DE102008030125A1 - Electronically stored text evaluating and searching method, involves assigning search words to neuron such that combining weights depends in frequency of search word in text with neuronal network, and calculating end activation of networks - Google Patents

Abstract

The method involves specifying multiple search items and generating two neuronal networks. The search words are assigned to a neuron of the network such that combining weights are kept constant with the neuronal network and the neurons are activated in dependence of the frequency of the search word in the text to be evaluated. The search words are assigned to the neuron such that combining weights depends in frequency of the search word in the text with another neuronal network and the activation of the neurons is kept constant, and end activation of both the networks is calculated. Independent claims are also included for the following: (1) a search and evaluation system for evaluation of electronically stored text (2) a computer program product for evaluation of electronically stored text.

Description

The The invention relates to a search and evaluation method for evaluation electronically stored texts by means of a computer system and neural networks using predeterminable or previously determined search terms, wherein each neuron activation and over a weighted connection associated with at least one other neuron , wherein the product of activation of the neuron with the compound weight is a Final activation forms. Furthermore, the invention relates to a search engine and evaluation system for the evaluation of electronically stored Texts, in particular for the application of the method, comprising a User interface for obtaining search terms, as well as a search engine interface to initiate a search using the search terms and to obtain the result of the search and the found texts.

It are various search and evaluation methods for the evaluation of electronically stored texts by means of a computer system known. Usually are doing using one or more keywords electronically stored texts searched for these terms and subsequently listed in result lists.

Also Neural networks are known in which each neuron activates and over a weighted connection associated with at least one other neuron is, where the product of the activation of the neuron with the compound weight forms an end activation.

adversely with the known search methods, it's just that the fact that a search term is included in the searched text the result list can be opened, however, there are none Evaluation of the texts in terms of their relevance to the used search terms. By not sorting in terms of relevance takes place, are sometimes unusable result lists with an unmanageable huge Number of hits generated, which are useless due to the list length.

The The object of the invention is to provide a method and a system for searching and evaluate electronically stored texts, by means of which a determination and representation of the relevance of a searched and evaluated text in relation to the used Search terms is enabled, in particular a sorting of the searched and evaluated To enable texts according to relevance.

These The object is achieved by a Method according to claim 1 and a system according to claim 8 solved. Advantageous developments of the method according to the invention and of the system according to the invention are in the dependent claims specified.

Especially advantageous in the search and evaluation method according to the invention for the evaluation of electronically stored texts by means of a Computer system and neural networks using predeterminable or previously determined search terms, each neuron containing a Activation and over a weighted connection associated with at least one other neuron is, where the product of the activation of the neuron with the compound weight is an end activation, it is that multiple search terms are given and at least two neural networks are generated, each one Keyword is assigned to a neuron of each network, wherein, in the first neural network, the connection weights are constant are and the activation of each neuron depending the frequency the search term is carried out in the text to be evaluated, and wherein in the second neural network the connection weights in dependence the frequency of the search term in the text to be evaluated and the activation is constant and a calculation of the respective end activations both networks takes place.

Under The concept of an electronically stored text is each electronically stored file containing data and / or strings to subsume. Under the term of a search term are here words or parts of a natural word Language or any other arbitrary string. Since also word parts of a natural Thus language can be used also any Trunkierungen in the search and evaluation allowed be, d. H. that used search terms especially at the beginning and / or can be provided with a placeholder at the end.

Preferably is for each search term is the arithmetic mean of both end activations determined. This mean can be used as a measure of the meaning of the term for the Analyzed file or vice versa as a measure of the relevance of the text in Reference to the search term and in particular for more Evaluations and / or results representations are used.

By the consideration Both end activations can be a reliable measure of the relevance of the analyzed File in relation to the search term used.

Preferably, a sorting of the search terms durchge after decreasing end activation This means that those search terms are given a larger bea-interpretation whose respective final activation is greater than those search terms with a smaller final activation.

In a preferred embodiment In the process, the search terms are by means of a previous search using a first search term determined corresponding additional terms.

Preferably is after the evaluation of several texts a ranking of the Texts are formed by relevance, with the relevance of the final activation one or both neural networks and / or from an average derived from the end activations of both neural networks.

Prefers are by means of a predetermined negative list certain terms of a search and / or evaluation removed. This can be trivial Results by previous exclusion of certain terms or String can be avoided.

Preferably become by a first application of the method corresponding search terms to extend the search, by means of which under renewed Application of the method iteratively one or more other pairs generated by neural networks and determines their end activations become.

By an iterative application of the method can on the one hand corresponding Search terms used in the context of the search parameters used primarily be identified and, secondly, the quality of the search and evaluation are greatly improved.

Especially advantageous in the search and evaluation system according to the invention for the evaluation of electronically stored texts, in particular for the application of the method according to the invention, having a user interface for obtaining search terms, and a search engine interface for initiating a search using the search terms and to get the result the search and the found texts, it is that the system has one Net generator, by means of which at least two neural Nets are generated, with each neuron having an activation and over a weighted connection associated with at least one other neuron is, where the product of the activation of the neuron with the compound weight forms an end activation, each search term corresponding to one Neuron is assigned to each network, and wherein a text evaluation unit is provided by means of which the absolute and / or relative frequencies The search terms are determined in a text and passed to the network generator be, by means of the network generator at the first neural Net the activation of each neuron depending on the frequency of assigned search term at constant connection weights and in the second neural network, the connection weights in dependence the frequency the associated search term is carried out at constant activation and determining the respective end activations, and having an output unit, by means of which the end activations of the neural Networks are output.

Preferably the system has a graphical user interface by means of and the end activations of the two neural networks with respect to the search terms used are displayed graphically or representable are.

hereby can the relevance of the analyzed texts and files as well as the relevance the search terms used to each other in an advantageous and easy understandable Be mapped way.

Especially Preferably, the system has a user interface by means of an explanation in a natural Language over the relevance of the search terms in relation to the evaluated text is output, the explanation by means of the text evaluation unit from the end activations of the two neural networks generated in relation to the search terms used and.

By the automatic generation of a declaration text in natural Language is almost a translation of the numerical analysis result into a directly understandable one Result from which the relevance of the search terms used can be deduced is.

Prefers By means of the output unit, a ranking list of a plurality of Evaluated texts are formed and output, with the rankings depending on the relevance of the texts in relation to the search terms.

Preferably the system has data storage means for storage and / or Caching of search terms and / or the found and / or evaluated texts and / or neural networks as well Network parameters and / or ranking lists. Due to the arrangement of volatile and / or permanent data stores, the various files, Parameters, results, etc. are always kept readily available.

Particularly preferred is an interface for obtaining it by a previous search averaging provided to the first search terms corresponding further terms, wherein using the further terms, the first and second formed neural network expanded and / or a third and fourth neural network is formed by means of the network generator. There is thus the possibility of an iterative application and / or in-depth and / or extended search by means of the system according to the invention.

Preferably a filter is provided by means of which a definable negative list of terms of a search and / or evaluation is removable, to avoid trivial results, such as the consideration an article or a verb of a natural language, thereby the result of the data analysis could be strongly distorted.

through the process of the invention and of the system according to the invention can thus in particular the search results of Internet search engines be evaluated, d. H. that the procedure and the system in particular in the form of a so-called meta-search engine, which is for a first Search on other, external search engines resorting to, operated can. However, a primary search function can also be integrated, so not first resorted to external search engines must be, but already the primary search is carried out independently.

at forms the illustrated embodiment the system according to the invention a meta search engine. A meta-search engine is a program that on existing search engines such as Google or Yahoo or the like. The aim of the search is the specification of Texts. Preferably, the method and the system in one Computer network such as the Internet running or integrated. The required interfaces for input and output of data are thus available.

If So a user one or more search terms in the system of the invention then at least one search engine, usually Google, will automatically commissioned with the first search. The evaluation system takes over the texts specified by the search engine and starts with his actual work, namely the evaluation of the found texts. The search strategy of z. Google is modified by a search statement of the program because Google is strictly based on the specifications of the user and not as usual at the frequency, with which certain texts have already been called. The present Meta search engine has the general goal of getting information out of the to filter and condense texts found by search engines.

Specifically, the system operates as follows:
A user of the system is initially asked to input to the subject for which he or she is searching for relevant texts several terms in an order that represents the relevance of the terms to the user. An interface is available for this purpose.

A second input method is that of the priority lists. Here every term can be at the same time a priority be given to define a first structure. It can thus For each term optionally sub-terms can be assigned. The representation the terms are in a tree whose leaves represent the concepts.

alternative or cumulatively However, the search terms underlying the analysis are also automatic from an electronic file and / or from a storage medium read out and accepted become. The term of the user is thus merely a synonym for one Specification at the start of the process.

through the process of the invention and of the system according to the invention Now, an artificial neural network is constructed from this list of terms. This represents the contextual structure of terms from the user's point of view and is referred to below as a user network or network agent 1. The units of this network agent 1 are the terms that the user pretending; the connections between the units are called numeric Values (weights) either according to the order of user input or according to the priority list inserted into the network.

Then assigns the system puts the texts found by the search engine into a ranking list. For this Ranking activates another agent called the search agent because it examines the texts according to certain criteria. Of the Search Agent orients its ranking by how representative the are user-defined terms for the respective text, namely a) to what percentage the terms relative to the total number of all words in the text, b) how often contain the terms relative to the total number of terms in a text and c) how equally distributed the terms occur in the text. For the Criterion c) applies that the higher the value, the more homogenous a text is the terms occur in the text. Then the system starts the evaluation the text, which is in the highest ranking in the ranking.

To the evaluation will list the title of the document, a Short description (the first 300 characters of the document), the search engine, which has delivered the result and the ranking value in the form of a green Bar.

Around The user can either view the original document click on the title or on the "Show" scrapbook. It can also each one automatic display. In all cases, a browser window opens, by presenting the document. Under the "Details" button can be detailed information about the document will be retrieved. These should give the user a complete To avoid a review of the document, d. h., he may be due to the information presented will decide if it makes sense for him is the document to be considered as a whole or not.

Analogous to construct the user network, the system is now designing the search terms two more network agents (network agent 2 and network agent 3), which are called text networks. These nets indicate which ones Meaning the search terms in the text have on the base their frequency in the Text. This frequency is translated into two network structures and the network is "activated", i. H. it is put into a dynamic. This is realized by that the units are provided with certain initial numerical values whereupon the dynamic process is set in motion. The Result of the dynamic processing of the network, namely the End-activation values of the terms, then the meaning of the terms in the text. This generation of meaning is in the form of visual graphics as well as a linguistic interpretation. In addition, can an analysis will be issued, to what extent the structured specification deviates from the result of the text analysis.

The Both network agents differ in that Network Agent 2 is the same Has connection values of 0.5 from and to all units. The dynamics of Network Agent 2 is created by the units with values be activated according to the frequency correspond to the terms in the text. At the network agent 3, the Values of the connections are set proportional to the frequency in the text, while the activation happens by equal numerical values. The respective ones End activations of the two network agents are displayed as curves, which correspond in the form of progression. Because the absolute end values usually differ, the arithmetic mean is calculated as Meaning is interpreted.

There such curves for untrained User usually are difficult to understand, the curves are in a colloquial Explanation translated.

The previously described program parts and operations make up the first Part of the overall system. The clarity half the individual components are again schematic shown:

Component list 1:

• a) Interface for the user to enter of own terms as well as for the outputs of the results to the user;
• b) "Search" tool for instruction to a search engine for a special form of search (= modification of the operations of Search engine);
• c) tool "Agent Constructions";
• d) network agent 1 (user network);
• e) analysis agent to construct a ranking, what the foundation for the Construction of the network agents is;
• f) Network Agents 2 and 3 (text networks);
• g) Tool "Graphics" for construction the visual representation;
• h) tool "translation" for construction a linguistic text.

There the system in addition to the user's search list by another search agent (search agent 2) creates a second list of terms that are contained in the text the user can decide if he is his original Extend the user network with new terms from this second list want. This can be done manually or automatically. In the first Case, the user indicates which additional terms from the second List should be inserted in the previous network agents. Resulting from it by the same method as described above, two more Network agents 3 and 4. In the second case, the system searches for specific ones Criteria, namely oriented to frequencies as in the ranking process, additional Terms out of the second list and constructs two more network agents 5 and 6. In both cases Thus, the new network agents consist of the previous network agents 2 and 3 as well as either manually entered or automatically selected additional terms. If desired, The user can also execute both options one after the other.

It Thus, the user does not perform an intervention during the procedure The method is rather iterative applicable, of course the previously generated analysis results as input variables of another Process can serve.

Depending on the selected option, the network agents are activated again and the result is output to the user as the meaning of the entire terms. The output to the user happens with the same tools as already described. With these extension procedures, the user can now proceed again or abort the analysis of the text. This Erwei concept is based on the experience that at the beginning of a search often only a few terms are available and that even intermediate results of the search can bring the user to expand his search terms by those additionally found by the program. When the user aborts the text analysis, the results are saved and the program starts the same process with the interpretation of the second text in the ranking list.

Around to ensure, that the procedure and system in the creation of the second list only takes up concepts that are for the text are relevant, are established "negative lists" in the program contain. These contain words the colloquial language that does not in the construction of the second list considered should be. These blacklists are standard and could be the System just added become. The main advantage is that the user even expand these negative lists.

Also the additional components just shown will be listed again.

Component List 2:

• i) search agent 2 for creating an extended list;
• j) file "negative list";
• k) file list 2;
• l) Tool "Option" for manual and / or automatic expansion of the network agents;
• m-p) Network agents 3. 4. 5 and 6;
• q) Results file of the first one Text.

If the user continues with the same procedures after a single extension of the network agents 1 and 2, additional components are created, but these are not listed in detail. For a schematic overview of all components see 1 ,

• a) Die Bedeutung von Begriffen aus der Sicht des Benutzers einerseits und für einen bestimmten Text andererseits wird jeweils als Ergebnis eines dynamischen Netzprozesses aufgefasst. Die Konstruktion der jeweiligen Netze wird durch eine neuartige Lernregel geleistet. Damit folgt die Grundlogik des Verfahrens im Wesentlichen den Prozessen, die im menschlichen Gehirn ablaufen. Derartige Überlegungen für die Konstruktion von Meta-Suchmaschinen sind bisher noch nicht realisiert worden.
• b) Das Programm kann insbesondere in dem Sinne streng am Benutzer orientiert sein, dass er nicht nur die anfänglichen Vorgaben eingibt. Insbesondere kann die Basis für den gesamten Ablauf des Programms die Sicht des Benutzers sein und nicht so sehr die Orientierung an einer objektiven Textstruktur – diese wird allerdings auch verwendet. Der Benutzer kann insbesondere die Netzerweiterung wieder durch seine eigene Sicht beeinflussen, was für ihn relevant ist oder nicht. Das Verfahren präsentiert dann auch die erweiterten Ergebnisse in Relation zur Sicht des Benutzers.

The principle new and particularly advantageous in this method and system consists in particular in the following aspects:

• a) The meaning of terms from the point of view of the user on the one hand and for a particular text on the other hand is understood in each case as the result of a dynamic network process. The construction of the respective networks is done by a novel learning rule. Thus, the basic logic of the procedure essentially follows the processes that take place in the human brain. Such considerations for the construction of meta-search engines have not yet been realized.
• b) The program may be strictly user-oriented, in particular in the sense that it not only inputs the initial specifications. In particular, the basis for the entire course of the program may be the user's view and not so much the orientation on an objective text structure - this is however also used. In particular, the user can influence the network extension again by his own point of view, which is relevant for him or not. The method then also presents the extended results in relation to the user's view.

Help functions for program usage

• • Innerhalb des Programms wurden statische Hilfetexte auf jedem Programmfenster eingefügt, die den Benutzer durch die einzelnen Programmschritte leiten.
• • Als Nachschlagewerk dient eine implementierte „Online-Hilfe”, in der, nach Stichpunkten geordnet, alle wesentlichen Bestandteile des Programms erklärt werden.
• • Zum schnellen Einstieg in die Bedienung des Programms dient ein Video-Tutorial, bei dem permanent die Programmoberfläche sichtbar ist und über gesprochene Erläuterungen alle Programmschritte in Lektionen unterteilt durchgearbeitet werden.

To assist program users, three basic help systems have been developed.

• • Within the program, static help texts have been added to each program window, guiding the user through the individual program steps.
• • The reference work is an implemented "Online Help", in which, arranged by bullet points, all essential components of the program are explained.
• • For quick access to the program, a video tutorial is used, in which the program interface is permanently visible and through spoken explanations all program steps are subdivided into lessons.

Of the The user is therefore audibly and visually addressed, resulting in different Areas of the brain are activated and greater learning success is possible. These three systems make a user optimal and continuous supported when using the program. Basically, in the development also paid attention to intuitive usability.

The advantages at a glance

• • For the beginner up One area: The extension concept is based on the experience that at the beginning of a search frequently only a few terms available are and that intermediate results already search the user for it can bring his Expand search terms by those additionally found by the program. Thus receives the user synonyms or additional Terms that make a search more effective.
• • For the expert: Due to the structured and advanced search, the user can post looking for specialized information.
• • Of the User can even extend the negative lists to unwanted search results (eg advertising, seminar or lecture announcements).
• • The relationships between the terms in a text can be viewed graphically.
• • Time savings through effective and user oriented research.

primary applications

• • For all, who must follow the latest developments in research and technology.
• • The Program can be used as a new form of e-learning. Depending on the level of knowledge can the information about various texts accessible be made.

Claims (15)

Search and evaluation method for the evaluation of electronically stored texts by means of a computer system and neural networks using predefinable or previously determined search terms, wherein each neuron has an activation and is connected via a weighted connection with at least one other neuron, wherein the product of the Activating the neuron with the connection weight forms an end activation, characterized in that a plurality of search terms are given and at least two neural networks are generated, each search term is assigned to each neuron of each network, wherein in the first neural network, the connection weights are constant and the Activation of each neuron takes place as a function of the frequency of the search term in the text to be evaluated, and wherein, in the case of the second neural network, the connection weights are dependent on the frequency of the search term in the text to be evaluated; tivation is constant, and a calculation of the respective end activations of both networks takes place. Method according to claim 1, characterized in that that for each search term is the arithmetic mean of both end activations is determined. Method according to claim 1 or 2, characterized that sort the search terms by decreasing final activation is carried out. Method according to one of the preceding claims, characterized characterized in that the search terms by means of a previous search using a first search term determined corresponding additional terms. Method according to one of the preceding claims, characterized characterized in that after the evaluation of multiple texts one Ranking of texts is formed by relevance, with relevance the end activation of one or both neural networks and / or out derived from an average of the end activations of both neural networks becomes. Method according to one of the preceding claims, characterized characterized in that by means of a predetermined negative list certain Terms from a search and / or evaluation are excitable. Method according to one of the preceding claims, characterized characterized in that by means of a first application of the method corresponding search terms are found to expand the search, by means of which, using the method again iteratively one or several other pairs of neural networks and their generated Final activations are determined. Search and evaluation system for the evaluation of electronic stored texts, in particular for the application of a method according to one of the preceding claims, having a user interface for obtaining search terms, and a search engine interface for initiating a search using the search terms and to get the result the search and the found texts, characterized in that the system comprises a network generator, by means of which at least two neural networks are generated, with each neuron one Activation and over a weighted connection associated with at least one other neuron is, where the product of the activation of the neuron with the compound weight forms an end activation, each search term corresponding to one Neuron is assigned to each network, and wherein a text evaluation unit is provided by means of which the absolute and / or relative frequencies The search terms are determined in a text and passed to the network generator be, by means of the network generator at the first neural Net the activation of each neuron depending on the frequency of the assigned search term at constant connection weights and in the second neural network, the connection weights in dependence the frequency the associated search term is carried out at constant activation and determining the respective end activations, and having an output unit, by means of which the end activations of the neural Networks are output. System according to claim 8, characterized in that it has a graphical user interface by means of which the end activations of the two neural networks in relation to the used search terms are graphically displayed. System according to claim 8 or 9, characterized in that it comprises a user interface by means of which an explanation in natural language about the relevance of the search terms in relation to the evaluated text is output, wherein the explanation by means of the text evaluation unit generated from the end activations of the two neural networks with respect to the search terms used and. System according to one of claims 8 to 10, characterized in that by means of the output unit a ranking list of a plurality of Evaluated texts are formed and output, with the rankings depending on the relevance of the texts in relation to the search terms. System according to one of claims 8 to 11, characterized that the system data storage means for storing and / or Caching of search terms and / or the found and / or evaluated texts and / or neural networks as well Network parameters and / or rankings. System according to one of claims 8 to 12, characterized that an interface to receive from a previous one Search determined further corresponding to the first search terms Conceived is provided, wherein using the other terms the first and second formed neural networks are expanded and / or expanded a third and fourth neural network by means of the network generator is formed. System according to one of claims 8 to 13, characterized a filter is provided by means of which a prescribable negative list of terms of a search and / or evaluation is removable. Computer program product comprising one on one Computer running Computer program comprising the method according to one of claims 1 to 7 performs if it is running on a computer.