JP2006235716A - Document filtering system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve efficiency in searching a document by extracting interesting information from a suitable document sample selected by a user oneself, and expressing interest to the document on the basis of the information in a differentiation structure. <P>SOLUTION: The system is provided with a differentiation structure generating part 111 for generating a differentiation structure of a plurality of clusters from a group of documents, a cluster representative point creating part 112 for creating a center gravity position of the cluster by using the document in each cluster, a similarity calculating part 121 for calculating a similarity of new document and each cluster representative point, and a recommendation ranking indicator calculating part 122 for converting a plurality of similarities to a single index. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a document filtering system, and more particularly to a document filtering system capable of supporting a task of searching for a document related to a specific interest.

  The conventional document filtering system is designed based on the idea that “interest in a paper can be expressed by only a few keywords, and the keywords can be easily selected by researchers”. In the case of this system, each researcher selects in advance several keywords that are likely to be included in relevant documents (articles) and registers them as the researcher's own interests in the system. However, since this method is simply the registration of keyword groups, it can be said that “the expression with the interest structure remaining undifferentiated”.

On the other hand, it extracts and accumulates terms from new documents published every day. A term is a word that characterizes a document. A Boolean match is made between the extracted term and the registered keyword selected by the researcher, and the degree of match is considered as an estimated value of the system for the degree of interest of the researcher. If this estimated value is greater than or equal to a predetermined value, it is determined as a conforming document, and if it is less than the value, it is determined as a non-conforming document. A system that presents only relevant documents to users is known. There are, for example, Patent Document 1 and Non-Patent Document 1 as documents showing this type of technology.
JP 2003-157273 A SLKDBollacker and C. Giles "A system for automatic personalized tracking of scientific literaure on the web" proc.of the fourth AGM Conference on Digital Libraries, pp.105-113 (1999)

   As described above, in the conventional document filtering system, since the conforming documents are selected by simple Boolean matching with undifferentiated structure keyword groups, there is a problem that there are many omissions in the selection of desired conforming documents and mixing of nonconforming documents. There is. These issues are thought to be due to lack of information for interest expression, difficulty of expression of interest by researchers themselves, and imperfect structure expression for interest.

  In general, it is difficult to represent a complicated structure of a document with only a few keywords. Even if the same keyword is included in two documents, if the frequency of appearance, co-occurrence terms, etc. are different, the contents of the documents may be different. This indicates that the document discrimination capability is not sufficient in the Boolean collation in which the similarity between documents is evaluated only by information on the presence or absence of several keywords.

  Also, researchers are not always aware of their own interests in a clear way. This is thought to be due to the nature of research activities, starting from a vague state and reaching clear knowledge after trial and error. In the first place, it is inherently difficult for researchers to express the “interest” that is ambiguous and difficult to grasp by using several keywords.

  In addition, there are various aspects of in-house researchers interested in academic documents. For example, there are various cases such as an interest in a problem to be solved, an interest in a technique for solving the problem, and an interest in a related secondary technology. Depending on the aspect you are interested in, the corresponding keywords are quite different. As a result, there is a problem that the keywords that express different aspects are unconsciously mixed in the selected keyword, and the focus of interest expression becomes ambiguous and the accuracy is lowered.

  FIG. 2 is a diagram showing the distribution of conforming papers and non-conforming papers, and FIG. 3 is a diagram illustrating an undifferentiated state of the interest structure. As shown in FIG. 2, it is considered that the academic document group including the conforming paper (white circle mark) and the non-conforming paper (black circle mark) is distributed on the vector space by the term vector as shown in FIG. At this time, as shown in FIG. 3, if it is determined whether the document is compatible or non-conforming based on the undifferentiated structure expression composed of several keyword groups, it is considered that the presentation of the conforming document or the mixing of the non-conforming document occurs.

  It is desirable for the user to display the ranking of the filtering results, but in the undifferentiated structure representation, as shown in FIG. 3, it is a “cobblestone mixed state” in which compatible and non-conforming documents are mixed, and the meaningful ranking display is I can't get it.

  The present invention has been made in view of these problems, and extracts interest information from a relevant document sample selected by a user such as a researcher, and expresses interest in the document in a differentiated structure based on this information. This improves the efficiency of document search.

  In order to solve the above problems, the present invention employs the following means.

  Differentiated structure generator that generates a differentiated structure by multiple clusters from a document group, cluster representative point generator that creates the center of gravity of a cluster using documents in each cluster, and the similarity between a new document and each cluster representative point A similarity calculation unit for calculating, and a recommended ranking index calculation unit for converting a plurality of similarities into a single index are provided.

  Since the present invention has the above-described configuration, it is possible to improve the efficiency of searching for a document of interest.

  A few keywords are not enough to structure the researcher's interests. The necessary and sufficient amount of information needs to be extracted from the relevant document sample. For this purpose, the inventors have used a vector space method for expressing a document as a vector using terms and frequency information contained in a conforming document (Inoue, Hashimoto: “Document Filtering Method Using Nonconformance Profile”, Information Processing Society of Japan, J42. -D2, 3, pp507-517 (2001) et al.

  In this method, different interests are not expressed by undifferentiated structures as in the prior art, but different interests are expressed by differentiated structures using different profiles. Thereby, it is possible to improve the accuracy of searching for a compatible document.

  When academic documents are mapped onto a vector space based on term vectors, it is considered that documents including many of the same terms among the conforming documents are distributed together as shown in FIG. Focusing on this point and clustering nearby documents by the clustering method, clusters corresponding to the number of researchers' interests are generated as clusters of matching documents. Depending on the clustering algorithm, some non-conforming documents may be included in each cluster, but the majority are composed of conforming documents. Therefore, a cluster representative point corresponding to the vector average of only the relevant documents in the cluster is obtained. This can be said to express the researcher's interest in a differentiated structure. Here, cluster representative points corresponding to the interests of these researchers are called profiles.

  FIG. 4 is a diagram for explaining an expression of the differentiated structure of interest composed of a plurality of clusters, and defines an appropriate index for expressing the similarity between each profile and a document. As shown in FIG. 4, many conforming documents are distributed in a region having a high degree of similarity, but most of the regions having a small degree of similarity are occupied by non-conforming documents.

  On the other hand, it is conceivable that an area with a medium similarity includes a part of a document that can be a hint for a new idea. Each region may overlap, but in the method of the present invention using the differentiated structure representation compared to the conventional method based on the undifferentiated structure, each interest is expressed separately by each profile. Relevant documents will be separated more sharply. Therefore, the ranking display of the filtering result can be realized by utilizing the sharp separation performance of each profile.

In general, it is more convenient for the user not to simply present the relevant documents selected by the system but to display the rankings in the order of recommended reliability. In the case of the conventional method, since only the similarity with a single profile of an undifferentiated structure is obtained, this similarity may be used as a ranking index. However, in the case of the differentiated structure profile of the present invention, since a plurality of similarities between a new document and a plurality of profiles are calculated, ranking cannot be performed as it is. Therefore, in order to display the ranking of new documents, it is necessary to convert a plurality of similarities into one index when selecting relevant documents. For this purpose, a reference threshold value θi corresponding to the power range for each profile is calculated by formula (1) when creating the profile.

  Here, i is a cluster number, θi is a reference threshold value of cluster i, j is a sample number in cluster i, Ni is the number of samples in cluster i, and Sim (pi, dj) is a representative point of cluster i. And the j-th sample dj in the cluster i. The reference threshold value θi is the minimum value among the similarities between the profiles and the elements in the cluster. In the next relevant document selection phase, a recommended ranking index is calculated based on the reference threshold value, and a new document ranking is presented based on the calculated index.

  Next, the distribution state of documents in the term vector space is considered.

FIG. 5 is a diagram for explaining the distribution of documents in the term vector space when one cluster is considered. Since the reference threshold value θi is determined by a limited learning sample, it should be a value larger than the optimum threshold value. If the degree of similarity is greater than the reference threshold, it can be said that the document is definitely relevant. However, since there is a conforming document that is smaller than the reference threshold, the recommended document is not recommended. Therefore, a correction threshold value δi (α) obtained by multiplying each reference threshold value by a number α of 1 or less is defined by Expression (2).

  This can be said to be a threshold value which is corrected by loosening the condition of the reference threshold value θi and slightly increasing the threshold value. If the document is presented when the similarity is equal to or greater than the correction threshold value, mixing of non-conforming documents slightly increases, but the omission of presentation of conforming documents can be reduced. When α is 1, the correction threshold value δi (α) is equal to the reference threshold value θi, and becomes the boundary value of the narrowest region in which only relevant documents are included. When α is 0, the correction threshold δi (α) is 0, and all documents are presented unconditionally. For this reason, α is usually a value between 0 and 1.

Next, the ranking of the relevant documents is presented using the sharp document separation characteristic obtained by the differentiated structure expression. In this case, first, the recommended ranking index fr (dk) is calculated by the formula (3) using the correction threshold value for each profile and the similarity of the new document, and the conforming document is determined by this index value fr (dk). To the user in descending order.

  That is, when a correction document δi (α) is sequentially expanded starting from the reference threshold value θi by sequentially reducing α from 1 for a new document, the degree of similarity with a profile that is a certain point becomes the correction threshold value. It should exceed δi (α). Fr (dk) according to equation (3) is the value of α at that time. The greater the recommended ranking index fr (dk), the greater the possibility that the document dk is a conforming document, and the smaller the recommended ranking index fr (dk), the greater the possibility that the document dk is a non-conforming document. If the value is medium, it may be a non-conforming document, but it may also be a new idea source document that is the source of new ideas.

  In this example, instead of simply presenting the filtered result to the user as in the conventional system, new public documents are presented to the user in descending order of the recommended ranking index fr (dk). As a result, users can select necessary documents according to their research phase.

  FIG. 1 is a diagram illustrating the overall configuration of the document filtering system according to the present embodiment. In the figure, 2 is a CPU, 3 is a memory used by the CPU, 4 is a keyboard as an input device, and 5 is a display. A bus 6 connects the storage device 7 and the CPU. The storage device includes a profile creation unit 11 and a matching document ranking unit 12 configured by a program.

  Reference numeral 111 denotes a differentiated structure generation unit that generates a differentiated structure including a plurality of clusters in which relevant papers having a high degree of similarity are grouped together. Reference numeral 112 denotes a profile (cluster representative point) creation unit, which creates a profile composed of the centroid positions of clusters using documents in each cluster. The profile represents a matching document included in each cluster, and can be considered as a virtual document located at the center of the cluster. A reference threshold value calculation unit 113 calculates the similarity between the representative point of each cluster and the elements in the cluster, and sets the minimum value among them as the reference threshold value. Reference numeral 114 denotes a storage device that stores a sample of a profile creation paper, 115 denotes a storage device that stores a profile, and 116 denotes a storage device that stores a reference threshold value.

  Reference numeral 121 denotes a similarity calculation unit that calculates the similarity between each profile and a new document that has been extracted and vectorized. For example, the new document is mapped to a vector space based on the terms extracted from the new document. The similarity between the mapped new document and the representative point (profile) of the cluster is calculated. 122 is a recommended ranking index calculation unit, and the ratio of the correction threshold to the reference threshold when the similarity with the representative point of the cluster when the threshold is a correction threshold that is sequentially expanded from the reference threshold exceeds the correction threshold. Calculated as a recommended ranking index. A ranking presentation unit 123 presents the document for which the index has been calculated to the user in descending order of the index (in descending order of similarity). Reference numeral 124 denotes a storage unit that stores a new document to be a ranking presentation target, and reference numeral 125 denotes a storage unit that stores a recommended ranking index.

  FIG. 6 is a diagram for explaining a process of generating a differentiated structure of interest. This process is initially performed only once at the beginning of filtering. In FIG. 6, each user first registers in the system a list of conforming documents already collected and held and a group of conforming documents digitized. At this time, a non-conforming document is also registered at the same time in order to determine the boundary between conforming and non-conforming (step 41). The stop word is removed from the word group included in the registered conforming and non-conforming document group, and the “base term” is selected. A “term vector” Drel (conforming) and Dirrel (non-conforming) are constructed with the “basic term” as a vector base and the occurrence frequency in each document of each term, that is, a TF (Term Frequency) value as a weight. As a result, the conforming / non-conforming document is positioned on the vector space by the term vector (step 42).

  Next, the conforming and non-conforming document groups positioned on the term vector space are clustered (step 43). In order to perform clustering, it is necessary to define the similarity between documents and between clusters. Similarities between documents include the simplest TF method that uses term frequency information, the log TFIDF method that weights terms that occur frequently in a specific document, and the SMART method with correction based on empirical parameters. . The inventors adopted the SMART method, which reported the best results among these methods.

  The similarity between clusters is defined as the similarity between documents belonging to each cluster. The clustering method includes a single link method and a complete link method depending on the difference in similarity definition between the clusters. The inventors adopted a complete link method that facilitates the formation of small clusters in order to capture the structure more faithfully. Note that in practice, up to one non-conforming sample is allowed in order to prevent a tiny cluster consisting of a single element from standing up. For this reason, a maximum of one non-conforming sample may be included in the cluster, but this is excluded when actually creating a profile.

A cluster is a collection of relevant documents that are close in vector space, that is, having a high degree of similarity. Therefore, one cluster should contain documents corresponding to the same interest. Next, a “base term” is selected for each relevant document in the cluster. A vector Prel (i) is constructed with the “base term” as a vector base and the weight of the average value of the appearance frequency TF (Term Frequency) value in each document of each term (see equation (4)).

  This vector Prel (i) is a profile generated from the i-th cluster. That is, as many profiles as the number of types of interest can be created. The profile represents a matching document included in each cluster, and can be considered as a virtual document located at the center of the cluster. Next, the reference threshold value θi is obtained for each cluster using the equation (1) (step 44). At this time, the similarity Sim is calculated by adopting the SMART method. The smaller the reference threshold, the larger the cluster area range. That is, it can be said that the smaller the reference threshold, the greater the specific gravity of the corresponding interest.

  FIG. 7 is a diagram for explaining the compatible document selection process. In this process, for example, every time a new document is acquired, the relevant documents are selected from the acquired documents and presented in a ranking. In FIG. 7, first, a document posted in a digital library or the like is regularly monitored, and if there is a new document, it is acquired. Since the document to be acquired is usually in the PDF (registered trademark) format, it is converted into a plain text format that is easy to process. The term is extracted from the new document and vectorized by the same method as the term vectorization of the matching document described above (step 51). Next, the similarity with each profile and the correction threshold are calculated for each user. As a similarity calculation algorithm, the same SMART method as that used in clustering is used. Further, the correction threshold δi (α) is calculated by multiplying the reference threshold by a numerical value α of 1 or less according to the equation (2) (step 52), and the equation (3) is calculated based on the calculated correction threshold δi (α). Is used to obtain a recommended ranking index fr (dk). Next, the acquired new document is presented to the user in the form of a document title list in descending order of the recommended ranking index fr (dk) (step 53).

  Next, the filtering performance of the document filtering system of this embodiment was confirmed by experiments. An academic sample experiment was created for filtering experiments. An evaluation experiment was performed along the aforementioned processing flow (FIGS. 6 and 7). The obtained effect will be described below.

  The subjects were three researchers, a researcher A with relatively little experience, a medium-sized researcher B active at the forefront of research, and an experienced veteran researcher C (hereinafter referred to as Ra, Rb, and Rc, respectively). Called). Among academic documents, about 10,500 of a series of documents in the field of information science for two years (2001/4 to 2003/3) were used as samples for filtering experiments. In order to limit the scope of the document and increase the processing efficiency, screening is performed in advance using keywords specified by the subject himself / herself. In the subsequent experiments, a screened sample is used (step 41).

  Subjects who have screened themselves are classified into two types: “conforming” and “non-conforming”. As criteria for conformance and non-conformity, the title, author, date of issue, abstract, keyword information included, and the full text will be presented to improve the efficiency of judgment work. As a result, 68 conforming samples and 177 nonconforming samples were obtained for the three people. In order to apply the multi-fold cross-validation method, the sample is divided into 6 parts. Using this divided sample, a relevant document selection experiment is performed.

  First, the screened samples are converted into term vectors (step 42), and the profile creation samples are clustered for a total of six experiments. The SMART method is used as an intercluster similarity calculation algorithm used in clustering, and the complete link method is used as a clustering method. A profile is created for each divided experiment based on the clustering result (step 43). Further, by calculating the similarity between the profile and the sample included in the cluster, a reference threshold value is obtained using equation (1) (step 44) and stored together with the profile.

  Next, a filtering experiment is performed using the filtering experiment sample (new document 124). For each subject, a maximum of 6 experiments are performed by multi-division cross validation. A correction threshold value is calculated based on the similarity between the relevant document selection test sample and each profile and the already calculated reference threshold value (step 52). Next, a recommended ranking index value is calculated according to equation (3) based on this correction threshold (step 53). Histograms were created by combining recommended ranking index values obtained in up to six experiments (see FIG. 8).

  Furthermore, a reproducibility / accuracy graph was created in order to quantitatively evaluate the separation accuracy of the conforming / nonconforming sample (see FIG. 9). In this graph, when the threshold value frmin of the recommended ranking index is changed from 0 to 1, the horizontal axis represents frmin, and the reproduction rate / accuracy when the sample dk satisfying fr (dk)> frmin is regarded as the recommended result is shown vertically. It is plotted on the axis. The recall and accuracy change depending on the value of frmin, but how much non-conforming document can be presented under the condition of 100% recall, that is, recommended leakage rate = 0%. Important in filtering systems.

  Based on FIG. 9, the reproduction rate, accuracy, and nonconforming document rejection rate when frmin = 0.55 were obtained for three subjects. The result is shown in FIG. In this table, the non-conforming document rejection rate is the ratio of the number of non-conforming samples mixed into the recommended target to all non-conforming samples. This is an important indicator in terms of how much unnecessary documents can be read.

  From FIG. 10, it can be seen that in Ra and Rb, non-conforming documents can be eliminated by 74% and 43.5%, respectively, while maintaining a reproduction rate of 100%. It can be said that mixing of non-conforming documents can be reduced to 1/2 to 1/3. This is considered to be the effect of introducing the differentiated structured profile. With the newly introduced recommended ranking index, the results can be presented to the user in the order of recommended reliability as shown in FIG. It can be seen from this graph that there are many relevant documents in the top ranking documents. Further, when the subjects were asked to check whether there was a document that could hint at a new idea for the medium document, it was found that there were actually one or two such documents for each subject.

  As described above, according to the present embodiment, first, a user's interest such as a researcher is clustered to cluster a user's interest, thereby expressing the differentiated structure of the user's interest using a plurality of profiles. Next, a recommended ranking index is calculated for the new document based on the similarity to each profile and the differentiated structure information, and the result is presented as a ranking. As a result, the user can efficiently search for a new document of interest.

It is a figure explaining the whole structure of the document filtering system which concerns on this embodiment. It is a figure showing distribution of a conformity paper and a non-conformity paper. It is a figure explaining the undifferentiated state of an interest structure. It is a figure explaining the differentiation structure expression of interest comprised from a plurality of clusters. It is a figure explaining distribution of a document on term vector space when thinking about one cluster. It is a figure explaining the production | generation process of the differentiation structure of interest. It is a figure explaining the selection process of a relevant document. It is a figure which shows the evaluation experiment result (conforming / nonconforming document number) for every test subject. It is a figure which shows the evaluation experiment result (reproducibility / accuracy) for every test subject. It is a figure which shows the evaluation-experiment result (recommended omission / nonconformity document exclusion rate) for every test subject.

Explanation of symbols

2 CPU
3 Memory 4 Keyboard 5 Display 6 Bus 7 Storage Device 11 Profile Creation Unit 111 Differentiation Structure Generation Unit 112 Profile (Cluster Representative Point) Creation Unit 113 Reference Threshold Calculation Unit 114, 115, 116 Storage Device 12 Relevant Document Ranking Unit 121 Similarity Calculation Unit 122 recommended ranking index calculation unit 123 ranking presentation unit 124, 125 storage device

Claims (8)

  Differentiated structure generator that generates a differentiated structure by multiple clusters from a document group, cluster representative point generator that creates the center of gravity of a cluster using documents in each cluster, and the similarity between a new document and each cluster representative point A document filtering system comprising a similarity calculation unit for calculating, and a recommended ranking index calculation unit for converting a plurality of similarities into a single index. The document filtering system according to claim 1, wherein
The differentiated structure generation unit includes a differentiated structure generation unit using a clustering method, and a recommended ranking index calculation unit that converts a plurality of similarities into a single index includes a reference threshold value calculation unit proportional to the size of each cluster, a threshold value A document filtering system comprising: a correction threshold calculation unit that calculates a correction threshold obtained by sequentially enlarging the reference threshold value, and a recommended ranking index calculation unit that calculates a ratio of the correction threshold to the reference threshold as a recommended ranking index. A differentiated structure generation unit that clusters a plurality of preselected documents into a vector space based on terms extracted from these documents and term vectors represented by the appearance frequency of the terms, and classifies the plurality of documents into a plurality of clusters. When,
A cluster representative point creation unit that calculates the center of gravity of each classified cluster as a cluster representative point;
A profile creation unit including a reference threshold value calculation unit that calculates a similarity between a representative point of each cluster and an element in the cluster, and sets a minimum value thereof as a reference threshold value;
Means for extracting a term from a new document and mapping the new document to the vector space based on the extracted term;
A similarity calculator for calculating the similarity between the mapped new document and the representative point of the cluster;
A document filtering system, comprising: a recommended ranking index calculation unit that converts the calculated similarity into a single index based on the reference threshold value of each cluster. The document filtering system according to claim 3, wherein
A document filtering system comprising a ranking presenting unit for ranking and presenting new documents according to the single index. The document filtering system according to claim 3, wherein
A document filtering system, wherein a plurality of preselected documents are screened according to preset terms and classified into conforming documents and nonconforming documents. The document filtering system according to claim 3, wherein
The recommended ranking index calculation unit calculates the ratio of the correction threshold to the reference threshold when the similarity with the representative point of the cluster exceeds the correction threshold when the threshold is a correction threshold that is sequentially expanded from the reference threshold. A document filtering system characterized by calculating as follows. The document filtering system according to claim 4.
A document filtering system, wherein the ranking presentation unit ranks and presents new documents in descending order of index values. The document filtering system according to claim 3, wherein
A document filtering system characterized in that similarity is calculated by the SMART method.

Images