Patent Application
20050171946
The present invention relates in a general manner to methods and systems for managing resources such as web pages accessible via the Internet, or any other types of documents, aimed on the one hand at improving the obtaining of resources that are “close” to given resources, in terms in particular of centers of interest for the user, and aimed on the other hand at allowing the user, in a particularly simple and intuitive manner, to effect associations between resources himself, especially so as to benefit therefrom during the obtaining of close resources.
The quantity of information potentially relevant for each individual is becoming such that the present procedures for storing and searching for information are scarcely adequate. Alongside systems making it possible to retrieve information organized explicitly (such as “favorites”) or by key words (via a search engine), it would be desirable to have available a method which spontaneously proposes context dependent relevant information.
Systems which provide relevant links (or rather “related links” to use the jargon) with respect to a current page visited on the web are known. Typically these systems comprise an extension to the Internet browser which communicates with a remote server which provides the relevant links as a function of the current page presented in the browser's main window. Typically these links are presented, in the form of a list of URLs, in a window adjacent to the browser's main window.
However, such systems are not extended to serve as associative memory.
An object of the present invention is to propose computer methods and systems for searching for resources (especially web pages, diverse computer documents) that are “close” to given resources (this notion of closeness being made explicit later), and methods for the associative management of resources.
In particular, the invention is aimed at characterizing information elements with respect to new pages which appear on the web, thus opening up the way to multiple new applications of dynamic management of content with respect to the user's browsing context.
More precisely, it is the aim of the invention that each information element be associated with links on relevant web pages which characterize it and which are automatically maintained up to date. It is thus possible to characterize nontexual information, such as photos, sounds and animations (in flash, etc.) and dynamically select the elements to be presented to the user as a function of the context of his browsing which is also characterized by sets of relevant web pages. This approach is suitable especially, but not exclusively, for magazines in the art of living, fashion and in all other areas of “taste” where it is difficult to characterize through key words the interest shown by the subscriber in an item of information (when for example it represents a piece of music, a piece of art, a culinary dish, etc).
Another object of the invention is to associate other targeted elements, such as targeted advertisements, with information elements, in exchange for an innovative associative memory service offered to surfers.
In particular, the aim is that, typically by means of an extension of their browser (extension downloadable from a given website), users can use the information elements of this site as “associative memory”. Thus, during the user's browsing, the most relevant element of the site with respect to the web page visited—as well as with respect to the browsing context—will be presented to him spontaneously; the user will then be able to drag and drop onto this element any resource from his computer, such as the icon of a file of the client station, or else the URL of a web page, so as to store it. Thereafter, each time he visits any web page which is relevant with respect to this element, the resource that he had stored will be presented to him spontaneously, together with the resources (such as advertisements) that the author of the element had himself associated with the element. The advertisements presented will thus correspond to the current centers of interest of the user and are provided in exchange for a new associative memory service.
The invention is aimed moreover at harnessing modern user interfaces to create, in a particularly simple and intuitive manner, associations between information resources (web pages, or document files) especially within the framework of the above objectives.
The invention proposes according to a first aspect a method for determining relevant additional resources with respect to a given set of starting resources, characterized in that it comprises the following steps:
The relevance score calculation performed in step c) comprises the calculation of a plurality of sums of citing resource relevance scores, each sum advantageously comprising only the relevance scores of the citing resources comprising a link to a given resource consisting of the candidate resource or a starting resource.
In a preferred manner, the above method also comprises the calculation of at least one sum of citing resource relevance scores, each sum comprising only the relevance scores of the citing resources comprising a link to one among a set of at least two given resources, this set comprising the candidate resource and at least one starting resource.
According to a second aspect, the invention proposes a method for determining relevant additional resources with respect to a given set of starting resources, characterized in that it comprises the following steps:
The invention furthermore proposes a system for browsing among information resources, each resource comprising at least one link activatable in a first mode by an input device so as to bring about access to another information resource designated by a resource identifier associated with this link, characterized in that at least certain resources comprise at least one link activatable in a second mode with the aid of an input device so as to send to an engine for searching for new information resources a search query containing the resource identifier associated with the link in question.
This system exhibits the following preferred but optional aspects:
The invention also proposes a system for searching for new information resources on the basis of existing information resources, characterized in that it comprises a search engine based on the analysis of links between the various resources and accepting as input a query comprising a series of resource identifiers, a means of selecting identifiers which is able to store a set of identifiers (URI) of resources selected one after the other by a user, and a user activatable query generating means for devising a query containing the set of identifiers previously selected destined for the search engine.
In a preferred but nonlimiting manner, the means of selection is able to store the identifiers selected in a remanent manner, in such a way that the means of selection can be implemented in a manner staggered over time with a view to the generation of one and the same query.
The invention moreover proposes a method of searching for new information resources on the basis of existing information resources, characterized in that it comprises the implementation of a search engine based on the analysis of links between various resources and accepting as input a query comprising a series of resource identifiers and in that it comprises the following steps:
There is also proposed a method of searching for new information resources on the basis of existing information resources, characterized in that it comprises the implementation of a search engine based on the analysis of links between various resources and accepting as input a query comprising a series of resource identifiers and in that it comprises the following steps:
According to a preferred aspect of the above method, each group of resource identifiers is represented by a graphical object on a display device of the user, and in that said signaling is carried out at least by change of appearance of this graphical object.
The invention furthermore proposes a method of managing resources in a computer system provided with a display screen and with an input device for cursor movement and actuation such as a mouse, each resource possessing a representation displayed on the screen in such a way as to be able to be moved with the aid of the input device, method characterized in that it comprises the following steps:
Certain preferred, but optional, aspects of this method are the following:
The invention also proposes a method for identifying on the basis of a text resource, part of said resource able to constitute a pertinent query for a search engine, characterized in that it comprises the following steps:
Advantageously, the step of implementing the method for distilling resources is performed only with text parts selected as prevalent, where the citing text parts are the text parts which comprise at least one word in common with the prevalent text part or parts, where a link is created from each citing text part to the prevalent text part or parts, where the text parts containing at least one word also contained in the citing text parts are identified, so as to form a group of co-cited text parts, and where a link is temporarily created from each citing text part to each co-cited text part with which said citing text part possesses at least one word in common.
The text parts are typically phrases.
According to another aspect, the invention proposes a method of managing information resources such as web pages in a computer system comprising a user station furnished with a display screen, each resource possessing an identifier (URI) allowing its access from the user station, method characterized in that it comprises the following steps:
According to certain preferred but nonlimiting aspects:
The invention moreover proposes a method for identifying information resources accessible via recent links (such as web pages), relevant with respect to at least one given resource, characterized in that it comprises the following steps:
According to yet another aspect, the invention proposes a method for allowing access by a user to relevant information entities from a starting information entity, each information entity being accessible via an identifier (URI), characterized in that it comprises the following steps:
Preferred, but nonlimiting aspects of the above method are as follows:
According to another aspect of the invention, a method for determining relevance scores of text units such as phrases in a textual document, comprises the following steps:
The invention also proposes a method for determining relevance scores of text units such as phrases in a textual document, characterized in that it comprises the following steps:
The invention also proposes a method for determining scores allocated to words or groups of words contained in text units such as phrases in a textual document, characterized in that it comprises a step which consists in adding up the relevance scores, determined by one of the methods above, of the text units in which said words are located.
FIGS. 1 to 7 of the appended drawings illustrate various steps implemented in the present invention.
Lexicon
Resource (or element): Information resource such as a web page, a part of a web page, a document, or an XML element. Each resource may itself consist of resources, thus forming a tree structure.
Current resource: Resource accessed by the user at the current moment during browsing (it is in particular the web page displayed in the main window of the browser).
URI (Uniform Resource Identifier): Resource address. Will sometimes be used as a synonym for URL (universal resource locator).
Link: URI placed in a resource. In general, by clicking on a link, the user can access the resource pointed at by it.
Cite (a first resource cites a second resource): the first resource possesses a link to the second resource.
Popular: Said of a resource which is accessed by a large number of users (for example on the web) from its URI.
Private resource: Resource that is not accessible by a large number of users (in particular which is not published on the web or is not widely known).
Associative storage: Addition of a link to a first resource, on a second resource, so as to be able to retrieve the first resource via the associative search method.
Associative search: In order to retrieve a first resource, access to a relevant resource with respect to a second resource to which a link to the first resource has been added.
Added link: URI inserted by the user into a set of associated links.
Proposed spot: Spot presented by the system by priority since it comprises the associated links that are most relevant with respect to the current context.
Spot: A spot is composed:
Domain of relevance of a spot: set of resources designated by associated links of this spot.
Given associated links: Associated links specified explicitly (by whoever creates or publishes the resource with which said set is associated, or else by whoever creates a spot for this resource).
Completed associated links: Associated links determined automatically (in particular by means of a relative distillation algorithm described in the present description).
Associated link score: Score of relevance with respect to the set of given associated links. This score may be calculated by a relative distillation algorithm such as one of those described in the present description.
Authority score: Relevance score of a resource with respect to a set of given associated links.
Hub score: Relevance score of a resource citing other resources, representing the relevance of the cited resources with respect to a set of given associated links.
Non-contextual score: Context independent relevance score.
Contextual score: Context dependent relevance score.
Noncontextual spot: With respect to a resource (or to a set of resources) in question: Spot whose associated links comprise the URI of the resource in question (or at least some of the URIs of the resources in question) with a score (or a mean score) that is greater than a given threshold or that is selected in such a way as to maximize it (cf. the spot selection procedure described in the present description).
Contextual spot: Spot whose associated links are the most relevant with respect to the context.
Context: Browsing context.
Spot server: Server on the Internet providing the association between associated link and spot.
Current spot server: Spot server to which the user is directly connected.
Relevant region of a resource: Part of a resource containing at least one relevant link and containing no nonrelevant link.
Methods of Associative Storage and Associative Search
[Vocabulary used:
First page=page stored by the user so that he can retrieve it easily;
Second page=page used by the user as storage medium (to store an association with the first page, which we shall subsequently refer to as “for storing the first page” for the sake of conciseness);
Current page=page presented at the current moment in the main window of the Internet browser.
These are for example web pages, however the first page may be a private resource such as a document (text, multimedia or other document) which belongs to him].
The system allows the user to add a link to a first page on any second page whatsoever (or in the vicinity of the second page; we shall subsequently use the expression “on the second page” for the sake of conciseness).1
1 The step consisting in adding a link in this manner, on a second resource, to a first resource (so as to be able to retrieve it by the method described in this report) is called associative storage.
The user accesses the pages by means of a browser furnished with the system specific extension (or via an intermediate web server). Adding a link can be done for example by a drag and drop: the user grabs a handle representing the first page and drops it onto the second page; for example the link added is then presented by the system as a vignette in the style of a “post-it” in the place where it was dropped, or in a window adjacent to the main window of the browser (or in a frame adjacent to the frame presenting the original web page). He can also drop it on an icon representing the second page (for example in his favorite links). The system then stores the relation with the user considered, the association between the link on the first page and the second page in question.
Thereafter, when the user accesses a page relevant with respect to the second page (or the second page itself), the URI2 of this added link to the first page is automatically presented to him.
2 As well as optionally other indications pertaining to the link added, such as the text or the graphical object which accompanies the added link, or else a simplified or miniaturized presentation of the first page itself.
Thus, to retrieve the first page, the user merely has to access any page whatsoever3 which is relevant with respect to the second page.
3 Said any page whatsoever is already or will have to be taken into account by the system. The user will thus prefer to choose a popular page to speed up the search. The system is furnished with a crawler the aim of which is precisely to take into account as many accessible pages (especially on the Internet) as possible which are of interest to the user.
More simply, in so far as:
Note that during the step of associative storage the user can increase his chances by adding a link to the first page on several second pages.
Furthermore, in so far as the relevance relations are symmetric, the added links are implicitly bi-directional. Furthermore, in the case where the current page is a private resource, the system can liken it to the second page(s) on which, as appropriate, the user had added a link to this private resource, and present the other first pages that he also added on this (these) second page(s).
The step of associative storage can be automated (or be computer aided). Specifically, the addition of a link to a first page on a second page can be (semi-) automated according to the following steps:
Note that as regards step I, various techniques for automatically extracting key words or main phrases of a text already exist.
The key words may also be extracted from the text in the following manner:
The two methods presented above may be combined by retaining from the key words selected only those which are located in the phrases selected. The complete method for extracting the key words from the text is then as follows:
As a variant, in so far as (one or) certain phrases of the text may be labeled as being prevalent, to determine the scores of the phrases, instead of the absolute distillation procedure it is possible to use the relative distillation procedure (described later) to determine the relevance score of the phrases with respect to said prevalent phrases.
Moreover, instead of actual phrases, it is possible to consider any kind of text parts or units. The method using relative distillation thus consists in determining relevance scores of co-cited “text units” (such as phrases):
The text units comprising at least one word in common with the prevalent unit (or set of units) are identified so as to form a group of citing text units. A link is created (temporarily) from each citing text unit to the prevalent text unit (or set of units).
The text units containing at least one word also contained in the citing text units are identified so as to form a group of co-cited text units. A link is created (temporarily) from each citing unit to each co-cited unit with which said citing unit possesses at least one word in common.
One of the methods, described later, of calculating relevance scores by the relative distillation procedure is then applied. The whole set of identifiers of the relevant text units constitutes the URIs of the query.5
5 The set of identifiers of the citing text units constitutes the set R−. The set of identifiers of the co-cited text units constitutes the set R−, and so on and so forth.
The implementation of the associative search system will now be described.
To present, to a user who accesses a current page, links on first pages, the system performs the following steps:
As a variant, during the associative storage, instead of adding on the second page a link to the first page, the user can overlay onto the second page or insert thereinto an annotation (or any resource such as an icon or other graphical object), which then plays the role of first page within the sense of the present method. In this case, during step c) of the associative search, the system presents the second page or pages which have been selected while also presenting their annotations (or the resource that has been added to them).8
8 In the remainder of the description, the expression link added on a second page is understood to mean that we include this typical case where there is a resource added to the second page.
To facilitate reading, the following 7 steps (see
To determine the relevance score of the second candidate pages with respect to a current page R (understand R here as current resource11), the system implements a method of “relative distillation” comprising at least one out of the following steps a and a′.
11 Since here the query is formed of a single page.
Step a:
The calculation of the relevance scores in steps a-3 and a′-3 may be performed by means in particular of one of the equations presented later in the “selecting the spots” section which moreover describes improvements to the method presented above. In particular the scores are sharpened by successive iterations. During these iterations, the hub pages in step a and the authority pages in step a′ also acquire relevance scores (hub scores and authority scores respectively). In addition to the second candidate pages (that is to say in addition to the URIs of the pages of R−+ in step a and/or of R+− in step a′) determined as described hereinabove, it is then also possible to include, in the resulting set provided at step b, the hub pages of step a and the authority pages of step a′ (since they now have relevance scores). Moreover the weights of the links between close pages13 are diminished so as to further improve the results.
13 To identify the closeness of the pages to the ends of the links the system additionally identifies the set of pages R− of the pages possessing at least one link to the pages R− and the set of pages R−+− of the pages possessing at least one link to the pages R−+− (see the “filtering” section).
The system can therefore select the second pages that are most (or sufficiently) relevant to step b and perform step c to present their added links to the user.
The results obtained by the relative distillation method may be stored (then maintained—see later the “maintaining the spots” section) with the aim of avoiding recalculating them during accesses to the current pages already processed. Thus, the system maintains, in a second memory, the scores of the second pages with respect to the current pages in the cases where these scores are greater than a given threshold. For a current page already processed, the response of the system is then almost immediate.
Stated otherwise, step a is modified as follows:
Step a′: Consult the second memory to ascertain whether the second pages most relevant for the current page have already been stored (and if these data in memory are sufficiently fresh), as appropriate go to step c, otherwise determine and store the relevance score of second candidate pages with respect to the current page.
As a variant, the system stores (then maintains—see later the “maintaining the spots” section) the necessary data without waiting for a user to access a current page; storage is triggered by the use, by the user, of a new second page (as associative storage medium).
By utilizing the fact that the relevance scores are reflexive14, the system starts from each second page to construct R− and R−+ and (R−+−) and/or R+ and R+− (and R+−+), calculates by relative distillation the relevance scores of all the potential current pages, and stores them in a second memory (this being an inverse memory able to provide, for each potential current page, the second relevant pages).
14 (i.e. the relevance score of a second page with respect to a current page is equal to the relevance score of this current page with respect to this second page)
Moreover, as already indicated, the system maintains a first memory containing the links added by user and second page.15
15 Note that, advantageously, the data in the second memory are not per user and may thus serve all the users.
Thus, when a user actually accesses a current page, the system selects from the second memory the second pages—from among the second pages used by this user as storage medium16—which have the highest relevance scores with respect to said current page, then retrieves (from the first memory) the links added by this user on these second pages.
16 (they are indicated in the first memory)
Stated otherwise, the method comprises the following steps17.
17 Steps m1 and m2 describe the associative storage method, steps a, b and c describe the associative search method.
For each new second page R (on which a user adds a link)18:
18 Step m1 is performed only for the new second pages, while step m2 is performed each time a second page is used by a user, whether or not it is new for the system.
Step m1: Perform at least one of steps m1-1 and m1-1′, then perform step m1-2:
Step m1-1:
Step m1-1′:
Step m1-2: store, in a second memory, the URIs of the pages having a sufficient relevance score with respect to R, in relation to R, in such a way that on the basis of the URI of each of said pages having a sufficient relevance score with respect to R it is possible to retrieve19 (the second page) R as well as said sufficient relevance score;
19 (As well as the other second pages, as appropriate, for which the relevance score of R is sufficient)
Step m2: (in parallel with step m1) store in a first memory, for each user and each second page, the added links that said user has added on said second page;
During access to a current page by a user:
(Step a is no longer necessary since the scores are already in memory).
Step b-m: Select from the second memory a certain number of second pages20, from among the second pages used by said user (that are indicated in the first memory), for which the relevance scores of said current page are the highest (if they exist);
20 Normally, in the second memory, the URIs of the second relevant pages with respect to a potential current page are already sorted by relevance score.
Step c (unchanged): retrieve from the first memory the links added by said user on the second pages selected in step b-m and present them to said user (with optionally the second pages on which they have been added and in a sorted manner).
The improvements presented later in the “selecting the spots” section will also be applied. In particular as the scores are sharpened by successive iterations, the hub pages in step m1-1 and the authority pages in step m1-1′ also acquire relevance scores (hub scores and authority scores respectively) and may thus be included in the resulting set provided in step m1-2 (in addition to the URIs of the pages of R−+ in step m1-1 and/or of R+− in step m1-1′). Moreover, here also the weights of the links between close pages are diminished so as to improve the results (see the “filtering” section).
With this latter method, the added links are presented almost immediately by the system in all cases, that is to say even when a current page is accessed by a user for the first time.
It was mentioned that during the associative storage step the user can increase his chances by adding a link to the first page on several second pages. He will now be allowed to form groups of second pages to which is added a link to the first page (the idea being that, as the first page may be of interest with respect to more than one center of interest of the user, the groups make it possible to class the first page with respect to distinct centers of interest, each group corresponding to a different center of interest).
Specifically, each time the user adds a link (to the first page) on a new second page, the group or groups of second pages that he had already formed, as appropriate, for the first page are proposed to him by the system and he can then choose one or more of these groups into which to insert said new second page, or otherwise he can create a new group formed of the single new second page.
At the same time he can also manipulate his groups more widely, such as for example delete a second page of a group, split a group into two, merge two groups, delete a group, etc. Finally, he can also duplicate a group so as to add thereto a link on another first page.
Each group is processed by the system as a relative distillation query. In a similar manner to the last method described21, for each query R (that is to say for each group of second pages) the system identifies and stores (then maintains—see later the “maintaining the spots” section) the potential current pages which have a sufficient relevance score, and thus forms an inverse memory able to provide, for each potential current page, the most relevant queries (that is to say the most relevant groups).
21 The difference is that here R represents a query formed of one or more resources whereas before R represented a single resource (a single second page).
Stated otherwise, the associative storage comprises the following steps:
(Step m1 is performed only for the queries not already known by the system or not sufficiently fresh, while step m2 is performed for all the users' queries, whether or not they are new for the system).
Step m1: Perform at least one of the steps m1-1 and m1-1′, then perform step m1-2:
Step m1-1:
Step m1-1′:
Step m1-2: Store, in a second memory, the URIs of the pages having a sufficient relevance score with respect to R, in relation to R, in such a way that on the basis of the URI of each of said pages having a sufficient relevance score with respect to R it is possible to retrieve22 R as well as said sufficient relevance score;
22 (From among the set of queries stored, as appropriate, for this page)
Step m2: (in parallel with step m1) store in a first memory, for each user and query, the added links (to first pages);
During access to a current page by a user:
Step b-m: Select from the second memory a certain number of queries, from among the queries (groups) used by said user as associative storage medium (that are indicated in the first memory), for which the relevance scores of said current page are the highest (if they exist);
Step c: retrieve from the first memory the links added by said user on the queries selected in step b-m and present them to said user, with optionally:
The improvements presented later in the “selecting the spots” section will also be applied. In particular as the scores are sharpened by successive iterations, the hub pages in step m1-1 and the authority pages in step m1-1′ also acquire relevance scores (hub scores and authority scores respectively) and may thus be included in step m1-2 (in addition to the URIs of the pages of R−+ in step m1-1 and/or of R+− in step m1-1′). Moreover, here also the weights of the links between close pages are diminished so as to improve the results (see the “filtering” section).
In step b-m, the system provides a set of selected queries. It would be advantageous to sharpen the selection in such a way as to present to the user (the) request or requests24 that are the most relevant with respect to the user's browsing context. This is what will now be described.
24 (With the first pages and the corresponding relevant links)
The history of a user's browsing is modeled with the aid of a “context stack”, here with each link (that may be presented to the user) is associated a relevance score at each browsing level, and when a link is nonexistent it is likened to a link whose score is equal to zero.
When the user clicks on a link and accesses a new page, the system adds a level to the context stack. On the other hand, when he clicks on the “back” command of his browser the system pops a level.
For a given link, the contextual score is an average of the noncontextual scores25 at each level of the context stack, these scores being weighted as a function of depth. So as not to have to recalculate all the scores each time, an exponential weighting is used, this implying that the contextual score at a certain level is the weighted average of the noncontextual score at this level and of the contextual score at the previous level.
25 (That is to say determined taking no account of the context)
Stated otherwise, for a given URI, s being the noncontextual score at the last level and r the contextual score at the previous level, the contextual score at the last level is: lambda.r+(1−lambda).s (lambda being a constant weighting between 0 and 1, in principle less than ½: the larger lambda is, the more important is the past).
Among the queries (that is to say the groups) selected in step b-m, the system selects those which are closest to the context, that is to say those for which the scores of the URIs stored in step m-2 are the closest to the contextual scores for the user in question. To determine the closeness of each request with the context, the system calculates the sum of the products, for each URI of the query, of the (noncontextual) score of the query with the contextual score for the user in question.
Step b-m is thus replaced by the following step b′-m:
Step b′-m: select from the second memory a certain number of queries, from among the queries (groups) used by said user as associative storage medium (and indicated in the first memory), for which the relevance scores of said current page are the highest (if they exist) and for which the relevance scores of the potential current pages are the closest to the contextual relevance scores.
We shall now describe a method, utilizing the system of cookies, for recognizing the user when he goes from one site to another, in such a way as to be able to maintain his context stack.
Let us recall that the cookies system allows servers of sites of an Internet domain (i.e. domain name or IP address) to recognize a user (that is to say his computer) when he accesses web pages belonging to one and the same Internet domain.
The method described here allows a server, which implements our method—it will be called a client server (CLI)—to recognize even users who browse from one site to another which do not form part of one and the same Internet domain, even though in their browsing these users pass through sites that do not implement our method.
To do this, three communication mechanisms are used:
There are three possible cases which are described hereinafter (see
New user for the client server and for the centralized server:
New user for the client server but not for the centralized server:
User already known to the centralized server and to the client server:
The method described above makes it possible to select the links to be displayed in the web pages as a function of the browsing context26. This is what will now be described.
26 (Or, as described above, to select the queries themselves; this being trivial, it is not described again)
Let us start from the situation where each query (the server which hosts it) possesses a set of initial URIs as well as the set of links that could be proposed to the user with their default scores: the noncontextual scores.
As already described, the contextual score is an average of the noncontextual scores, weighted as a function of depth, at each level of the context stack. Thus, ri being the noncontextual score at the last level and {tilde over (r)}i the contextual score at the previous level, its value after having followed a link is: {tilde over (r)}i♦λ{tilde over (r)}i+{overscore (λ)}ri27.
27 Thus giving {tilde over (r)}i={overscore (λ)}n=0d−1λnri,n+λdri,d with d the depth of the root and ri,n the score of page Pi at depth n.
The links presented to the user are those which have the largest contextual score.
The context stack can be displayed in the URS frame (the first frame) introduced above. Thus the user can see which pages are the ones that were involved in the calculation of the pages to be displayed. He can click elements of the stack to climb back up the levels, and an “Erase” button makes it possible to empty the context stack.
The context stack is stored, for each user, in the centralized server (URS), with the user's identifier. Thus, each time a user opens a page at a client server (CLI), the latter, having obtained the user's identifier, will give URS the noncontextual scores28, which will respond with the contextual scores after having performed the weighted average described above29. The server of the client site may then display in the page the links which have the best score.
28 To avoid unnecessary traffic it is possible to select the pages to be sent, taking only those that have a score greater than a certain threshold, for example half the threshold required in order for a page to be displayed to the user
29 This is performed within the framework of step 6 described above.
The steps are thus as follows (see
It may be beneficial on the one hand to group the links in various parts of the pages, or even to hierarchize the parts, that is to say to allow parts to contain subparts, in addition to links. Here are the changes that this involves:
A more complete example (see
Here therefore is what happens when the user, already in a particular context (for the page cl/com/main.html), clicks on a link http://CLI.com/index.html?part=1 (part=1 signifies that the user has clicked in part 1). It is assumed that the client server CLI does not yet know the user:
The concept of user can in reality encompass several users who share added links (and the groups which serve them as support). Of course, a finer organization of the users according to the added links that they share is possible.
We shall now describe the case where an end user subscribes to a provider user so that, according to the context, the system proposes the groups and first pages (in the sense of the groups and first pages described hitherto) created by the provider user to the end user. The first pages may in particular be advertisements which (by virtue of the capabilities of the system as hitherto) are automatically selected with respect to the context.
The groups created by the provider user and proposed by the system to the end user are called “spot”.
The provider user manipulates and utilizes the spots as described hitherto for the groups of second pages.
The end user can use a spot as storage medium by making a personal version thereof and adding thereto a link to a first page (this is described later).
The main advantage of this approach is to afford the possibility of creating new spots (and the expensive calculations of scores that they involve) to certain users only (namely the provider users) and to offer the function of storage/associative search by way of pre-existing spots (which is not expensive in terms of machine resources) to all users.
Spot
The system that we shall now describe provides relevant links (also known as “related links”, see above the “state of the art” section). However, rather than searching for relevant links directly, our system searches firstly to see whether there exists a spot—or reference resource—whose associated links are sufficiently close to the current resource or to the browsing context of the user. If such is the case, the system returns the spot(s) whose associated links are the closest, as well as its associated links offered in the guise of relevant links.
Typically the spot is proposed in a window adjacent to the main window of the browser, like the existing systems providing “related links”, however in contra-distinction to these existing systems
Furthermore, presenting the end user with relevant links by way of spots offers advantages per se, such as prompting to click in order to access the reference resource (that is to say the page presenting the spot).
Let us now examine a few typical storage/associative search scenarios implementing spots.
First Scenario of Use:
The provider user creates a new resource or chooses an existing resource (for example a web page which he wishes to access, or a particular element contained in a page . . . ) so as to make thereof the reference resource of a new spot.
To do this, he allocates it at least one given associated link pointing to a popular page.
The system completes the set of associated links34 (as described in the “selecting the spots” section).
34 This is the equivalent of the second memory described in the previous section.
Thus, in the future, each time an end user accesses a resource pointed out by one of the links associated with this spot, this spot may35 be proposed to him. Also, as described in the subsequent two scenarios of use, end users may then use this new spot as storage medium (in a manner analogous to the use of a second page or of a group of second pages, described above).
35 It will not necessarily be this spot that is proposed but rather, among all the spots whose associated links point to resources forming the current context, the spot in which these associated links have the highest relevance scores (or the spots in which these associated links have the highest relevance scores). The selection of the spot (or spots) is described in the “selecting a spot” section.
The creator of this spot thus has the advantage not only of putting it to his own use but also of seeing it proposed to end users. As a link on the reference resource (prompting the user to click) is included in the presentation of the spot, the reference resource is thus promoted to the end users. Moreover, its added links (such as advertisements) on this spot will be presented to the end users.
Second Scenario of Use:
On the web the end user “lands” on a first page (or other type of resource) that is so interesting that he would like to store it in order to be able to retrieve it easily and land back on it spontaneously when he accesses resources that are relevant with respect to it.
Let us assume that no spot is spontaneously proposed by the system for this page.36
36 In the converse case, on (his personal version of) this spot, the user will directly add a link to this first web page. Note however that this action is not strictly necessary. Specifically, already without doing anything the user will have to retrieve this first page by visiting a close page that is not very popular (in the guise of relevant link associated with this same spot or with a neighboring spot). However, by doing this action the user has the extra advantage of being able to retrieve it in the guise of link added explicitly by him, that is to say in such a way that it is made evident.
The user visits a (at least one) second page, which is relevant with respect to the first,
In the future, this added link will then be presented to him spontaneously each time that this same spot, or that a close spot, is proposed to him for the current context of his browsing.
Third Scenario of Use:
The end user wishes to store a private resource (such as a document which belongs to him and which is not published on the web). The private resource here plays the role of first page.
He accesses a (second) page which is relevant with respect to his private resource (and which preferably is popular, or for which he knows that a spot is proposed) and he adds thereto a link to his private resource (that is to say he inserts this link into his personal version of the spot proposed for this second page).
Optionally, to reinforce his action, he will also add a link (to his private resource) on yet (other spots which are proposed to him for) other second pages that he finds relevant with respect to his private resource.
In the future, a link to his private resource will be presented to him spontaneously each time that one of the spots that was proposed to him for the second page or pages, or that a close spot, is proposed to him for the current context of his browsing.
Thus, in the last two scenarios above, a link to the first page is presented to the user spontaneously each time that he visits pages in the domain of relevance covered by the spots proposed for the second pages37.
37 And insofar as the second pages were chosen by the user because according to him they are relevant with respect to the first page, and the relevance relation is transitive at this level, a link to the first page is presented to the user spontaneously each time he visits pages which according to him are in the domain of relevance of the first page!
Selecting the Spots
Before the spot(s) selection step proper, the system must obtain the set of “completed associated links” from the set of “given associated links” (which are given by the provider user, as described in the first scenario of use).
Completing the Associated Links:
The set of resources pointed at by the given associated links is the query R.
The calculation of the completed associated links is performed by means of the “relative distillation” method, comprising the following steps:
The calculation of the scores in step 3 may be performed by calculating, for each resource of R−+, the ratio between
The authority scores are normalized (in such a manner that their sum becomes equal to 1).
The authority scores having been obtained, they can be put to use to allocate hub scores to the elements of R−:
Iteration restarting from step 3: the hub scores having been obtained, they can be put to use to sharpen the calculation of the authority scores. Step 3 then takes account of the hub scores so as not to consider all the elements of R− on an equal footing (the resources of R− pointing to resources having a higher authority score will thus have a greater influence). The cardinalities used to calculate the authority scores are thus replaced by weighted cardinalities. That is to say each hub resource, instead of counting for one, counts proportionately to its hub score. (The equations are detailed later).
Step 3 then includes the taking into account of the resources of R−+− pointing to the resources of R−+ having the largest authority scores, in addition to R− (a method optimizing the way in which R−+− is taken into account is described later).
After step 3, we can optionally perform step 4 again, and so on and so forth until convergence, that is to say until the difference between the results obtained in the last iteration and those obtained in the previous iteration are negligible (in general, fewer than 10 iterations are sufficient).
Variant for step 2: to form R−+, instead of taking all the links contained in the resources R− the system will take only the links located in the relevant regions of the resources of R−. As these relevant regions can be determined only onward of the moment at which the hub scores of the links that they contain are known, this variant will be implemented only onward of the first iteration, that is to say after having performed step 4 the system will iterate restarting from step 2 rather than from step 3.
Variant for Step 3:
With each link possessed by a resource of R− (or of R−+−) is associated a weight equal to the complement of the closeness of the two resources connected by this link. Thus, the links connecting two close resources will be weakened. Thus the importance of the links between the resources which mutually promote one another (for example because it form part of one and the same web site and mutually cite one another) is thus decreased. Once the links are thus weighted, the system calculates the authority scores, not now by using the sum of the hub scores, but the sum of the hub scores multiplied by their weights (this is detailed and illustrated by an example later).
The closeness of the two resources connected by the link in question is obtained by calculating the ratio between
It is also advantageous to perform the same algorithm downstream, that is to say by calculating the hub scores of the resources of R+− (which downstream cite the same resources as the query).
The downstream algorithms are identical to those upstream except that B (backward) is replaced by F (forward) and vice versa39, and − is interchanged with + (e.g. R−+ is replaced by R+−).
39 B(Ri) is the set of URIs of the pages having a link to the page Ri. F(Ri) is the set of URIs of the pages to which Ri has a link.
Consideration will also be given, advantageously, to the hub resources upstream and the authority resources downstream, in such a way that the hub pages in step m1-1 and the authority pages in step m1-1′ also acquire relevance scores (hub scores and authority scores respectively) and may thus be included in the resulting set provided at step m1-2 (in addition to the URIs of the pages of R−+ and/or of R+−).
By completing the associated links of each new query (spot) introduced, the system forms an inverse memory able to provide, for each potential current resource corresponding to an associated link, the most relevant queries (that is to say the most relevant spots).
Stated otherwise, the associative storage now comprises the following steps:
The relative distillation method will now be detailed.
The essential idea of the calculation of the relevance score (of a web page P2 with respect to a given web page P1) is as follows41:
41 Hereafter, we shall assume that P1 and P2, (or Pi, Pj, etc) are web pages, although the methods described are far more general, as has already been mentioned. For example, it should be noted that instead of utilizing the hypertext links and the queries as mentioned hereinabove, the system may be based on analysis of the traces of the cutting and pasting of information fragments performed by the users (within the framework of creating and manipulating information resources), so as to automatically suggest other fragments which might enrich these resources. These traces may in fact be likened to links. For example, when part of a web page is copied into a document, the system is capable of deducing therefrom and of storing the existence in the document of a link to the web page, and the same mechanisms described here may then be applied. Moreover, the method described here may advantageously be applied by likening the links from one resource to another resource, to links from a user to a resource that he likes (that is to say to a resource which interests him). It is thus possible to determine the quantity of common reasons (between several resources) to be liked by users. This can in particular serve to categorize these resources.
The relevance of a page with respect to a set of pages may be defined by the “quantity of common reasons” to be interested in all these pages.
Algebraic calculations make it possible to obtain equations giving the quantity of common reasons between several pages. This quantity (or closeness, or else homogeneity) is denoted x, subscripted with the pages concerned; the probability of being linked to a certain page Pi is denoted pi; the probability of being linked to at least one page out of Pi, Pi, . . . , Pn is denoted pij . . . n:
and so on and so forth (all the subsets of odd size in the numerator, and the others in the denominator)43.
43 The bars above indicate complements, and pø, the probability of liking at least one page of an empty set, is a constant equal to zero; it is present in the equation for reasons of consistency.
This equation may be denoted more compactly thus:
The probabilities concerned above involve the number (the count) of pages of R− which contain a given link or a link from among a set of given URIs (to pages of R−+). It would be beneficial to weight this number by the quality of citation (hub score, described later) of each page which contains such a link.
It would thus be desirable for a page of R− citing more better pages (of R−+) to be regarded as being of better quality of citation, and for in return a higher weight to be given to it within the framework of the calculation of the scores44 of the pages that it cites (R−+), the scores of the pages of R− and those of the pages of R−+ mutually influencing one another in an iterative approach (bipartite reinforcement) which converges45.
44 Recall that here one is dealing with relevance scores with respect to the query, in contradistinction to the state of the art which makes it possible to determine a score of quality “in the absolute”.
45 Note that the calculation of the relevance score of a page of R−+ may result in a negative value (that we will then neutralize; this is described later). Specifically, certain pages may not only be close to the query, but even be antagonistic with respect to it (the fact of being of interest thereto decreases the chances of liking the pages of the query and vice versa).
The number of pages of R−+− citing each candidate page (that is to say of R−+) also comes into the calculations. However, it is expensive to take them into account.
Hence, the results will be approximated by considering only those which cite the candidate pages having a good score, this score being calculated firstly by considering only R− and subsequently by extending this set to R−+− gradually. To calculate the relevance score of a candidate page, instead of taking the result of the equation for the quantity of reasons directly, it is preferable
After a first iteration, in the citing pages the system can
As the links in question are located under nodes of a typically tree-like document structure (such as in HTML in particular), to determine a relevance region it suffices to take the (minimal) nodes which encompass all the good links and to take away from them the (maximal) subnodes which contain a bad link (score too low, or URI explicitly refused) and which contain no good link (sufficient score). The algorithm makes it possible, having a homogeneous set (having sufficient homogeneity) of URIs associated with close pages, to obtain a list of URIs of pages which are relevant in regard to this set. The way in which this algorithm may be utilized to obtain a set of relevant pages for an inhomogeneous set will be described later.
As input, this algorithm takes
We have: K−⊂H⊂A− and T∩K=ø. (E being a set of URIs,
The upstream homogeneity wS of a set S is defined as follows:
Stated otherwise, ljP is equal to 1 if there is a link
This signifies quite simply that ap is the total of the hub scores of the pages (of H) which point at at least one page of P (P being the current subset of S which is considered).
For each existing link lji, it is possible to associate with it a weight as a function of the closeness of the pages Pi and Pj and thus to improve the result—see later.
Here, since ∀PiεK we have ri+=WK (the relevance is the same for all the pages Pi of K), the relevance score ri+ has to be calculated only once for the pages of K (besides, it will already be calculated during the procedure for chopping the query R into subqueries (kernels) K, and will therefore already be known on entry to the procedure).
Let
be the global variation of the relevance score.
If δ<ε(ε>0 being a margin of error) we assume convergence has occurred and the method stops. Otherwise, the method continues.
Initially, so as to process only a reduced number of pages, the relevance scores may be calculated on the basis of R− (if we took H=R−). Hence, this will only be an approximation. Specifically, for the scores to be correct, they have to be calculated based rather on H=R−+−. However, as the construction of R−+− is relatively expensive, we shall take only a subset: for R−+− we shall take only the pages pointing at the pages of A which have a good score.
Thus49, a subset will be added before the end of step 2.a:
49 Several procedures may be used; here we present the preferred one.
2.a. 1. In the case where the score ri+ of the current page (Pi of A) is sufficient50, ri+ is recalculated after having inserted the new pages of B(Pi) into H.
50 (That is to say greater than a chosen threshold; this threshold can be dependent on the current cardinality of H, specifically, the closer we get to R−+− (e.g. Hfinal), the more chance the calculated score has of being correct)
H|→B(Pi)∪H
We introduce an authority score for the pages of A and the equation ri+ is r=wi∪K·ai (rather than r=wi∪K). The new coefficient ai will make it possible to weaken the pages that are not very reliable (because they are not very popular). Furthermore, the equation will be more consistent insofar as the relevance score will no longer be the same for all the pages of the query.
The procedure is now as follows:
For each existing link lji, it is possible to associate therewith a weight dependent on the closeness of the pages Pi and Pj and to thus improve the result. This makes it possible to decrease the importance of the links between pages which mutually promote one another. Typically one thus succeeds in filtering for example the links of the “abstracts” and other “menus” which, repeatedly, are located in all the pages of a site.
The basic idea consists in weakening the links connecting two pages that we know to be close, by assigning a weight to each link, which weight will be equal to the complement of the closeness of the two connected pages (the greater the closeness, the more the link must be weakened). Once the links have thus been weighted, it is possible to calculate the homogeneity of a set of pages using the sum of their weights, rather than the number of citing pages.
In point 3.a of the algorithm, in the definition of the authority score we replace
Explanations:
Stated otherwise, if there is at least one link
To determine the closeness xji, we can take the equation (already described) for the quantity of common reasons:
The number of pages pointing at page B is equal to 0.9+0.1+0.3+0.5=1.8
The number of pages pointing at A or B (NpAB) is equal to 0.9+0.2+0.9+0.8+0.3+0.5=3.6
Thus, if we assume that |H|+h=100, the calculation of the closeness of A and B gives:
The filtering described above uses a weight {overscore (xji)}. Since we now have the scores51 of the citing pages, we can optionally improve the method by taking
as weight (instead of {overscore (xji)}), where hj is the score of the citing page (weakening a link) (from a citing page Pj to a cited page Pi) further when the score of the citing page Pj is low.
51 (Whether absolute or with respect to the query)
It should be noted that in order to calculate the closeness xji between two connected pages Pi and Pj, instead of using the equation for the quantity of reasons as illustrated hereinabove, it is possible to calculate the ratio between:
We provide the system with a set R of pages and possibly a set of pages Rx of pages that we do not explicitly want (R∩Rx=ø). The system will identify within R at least one group of “homogeneous” pages and will launch a separate sub-query on this or each group. These groups are called “kernel”. To form the response, we shall then take a combination of the scores obtained. This method thus comprises the following steps:
As a variant, instead of basing ourselves on the homogeneity equation
as described hitherto, the relevance scores calculation example method can be based on another homogeneity equation, such as for
in which the ensemble cardinalities (represented between vertical bars) are replaced by the total of the hub scores of the pages in question53.
53 We can say that the cardinalities are replaced by “weighted cardinalities”, the weights being the hub scores.
Downstream Processing:
Instead of searching for the good pages in relation to those of a kernel from among the pages that are cited in common with them it may be beneficial to perform the same algorithms in the other direction, i.e. by searching among the pages which cite the same pages as the kernel, or even to perform both and to calculate an arithmetic average.
The downstream algorithms are identical to those upstream except that B is replaced by F and F is replaced by B, and − is interchanged with + (for example R−+ is replaced by R+−).
The upstream and downstream methods may advantageously be integrated in the following manner: after the upstream processing (possibly even after each upstream iteration), with the candidate pages (R−+) having obtained a sufficient relevance score, we associate downstream a set of extra pages (“artificial pages”) whose cardinality is dependent on said relevance score. Each artificial page is also cited by (at least) one page of the query. The scores of these good pages (of R−+) found upstream54 are thus given downstream an “advantage”, and consequently the scores of the pages (of R+−+) cited as appropriate by these good pages are also indirectly given an advantage.
54 Note that, advantageously, this is done without amalgamating the relevance scores upstream and downstream.
And conversely, after the downstream processing (possibly even after each downstream iteration), the same method is applied symmetrically upstream. Thus the good pages of R+− are favored, as are indirectly the pages (of R−+−) which cite them, as appropriate.
By not amalgamating the scores upstream (of the pages R−+) with the scores downstream (pages R+−) it is possible to dissociate them in the calculations. In particular, the influence of the scores obtained downstream can be decreased in the upstream processing or vice versa.
Moreover, by virtue of this idea of “artificial pages”, the present method may be applied as a complement to the existing procedures of the prior art. Specifically, once the scores have been obtained for each page, the respective numbers of citing and cited pages can be modified artificially before applying these procedures.
It is possible to trek (known as “crawling”) the web by following the links (upstream and downstream) around the previously cited pages of the 7 sets, exploiting the addition of the artificial pages to advantage the web pages linked to the pages which are more relevant with respect to the query.
Insofar as the pages having the best scores are presumed to be relevant to the user (and insofar as the relevance is transitive), the methods described here will be able to be applied recursively thereto to discover yet other relevant pages. It is thus possible to trek the web based on the user's query.
A system implementing the relative distillation method described hereinabove is able to receive a search query composed of a set of URIs making it possible to access information resources such as web pages and to provide in response the URIs (or directly the pages) which are presumed to be the most relevant with respect to said query.
The query being composed for example of the favorite links of the user and the goal of the system being for example to monitor the web around these links and to notify the user when new interesting pages appear therein, either by “Push” technology at the initiative of a server, or by “Pull” technology at the initiative of the user.
The user can of course provide the system directly with a set of URIs, nevertheless, other means may also be offered to him to assist him in the preparation and submission of a search query.
To trigger the execution of a search query from a hypertext link located in a page, the user can use any one of the devices from among the following:
Each of these devices can advantageously make it possible to execute said search query in addition to (in parallel with) access to the page designated by the link in question. The result of the search query will for example be displayed in a second window (new instance of the browser) or else in a subwindow of the browser55.
55 In a manner analogous to the subwindow existing today for favorites, this subwindow may be adjacent to the main subwindow in which the page containing the link that the user has clicked was displayed and in which the page accessed by the act of clicking on this link is subsequently displayed.
As a supplement to the link selected, other URIs may be added routinely into the search query56. They may in particular be:
56 Specifically, one of the essential advantages of the system is to be able to operate (find the relevant information resources) even if the search query is composed of a plurality of URIs.
We shall now describe how the user can prepare a query composed of several links that he gleans in the course of his browsing.
a) Displaying of the Current Query Under Preparation
Instead of triggering a search query directly, the user's action (as described above, for example the act of clicking on a link with the right-hand button and choosing the appropriate option) triggers the displaying of an accessory page in which:
Thus the user can prepare a query gradually, by selecting links one after the other59 during his browsing60 and thereafter send a query composed of several URIs.
59 (In one and the same page or in different pages)
60 (During one and the same browsing or staggered over time)
Said accessory page may additionally contain drop-down graphical objects (such as for example directories, records, folders, or similar metaphor) representing queries under preparation other than the query in progress. The user can thus choose the query or queries which will be enriched by the new link that he has just selected.
Following the preparation of a query from a URI corresponding to a hypertext link in a page (as described above), the already existing queries which, as appropriate, contain this URI are optionally presented to him.
Advantageously, said accessory page may be composed of two parts. One of these parts contains the elements described hereinabove (that is to say the elements of the query under preparation). The other part presents the content of the page designated by the link selected by the user.
For example, if the user clicks on a link while the page is in the state where all clicks trigger the displaying of the current query under preparation (or with the right-hand button of the mouse, etc.), the server returns said accessory page to it, which thus comprises:
Thus, the use of the system represents an important advantage with respect to conventional browsing around the web: the user receives not only the page designated by the link that he has clicked (this is conventional web browsing), but at the same time he benefits from the possibility of sending a query (containing several URIs) to obtain yet other resources relevant in relation to this page.
As a variant, said accessory page is returned after fast (or even restricted61) execution of the search query in the course of which the link clicked was added.
61 In the case of a query regarding pages already crawled, the system can directly return the relevant URIs (or pages) already known and return the rest of the results later on.
The second page then directly contains a part of the result62. The user then receives not only the page designated by the link that he has clicked, but in addition he benefits directly from other resources relevant in relation to this page.
62 (For example in the form of a list of URIs or a set of vignettes representing these pages in miniature)
More advantageously still, said accessory page may be displayed in a subwindow63 adjacent to the main subwindow of the browser. This adjacent subwindow opens in response to the action of the user who desires the displaying of the query under preparation (that is to say said accessory page).64
63 (Analogous to the favorites subwindow of the current browsers)
64 Note that, in parallel with the displaying of the query under preparation, the server can advantageously already begin to trek the web (crawling)—that is to say construct R−, R−+, R−+−, R+, R+− and R+−+ as already described—around the link selected.
The query under preparation can thus be displayed in parallel (asynchronously) with the displaying of the page designated by the link clicked; the latter page being displayed (independently) in the main subwindow.
The result of the search query can thereafter be presented in the same adjacent subwindow.
As mentioned previously, a (partial) result may possibly be returned after partial or restricted execution of the search query in progress, to which query the link clicked was added. The adjacent subwindow then directly presents a fast search result (which will possibly be supplemented subsequently).
b) Result of the Execution of a Search Query
For each search query, the server can return the results directly (for example returned from the HTTP query) or later on (for example by email).
The server returns the URIs (resulting from a query) in a page exhibiting the same structure as said accessory page (or said query under preparation), namely:
The page returned also presents the other queries (from the same user) in the form of drop-down graphical objects, as already described. Their presentation may be hierarchized according to their relevance with respect to the link clicked (according to the relevance calculation methods described later).
The page returned presents means of control allowing the user to create new queries and to delete existing queries. Of course, the user can cut and paste URIs from existing queries or from any other resource. Also, when the result of a query is returned by the server, the user can shift (hive off) the URIs received into other queries. Each query is individually accessible by means of its own URI.
Maintaining the Spots
Described hitherto are several methods that use the relative distillation procedure, starting from a query (e.g. the given associated links of a spot) composed of a set of URIs, to determine and store relevant URIs (e.g. the completed associated links of a spot) with respect to this query, together with their relevance scores. These stored results are obtained on the basis of counting links located in the resources of the sets R−+, R−+−, R−+−, R+−−, R+−+, R+−+−68 etc. which are themselves stored at least in part. Now, these sets vary over time (and the links located in the resources constituting these sets also vary). The stored data must therefore be kept up to date and the calculations must be redone when the data that they take as input vary significantly.
68 R−+−, R+−, and R+−+ are in particular used to calculate the closeness of linked resources, and to filter, as described above, by taking the complement of this closeness as weighting for the counting of the links in question.
Moreover, it is desirable to disclose new relevant resources even before links pointing to them appear on the web. A method making it possible to do so will now be described.
For each query (for example for each spot),
The similarity of a resource with respect to other resources is determined by comparing their contents. Described hereinbelow is the way to determine the similarity as a function of the distribution of the words in the resources in question.
Time-Dependent Authority Score:
Each new authority resource has a hypertext authority score (aht) and a similarity authority score (as). Let τ be the ratio between
Thus τ is used as a weighting to go gradually from a similarity score to a hypertext score and the formula for the global score is
(with τ′=1−τ).
As the distribution of the words of a new resource varies in principle less than the hypertext links which point to it, as is considered to be constant while aht must be updated over time. Thus the score as must be calculated at the moment at which the new resource is discovered, and for all the queries for which it is in a relevant region, until it becomes old (thus if a link to this resource appears in a relevant region after it has become old, then its similarity with the resources of said second set will not be determined).
Similarity:
An absolute distillation algorithm will be used to determine the score as of each new resource.
The known method of absolute distillation over a set of nodes connected by links (thus forming an oriented graph) comprises the following steps:
In addition, here the links are weighted by the similarities of the resources in question with respect to the distribution of their words. Steps 2 and 3 are replaced by the following:
The weight of the similarity link between two resources is equal to the scalar product of their distributions of words (that is to say to the sum, for each word located in the two resources, of the product of the frequencies of this word in these resources; the resulting sum is a number between zero—case where there is no word in common—and 1—case where the two resources have the same content) after having removed the nonpertinent words (“stop words”).
It should be noted that the similarity links thus obtained are bidirectional.
Thus, the absolute distillation can thus be performed over the set of resources comprising:
The methods described above also make it possible to select, from among a set of extra resources, a resource which is the most relevant with respect to a starting resource.
Accordingly, the following three steps are implemented:
Such a method makes it possible in particular to dynamically generate the content of web pages published as a function of context.
| Number | Date | Country | Kind |
|---|---|---|---|
| 02/00341 | Jan 2002 | FR | national |
| 02/05751 | May 2002 | FR | national |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/FR03/00089 | 1/13/2003 | WO |
