DE10160920B4 - Method and device for producing an extract of documents - Google Patents

Abstract

A method for automatically generating an extract for a document, wherein
- the document is part of a document collection electronically available for an information procurement system,
A user interacts with the information gathering system, wherein
- the information procurement system offers the user certain documents from the collection and
The user selects certain of the offered documents to display them,
The information acquisition system logs in chronological order which documents the user has selected in which order,
- the user specifies the document for which the extract is to be generated,
If, in each case, two documents selected consecutively according to the chronological order in chronological order have a high degree of similarity, the earlier document is rated as relevant,
- document parts of the selected documents are identified,
- evaluations of the identified document parts are calculated,
- whereby the evaluation of at least one part of the document as a function of a measure ...

Description

The The invention relates to a method and a device for automatic Generation of an extract for at least one document. This document is part of a collection electronic documents for an information procurement system is available.

"Document" is below a generic term for Texts, graphics with texts, tables, data collections and the like. The documents are given in textual form in electronic form. One For example, the document in the collection is from a word processing system or a database system or by capturing all pixels a paper-based document. A document of the collection can also be used as a set of html or xml pages on the internet or a Intranet available. The term "extract" is used below as a generic term for a list of keywords or for a summary that uses the document or documents describe.

One Extract describes the document and helps a user, for example to decide whether the document is relevant to him or not, if the user is on a corporate intranet with a document collection looking for information and not having the time, every document itself to read. An extract supported z. B. a full-text search on the Internet with known search engines, which often deliver a lot of documents. One reason for this is that the same Term in different meanings both in relevant as well occurs in non-relevant documents. On the other hand delivers a single search with a well-known search engine often therefore unsatisfactory Results, because for same situation in different documents different Terms are used, eg. B. conjugated word forms, synonyms and abbreviations. Therefore, the must Users perform multiple searches to let.

In DE 195 26 264 A1 discloses a method for generating keywords for the classification of documents. The keywords are in DE 195 26 264 A1 Descriptors called the documents texts. The word forms of the words of the document are decomposed into shorter word forms and possibly word parts, and the keywords are formed from the word forms and word parts remaining after the decomposition.

In K. Zechner: "A Literature Survey on Information Extraction and Text Summarization ", among others Procedure disclosed to a summary for a single electronic available Produce document. From J.G. Carbonell, J. Goldstein, The use of MMR, Diversity-Based Reranking for Reordering Documents and Producing summaries, Proceed. SIGIR '98, 1998, Melbourne, Australia, a method is known in which document parts, especially sentences, be evaluated and the abstract composed of highly rated sentences becomes. Therefor become terms, especially words or parts of words, with statistical Method evaluated. Other methods, e.g. In K. R. McKeown, J. Robin, K. Kukich: "Designing and Evaluating a new revision-based model for summary generation ", Information Processing and Management Vol. 5, 1995, presented methods Information extraction and methods for generating natural language Texts to create a summary.

All consider these procedures only one document at a time, namely only the document for which the Extract should be produced. Therefore, they can change the context of the document do not involve. Often, however, can only be decided out of context which information is really contained in the extract have to. For example, some words have different meanings in different contexts, eg. Eg "bank" or "program". If in the production of an extract this context is not taken into account, has to be User of the information acquisition system from the extract the filter out information of interest.

From M.-R. Amini: "Interactive Learning for Text Summarization," Proceed, PKDD 2000, Workshop on Machine Learning and Textual Information Access, pp. 10-18, teaches a learning method that generates summaries and improves results through user queries In the method, those sentences of a text are identified which receive the highest ratings on the user request.The rating of sentences is automatically performed using a tf / idf term weighting method.At each document, the user becomes the r highest rated sentences of the document In the second step, the user provides feedback to the system by checking the automatic selection and rating of sentences and modifying them as needed used to improve the automatic assessment.

satisfactory At most, this method can produce sufficient results and to make good reviews of the user. The feedback is for connected the user with considerable effort. The need the feedback can lead to erroneous results.

J. Goldstein, V. Mittal, J. Carbonell, J. Callan:
"Creating and Evaluating Multi-Document Sentence Extract Summaries," Proceed.ACM Conf., Information and Knowledge Management (CIKM 2000), McLean (VA), 6-11-11, 2000, pp. 165-172 discloses a method For example, to create a summary of sets of multiple documents from a given collection of documents, for example, the user performs a search on the Internet by using a search engine to submit a search query to find sites on the Internet In this example, the result of the user request to the Internet provided by the search engine is.

According to the in J. Goldstein et al. disclosed methods are the documents of the Collection disassembled into document parts. The document parts can be sentences or Be sentence parts. First those parts of the document are identified that are relevant to the Document collection and the user's search query. Out Some of these relevant document parts are selected as follows: Step by step is for every relevant, but not yet selected part of the document a rating according to the procedure "maximum marginal relevance-multi-document "(MMR-MD) the highest evaluated document part is additional selected. Such a selection step is repeated until a termination criterion Fulfills is, for example, up to a predetermined number of relevant document parts selected has been. The abstract is generated by the selected document parts sorted and put together.

A document part p is here according to the calculation rule Evaluation (p) = λ * sim_1 (p) - (1 - λ) · sim2 (p) rated, where
sim_2 (p) = max {sim_2 (p, p_sel) | p_sel already selected}.

Sim_1 (p) is the higher the more similar the document part p of the request is the higher the information content of p and the newer is the document containing p. Sim_2 is the higher the more similar p already selected Document part is. Because sim_2 is subtracted, it is avoided that several similar and thus redundant Document parts in the abstract occur.

The in J. Goldstein et al. In particular, the method disclosed is time-consuming and thus time-consuming, if the collection is extensive is, for example, if the search query several hundred documents provided. Not automatically considered which of these many documents the user considers relevant has classified and which not.

In JG Carbonell & J. Goldstein: "The Use of MMR, Diversity-Based reranking for reordering Documents and Producing Summaries" Proceed 21 ^St Annual boarding ACM SIGIR Conf Research and Development in Information Retrieval, Aug. 24 ^th -... 28 ^th , 1988, Melbourne, Australia, pp. 1-2, a method is disclosed for generating an extract for a document belonging to a document collection, one document at a time being determined from the document collection, on the one hand as similar as possible to the request of a searching user, on the other hand as dissimilar as possible to the previously determined documents of the collection This is repeated until a termination criterion is met, and the documents are presented to the user as a result of his query In order to generate a summary of a document, the documents matching the user request are included in the document part e decomposed and evaluated the document parts according to the similarity measure.

As a user's evaluations are included in the production of the extract not described.

The invention is based on the object, starting from J. Goldstein, V. Mittal et al. to provide a method of the type mentioned above, which evaluates document parts in the generation of the extract and automatically takes into account in the document part evaluation, which documents of the collection of the user as relevant without requiring additional feedback or other information from it.

The The object is achieved by a method according to claim 1, a device according to claim 11 and a computer program according to claim 13 or claim 14 solved.

The Invention discloses a method for automatically generating a Extract for a document. The document is part of an information procurement system electronically available Collection of documents. A user interacts with the information procurement system. In this interaction, the information procurement system provides to the user certain documents from the collection. The user choose the documents offered to them to view them to let.

The Information procurement system logs a chronological order. This logged chronological order is the order in which the user has selected documents of the collection.

Of the User specifies the document for which the extract is generated should. Automatically determines which of the documents for the user are relevant. This is when, if any, two according to the logged sequential order selected documents a high degree of similarity have, earlier in time Document is considered relevant.

• – eines Maßes für die Ähnlichkeiten des Dokumentteils zu den relevanten unter den ausgewählten Dokumenten und
• – eines Maßes für die Ähnlichkeiten des Dokumentteils zu den nicht relevanten unter den ausgewählten Dokumenten.

Document parts of the selected documents are identified. Ratings of the identified document parts are calculated. In this case, the evaluation of at least one document part is calculated as a function of two similarity measures, namely

• A measure of the similarities of the document part to the relevant ones among the selected documents and
• A measure of the similarities of the document part to the non-relevant ones among the selected documents.

In the function goes the similarities to the relevant and the similarities to the non-relevant documents in various ways.

document parts become dependent selected from their ratings. The extract is selected from Composed of document parts.

The Method provides inputs of the user in the information procurement system to log him in the search for information anyway makes, namely the selection of specific documents.

method According to the prior art ask the user feedback z. In the form of an evaluation of the search results. This survey is time consuming, is often perceived by users as annoying and unnecessary and is error prone because Sometimes users can not name what they really care about especially if they are interviewed in retrospect. The inventive method Saves this survey by adding user input to the information gathering system, which he does anyway, logged and for the determination of further documents be evaluated. additional User input does not require the inventive method. The user In particular, he does not need to be asked which documents he wants assessed as relevant and irrelevant to him. These Rather, information is automatically extracted from user input won.

By the inventive method An extract is generated automatically, which is sent to the respective user and its interests is. This is achieved because logged entries of the respective User considered become. The method can therefore be with the same document and different To give users different results. In particular, be thereby automatically distinguishing different meanings of words. For example, if a user is looking for financial services documents interested, the procedure ensures that documents are classified as relevant in which the word "bank" is meaningful as a financial services provider, but no documents with one Bank in meaning as a seat.

Is possible even that in the produced according to the invention Extract for the document a keyword or a phrase or phrase occurs in the document itself does not occur at all.

The Use of the user inputs has the further advantage of being preferred exclusively Information to be used on a network subscriber computer (Client) available are. This can be the procedure independent from a specific internet service program (web server) or realize a specific network central computer (server). The User input can be about it Be assigned to a specific user, so that this method of the invention from different network subscriber computers run out of can let. The Internet service program or the network central computer can even be exchanged, without renewed Inputs of the user are needed.

The inventive method considered multiple documents, but is still much less computationally as a method of the prior art. This advantage is particular achieved by first other documents of the collection using the logged User inputs are determined and only the selected others Documents in the place of all documents of collection in detail and with elevated Computational effort to be investigated. Often the collection includes several hundred Documents, as further documents but z. B. only ten or twenty documents selected.

Of the Extract can refer to a single document. This document is for example specified by a user. A user can also specify several documents, and the extract produced relates on all these given documents.

Is possible but also that one only extract for the document and the selected ones additional documents is generated (claim 3). A training sees before that Document, for that the extract is produced is not predetermined, but one the selected one other documents.

Of the Extract includes For example, a list of keywords (claim 7), and several or all of the identified and evaluated parts of the document are words, eg. B. words of the German or English language or distinctive abbreviations. As keywords be z. B. selected those words whose rating a predetermined Barrier crossed, or the s highest evaluated words, where s is a predetermined number. A list of keywords is shorter and thus to grasp faster than other forms of an extract. To the document z. B. to sort into an electronic catalog, keywords are needed.

According to claim 8 includes the extract a summary, and to the identified and evaluated parts of the document full Sentences or Sections. The summary is for example from the s am highest rated sentences where s is a predetermined number and the order the sentences in the summary equal to the order in the documents the collection is. Such a summary is often better understandable as a mere Collection of keywords. A Training provides that all identified and evaluated document parts are sections and the abstract consists of a single section.

Around to record which documents the user has selected so far, Preferably, the user click path is determined. The user click path consists of references to the documents previously selected by the user the collection, preferably sorted in chronological order, or from suitable representations of these documents.

Preferably is for the generation of the extract a query from a network subscriber computer transmitted to a network central computer. The search query comprises suitable representations of the chosen other documents or references to the selected further documents (claim 4). For example, in data stores of the network host the documents of the collection are stored. By this configuration it is possible to reuse known information acquisition systems. In contrast to known information procurement systems exists the search query does not come from search terms entered by the user, but from an automatically generated search query. The search query let yourself generate information stored on the network subscriber computer, so that one previous information request from the network subscriber computer to the network central computer is not required. This embodiment saves the user Working time and waiting time in front of his network participant computer.

The inventive method Provides that identified part of the document depending on a measure of similarity between the document part and the selected further documents automatically Is evaluated.

The Embodiment according to claim 5 provides an additional influence on the Evaluation of the identified document part. For the calculation the document part rating be additional Weightings of those selected further documents, in which the document part occurs, calculates and uses. In order to let yourself automatically a differentiation among the selected others Make documents, what the meaningfulness of the extract produced further increased.

The Continuing education of this embodiment according to claim 6 provides, not only between similar and not similar Documents, but also a measure of similarity between the document and the selected additional document to calculate. This measure will for the Calculation of the weighting of a selected additional document used. For example, that's the measure of similarity a number between 0 and 1 that is used directly as a weight becomes.

According to claim 9, an extract is produced which comprises a summary. For this will be Document parts identified by adding documents of the collection in sentences or Sections are disassembled. A list of keywords is generated, for example by a method according to claim 7. The generated summary comprises those identified document parts that have at least one keyword contain. The summary thus produced comprises less Sentences or paragraphs as a summary produced in a different way, and present the sentences Keywords isolated in their context instead.

at the embodiment according to claim 10, a summary is generated, the sentences and / or Sections exists. The at least one evaluated document part is a sentence or section. For the Calculation of its rating will be a rating of at least one Word contained in the sentence or section. The Rating of this word depends from a measure of similarity between the word and the selected one from further documents.

• – Mittel zum Auswählen von weiteren Dokumenten der Kollektion,
• – Mittel zum Protokollieren von Benutzereingaben in das Informationsbeschaffungssystem,
• – Mittel zur Bewertung von Dokumentteilen in Abhängigkeit von den protokollierten Benutzereingaben,
• – Mittel zum Auswählen von Dokumentteilen in Abhängigkeit von ihren Bewertungen
• – und Mittel zum Zusammensetzen des Extrakts aus ausgewählten Dokumentteilen.

Claim 11 provides an apparatus for performing the method according to one of claims 1 to 10. This device includes

• - means for selecting further documents of the collection,
• Means for logging user input into the information procurement system,
• Means for evaluating document parts in dependence on the logged user input,
• - Means for selecting document parts depending on their ratings
• And means for assembling the extract from selected parts of the document.

Preferably if the device is part of a data processing system, the at least temporarily as a network subscriber computer with at least a network central computer is connected (claim 12). This network central computer includes data storage in which Documents of the collection are stored. The network subscriber computer is preferred over the Internet or an Inteanet with the network central computer or the network central computers.

claim 13 provides for a computer program. This computer program will processed in a computer and thereby performs a method one of the claims 1 to 10 off.

The The computer program product of claim 14 is on one of Computer readable medium stored. It includes readable by a computer Program means that cause the computer to follow a procedure one of the claims 1 to 10.

In the preferred embodiment is that going? inventive method on a data processing system, eg. As a PC from, as Network subscriber computer with a network central computer or a set of Network central computers is connected. In data stores this Network central computer are documents of the collection included, so that the network participant computer can access all documents. To the network participant computer with the network central computer or the network central computers Connected to, internet technology is preferred with at least an internet service program running on the network hosts expires used. The network central computer belong for example, to the Internet or to a corporate intranet, you can Communicate with multiple network subscriber computers. Preferably be for the extraction of the extract exclusively uses information the on the network participant computer available are. Thus, the method works independently of a particular network host.

The Information procurement system in which the inventive method expires runs on the network subscriber computers or the network central computer (s). A network subscriber computer sends jobs, eg. B. inquiries, to a Network central computer, and the commissioned network central computer sends answers, eg. Eg a specific document of the collection, to the network subscriber computer.

One User interacts with the information procurement system. there The information procurement system provides certain documents of the collection to the user, and the user chooses among the certain documents to be displayed. Essential entries of the user are logged, eg. B. which documents he has selected in which order. At the request of the user will be an extract of documents of the collection generated. Preferably, he specifies a document for which Extract under consideration of the previous user behavior is generated. This predetermined Document is referred to below as the first document D_0. Is possible but also that the User specifies multiple documents or that no document specified is, but the extract refers to additional documents.

The documents of the collection are preferably represented using a vector space model. Vector space models are from CD Manning, H. Schütze:
"Foundations of Statistical Natural Language Processing", MIT Press, Boston, 1990. An indexing language, that is a set of M strings, is given, and the strings of the indexing language are called indexing terms in the following.

• – Worte, vorzugsweise Worte in der Dokumenten-Kollektion. Da verschiedene Flexionsformen als unterschiedliche Zeichenfolgen behandelt werden, wird die Indexierungssprache groß, und die Übereinstimmung zwischen verschiedenen Flexionsformen eines Worts wird nicht oder nur mit zusätzlichem Rechenaufwand erkannt.
• – Wortstämme, die mittels linguistischer Regeln aus den Worten abgeleitet werden. Für jede Sprache, u. U. sogar für jeden Dialekt einer Sprache, sind erneut linguistische Regeln aufzustellen.
• – Zeichenfolgen der Länge n, sogenannte n-Gramme. In der Praxis werden Quadgramme, das sind Zeichenfolgen der Länge n = 4, und Trigramme, das sind Zeichenfolgen der Länge n = 3, verwendet. Diese Ausgestaltung hängt nicht von der Sprache ab und ist robust gegenüber Flexionsformen und Tippfehlern. Jedoch haben viele n-Gramme keine anschauliche Bedeutung für den Benutzer.
• – Phrasen, also feste Abfolgen von Worten. Die Menge von Phrasen muß bei dieser Ausgestaltung für jede Anwendung, z. B. jedes Themengebiet oder gar jede Kollektion, neu definiert werden.

The indexing language can be built from the following types of strings:

• - Words, preferably words in the document collection. Since different inflectional forms are treated as different strings, the indexing language becomes large, and the match between different inflectional forms of a word is not recognized or only with additional computational effort.
• - Word stems derived from the words using linguistic rules. For every language, u. Even for each dialect of a language, again linguistic rules are set up.
• - strings of length n, so-called n-grams. In practice, quadgrammes, which are strings of length n = 4, and trigrams, that is, strings of length n = 3, are used. This embodiment does not depend on the language and is robust against inflectional forms and typos. However, many n-grams have no illustrative meaning to the user.
• - Phrases, ie fixed sequences of words. The set of phrases must in this embodiment for each application, for. For example, every topic or even every collection can be redefined.

Prefers Words, quadgrams or trigrams are used as strings of the indexing language.

each Document becomes according to the vector space model by a vector of length M represents, where M is the number of indexing terms. The jth element of the vector V_i for the document D_i gives a weighting for the meaning of the indexing term j in document D_i. For example, the weighting is either 0 or 1. To save space, are used by a vector for a document only the indexes and entries those vector elements which are not equal to 0 are stored.

As weights, not only the values 0 and 1 are preferably used, but the weights determined by the tf / idf method. The weighting G (t, i) of a character string t with respect to the document D_i becomes the tf / idf method according to the formula G (t, i) = tf (t, i) · idf (t) certainly. Here, the string frequency tf (t, i) indicates how many times the string t occurs in the document D_i. The inverse document frequency idf (t) is the greater the rarer the term t occurs in the document collection. It is equal to N / df (t), where N is the number of documents in the collection and the document frequency df (t) is the number of documents in which the string t occurs.

If a string t does not appear in any document in the collection, G (t, i) = 0 is set.

Words that occur frequently in a document and do not say anything about the content of the document are called function words. Function words include pronouns, articles, and auxiliary verbs. For Function words are preferably not weighted and are disregarded in identifying and evaluating parts of documents, especially keyword searches. The automatically determined and not taken into account function words are summarized in a so-called stop word list.

Out C. J. van Rijsbergen: "Information Retrieval, "Butterworth Press, London 1979, is a method of automatic detection the stop word list known. This method is preferred on the Transfer collection of documents. Function words are approximately Poisson-distributed in the document. Of the Parameter λ = λ (w, i) = tf (w, i) / N of the Poisson distribution is equal to the relative abundance of the function word w in the collection of documents where N is the Number of documents in the collection is. To decide if a word w is a function word becomes the existence of a Poisson distribution checked.

According to the invention determined next to the first document D_0 more documents of the collection and evaluated to produce an extract. An embodiment The invention provides, in this case for documents of the collection to look the same to the first document D_0. From U. Bohnacker, L. Dehning, J. Franke, I. Renz, R. Schneider: "Weaving Intranet Relations - Managing Web Content, "Proceed. RIAO 2000: Content-Based Multimedia Information Access, Paris, 12.-14. 4th 2000, pp. 1744-1751, is a procedure known to provide in a given collection documents, which are similar to a given document to determine. Once in advance will be for every pair of documents (D_1, D_2) of the collection the similarity between D_1 and D_2 rated. If a user selects a document D_l and according to D_1 similar documents to search the collection as similar evaluated documents of the collection and displayed to him. The user request will be answered very quickly. Prefers For example, U. Bohnacker et al. applied known method to automatically look similar to D_0 To search documents.

For each Indexing term, that is for each string of the indexing language becomes as described above a weighting of the indexing term with respect to each similar one Document of the collection determined and evaluated, preferably according to the tf / idf method. The individual weights of the indexing term will be used to evaluate the indexing term in terms of Collection summarized. Let G (t, i) be the weighting of the indexing term t with respect to the Document D_i. Let B (t, Koll) be the evaluation of the indexing term determined according to the invention t regarding the collection. Be D_1, ...

D_k the documents of the collection similar to the first document D_0. The rating B (t, Koll) is determined either by the rule B (t, Koll) = G (t, 0) + G (t, 1) + ... + G (t, k) or by regulation
B (t, coll) G (t, 0) · sim (0, 0) + G (t, 1) · sim (0, 1) + ... + G (t, k) · sim (0, k) certainly. Here sim (0, i) is a measure of the similarity between the documents D_0 and D_i (i = 0, 1, ..., k). In both cases, only D_0 and similar documents to D_0 are used for the production of the extract.

The Rating B (t, Koll) is determined as described above for each indexing term. So M scores are determined, where M is the number of indexing terms is. This forms a cluster indexing term vector of length M. Indexing terms contained in multiple documents are preserved usually a higher one Valuation as such, which are contained only in one document.

Preferably, an extract in the form of a list of keywords is generated using the cluster indexing term vector. This is preferably done in one of the following two ways:
The first type: The cluster indexing term vector is sorted in descending order according to the evaluations B (t, Koll) of the indexing terms relative to the collection. Keywords are the highest ranked indexing terms used. This method can be used in particular when words are used as indexing terms. Or keywords used are the words of documents containing the highest-ranked indexing terms.

The second way: For each word w of the first document D_0 and the scores B (t, Koll) are added to the collection of all indexing terms contained in the word w. The words of the stop The word list is disregarded here. Let w be a word in D_0 that does not belong to the stop-word list, and let t_1, ..., t_l be the indexing terms that appear in the word w. Then, the score B (w, coll) of the word w with respect to the collection is calculated according to the rule B (w, coll) = B (t_1, coll) + ... + B (t_l, coll) certainly. Those r words of the document D_0 which receive the highest scores B (w, Koll) are designated as keywords. This embodiment is particularly useful for trigrams and quadgrams as indexing terms. In a training, the evaluations B (w, Koll) are generated not only for the words of the first document but also for the words of the similar documents. The list of keywords, which in turn is generated by selecting the highest ranked words, may then also contain words that do not appear in the first document D_0 but in similar documents to D_0.

A Further embodiment of the invention provides, the so-called user click path record and for to use the production of the extract. In the user click path are those documents of the collection listed so far the User selected for display has, for. For example, after being sent to him by the information procurement system Selection were offered. This listing includes z. B. References to the documents or representations of documents according to the vector space model. The user click path represents thus the profile of the user in his search for documents. The user click path is preferred in chronological order sorted in which the user has selected documents as the last selected document is the last of the user's click path. For example, the Internet or intranet address of a document in the form of a unique Identifier, e.g. As a so-called "Uniform Resource Locator" (URL) determined and locally stored in the network subscriber computer, whereby the User click path on a network subscriber machine is available. The user click path is z. B. in so-called "cookies" an Internet access program stored. The user click path includes a list of identifiers all of the user selected Documents. Preferably, the user's clickthrough path is called a query from a Network subscriber computer sent to a network central computer. By this configuration no information is needed which are stored in the network central computer. Rather enough it to log the user click path on the network participant machine and save.

bestimmt.By way of example, the generation of a list of keywords as the extract will be described below. Let D_1, ..., D_k be the documents of the user-click path. One of these documents may in this example be the same as the first document D_0, e.g. For example, the last document of the user click path. For the indexing terms, as described above, a weighting G (t, i) of the indexing term t with respect to the document D_i is determined. From these weights keywords are preferably determined in one of two ways:
Embodiment 1: A score B (t, BKP) of the indexing term t with respect to the user click path BKP and the first document D 0 is determined as described above. Included in this evaluation are the k weights G (t, D_1), ..., G (t, D_k) of the indexing term t with respect to the k documents of the user click path generated according to the vector space model. For example, G (t, D_i) = tf (t, D_i) * idf (t), and G (t, BKP) becomes according to the rule

certainly.

bestimmt. Eine weitere Ausführungsform ist

As described above, then, from the scores B (t_1, BKP), ..., B (t_l, BKP) of the indexing terms t_1, ..., t_l included in a word w, a score B (w, BKP) of the word w generated. The l ratings of the indexing terms t_1, ..., t_l contained in a word w are combined into a score B (w, BKP). The rating B (w, BKP) is preferably according to the regulation

certainly. Another embodiment is

In In both calculation rules, n is the number of characters in the word w. Word ratings are for Words of the first document D_0 and the documents of the user click path. The the highest rated words are considered keywords used.

berechnet, wobei n die Anzahl von Zeichen im Wort w ist.Embodiment 2: In a first step, the keywords G (t, D_i) are used to calculate separate keywords for the documents D_1,..., D_k of the user click path, ie without the complete user click path for this keyword determination a document is used. For the determination of the keywords above-described methods are used analogously, z. By determining, for a word w in the document D_i, the indexing terms t_1,..., T_l contained in the word w and assigning their weightings G (t_1, D_i),..., G (t_l, D_i) to a rating B (FIG. w, D_i). For example, B (w, D_i) becomes according to the rule

where n is the number of characters in the word w.

For each Document D_i (i = 1, ..., k) become r_i keywords certainly. Let B (w, i) be the single evaluation of a keyword w respect of the document D_i (i = 1, ..., k). From the individual reviews of Keywords in terms of The documents will be reviews of keywords regarding the Collection determined by adding the individual ratings.

An example of Embodiment 2: If the word w is a keyword for the documents D_1, D_2 and D_k (with k> = 3), then the overall score is B (w, BKP) = B (w, 1) + B (w, 2) + B (w, k). Another example:
If w is keyword only for document D_3 and no other user path document, the overall score is B (w, BKP) = B (w, 3).

These Configurations lead usually with no further provision for words to be highly valued, Describe the topics for the user is relevant or interests him. Because many documents of the user's click path relate to topics that the user interested and / or for him are relevant. If the user is interested in a Document encounters will he often after similar Search or search documents. Words received in this document therefore a high rating, some become key words. On the other hand he hits on one for him uninteresting document, he will not request a similar document, and receive the words appearing only in the uninteresting document a low rating.

The just described embodiments can be used to key words for a single first document D_0 considering the context to determine. It is also possible, a set of keywords for many Create documents, such as the user click path documents. The generated keywords describe the essential content of the documents selected by the user.

A Training the invention provides the evaluation of a word as described in the previous two sections, but in addition To determine reviews of the k documents of the user click path and to use in the calculation of word scores.

According to one Design of this training, the higher the value of a document, the later it occurs in the chronologically sorted user click path. The document that the user first selected will receive the lowest rating and the last one selected the highest.

• • Notiert wird, für welche Dokumente der Benutzer eine Suche nach ähnlichen Dokumenten forderte. Aus U: Bohnacker et al. ist bekannt, wie ein Informationsbeschaffungssystem automatisch eine Suche nach ähnlichen Dokumenten durchführt. Diese Dokumente werden als relevant eingestuft.
• • Die Ähnlichkeit zwischen jeweils zwei Dokumenten der Kollektion wird wie oben beschrieben vorab ermittelt, z. B. durch das aus U. Bohnacker et al. bekannte Verfahren. Wenn im Benutzer-Klickpfad zwei nacheinander folgende Dokumente D_i und D_j ähnlich sind, also ein hohes Maß für die Ähnlichkeit haben, so wird das zeitlich frühere Dokument D_i als relevant bewertet.

In a further embodiment, documents are evaluated differently in terms of their relevance to the user when calculating the document evaluations. The relevance is automatically determined, whereas so far only methods are known to determine the relevance by questioning the user and evaluating his feedback. Preferably, one of the following two methods is used to automatically determine the relevance of a user click path document:

• • It is noted for which documents the user requested a search for similar documents. From U: Bohnacker et al. It is well known how an information retrieval system automatically performs a search for similar documents. These documents are considered relevant.
• • The similarity between every two documents in the collection is determined in advance as described above telt, z. B. by the U. Bohnacker et al. known method. If two consecutive documents D_i and D_j are similar in the user click path, ie have a high degree of similarity, the earlier document D_i is evaluated as relevant.

bestimmt. Hierbei bezeichnen D[R] und D[N] die Menge der relevanten bzw. nicht relevanten Dokumente im Benutzer- Klickpfad und R bzw. N die Anzahl der Dokumente in D[R] bzw. D[N]. Aus den l Bewertungen B(t_j, D_l), ..., B(t_j, D_k) wird eine Bewertung B(w, BKP) des Worts w bestimmt. Hoch bewertete Worte werden wiederum als Schlüsselworte verwendet.The way in which classifications of the additional documents determined are automatically used in relevant and irrelevant documents is again described using the example of an extract in the form of a list of keywords. Because of their relevance, ratings for the user click path documents are determined and used to calculate word scores. An embodiment provides for a word w of the first document D_0 to determine the indexing terms t_1,..., T_l of the indexing language that are contained in the word w. For such an indexing term t_j (j = 1, ..., l), as described above, k individual scores B (t_j, D_1), ..., B (t_j, D_k) of the indexing term t_j with respect to the k documents D_1, ..., D_k of the user click path and a single score B (t_j, D_0) with respect to the first document D_0. The overall score B (t_j, BKP) of the indexing term t_j with respect to the user click path BKP and the first document is determined according to the formula

certainly. Here, D [R] and D [N] denote the amount of relevant or irrelevant documents in the user click path and R and N respectively the number of documents in D [R] and D [N]. From the l judgments B (t_j, D_l), ..., B (t_j, D_k), a score B (w, BKP) of the word w is determined. Highly valued words are again used as keywords.

• 1. Das erste Dokument D_0 wird in Sätze zerlegt.
• 2. Die Sätze des ersten Dokuments werden bewertet. Um die Bewertung eines Satzes s zu berechnen, werden die Bewertungen von im Satz s enthaltenen Worten verwendet.
• 3. Aus den hoch bewerteten Sätzen wird eine Zusammenfassung des ersten Dokuments erzeugt.

After key words for the first document D_0 have been determined according to the invention, these keywords can be used to generate a summary for the first document. The following steps are preferably carried out for this purpose:

• 1. The first document D_0 is decomposed into sentences.
• 2. The sentences of the first document are evaluated. To compute the score of a sentence s, the scores of words contained in the sentence s are used.
• 3. A summary of the first document is generated from the high-valued sentences.

• – Ein Satz endet mit einem Satztrennzeichen (!, ? und.).
• – Ein Satz beginnt am Textanfang oder nach einem Satzende-Zeichen.
• – Befindet sich vor einem Punkt eine Abkürzung, so markiert dieser Punkt nicht das Satzende.
• – Befindet sich ein Punkt unmittelbar nach einer Zahl zwischen 1 und 31, so markiert dieser Punkt ebenfalls kein Satzende.
• – Befindet sich ein Punkt direkt nach einer Ziffer und ist das nächste Zeichen, das kein Leerzeichen ist, ebenfalls eine Ziffer, dann markiert der Punkt ebenfalls kein Satzende.
• – Befinden sich mehrere Satztrennzeichen unmittelbar hintereinander, so wird nur das letzte Satztrennzeichen berücksichtigt.
• – Ein Satzende-Zeichen kennzeichnet dann das Ende eines Satzes, wenn darauf mindestens ein Leerzeichen gefolgt von einer Ziffer oder einem Großbuchstaben folgt.
• – Befindet sich ein Satzende-Zeichen am Ende des Textes, so kennzeichnet dieses auf jedem Fall das Ende eines Satzes.
• – Ein Satz muß mindestens drei Worte haben.
• – Ein Satz darf eine vorgegebene Anzahl von Worten oder Zeichen nicht überschreiten.

To perform step 1, that is, breaking the first document into sentences, it is preferable to use the following syntactic rules that apply to many languages:

• - A sentence ends with a sentence delimiter (!,? And.).
• - A sentence starts at the beginning of the text or after the end of a sentence.
• - If there is an abbreviation before a point, this point does not mark the end of the sentence.
• - If a point is immediately after a number between 1 and 31, this point also does not mark a sentence end.
• - If a point is directly after a number and the next character, which is not a space, is also a number, the point also does not mark a sentence end.
• - If there are several record delimiters immediately after one another, only the last record delimiter is taken into account.
• - An end-of-sentence character marks the end of a sentence if it is followed by at least one space followed by a digit or a capital letter.
• - If an end-of-sentence character is at the end of the text, this always marks the end of a sentence.
• - A sentence must have at least three words.
• - A sentence may not exceed a given number of words or characters.

These Rules are applied to the first document D_0 and deliver a list of sentences.

For step 2, namely the calculation of sentence evaluations, preferably only keywords which are contained in the first document are used. Their ratings are combined into a sentence rating. Let's be a sentence of the first document. If this sentence does not contain a keyword, it receives the lowest score, e.g. For example, the number 0. Otherwise, w_1, ..., w_k are the k keywords of the sentence s (k> = 1). Let B (w_1), ..., B (w_k) be the k ratings of these keywords, which were determined according to one of the above-described embodiments of the invention. Depending on the configuration, let S be the number of words (keywords and other words) or the number of characters of the sentence s. Preferably, the score B (s) of the sentence s is determined according to one of the following rules:

Thereby, that in the rating factor k, ie the number of keywords, is included In the sentence, a phrase with multiple keywords is usually rated higher as one with only one keyword, too if this gets a high word rating.

At the Step 3, generating the summary of sentences the sentences of the first document prefers descending to those in step 2 calculated sentence evaluations sorted. The summary will from the m highest rated sentences of the first document D_0. The order of the sentences is in the generated part summary the same as in the first one Document D_0. Preferably, an upper bound for the number given the character of the summary, z. For example, a maximum of 500 characters. The number m of the selected Sentences of Summary is determined so that the m sentences together no longer sign than the given limit. Another embodiment sees to directly specify the number m of sets to be selected.

A Training of the invention provides, instead of sentences rather Evaluate sections of the first document and the abstract composed of highly rated sections. The abstract becomes analogous to the three steps described above by three steps generated.

berechnet. Der Abschnitt a enthält hierbei die k Schlüsselworte w_1, ..., w_k und insgesamt A Worte.In order to carry out step 1, namely the decomposition of the document into sections, the formatting of the first document is preferably taken into account, eg. By searching for empty lines or by applying syntactic rules. In document description languages, such as HTML or XML, sections are identified by specific strings, such as text strings. For example, <P> and </ P>. In step 2, the evaluation of a section a is preferred according to the formula

calculated. The section a contains the k keywords w_1, ..., w_k and a total of A words.

A another embodiment provides that a summary exclusively from the highest evaluated section of the text is generated. For this is a upper barrier for the number of characters of the summary given. The first The document is divided into sections so that each section consists of complete, in the first document consists of successive sentences and each section highest has as many characters as the upper bound. Every section is evaluated according to one of the embodiments described above, for what Reviews of keywords included in the section can be used. The summary is equal to the highest rated section.

A Another embodiment of the invention discloses a method to with Help the user click path directly extract in the form of a Generate summary without first identifying keywords. Preferably As described above, the user's click path becomes a query as a query from a network subscriber to a network host sent to determine the other documents. A training provides only those documents of the user click path to consider the generation of a summary automatically have been classified as relevant, preferably after one of the embodiments described above. The abstract for the first document D_0 is generated by evaluating the sentences of the first document. The summary is through Composition of highly valued sentences generated.

The evaluation of a sentence s is calculated using the user click path - or only the relevant documents of the user click path. Let D_1, ..., D_k be the documents of the user click path used for the calculation of the evaluation - either all or the documents determined to be relevant - which will be used as further determined documents in the following. In the evaluation of the sentence s, evaluations of indexing terms, which are contained in the sentence s as well as the k documents D_1,. D_k the other documents detected occur. Let t_1, ..., t_l be these indexing terms. These indexing terms belong to a given indexing language, preferably quadgrammes, trigrams, words or root words.

An embodiment provides that words of the sentence, in particular keywords, are evaluated and function words of a stop word list are not taken into account. A score B (t_j, s, BKP) of the indexing term D_j with respect to the set s and the other identified documents of the user click path BKP is determined (j = 1, ..., l). The sentence score B (s, BKP) is determined according to an embodiment from these l Indexierungsterm evaluations by adding, ie according to the rule

These Prescription is considered Euclidean measure of the similarity between the sentence and that treated as a search query of the information procurement system User Click path BKP designates.

berechnet.The rating B (t_j, s, BKP) of the indexing term t_j used in the above calculation rule is preferred according to the rule B (t_j, s, BKP) = G (t_j, s) · G (t_j, BKP) calculated. The valuation of the sentence s will therefore be in accordance with the requirement

calculated.

The weights G (t_j, s) and G (t_j, BKP) are both calculated using one embodiment of the tf / idf method, namely according to the rules

wird in dieser Ausgestaltung als inverse Dokumenthäufigkeit des tf/idf-Verfahrens verwendet.Here, the term frequencies tf (t_j) and tf (t_j, BKP) indicate how often the indexing term t_j occurs in the set s or in the k further determined documents D_1, ..., D_k, and df (t_j) determined the number of others in which the indexing term t_j occurs at least once. The factor

is used in this embodiment as the inverse document frequency of the tf / idf method.

berechnet. Die Gewichtungen des Indexierungsterms t_j werden wie oben beschrieben berechnet.On the other hand, according to a preferred embodiment, a normalization is performed and a cosinus measure is used as a measure of similarity between the set s and the additional documents determined. The sentence rating will then be according to the regulation

calculated. The weights of the indexing term t_j are calculated as described above.

A development of this embodiment provides to first generate an extract in the form of a summary with sentences or sections and then to identify keywords in this summary. As a result, fewer keywords than other methods are determined, because as Keywords only words of the summary come into consideration. The key words are preferably those words of the summary which have contributed most to the evaluation of the sentences selected for the abstract.

berechnet. Hierbei geben tf(w, s) und tf(w, BKP) an, wie oft das Wort w in dem Satz s bzw. in den k weiteren ermittelten Dokumenten auftritt. Die Anzahl von weiteren ermittelten Dokumenten, in denen das Wort w mindestens einmal vorkommt, wird mit df(w) bezeichnet.Preferably, evaluations of words are generated with respect to the individual sentences of the summary, and then aggregated evaluations of the words are generated. For example, the overall score B (w, s, BKP) of a word w with respect to the sentence s of the summary and with respect to the further determined documents D_1, ..., D_k of the user click path including the first document is determined according to FIG

calculated. Here tf (w, s) and tf (w, BKP) indicate how often the word w occurs in the set s or in the k other documents found. The number of further detected documents in which the word w occurs at least once is denoted df (w).

Claims (14)

Method for the automatic production of an extract for a Document, where - the Document part of a for an information procurement system electronically available document collection is - one User interacts with the information procurement system, wherein - the information procurement system offers the user certain documents from the collection and - the user select from among the offered documents to display them to let, - the Information procurement system logged as chronological order, which documents the user has selected in which order, - the user the document pretends for that the extract should be produced, - then, if any two according to the logged sequential order selected documents a high degree of similarity have, earlier in time Document is rated as relevant, - Document parts of the selected documents be identified, - Reviews the identified document parts are calculated, - in which the evaluation of at least one part of the document as a function of a measure for the similarities of the document part to the relevant ones among the selected documents and a measure for the similarities of the Document section on the non-relevant of the selected documents is calculated and the similarities to the relevant and the similarities to the non-relevant documents in various ways in the To take action, - Document parts dependent on selected from their ratings be and - of the Extract selected from Document parts is composed. Method according to Claim 1, characterized, that one Document of the collection is then rated as relevant, if the user for this document a search for similar Caused documents. A method according to claim 1 or claim 2, characterized characterized in that Extract as a single extract for the document and the selected others Documents is generated. Method according to one of claims 1 to 3, characterized, that for the generation the extract a query from a network subscriber computer transmitted to a network central computer becomes, where the search query the selected other documents or References to the selected further documents. Method according to one of claims 1 to, characterized, that weights selected by those other documents in which the identified document part occurs, calculated and additionally for the Evaluation of the identified document part. Method according to claim 5, characterized, that - a measure of similarity between the document and the selected additional document - calculated and for the calculation of the weights of the selected further documents is used. Method according to one of claims 1 to 6, characterized, that - several or all of the identified and evaluated parts of the document words are - and the extract produced comprises a list of keywords. Method according to one of claims 1 to 6, characterized, that - several or all of the identified and evaluated document parts phrases or Sections are - and the extract produced comprises a summary. Method according to claim 8, characterized, that a List of keywords of the document is generated, and the generated summary includes identified and evaluated document parts, the at least one keyword contain. Method according to claim 8, characterized, that for the rating of the identified document part, a rating of at least one Word contained in the document part is used, in which the word rating from a measure of similarity between the word and the selected one depends on other documents. Apparatus for carrying out the method according to one the claims 1 to 10, characterized in that - means for selecting other documents of the collection, - Means for logging user input to the information procurement system, - Medium to evaluate document parts depending on the logged ones User input, - Medium to choose dependent on document parts from their reviews - and Means for assembling the extract from selected parts of the document includes. Device according to claim 11, characterized, that the Device is part of a data processing system, the at least temporarily performs the function of a network subscriber computer, in which the network subscriber computer with at least one network central computer, which includes a data store with documents of the collection is. Computer program that works in a computer is thereby a method according to one of claims 1 to 10 executes. Computer program product on one of one Computer readable medium is stored and stored by a computer has readable program means that cause the computer to to carry out a method according to one of claims 1 to 10.