Full-Text XML search is a difficult problem due to its structural and textual constrains. The latter being individually well explored topics, lack mechanisms of integration with each other. Different approaches were proposed to cope with this problem. A technique of personalization usage to enhance one of these methods is given. The usage of direct or indirect results assessment is presented.
During the recent years XML has been widely applied and accepted as a standard for data integration and exchanging. As the popularity was growing, so did the range of applications of this format and formidable collections of text documents, represented in XML appeared. This, in turn, led to the actualization of a textual search in XML documents.
XML documents can be divided into following categories: • Data-Centric • Document-Centric
The former are data in a classical interpretation:
results of mathematical calculations, experimental
measurements, etc. They are processed by standard
XML query languages like XPath and XQuery [
New approaches [
• Tagged search, where term receives additional information about a tag, where it should be found.
•
Navigation approach. This approach attempts to enrich navigational languages with some means to allow textual search.
• A mixture of XQuery-like language and keyword search. The goal of these systems is to add a feature of textual search to XML query language, while leaving untouched its initial capabilities to search non-textual data.
A full overview of this classification, and an
extensive list of example systems can be found in [
• The first method [
• Second method employs the idea of relaxations
[
There are a lot of exterior to XML evaluation
and information retrieval techniques dedicated to
improving quality and speed of search using:
onthologies, statistics etc [
was proposed. Its architecture is shown on Fig. 1. Personalization occurs in two modules: the personalization module, which performs query rewriting and the ranking module, which employs different mechanisms of ranking corresponding to different user profiles. The system uses a repository of user profiles, which contain rules specific to a given user. When the query is formed, the system rewrites it according to these rules. Contextual hints, concerning the information field, represented in a form of rules, are fed into the aforementioned system. This additional information could be, for example user's geographical location, his age, interests. These data could change the nature of search and supply more relevant answers.
In this paper I present the idea of application of these mechanisms to user's awareness of a document structure. The proposal is to extract (from query history) and keep some facts about user’s knowledge of document. Query evaluator would utilize these facts, performing the rewriting. 2 Model
The architecture of the system partially reuses the design of the personalized query engine from the previous chapter; it is shown in Figure 2. Let’s see how query flows through the stages mentioned in the diagram. When the query reaches into personalization module it is copied and rewritten according to applicable transformations of its tree pattern. These applicable transformations are acquired from the repository. Then, these queries are evaluated; their results are mixed according to the personal ranking scheme, and presented to user. By choosing the relevant documents user can notify the system about effectiveness of existing transformations, or of the necessity of new ones. The result analyzer is responsible for dealing with user feedback from delivered results.
Our mechanism of the query rewriting will be
based on a modified well-known technique of query
rewriting called relaxations [
• Axis generalization. This relaxation is a modified source query, where parent-child condition is substituted with ancestor-descendant condition. Consider such an example query (in XPath notation): /a/b. The relaxed query is the /a//b.
• Leaf deletion. An idea of constraint dropping is the basis of this relaxation. Its result is a set of queries, each one of them has some leaves clipped out.
• Subtree promotion. It is possible to get relevant results from a modified query, which has some of its nodes lifted up the hierarchy. An extreme case of this structural promotion is the leaf promotion. The example of application this relaxation to /a/b[./c] is /a[./c]/b.
• Contains promotion. Another case of promotion is the textual constrains promotion. This operation takes the constraint of some node and then moves it to its parent, resulting in a new relaxed query. Given a query like this a/b/c[“textual constraint”] a possible relaxation would be a a/b[“textual constraint”]/c.
We would change the sense and the notion of relaxations. To distinguish the sense, we will call our modification of relaxations – transformations. The above-listed relaxations are designed by their in-system meaning, we would aim to redesign them guided by more explicit user’s specificity.
Possible transformations could be divided into two classes:
• Structural transformations The aim of these transformations is to determine user awareness regardless of textual content.
Here is an example of transformations: the leftmost tree is the initial query, the solid line marks an ancestor-descendant relationship and the thick line is a parent-child relationship. Then, we have a list of some possible transformations, which became less and less strict.
•
Their aim is to determine user's awareness of the subject in general. This class of transformations guides the dissemination of textual constraints. For example, having a bush representing a query, a set of acceptable transformations will consist of trees, the content predicates of which are "lifted up" to parent nodes.
Figure 4 shows an example of possible transformations with textual constraints. The leftmost tree pattern in the upper row is the initial query. The rightmost tree pattern shows the least specific query which could supply user with relevant results, which can’t be found by initial query. Thus, it could relieve user from additional work: for example, searching for a few related facts (the example shows a book of calculus theorems, some of which are proved on the basis of others) in a mathematical book, without knowing their hierarchical relations, one can write a dozen of queries with irrelevant results or even, nothing.
The core of this system is the module Result Analyzer and its subsystem called history. The history contains user’s queries, system response (query results), and users response (it may consist of the chosen item, a set of chosen items, users time to examine a particular result or some other behavior which can be a hint of success or failure).
Now we describe some important history properties: 1) Storage of a query, results and user's judgment.
The main role of the history is to create all necessary conditions for the fact analyzer and supply all possible transformations.
2) History management
History must include control interface which should allow flexible customization of the acquisition and usage of rules, which represent user's knowledge (through transformations)
3) Dynamical nature of history: updates and discards.
User's knowledge is not static. During the work with the document, a user acquires the knowledge about its schema. Also, he could get knowledge about a textual content of some subset of nodes and start using it very effectively. To enhance quality of search, this model should update its facts about user's knowledge, deleting or substituting the former. Or there exist another kind of cases - when a user didn't used the system for a long time, one can assume that his knowledge has reduced, so the system has to mark this by discarding the related data. Also, history should cope with dynamicalness of the source data. That means that adequate responses to updates and in a lesser extend to deletes of the underlying XML source data are required. This situation with the preference of updates upon deletes is generated by the nature the texts: it is not likely that information would be deleted.
4) Hierarchy and compositionality of facts.
The facts of user awareness could be combined with each other, effectively giving new knowledge. Conversely, the system may have more general knowledge of user awareness, but the query involves some specific information, which should be deduced.
Further work includes the evaluation of existing fact manipulation models, experiments with them. The aim is to find appropriate model and tailor it for XML retrieval needs, or invent it. Also, this model should allow subsistent level of self-manipulation, to provide sufficient level of freedom of model altering to user needs. This model lacks specific ways of determining user response, and they should be introduced. The model will also undergo some refinement caused by alternation of the relaxation model: that not all kinds of transformations can be derived from existing systems, and not all relaxation will fit into this model due to different purposes of these systems.