile media, and removable and non-removable media.
The system memory 506 includes the computer readable media in the form of non-volatile memory such as read only memory (ROM) 512, and/or volatile memory such as random access memory (RAM) 510. A basic input/output system (BIOS) 514, containing the basic routines that help to transfer information between elements within the computer 502, such as during start-up, is stored in the ROM 512. The RAM 510 typically contains data and/or program modules that are immediately accessible to, and/or presently operated on, by the processing unit 504.
The computer 502 may also include other removable/non-removable, volatile/non-volatile computer storage media. By way of example,
The disk drives and their associated computer-readable media provide non-volatile storage of computer readable instructions, control node data structures, program modules, and other data for the computer 502. Although the example illustrates a hard disk within the hard disk drive 515, a removable magnetic disk 520, and a non-volatile optical disk 524, it is to be appreciated that other types of the computer readable media which can store data that is accessible by a computer, such as magnetic cassettes or other magnetic storage devices, flash memory cards, CD-ROM, digital versatile disks (DVD) or other optical storage, random access memories (RAM), read only memories (ROM), electrically erasable programmable read-only memory (EEPROM), and the like, can also be utilized to implement the exemplary computer environment 500.
Any number of program modules can be stored on the hard disk contained in the hard disk drive 515, magnetic disk 520, non-volatile optical disk 524, ROM 512, and/or RAM 510, including by way of example, the OS 526, one or more application programs 528, other program modules 530, and program data 532. Each OS 526, one or more application programs 528, other program modules 530, and program data 532 (or some combination thereof) may implement all or part of the resident components that support the distributed file system.
A user can enter commands and information into the computer 502 via input devices such as a keyboard 534 and a pointing device 536 (e.g., a “mouse”). Other input devices 538 (not shown specifically) may include a microphone, joystick, game pad, satellite dish, serial port, scanner, and/or the like. These and other input devices are connected to the processing unit 504 via input/output interfaces 540 that are coupled to the system bus 508, but may be connected by other interface and bus structures, such as a parallel port, game port, or a universal serial bus (USB).
A monitor, flat panel display, or other type of computer display 200 can also be connected to the system bus 508 via an interface, such as a video adapter 544. The computer display is able to display such outputs from the structured document retrieval system 100 as shown in
Computer 502 can operate in a networked environment using logical connections to one or more remote computers, such as a remote computer device 548. By way of example, the remote computer device 548 can be a personal computer, portable computer, a server, a router, a network computer, a peer device or other common network node, game console, and the like. The remote computer device 548 is illustrated as a portable computer that can include many or all of the elements and features described herein relative to the computer 502.
Logical connections between the computer 502 and the remote computer device 548 are depicted as a local area network (LAN) 550 and a general wide area network (WAN) 552. Such networking environments are commonplace in offices, enterprise-wide computer networks, intranets, and the Internet.
When implemented in a LAN networking environment, the computer 502 is connected to a local network 550 via a network interface or adapter 554. When implemented in a WAN networking environment, the computer 502 typically includes a modem 556 or other means for establishing communications over the wide network 552. The modem 556, which can be internal or external to the computer 502, can be connected to the system bus 508 via the input/output interfaces 540 or other appropriate mechanisms. It is to be appreciated that the illustrated network connections are exemplary and that other means of establishing communication link(s) between the computers 502 and 548 can be employed.
In a networked environment, such as that illustrated with the computer environment 500, program modules depicted relative to the computer 502, or portions thereof, may be stored in a remote memory storage device. By way of example, remote application programs 558 reside on a memory device of the remote computer 548. For purposes of illustration, application programs and other executable program components such as the operating system are illustrated herein as discrete blocks, although it is recognized that such programs and components reside at various times in different storage components of the computer 502, and are executed by the data processor(s) of the computer 502. It will be appreciated that the network connections shown and described are exemplary and other means of establishing a communications link between the computers may be used.
Various modules and techniques may be described herein in the general context of the computer-executable instructions, such as program modules, executed by one or more computers or other devices. Generally, program modules include routines, programs, control objects 650, components, control node data structures 654, etc. that perform particular tasks or implement particular abstract data types. Typically, the functionality of the program modules may be combined or distributed as desired in various embodiments.
An implementation of these modules and techniques may be stored on or transmitted across some form of the computer readable media. Computer readable media can be any available media that can be accessed by a computer. By way of example, and not limitation, computer readable media may comprise “computer storage media” and “communications media.”
“Computer storage media” includes volatile and non-volatile, removable and non-removable media implemented in any process or technology for storage of information such as computer readable instructions, control node data structures, program modules, or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by a computer.
“Communication media” typically embodies computer readable instructions, control node data structures, program modules, or other data in a modulated data signal, such as carrier wave or other transport mechanism. Communication media also includes any information delivery media. The term “modulated data signal” means a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, communication media includes wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared, and other wireless media. Combinations of any of the above are also included within the scope of computer readable media.
The structured document retrieval system 100 as described above provides an effective way to retrieve elements in smaller granularity than the whole document. A very important application of the structured document retrieval system 100 is to satisfy user's query needs by proper elements in any granularity.
In this disclosure, a hierarchical indexing mechanism for structured documents is described, as well as a scalable structured document retrieval method that operates based on the hierarchical indexing mechanism. An index term is propagated to an upper level element in the tree structure if it represents a more general concept judged by comparing its statistical information in a specific level element with other peer terms' values. Thus in the same structure of the document, index terms are distributed across the whole tree. Each element has a list of index terms which can best represent the concept of the element. The scalable retrieval method is dedicated to provide users with the most satisfying elements in any level.
Experimental results show that the structured document retrieval process as shown in
Although the systems and methods have been described in language specific to structural features and/or processes, it is to be understood that the invention defined in the appended claims is not necessarily limited to the specific features or processes described. Rather, the specific features are disclosed as preferred forms of implementing the claimed invention.