JP4967133B2 - Information acquisition apparatus, program and method thereof - Google Patents

Description

  The present invention relates to a search for related content information and an information acquisition apparatus that aggregates them.

Many information search systems including Web search are based on keyword search, and a user obtains a search result by giving a search keyword to the system. Since the search is performed only on simple word matching, the search results are often enormous and there are also many noises. At present, the user has to search for text that meets the requirements from the search result of the boulder chaos with a huge amount. Therefore, it takes a lot of labor for the user to obtain matched information from the search results. In addition, a situation in which it is desired to find related information during keyword search may occur frequently. In recent years, with the spread of information equipment, various types of information have been digitized, and now that a large amount of information can be browsed at any time, the task of efficiently selecting the necessary information from that is rapidly digitizing information. It has become difficult while proceeding to.
Therefore, Japanese Patent Laid-Open No. 9-185632 discloses an information editing / processing method for searching a plurality of databases to acquire desired information and editing / processing the information into a format desired by the user.

The information editing / processing method of the background art includes a search instruction for target information estimated to be held by at least one remote database, an instruction for editing / processing the searched information, and an instruction for determining the output form of the edited information. , A second process for generating a command for searching the at least one database based on the search instruction, searching the database, and at least one object obtained by the second process In accordance with the instruction of the first process for the information, the third process of editing and processing, and according to the output form determination instruction of the information for the target information edited and processed by the third process, It is characterized by comprising a fourth process in which it is converted into a predetermined output form and output by a method that can be perceived as visual, auditory or other senses. In the third process, the search results displayed in the fourth process are sorted in the output order designated by the user. As the output order, one of information relevance order, information generation time order, and search order is used.
JP-A-9-185632

  As described above, according to the information retrieval / editing method and apparatus of the background art, since the relation between the obtained information is conscious, the related information can be provided close to each other to the user. In addition, since information obtained from a plurality of databases is handled in a uniform, it is possible to appropriately provide information obtained from different databases regardless of time order or relevance order.

  However, even if the provided information satisfies the conditions for the topics, the quality or quantity of the information contents may not always sufficiently satisfy the user's request. Therefore, there is a problem that information search must be performed again when the information desired by the user is insufficient.

  It is also possible to compare and deepen understanding of information such as documents created from multiple viewpoints for a topic, but that work also increases the amount of information. It has become difficult as you go. In addition, regarding the related information, there is also a problem that the burden on the user increases because repeated portions of the information are repeatedly acquired.

  The present invention has been made in order to solve the above-mentioned problems, and by comparing and organizing related information in a search device for finding follow-up information, it efficiently obtains necessary information and meets the user's request. An object of the present invention is to provide an information acquisition device that can output in a format.

An information acquisition device according to the present invention includes a combination of feature vector generation means for generating a feature vector of weighted search information and a combination of feature vectors of all search information in the information acquisition device that has acquired at least one search information. Information similarity calculation means for calculating similarity, similarity matrix calculation means for calculating a numerical similarity matrix obtained by the information similarity calculation, numerical analysis of the similarity calculation result, and feature vector Eigenvector creation means for obtaining an eigenvector of the maximum eigenvalue of the search information, question vector creation means for creating a question vector of the inquiry content of the search information, and a calculated value of the cosine of the feature vector and the question vector multiplied by the value of the eigenvector. a search order determination means for determining the search order of the related information, part of speech engagement of contained in the text of retrieved information And dependency analysis means for parsing the relationship, and the noun extracting means for extracting a noun in clause according to clause containing the verb of each sentence, the similarity S ₁ and noun set between nouns alone of the extracted noun of a noun set similarity comparison calculation means for calculating a similarity S _2, and noun see matching ratio calculating means for calculating a similarity S ₃ ratio that matches the display of the extracted noun, the similarity S ₂ A sentence similarity calculating means for calculating the sentence similarity S by adding the similarity S ₃ , a sentence type selecting means for selecting according to the sentence type of the search information, and the sentence similarity calculating and related information of the sentence type selected. Output the integrated contents and the search results scored by the search order determining means, and output an input form for inputting the appropriateness of the search results and / or the parameter weighting degree. And a correction means for correcting the feature vector and the question vector based on information input to the input form, wherein the search order determination means is the feature vector and the question corrected by the correction means. The search order of the related information is determined again based on the vector, and the output means outputs the search results scored by the search order determination means, including the temporal relationship between the search information. Is . Here, “information” includes, for example, a sentence, a document, an article, an image, a sound, and the like. The “sentence type” is, for example, an overlapping part, a unique part, a supplementary explanation, or the like. In addition, the summary, schedule, reason, analysis, supplementary explanation, mode / hearing, ratio / inference, etc. shall be included.

Thereby, a feature vector of weighted search information is created, a similarity of a combination of feature vectors of all search information is calculated, a numerical similarity matrix obtained by the similarity calculation between the information is calculated, Analyze the similarity calculation result numerically, find the eigenvector of the maximum eigenvalue of the feature vector, create the query vector of the query content of the search information, and multiply the calculated value of the cosine of the feature vector and the question vector by the value of the eigenvector The search order of the related information required is determined, and the contents of the related information for which the search order has been determined are integrated and output, so information having a high similarity to the information is preferentially selected from the already searched information However, when considering a virtual user who reads with a probability corresponding to the degree of similarity, which information is reached when the steady state is reached after an infinite time. Searching relevant information while judging that corresponding to the probability of whether it is possible to obtain the necessary information for the user.
In addition, the dependency relation of the part of speech included in the sentence of the retrieved information is analyzed, the noun in the phrase related to the phrase including the verb in each sentence is extracted, and the similarity between the nouns of the extracted noun S ₁ and the similarity S _{2 of the} noun set are calculated, the similarity S ₃ of the proportion of the extracted nouns that coincide with each other is calculated, and the sentence similarity S ₃ is calculated by adding the similarity S ₃ to the similarity S _2. Calculate and select the search information according to the sentence type, and integrate and output the contents of the related information selected by the sentence similarity calculation and the sentence type, so the searched related information is duplicated, unique, and supplementary. Related information can be acquired in a state where information contents such as explanations are arranged. In addition, it is possible to reduce the burden on the user due to repeated display of duplicate information included in a huge amount of related information, and to efficiently organize the information on specific parts and supplementary explanations according to the user's request. Can be obtained by format.
Further, the search results scored by determining the search order are disclosed, the suitability of the search results is input, the features and the question vectors are corrected, and the contents of the search information are output. The feature vector and the question vector can be generated so as to approach the information determined to be compatible and away from the information determined to be incompatible. By applying this repeatedly, the search result desired by the user can be obtained with certainty.

If necessary, the information acquiring apparatus according to the present invention, the feature vector generating means includes a morphological analysis unit for performing morphological analysis of a sentence retrieval information, TF calculation to calculate word and the number of occurrences TF of the word for each information Means, an IDF calculation means for calculating the document frequency DF and its IDF for all words, a word weight calculation means for calculating a word weight for each word of each information using the TF-IDF method, and each document from the word weight in which and a feature vector generating means for generating a feature vector.

  Thereby, the feature vector creating means for creating the feature vector of the weighted search information performs the morphological analysis of the sentence of the search information, calculates the word and the appearance frequency TF of the word for each information, and the document frequency for all the words. DF and its IDF are calculated, word weights are calculated for each word of each information using the TF-IDF method, and feature vectors of each document are created from the word weights. To search for information, it can be used as an index for judging the importance of related information, and numerical calculation of the relationship between a document and a word is performed, and a high value is reflected in a feature vector. Thus, related information more important for the user can be acquired.

An information acquisition program according to the present invention includes a feature vector generation unit that generates a feature vector of weighted search information in an information acquisition program that causes a computer to function as an information acquisition device that acquires at least one search information; Information similarity calculation means for calculating the similarity of a combination of information feature vectors, similarity matrix calculation means for calculating a numerical similarity matrix obtained by the information similarity calculation, and the similarity calculation result as a numerical value Eigenvector generating means for analyzing and obtaining the eigenvector of the maximum eigenvalue of the feature vector, question vector generating means for generating a query vector of the inquiry content of the search information, and multiplying the calculated value of the cosine of the feature vector and the question vector by a numerical value of the eigenvector to determine the search order of the relevant information required Te search order decision hand , Dependency analysis means for analyzing the dependency relation of parts of speech included in the sentence of the retrieved information, noun extraction means for extracting a noun in a phrase related to a phrase including a verb in each sentence, and a single unit of the extracted noun noun see matching rate to calculate the similarity S ₁ and noun noun set similarity comparison calculation means for calculating a similarity S ₂ of the set, the similarity S ₃ ratio that matches the display of the extracted noun between noun A calculation means, a sentence similarity calculation means for calculating the sentence similarity S by adding the similarity S ₃ to the similarity S ₂ , a sentence type selection means for selecting by search sentence sentence type, the sentence similarity calculation and the sentence The contents of the related information selected by type and the search results scored by the search order determining means are output, and the suitability of the search results and / or the degree of weighting of the parameters are input. An output means for outputting an input form, and a computer functioning as a correction means for correcting the feature vector and the question vector based on information input to the input form, wherein the search order determination means causes the correction means to The search order of the related information is determined again based on the feature vector and the question vector corrected in this manner, and the output means determines the search results scored by the search order determination means as the time interval between the search information. This is output including general relevance .

In the information acquisition method according to the present invention, a computer of an information acquisition device that has acquired at least one search information generates a feature vector of weighted search information, and a feature vector of all search information A similarity calculation step between information for calculating a similarity of a combination, a similarity matrix calculation step for calculating a similarity matrix of numerical values obtained by the similarity calculation between the information, and a numerical analysis of the similarity calculation result, An eigenvector creating step for obtaining an eigenvector of the maximum eigenvalue of the feature vector, a question vector creating step for creating a question vector of the inquiry content of the search information, and multiplying a calculated value of the cosine of the feature vector and the question vector by a numerical value of the eigenvector a search order determination step of determining a search order of the relevant information sought, the search of A dependency analysis step for analyzing the dependency relationship of parts of speech included in the sentence of the information, a noun extraction step for extracting a noun in the phrase related to the phrase including the verb in each sentence, and a single noun of the extracted noun noun see matching ratio calculating a similarity comparison calculation step between nouns set to calculate the similarity S _2, the similarity S ₃ ratio that matches the display of the extracted noun similarity S ₁ and noun set between nouns A calculation step, a sentence similarity calculation step for calculating the sentence similarity S by adding the similarity S ₃ to the similarity S ₂ , a sentence type selection step for selecting by search sentence type, and the sentence similarity The calculation and the sentence type selected related information content and the search result scored by the search order determination means are output, and the suitability of the search result and / or the parameter Determining the search order by executing an output step of outputting an input form for inputting the degree of finding, and a correcting step of correcting the feature vector and the question vector based on information input to the input form The step determines again the search order of the related information based on the feature vector and the question vector corrected by the correction means, and the output step displays the search results scored by the search rank determination means. , Including the temporal relationship between the search information .

  Here, the present invention can be implemented in many different forms. Therefore, it should not be interpreted only by the description of the following embodiment. In the embodiment, the apparatus will be mainly described. However, as is apparent to those skilled in the art, the present invention can also be implemented as a program usable on a computer. Further, the present invention can be implemented in hardware, software, or software and hardware embodiments. The program can be recorded on any computer-readable medium such as a hard disk, CD-ROM, DVD-ROM, optical storage device, or magnetic storage device. Furthermore, the program can be recorded on another computer via a network.

[1. Hardware configuration]
FIG. 1 shows a hardware configuration diagram of an information acquisition apparatus according to an embodiment of the present invention. The computer 1 includes, for example, a CPU (Central Processing Unit) 2, a main memory 3, a HDD (Hard Disk Drive) 4, a video card 5, a mouse 6, a keyboard 7, an optical disk 8, and the like. A database or the like can be connected as necessary.

[2. Block configuration]
FIG. 2 shows a block diagram of the information acquisition apparatus according to the embodiment of the present invention. The present invention mainly includes an input unit 10, a follow-up information search unit 20, an information content integration unit 30, and an output unit 40. Here, the follow-up information search unit 20 includes a vector creation unit 21, a search order determination unit 22, and a conformance / nonconformity determination unit 23. Furthermore, the vector creation unit 21 includes a morpheme analysis unit 211, a TF calculation unit 212, an IDF calculation unit 213, a word weight calculation unit 214, a feature vector creation unit 215, an inter-information similarity calculation unit 216, a similarity matrix calculation unit 217, An eigenvector creation unit 218 and a question vector creation unit 219 are included. The information content integration unit 30 includes a dependency analysis unit 31, a noun extraction unit 32, a noun set similarity comparison calculation unit 33, a noun display match ratio calculation unit 34, a sentence similarity calculation unit 35, and a sentence type selection unit 36. A summary sentence creation unit 37 and an article aggregation unit 38.
First, the article data input by the input unit 10 is sent to the vector creation unit 21 in the follow-up information search unit 20 for processing. Here, as an example of an embodiment of the present invention, a newspaper article is taken as an example of a search target, and the contents of each component will be described in detail below.

[2.1 Follow-up information search]
[2.1.1 Vector space model]
In the vector creation unit 21, article data is first processed by the morphological analysis unit 211, the TF calculation unit 212, the IDF calculation unit 213, the word weight calculation unit 214, and the feature vector creation unit 215. Here, a vector space model (vector-space model) associates a multidimensional vector with individual data as a feature quantity for expressing the properties of individual data to be searched. This is a method of searching for something similar to an inquiry (question) by defining similarity in the meantime. Now, it is equipped with n-number of attributes as a feature of the search target, the i-th attribute and w _i. Consider that the vector of (Formula 1) corresponds to the jth data. If these vectors are linearly independent, an n-dimensional vector space is defined. In the vector space defined in this way, the feature vector of the jth data is

のように表すことができる（ｄ_jiはのｗ_iに対する値）。
ベクトル空間モデルにおける検索システムへの問い合わせ（質問）もベクトルで表される。ｎ次元のベクトル空間に対するその質問ベクトルは

(D _ji is a value for w _i of).
Inquiries (questions) to the search system in the vector space model are also expressed as vectors. The query vector for an n-dimensional vector space is

のように表すことができる（ｑ_iは質問ベクトルのｗ_iに対する値）。

(Q _i is a value for w _i of the query vector).

  The search is performed by calculating the similarity between the feature vector of (Expression 1) to be searched and the query vector of (Expression 2). The method of obtaining the feature vector differs depending on the purpose of the search and the type of data that is the object. For example, if the search target is document data, each element of the vector is weighted based on the appearance frequency of the word, and if the search target is an image, the density and color data for each pixel can be used. In newspaper articles, one article is assigned to one vector, and the TF-IDF of a word in the article is used for weighting the vector.

[2.1.2 Word weighting]
In order to search for a target document in the information search, a numerical calculation of the relationship between the document and the word is performed, and a document having a high value is set as a candidate. The evaluation value used there indicates how many important words are included in the document. The TF-IDF method described below is used to weight how important words in a document are. This method pays attention to the property of the following two keywords.
(1) Words that appear frequently and frequently in a document are important (2) Words that appear only in a small number of documents are important Word Frequency (TF) Word t appears frequently in document d If it appears, t characterizes d well. A scale based on this idea is the word appearance frequency, tf (Term Frequency). The appearance frequency tf (d, t) of the word t in a document d is defined by the following equation, and is calculated by the TF calculation unit 212.

ｆｒｅｑ（ｄ，ｔ）：文書ｄにおける単語ｔの出現頻度。

freq (d, t): Appearance frequency of the word t in the document d.

  It is an important property that the document frequency (df) tf is large, but it cannot sufficiently characterize the document. For example, in a Japanese document, the particle "ha" appears frequently in any document, but it is clear that it does not characterize a particular document. Therefore, the property that the word t appears only in a small number of documents in the document set to be searched is important. The number of documents in which the word t appears and the document appearance frequency df (document frequency) are defined by the following expression.

ｄｆｒｅｑ（ｔ）： 単語ｔが出現する文書数
ｄｆが小さいことが単語ｔの文書を特徴付ける能力が高いことを表すので、実際にはこの逆数をｌｏｇと文書集合中の文書総数Ｎにより正規化したｉｄｆ（ｉｎｖｅｒｓｅ ｄｏｃｕｍｅｎｔ ｆｒｅｑｕｅｎｃｙ）を用いる。

dfreq (t): number of documents in which the word t appears Since a small df indicates that the ability to characterize the document of the word t is high, this reciprocal is actually normalized by log and the total number N of documents in the document set. idf (inverse document frequency) is used.

Ｎ：文書の数
ｆｒｅｑ（ｄ，ｔ）：文書ｄにおける単語ｔの出現頻度

N: number of documents freq (d, t): appearance frequency of word t in document d

  The IDF calculation unit 213 first calculates idf after obtaining df. For the weighted word t by TF-IDF, by calculating the product of the number of times the word appears in the document and the proportion of the word in the entire document, the importance of the word is determined according to the importance of the word and its appearance frequency. The purpose is to express sex. If the word t has the properties of both tf and idf, that is, if tf is large and df is small, it can be said that the word t truly characterizes the document d. This concept is expressed as a numerical scale by weighting by TF-IDF. The weight w (t, d) of the keyword t in the document d is defined as follows, and is calculated by the word weight calculation unit 214.

ｄｆｒｅｑ（ｔ）： 単語ｔが出現する文書数
そして、特徴ベクトル作成部２１５において、これらの求められた数値を利用して特徴ベクトルを作成する。

dfreq (t): The number of documents in which the word t appears, and the feature vector creation unit 215 creates a feature vector using these obtained numerical values.

[2.1.3 Similarity]
The data processed by the feature vector creation unit 215 and the question vector creation unit 219 is sent to the search order determination unit 22. Here, in the vector space model, similarity between vectors must be defined in order to perform a search. There are various measures of similarity, but here the cosine between vectors is used.
In this method, a cosine value between two vectors is used as the similarity. The similarity sim (D, Q) between the feature vector D and the question vector Q is as follows.

ｓｉｍ（Ｄ，Ｑ）の値は０以上１以下であり、１に近づくほど類似度が高くなる。検索順位決定部２２では、余弦、いわゆるコサイン相関値を用いた類似度評価を行う。

The value of sim (D, Q) is 0 or more and 1 or less, and the closer to 1, the higher the similarity. The search order determination unit 22 performs similarity evaluation using a cosine, so-called cosine correlation value.

[2.1.4 Relevance feedback in vector space model]
When a search result is obtained in the search order determination unit 22, it is processed by the output unit 40. Rather than obtaining the final result in a single search, a new question is generated based on user feedback on the result, a repeated search is performed, and the search result is gradually brought closer to the result desired by the user. Do. That is, the suitability data for the search result of the user is input again at the input unit 10.

  In the relevance feedback in the vector space model, a question formula for the apparatus is also expressed by a multidimensional vector created by the question vector creation unit 219. The result of the question is obtained as a set of data obtained as a result of calculating the similarity between the question vector created by the question vector creation unit 219 and the feature vector of the data created by the feature vector creation unit 215. The data with higher similarity is considered to be more appropriate as an answer to the question, and the data included in the search result is presented to the user with a ranking of how much it matches the question. The user returns feedback from the presented search results.

[2.1.5 Feedback from users]
The suitability data of the input search result is sent to the conformity / nonconformity determination unit 23 and sent to the question vector creation unit 219 and the word weight calculation unit 214.
As feedback from a specific user, the most common is the presentation of positive examples and negative examples of results. In addition, there are various types such as those that feed back only positive examples and those that rank and rank the degree of validity. In addition, the method of reflecting the feedback from the user in the search can be roughly divided into the following two.
(1) Question Vector Correction (Query Vector Movement) corrects and transforms the vector of a search question to bring it closer to the features of the positive example and away from the negative example.
(2) Reweighting (Feature Re-weighting) adjusts the weight to correspond to the feature vector according to the user's needs. That is, weights are dynamically changed to emphasize the dimensions that are convenient for searching positive examples and reduce the impact of negative examples. The present invention utilizes both of these feedbacks.

[2.1.6 Modification of question vector]
As a method to bring the search results closer to what the user wants based on the results of user feedback, the question vector is generated so that the question vector is closer to the article that the user has determined to be relevant and away from the article that is determined to be non-conforming. I will do it. By applying this repeatedly, it is possible to gradually obtain a search result desired by the user. For this purpose, the Rocchio feedback method is often used, and the Rocchio equation is given as follows.

Q _i is a question vector used in the previous search, and Q _{i + 1} is a newly generated question vector. R is a feature vector for the document D _j determined to be conforming, and N is a feature vector for the document determined to be incompatible. R _n and N _n are the number of conforming documents and the number of nonconforming documents, respectively. α and β are variables for conforming documents and nonconforming documents, respectively. If the value of α is high, the change by the conforming document is emphasized, and if the value of β is high, the change by the nonconforming document is regarded as important. As a result of the relevance feedback, query position can be considered as moving from Q _i to Q _{i + 1.} Here, the data in the feature vector creation unit 215 is also processed in the question vector creation unit.

[2.1.7 Importance evaluation taking into account state transition probabilities]
The inter-information similarity calculation unit 216, the similarity matrix calculation unit 217, and the eigenvector creation unit 218 process the data from the feature vector creation unit 215 with the following contents.
FIG. 3 is an explanatory diagram of article importance evaluation based on similarity between articles of the information acquisition apparatus according to the embodiment of the present invention. PageRank is a Web page group linked by hyperlinks on www, based on the recursive relationship that “the pages linked from many high-quality pages are still high-quality pages”. This is a theory for evaluating the importance of a page, and the importance of the page obtained thereby. By using PageRank, it is possible to quantitatively determine which page is most important when there is a cross-reference relationship such as a hyperlink structure.

  FIG. 3A shows a conceptual diagram of PageRank. The basic method for calculating PageRank will be described using this figure as an example. First, all Web pages each have a value of PageRank. This value is evenly distributed to the previous page to which the page is linked. Taking FIG. 3A as an example, a page having a value of 100 in the figure has a link to two pages, so the value of 100 that this page has is divided into two links. Given to you. That is, each linked page gets 50 values.

[2.1.8 Article importance evaluation based on similarity between articles]
Whereas PageRank handles the weight of links between pages equally, in the present invention, weighting is performed based on the degree of similarity to each article. In this way, the newspaper article group is considered as a graph structure connected by links according to the strength of the relevance. Among them, an article linked as having higher relevance is obtained by calculating an eigenvector at the maximum eigenvalue of the transition probability similarly to PageRank. FIG. 3 (b) shows the degree of similarity between newspaper articles, and the calculation method will be described below.
First, when the number of articles is N, the inter-information similarity calculation unit 216 calculates the N × N N-order square matrix and the similarity between the articles as elements, and the similarity matrix calculation unit 217 calculates the similarity. Create a matrix. A matrix A as a result of obtaining a similarity matrix with respect to FIG.

次に、各記事、すなわち各列について合計が１になるように正規化し行列Ａを状態遷移確率行列Ｍとする。図３（Ｃ）は、新聞記事間の類似度による重み付けを行った遷移確率を示す。このときの記事間の関係は図３（Ｃ）のように示される。

Next, normalization is performed so that the sum is 1 for each article, that is, each column, and the matrix A is set as a state transition probability matrix M. FIG. 3C shows transition probabilities that are weighted according to the similarity between newspaper articles. The relationship between the articles at this time is shown as in FIG.

行列Ｍの状態遷移確率行列から、固有ベクトル作成部２１８が、最大固有値の固有ベクトルを計算した結果を図４に示す。

FIG. 4 shows the result of the eigenvector creation unit 218 calculating the eigenvector of the maximum eigenvalue from the state transition probability matrix of the matrix M.

  FIG. 4 is an example of calculating the importance of articles in the information acquisition apparatus according to the embodiment of the present invention. FIG. 4 shows that an article that is referenced by a higher weight from more articles has a higher score. The score of FIG. 4 is infinite when considering a hypothetical user who preferentially selects an article having a high similarity to the article from the currently viewed article and proceeds with reading with a probability corresponding to the similarity. This corresponds to the probability of which article will be reached when a steady state is reached after a period of time. In other words, it is a score that the user is interested in the article as being similar and easy to reach, and links with high weights as to which articles in the article group have similarity from many articles It can also be said that the score indicates whether or not

  By using the score obtained as described above together with the score calculated for the similarity on the vector space model, the user's interest is taken into account by the similarity to the question vector, and as a result, a representative article is included as a result. Can show. As a result, the user obtains an article suitable for use in the determination of the search result, and tries to perform a feedback search efficiently.

[2.2 Information content integration system]
The information data of the follow-up articles obtained by the follow-up information search unit 20 is subjected to information data sorting / classification processing in the information content integration unit 30. The determination based on the overlapping part, the unique part, each category of supplementary explanation, the similarity for performing categorization, and the sentence type will be described below.

[2.2.1 Category setting and classification]
In content integration, multiple newspaper articles are classified into sentence categories, and by combining them, aim to output in a format that meets the needs of users. Therefore, each category at that time must be in line with the advantages of expanding the user's choice of information acquisition and comparing multiple newspaper articles. Therefore, a total of three categories are set: a duplicate location that is common to each article, a unique location that is unique to each article, and a supplementary explanation that is supplementary content in the article. Note that, as a definition of correspondence between sentences in overlapping portions, the topic of the sentence to be compared is completely or partially included in one sentence.
In order to classify the target article into three categories of “overlapping part”, “unique part”, and “supplementary explanation”, two of “similarity in sentence units” and “sentence type” are used as judgment criteria.

With “similarity in sentence units”, the dependency analysis unit 31 analyzes the article data. The noun extraction unit 32 and the noun set similarity comparison calculation unit 33 use the distance between the nouns in the noun set with the verb in each sentence as a key and the notation, and the noun display match The sum of the proportions of nouns having the same notation in the noun word using those results obtained by the proportion calculator 34 is set as the value of the result calculated by the sentence similarity calculator 35.
In “sentence type”, the sentence type selection unit 36 performs sentence typing on each sentence. In the following, detailed description regarding each determination will be described.

[2.2.2 Sentence similarity between multiple newspaper articles]
Sentence similarity is used as a standard for selecting overlapping and unique locations. A method for calculating the similarity will be described below. In general, syntactic similarity and semantic similarity can be considered as an index of sentence similarity. In the similar sentence search, “dependency on a verb” is used to obtain the similarity of the syntax structure. In addition, in order to obtain a semantic similarity, “semantic attributes of nouns in phrases directly related to verbs” and “ratio of noun notation matching” are used. The search for similar sentences is performed in the following four steps.
(1) Detection of nouns in phrases related to phrases containing verbs (2) Comparison of similarities between noun sets using the verb extracted in (1) as a key (3) Noun notation using the results of (2) Ratio of matching (4) Similarity calculation using results of (2) and (3)

[2.2.2 Extraction of nouns in phrases related to phrases containing verbs]
FIG. 5 is an example of extraction of nouns in a phrase related to a phrase including a verb in the information acquisition apparatus according to the embodiment of the present invention. In the case of FIG. 5, the nouns A, B, C, D, G, H, I, 7, 8, and 9 in the shaded portion are extracted. Such processing uses a dependency on a verb included in the sentence, and can be performed using, for example, a Japanese dependency analyzer cabocha.

[2.2.3 Comparison using noun conceptual relationships in clauses involving verbs]
In order to put partial duplication into the similarity information between sentences by the Japanese morphological analyzer cabocha, the nouns in the clauses related to the clauses containing the verbs in each sentence are extracted, and a noun set for each verb is created. . In this case, the similarity between each noun in the set is obtained from the distance between the concepts by using an EDR electronic dictionary for those having different notations, and the value is maximized for those having the same notation. And the similarity of noun sets is calculated from the similarity of each noun single, and takes the value with the highest similarity among them. As a method of measuring the similarity of nouns, a method of obtaining the similarity from the position of the common parent attribute on the semantic attribute system and the length of the path between the two semantic attributes can be considered. In general, however, nouns can be assigned multiple semantic attributes. Therefore, in order to obtain the similarity of nouns, it is necessary to uniquely determine which semantic attribute is used as a noun from context information. In the present invention, without going into the problem of ambiguity, “word similarity calculation method using EDR electronic dictionary” [Reference: Tsuji et al., Information Processing Society of Japan Report NL-93-1, pp1-6. ], The noun similarity is comprehensively obtained from a plurality of semantic attributes assigned to the noun, which is a method proposed in the above. The reason why the nouns in the clauses related to verbs are particularly taken as an index of similarity is because of the difference in the number of topics included in one sentence.

[2.2.4 EDR Electronic Dictionary]
FIG. 6 is a structural diagram of an EDR dictionary of the information acquisition apparatus according to the embodiment of the present invention. The EDR electronic dictionary is developed for advanced language processing by a computer and is composed of several large-scale individual dictionaries such as a word dictionary. A dictionary is suitable to be called a Japanese and English vocabulary knowledge catalog that integrates a conceptual system (thesaurus) describing the synonyms of concepts defined in a word dictionary and a corpus database (collection of example sentences) as a dictionary description authority. This is an electronic dictionary for machine processing. The development policy is to eliminate the linguistic bias as much as possible and maintain the ease of tuning up to various applications. The EDR electronic dictionary is composed of a word dictionary, a bilingual dictionary, a concept dictionary, a co-occurrence dictionary, a technical term dictionary, and an EDR corpus.

  In the present invention, a concept dictionary and a Japanese word dictionary are used to examine the distance between noun concepts. The Japanese word dictionary is a word dictionary having a vocabulary of about 260,000 words. The basic role is to describe the correspondence between words and concepts (meaning), and to give grammatical characteristics when this correspondence holds. The concept dictionary describes knowledge about 410,000 concepts introduced as word meanings in the word dictionary, and is divided into a concept system dictionary and a concept description dictionary depending on the type of information. The concept system dictionary describes the upper and lower relations between about 410,000 concepts. A superordinate relationship is an inclusive relationship between concepts, and can be regarded as a kind of thesaurus. The concept description dictionary is a dictionary that describes the semantic relationships (actors, tools, places, etc.) between concepts that co-occur in a sentence (2 items).

[2.2.5 Process for calculating similarity between noun sets]
Processing for calculating the similarity between noun sets will be described below. Nouns are compared, and those with the same notation are calculated with a similarity of 1 as the maximum value. If there are other different notations, a comparison using a concept dictionary is performed. The procedure for comparing nouns using the concept dictionary is shown below. First, a concept identifier representing the concept of a noun is extracted from the Japanese word dictionary, and a semantic attribute list is obtained from the concept dictionary by using it. Next, based on the semantic attributes of both nouns, the relationship between the nouns is determined by “recognizing optional case elements by comparing similar sentences” [Reference: Shinohara et al., Information Processing Society of Japan, NL-139-14, pp101-108. ] Are classified into synonymous relations and similar relations.

FIG. 7 is a synonym / similarity diagram between nouns of the information acquisition apparatus according to the embodiment of the present invention. FIG. 7A shows the synonymous relationship. FIG. 7B shows a similar relationship. Based on these two relationships, the similarity between nouns using different notations using the concept relationship is obtained. The similarity a of the synonymous relationship and the similarity b of the similar relationship are respectively obtained by the following equations. For each formula, Shinohara et al.'S formula for conceptual relationships between nouns is adopted. In addition, Shinohara et al. Define the relationship that the concept identifier of the same relationship is the same, but since the concept is defined in detail in the EDR electronic dictionary, the same relationship is adopted. Absent.
Similarity a

Ａ_n ： 名詞ｎの意味属性数（ｎ＝１，２）
Ｄ_a ： 重複する意味属性数
類似関係の類似度b

A _n : number of semantic attributes of noun n (n = 1, 2)
_Da : Similarity b of the number of overlapping semantic attribute similarity relations b

Ｎ_n ： 名詞ｎの意味属性数
Ｎ_ni： 名詞ｎの意味属性ｉの上位概念数
Ｄ_ij：意味属性ｉ，ｊの上位概念の重複数
求めた類似度ａ、ｂを使用し、次式により概念間の距離による名詞同士の類似度Ｓ₁を求める。
概念間の距離による名詞同士の類似度Ｓ₁

N _n : Number of semantic attributes of noun n N _ni : Number of superordinate concepts of semantic attribute i of noun n D _ij : Duplicate superordinate concepts of semantic attributes i and j Using the obtained similarities a and b, A similarity S ₁ between nouns is calculated based on the distance between concepts.
Similarity between nouns based on the distance between concepts S ₁

以上より、単体の名詞間の類似度を求める。そして、以下にそれらを利用した動詞をキーとした名詞集合間の類似度の算出方法を述べる。

From the above, the similarity between single nouns is obtained. A method for calculating the similarity between noun sets using verbs as keys is described below.

FIG. 8 is an example of similarity calculation between noun sets of the information acquisition apparatus according to the embodiment of the present invention. From the dependency information obtained by cabocha, the noun phrase in the clause related to the clause containing the verb is extracted as a set using the verb as a key. When calculating the similarity between the noun set 1 and the noun set 3 in FIG. 8, when the article 1 is considered to be the subject, the one with the higher similarity between the noun A and the nouns F and G is assigned to the noun A 1. The nouns are similar, and here they are nouns F. Similarly, the noun B also has a higher similarity between the nouns F and G and is a noun similar to the noun B. When the number of nouns on the subject side is n, the similarity between the nouns A and F is S _AF , and the similarity between the noun and the noun G is S _BG , the similarity between the noun sets is S _{2. 2} is as follows.

同様に、名詞集合２と名詞集合３を比較し、集合間の類似度を求める。そこで名詞集合１と名詞集合３、名詞集合２と名詞集合３の類似度をそれぞれ比較し、値が高い方を動詞をキーとする名詞集合間の類似度とする。ここでは名詞集合１と名詞集合３の類似度Ｓ₂ とする。

Similarly, the noun set 2 and the noun set 3 are compared to determine the similarity between the sets. Therefore, the similarity between the noun set 1 and the noun set 3 and the noun set 2 and the noun set 3 are compared, and the higher one is set as the similarity between the noun sets using the verb as a key. Here, the similarity S ₂ between the noun set 1 and the noun set 3 is assumed.

[2.2.6 Matching nouns]
In the previous process, using the distance between noun concepts, a noun set was selected for each sentence using the verb with the highest similarity as a key. Here, the ratio of nouns with the same notation in the noun word in the sentence other than the noun set selected there is calculated. The similarity S ₃ is shown below.
S ₃ by percentage of notation

Ｄ_ij：文ｉ と文ｊ の動詞に係る文節以外の部分の名詞の内の一致した数
Ａ_i： 文ｉ中の動詞に係る文節以外の部分の名詞の数
Ａ_j：文ｊ中の動詞に係る文節以外の部分の名詞の数

D _ij : Number of matched nouns in the part other than the clause related to the verbs in sentence i and sentence j A _i : Number of nouns in the part other than the phrase related to the verb in sentence i A _j : Verb in sentence j Number of nouns in parts other than clauses related to

[2.2.7 Calculation of similarity]
The sentence similarity S is as follows according to S ₂ and S ₃ described above.
Sentence similarity S

[2.2.8 Selection by sentence type]
Proposed a new method for selecting duplicated sentences, specific sentences, and supplementary explanations based on the sentence type that considers the characteristics of newspaper articles for each sentence as a method for selecting duplicated parts using the characteristics of newspaper articles. To do. With this method, it is possible to newly adopt words and expressions specific to newspaper articles as selection indexes.

[2.2.9 Types of sentence types]
In addition, each sentence type has five sentence types, including summary, schedule, reason, analysis, and supplementary explanation, from the characteristics of newspaper articles, among many sentence types defined in the conventional summary processing, “Synthance and Meaning 2” [Reference: Hideo Teramura: Kuroshio Publishing] is an example of the implementation of the present invention based on the general mood and the above data analysis. . The following is a description of the five sentence types of summary, schedule, reason, analysis, and supplementary explanation, as well as the features and judgment criteria of the two sentence types: mode / hearing, ratio / inference.

[2.2.10 sentence type: summary, schedule, reason, analysis, supplementary explanation]
We will describe the criteria and characteristics of the five sentence types: abstract, schedule, reason, analysis, and supplementary explanation. In addition, regarding the judgment criteria, the assertive expression and the expression representing the date and time are # [dantei] and # [nitizi? ] (? = 1 or 2 or 3). The assertive expression is set as follows as a result of analyzing experimental data articles. It should be noted that the notation format of the following criteria is in accordance with the expression format of the regular regular expression.

また、日時を表す表現については、時間を表す部分（ｎｉｔｉｚｉ１）と季節や日付を表す語句（ｎｉｔｉｚｉ２）、そしてそれらに付随する語（ｎｉｔｉｚｉ３）に大きく分けて設定する。

The expression representing the date and time is roughly divided into a time part (nitizi1), a word and phrase (nitizi2) representing the season and date, and a word (nitizi3) associated therewith.

Of these, the part (nitizi1) representing time and the word (nitizi2) representing season and date are set based on 200 articles of experimental data as generally considered. In addition, the words (nitizi3) accompanying them are described in the analysis of experimental data and “Basic techniques of natural language processing” [reference: Hiromura Nomura, The Institute of Electronics, Information and Communication Engineers, pp246]. This is set with reference to the space-time relational group in the 7 case particles.
The abstract is a sentence that is thought to contain the entire summary as the first sentence in a newspaper article. Judgment criteria are the first sentence in the article as the summary sentence type.

  The schedule is a sentence type attached when the sentence represents the schedule of the event, such as the date and time of the event that the sentence will be performed. Judgment criteria are as follows based on the result of manually classifying the above training data.

理由はその文が理由を述べている場合に付けられる文タイプである。判断基準は前述のトレーニングデータを人手で分類した結果から以下のように作成する。

The reason is the sentence type that is given when the sentence states the reason. Judgment criteria are created as follows from the result of manually classifying the above training data.

分析はその文が記者の観点からみた出来事に対する分析、意見の場合につけられる文タイプである。判断基準は前述のトレーニングデータを人手で分類した結果から以下のように作成する。

Analysis is a sentence type that is attached to the case where the sentence is analyzed from the viewpoint of the reporter. Judgment criteria are created as follows from the result of manually classifying the above training data.

The supplementary explanation is a sentence type attached when the sentence is a supplementary explanation of the article. As a result of analyzing the aforementioned training data article, there are the following types of catching explanations.
(1) Introduction of the characters' characteristics in the article (2) Talk of experts and related parties about the events in the article (3) Additive explanation of the article content (technical terms, etc.)
In the supplementary explanation, we focused on the special symbols used in the article. Therefore, the sentence type is set with the portion after the symbol appears as the catching explanation. Special symbols used for determining the supplementary explanation are “◯”, “<”, “>”, “*”, “◇”, “◆”.

[2.2.11 Sentence Type: Concerning Mode / Hearing, Ratio / Inference]
The reason for adoption and criteria for the two sentence types described below, such as mode / hearing and ratio / inference, will be described. The state / hearing is a sentence type that combines the state and hearing. A state is a sentence type that conveys the existence of objects and the state of behavior. In addition, hearing is a type of sentence that is heard and heard from a person, not directly. Judgment criteria are shown below.

比況・推量は比況と推量が合わさった文タイプである。比況とは動作・状態などをほかのものにたとえて表すような文タイプである。また推量とはある根拠・理由や、確かな論理的要請などに基づいて、込み入った事情や人の心の中などをおしはかっているような文タイプである。判断基準を以下に示す。

Ratio / estimation is a sentence type that combines ratio and estimation. The ratio is a sentence type that expresses actions / states to other things. Inference is a type of sentence that makes use of complicated circumstances and people's minds based on certain grounds / reasons and certain logical requests. Judgment criteria are shown below.

  These two sentence types are excerpted from the general mood in the secondary mood auxiliary verbs described in "Japanese Synthance and Meaning 2" above. According to the above-mentioned “Japanese Synthance and Meaning 2”, the syntactic element of “mood” is the part that expresses the attitude and attitude of the speaker to bring out the thing in front of the opponent in various realities. . In the above description of “Japanese Synthance and Meaning 2”, the mood of the positive statement, the mood of the general statement, and the mood of the explanation are given as the mood. Here, the sentence type is the state / hearing, ratio / inference in the mood of the outline that is considered to be effective in expressing various subjectivity of the speaker as the sentence type of the newspaper article. Aspects, hearings, ratios, and guesses are individually listed alongside other forms, such as Nai, (La) Lel, (Sa) Cell, Thailand, Ta, etc. Has been described in its usage. That is, "Lashii" attached to the verb conjunctive form and adjective stem represents "(foundation) guess" and "Yoda" represents "hearing". Depending on the grammar book, “Yoda” or “Soda” of hearsay shall have the formal word “da” attached to the formal word, and may not be recognized as an auxiliary verb. There is also a view that “soda” in the form is suffixed with “da”. Teramura treats these in an integrated manner as sharing certain syntactic features and certain (greatest common divisor) meanings. A few forms deviate from the syntactic features, but they are included in this view in terms of semantic features. In addition, “Kamoshirenai”, “Kamowa Karanai”, “Nichiganai”, “Toiu”, etc., which are not treated as auxiliary verbs in ordinary grammar books, are connected with the ones that were originally particles, verbs and auxiliary verbs. From the above syntactic and semantic features, it is also included in the outline auxiliary verb. For these reasons, the general mood, aspect, hearing, ratio and guess are examples of sentence types.

[2.2.12 Sentence type application priority]
Although seven kinds of sentence types are set as described above, there are many sentences that also serve as a plurality of sentence types depending on the sentence. At that time, it is necessary to take into account the priority order for setting the sentence type. Therefore, the result of setting a sentence type for each sentence by making a prototype of an apparatus in which the sentence type priority order is not implemented for the above-described experimental data 200 articles is shown. FIG. 9 is a ratio for each sentence type of the information acquisition apparatus according to the embodiment of the present invention. Of these, since the abstract is a summary of the entire article, it is considered that about one sentence is set for each sentence, so the priority is the highest. In addition, the supplementary explanation also represents the supplementary explanation. Prioritize the other sentence types (schedule, reason, analysis, mode / hearing, ratio / inference) from those with the fewest occurrences. As a result, the priority order is as follows.
"Summary = Supplementary explanation> Ratio and guessing>Reason>Analysis>Schedule> Mode / hearing"

[2.2.13 Duplicate sentence, unique sentence, and supplementary explanation output by sentence type]
FIG. 10 is a categorization of overlapping parts, unique parts, and supplementary explanations of the information acquisition apparatus according to the embodiment of the present invention. This is the flow up to the output of duplicated parts, unique parts, and supplementary explanations by sentence type. When there is no same sentence type as shown in FIG. 10, duplicate sentence selection using the similarity calculated in the previous step is performed. In addition, for the summary and supplementary explanation, the target sentence type is extracted in advance from the properties described above.

[2.2.14 Aggregation between multiple articles]
The data sent from the sentence type selection unit 36 is sent to the summary sentence creation unit 37 or the article aggregation unit 38.

[2.2.14.1 Aggregation between two articles]
Aggregation between two articles is an aggregation of two articles, which is a combination of the unique part of the second article and the duplicate part and unique part of the first article. As a result, it is possible to aggregate items that overlap between two articles or supplemental items.

[2.2.14.2 Aggregation between three articles]
When there are three articles, three aggregates of two articles are created. Assume that three articles are article A, article B, and article C. Since the follow-up articles are classified, the first two articles A and B viewed in chronological order are basically collected. This aggregation and articles B and C are aggregated one by one, and sentences not included are added as aggregation. Thereby, the contents about the articles A, B, and C can be aggregated. In the aggregation of articles A and C, if an article that is determined to be an overlapping part of C is included in the aggregation, it is deleted to delete it, thereby creating a new aggregation. Of course, by applying this method, three or more articles can be aggregated at once.

[2.2.14.3 Aggregation taking into account the degree of association between articles]
In the above method, aggregation is performed taking into consideration the degree of association between articles. The degree of association between articles is obtained by representing each article as one vector and comparing the vectors. Although the degree of relevance between articles is obtained when searching, since it is a follow-up article, the first article and the second article, the second article and the third article are connected, but the first article The connection between the article and the third article may be weak. When the degree of association between articles is low, the comparison of articles A and C is not performed, so that unnecessary processing is not required.

[3. Operation]
While viewing the results, the user changes the importance of the article by selecting whether the article is relevant or not, or by modifying parameters. The system gets it as feedback, corrects the question vector, ranks it, and returns the search results to the user again.
FIG. 11 is a follow-up article search flow sheet of the information acquisition apparatus according to the embodiment of the present invention. Data is input (S100). The morphological analysis unit 211 performs part-of-speech decomposition by morphological analysis, and extracts only nouns (normal, unique, and strange) (S110). Here, a morphological analyzer teacup can also be used. The TF calculation unit 212 registers the word and the number of appearances (tf) of the word for each article in the article database, and calculates the TF (S120). The IDF calculation unit 213 calculates the document frequency (df) and its IDF for all words registered in the article database (S130). The word weight calculation unit 214 calculates an evaluation value for each word of each article using the TF-IDF method, and calculates the word weight (S140). The feature vector creation unit 215 creates a feature vector of each document from the evaluation value (S150). That is, each article is a vector having TF-IDF values of all nouns appearing in the article as elements. The inter-information similarity calculation unit 216 calculates the similarity for all article vector combinations (S160). The similarity matrix calculation unit 217 calculates a similarity matrix (S170). The eigenvector creation unit 218 passes the result to the numerical analysis program Octave, and calculates the eigenvector of the maximum eigenvalue of the vector to calculate the reference importance of each article (S180). The question vector creation unit 219 creates a question vector (S190). The search order determination unit 22 determines the search order by calculating the similarity for the search result (S200). Here, the score of an article D _i is obtained by multiplying the score sim (D _i , Q) based on the similarity with the question vector Q from the user and the importance obtained from the eigenvector.

μ_i：固有ベクトルにより求められた重要度

μ _i : Importance obtained from eigenvector

  The output unit 40 discloses the result of the scoring by the ranking determination unit to the user (S210). The user inputs appropriateness of the search result to feed back the search result (S220). The conformance / nonconformity determination unit 23 determines whether or not the search result is conformity (S230). If it is determined that the search result is not suitable, the feature and the question vector are corrected (S240). When it is determined that the search is appropriate, the information contents are integrated (S250). Information content integration will be explained below.

FIG. 12 is a flow sheet (2) of article content integration of the information acquisition apparatus according to the embodiment of the present invention. The searched article is input (S251). The dependency analysis unit 31 performs dependency analysis in order to use the dependency related to the verb included in the sentence (S252). The noun extraction unit 32 extracts a noun in the phrase related to the phrase including the verb in each sentence (S253). The noun set similarity comparison calculation unit 33 calculates a similarity S ₁ between single nouns (S254). Also, nouns collection Similarity comparison calculation unit 33 calculates the similarity S ₂ between noun set (S255). The noun display match ratio calculation unit 34 calculates the similarity S ₃ of the noun display match ratio (S256). The sentence similarity calculation unit 35 calculates the sentence similarity S from the similarity S ₂ and the similarity S ₃ (S257). The sentence type selection unit 36 selects a sentence type (S258). The summary sentence creation unit 37 summarizes the article content (S259). The article aggregating unit 38 aggregates article contents (S260). Here, article contents can be aggregated regardless of whether they are summarized. The output unit 40 outputs a content integrated article (S261). The search results, content integrated articles, and specific contents of the output will be explained below.

[4. Visualization method]
The information acquisition apparatus according to the embodiment of the present invention arranges information of follow-up articles so that the user can easily see using a function for classifying information, and displays an interface corresponding to the function. The details will be described below.
FIG. 13 is an execution diagram of the information acquisition apparatus according to the embodiment of the present invention. Information visualization is indispensable as a related technique for promoting interaction in information retrieval. Visualization not only allows the data presented by the system to be efficiently communicated to the user, but also allows the user more flexible access to the data. A is a related article title display, B is an article content display, C is a radar chart, D is various commands, and E is an association display between articles.

  FIG. 14 is an interface for search expression expansion / question vector selection of the information acquisition apparatus according to the embodiment of the present invention. FIG. 14A is a title display of the related article in FIG. 13A and shows an interface for search expression expansion. In the search expression expansion, the search expression is expanded based on the result of feedback whether the user is a conforming document or a non-conforming document with respect to the result of the retrieval once, and the retrieval is performed again. Therefore, feedback can be performed on the result of the user search. It is possible to select the conformity / non-conformance of each article so that input of conformance / non-conformity is accepted on the screen displaying the search result list. As shown in FIG. 14A, a list of article titles as search results is displayed on the user screen as document search results. When the user presses the button, the text is displayed in the text display area in FIG. 13B, and it is possible to select conformity / non-conformity while confirming.

  FIG. 14B shows an interface for question vector selection, which is the radar chart of FIG. The question vector is selected by showing the similarity between the question vector and the original article vector using a radar chart. It is arranged so that the similarity to the question vector is lower as it goes to the center of the radar chart, and the similarity is higher as it goes outward. By displaying in this way, articles that diverge the question vector appear concave. The query vector is refined by removing such articles from the question vector or by dragging the vertex of the radar chart with the mouse to correct the weight for the article. The relational calculation of the relation and the degree of relation is performed by comprehensively determining the degree of similarity in TF / IDF, vector space, the degree of similarity by a statistical method, the degree of importance of an article by PageRank, and the like. The radar chart is used to control each related calculation. Due to screen size limitations, only two radar charts are displayed. These radar charts can be replaced with any of the above-described determination calculations. Also, the above four can be displayed with a reduced size. Relevant calculation control is TF / IDF, similarity determination in vector space, similarity determination by statistical method, page rank determination of importance of articles by changing the weighting when comprehensive determination, In the related calculation, the weights of those calculation elements are changed. These controls are performed using an input means such as a mouse on the screen. The axis of the radar chart is a calculation element for TF / IDF, determination of similarity in a vector space, determination of similarity by a statistical method, and calculation calculation of importance of an article by PageRank. The closer the value on the axis is to the circumference, the greater the value of the weight, and the more important the calculation element. The closer the value on the axis is to the center of the circle, the less important the calculation element is. There is also a radar chart for comprehensive judgment. The axes at this time are TF / IDF, determination of similarity in a vector space, determination of similarity using a statistical method, and determination calculation of importance of an article using PageRank. The value on the axis is a weight for controlling each importance. By providing the above interface, it is possible to perform narrowing down and the like.

FIG. 15 is an interface for displaying a search result and finding a follow-up article in the information acquisition apparatus according to the embodiment of the present invention.
In FIG. 13E, there is an area for graphically displaying a search result which is a relation display between articles. Time is taken on the X-axis, the score for the search vector is taken on the Y-axis, and points representing articles are arranged in the space. Also, by clicking on a point representing an article, the article text is displayed in the text display area of B in FIG. Each article is linked to the strength of relevance to the article. The strength of relevance of each link can be identified by the thickness and type of the line. In FIG. 15, a strongly related item is indicated by a solid line, and a weakly related item is indicated by a dotted line. Here, each link can be color-coded according to strength. In addition, since the index for the user's overall interest is expressed by the Y-axis coordinates, and the index for the user's local interest is expressed by a link connecting the articles, for example, the user has Y coordinates as much as possible. By reading through articles that are large and linked to the currently viewed article with a link of a strong color, it is possible to discover articles with follow-up information from related articles.

  It is also possible to consider the time history of the article content. In the graph of the screen shot of FIG. 13, since a newspaper article is taken as an example, the horizontal axis is the date. The vertical axis indicates the importance of an article. The importance is higher as it is displayed at the top, and the importance is lower as it is displayed at the bottom. A point on the graph represents an article, and a line connecting the points indicates “related”. When the point is clicked, the article content is displayed in the window B in FIG. As for the line connecting the points, the solid line indicates the highest degree of association, and the dotted line indicates the lowest degree of association. Note that a highly relevant line can be colored red and a less relevant line can be colored yellow. Furthermore, the coloring of this line can be made to correspond to the spectral change from red to blue in order to increase the visibility.

  FIG. 16A is a diagram of an interface for displaying a plurality of article texts in the information acquisition apparatus according to the embodiment of the present invention, and (b) a display method when a new article text is displayed. In the article body display area of B in FIG. 13, there is an interface for simultaneously displaying a plurality of designated article bodies. When the user left-clicks on an article for which a text is to be displayed, the text of the specified article is displayed in an area in which the article text is not displayed among a plurality of display areas. If the three display areas are filled, the left area is initialized, and then the article text in the right area is displayed. The specified article text is displayed in the right area.

  FIG. 17 is a headline display area for article selection assistance of the information acquisition apparatus according to the embodiment of the present invention. For a plurality of newspaper articles listed by A, D in FIG. 13 compares the area displaying the headline of the article with a plurality of articles, calculates relevance, summarizes, aggregates, Provide buttons for various commands to search and return to the original article from a summary or aggregated state. FIG. 17 shows an example of three windows, but any number of windows may be used. Also, when the user right-clicks on an article, the article headline is displayed. This makes it possible to search for an article to be viewed without changing the text of the displayed article.

  FIG. 18 is an interface that displays a result of collecting articles of the information acquisition apparatus according to the embodiment of the present invention. In the method of collecting articles, a plurality of articles are regarded as one document, and sentence syntax analysis and a summary sentence are created. In the sentence syntax analysis, semantic paragraphs are created, and the connection structure of semantic paragraphs is created, the sentence structure is reorganized without breaking the connection relations, and the flying structure analysis of the semantic paragraphs is also performed. Evaluate important phrases in the descriptive form from semantic paragraphs analyzed by sentence structure, extract weighted phrases, extract phrases that consider word similarity, and complete words to complete them Create a sentence. It is also possible to summarize only each document without aggregation and attach a classification tag if necessary, and use it for document classification. The classification tag is created from, for example, an important word obtained by calculation of TF / IDF, created from a vector at the time of vector space calculation, or has a high weight value used for similarity calculation by a statistical method. It can be created from important elements.

  FIG. 19 is an example of a related article search result of the information acquisition apparatus according to the embodiment of the present invention. A related article is searched by designating the top of the article list of A. A list of related articles from which the second and following articles were searched. What is specified at the start of searching for related articles may be an article (the one at the top of the article list of A), a freely described sentence, or a combination of keywords. Relevance and relevance calculations are done using system defaults when the user controls described below are not in effect. In E, a graph of the relation between articles and the degree of relation is displayed. In B, when the display is designated by A or when a point in the graph of E is clicked, the contents of those articles are displayed. Any number of B sub-windows may be used. In the screenshot, three subwindows are displayed. As shown in the selection button at the top of B, at present, the number of sub-windows can be selected from one, two, and three. If the number of sub-windows is further increased, the sub-windows will become narrower and more difficult to read. The radar chart displays default values.

  FIG. 20 is an example summarizing each article of the information acquisition apparatus according to the embodiment of the present invention. A summary of each article of B is displayed. When calculating the relationship and degree of relevance, the shorter the article length, the faster the calculation speed. In addition, since the summary is focused on important information of articles, the quality of the aggregation process is improved.

  FIG. 21 is an example in which three summary documents of the information acquisition apparatus according to the embodiment of the present invention are integrated into one. A summary of the three B summary documents is displayed in an independent window. Aggregation can also be performed on articles that are not summarized.

  FIG. 22 is an example of examination of the aggregation result of the information acquisition apparatus according to the embodiment of the present invention. The consolidated document is moved to the rightmost window of B, the consolidated result is examined, and the middle and leftmost articles are newly selected and displayed as necessary. For these articles, recalculate the relationship and degree of association, search again based on the results, and display the new relationship and degree of association as a graph. As a result, the graph of E is updated. Repeat these operations as needed to get the final aggregation. This aggregation is the result of information retrieval. That is, the so-called information search result is output as one document. This is quite different from many current information retrieval systems that use a list of URLs as an output for information retrieval.

  Although the present invention has been described with the above-described embodiments, the technical scope of the present invention is not limited to the scope described in the embodiments, and various modifications or improvements can be added to these embodiments. And embodiment which added such a change or improvement is also contained in the technical scope of the present invention. This is apparent from the claims and the means for solving the problems.

It is a hardware block diagram of the information acquisition apparatus which concerns on embodiment of this invention. It is a block block diagram of the information acquisition apparatus which concerns on embodiment of this invention. It is explanatory drawing of the importance evaluation of the article | item by the similarity between articles of the information acquisition apparatus which concerns on embodiment of this invention. It is an example of importance calculation of the article of the information acquisition device according to the embodiment of the present invention. It is an example of the extraction of the noun in the clause which concerns on the clause containing the verb of the information acquisition apparatus which concerns on embodiment of this invention. It is a structure figure of the EDR dictionary of the information acquisition device concerning the embodiment of the present invention. It is a synonym and similarity relationship figure between nouns of the information acquisition apparatus which concerns on embodiment of this invention. It is an example of similarity calculation between noun sets of the information acquisition apparatus according to the embodiment of the present invention. It is the ratio for every sentence type of the information acquisition apparatus which concerns on embodiment of this invention. It is a categorization of the duplication location of the information acquisition apparatus which concerns on embodiment of this invention, a specific location, and supplementary explanation. It is a flow sheet of the follow-up article search of the information acquisition apparatus which concerns on embodiment of this invention. It is a flow sheet of article content integration of an information acquisition device concerning an embodiment of the present invention. It is an execution figure of the information acquisition device concerning the embodiment of the present invention. It is an interface for search type expansion and question vector selection of the information acquisition apparatus according to the embodiment of the present invention. It is an interface for search result display and follow-up article discovery of the information acquisition device according to the embodiment of the present invention. It is a figure of the display method in the case of displaying the interface for the some article text display of the information acquisition apparatus which concerns on embodiment of this invention, and the article text newly. It is a headline display area for article selection assistance of the information acquisition device according to the embodiment of the present invention. It is an interface which displays the result of having collected the articles of the information acquisition device concerning the embodiment of the present invention. It is an example of the related article search result of the information acquisition apparatus which concerns on embodiment of this invention. It is an example which summarized each article of the information acquisition apparatus which concerns on embodiment of this invention. It is an example which consolidated the three summary documents of the information acquisition apparatus which concerns on embodiment of this invention into one. It is an example of examination of the aggregation result of the information acquisition apparatus which concerns on embodiment of this invention.

Explanation of symbols

1 Computer 2 CPU
3 Main memory 4 HDD
5 Video Card 6 Mouse 7 Keyboard 8 Optical Disc 10 Input Unit 20 Follow-up Information Search Unit 21 Vector Creation Unit 22 Search Order Determination Unit 23 Conformity / Nonconformity Determination Unit 30 Information Content Integration Unit 31 Dependency Analysis Unit 32 Noun Extraction Unit 33 Noun Inter-set similarity comparison calculation unit 34 Noun display match ratio calculation unit 35 Sentence similarity calculation unit 36 Sentence type selection unit 37 Summary sentence creation unit 38 Article aggregation unit 40 Output unit 211 Morphological analysis unit 212 TF calculation unit 213 IDF calculation unit 214 Word weight calculator 215 Feature vector generator 216 Information similarity calculator 217 Similarity matrix calculator 218 Eigenvector generator 219 Question vector generator

Claims (4)

In an information acquisition device that acquires at least one search information,
Feature vector creating means for creating feature vectors of weighted search information;
An inter-information similarity calculating means for calculating the similarity of a combination of feature vectors of all search information;
Similarity matrix calculation means for calculating a numerical similarity matrix obtained by the similarity calculation between the information,
Eigenvector generating means for numerically analyzing the similarity calculation result and obtaining an eigenvector of the maximum eigenvalue of the feature vector;
Question vector creating means for creating a question vector of the inquiry content of the search information;
Search rank determining means for determining a search rank of related information obtained by multiplying a calculated value of the cosine of the feature vector and the query vector by a numerical value of an eigenvector;
Dependency analysis means for analyzing the dependency relationship of parts of speech included in the sentence of the retrieved information;
A noun extraction means for extracting a noun in a clause related to a clause including a verb in each sentence;
A noun set similarity comparison calculating means for calculating the similarity S ₁ between the nouns of the extracted nouns and the similarity S ₂ of the noun sets;
A noun display coincidence ratio calculating means for calculating the similarity S ₃ of the proportion of the extracted noun display coincidence;
Sentence similarity calculating means for calculating the sentence similarity S by adding the similarity S ₃ to the similarity S ₂ ;
Sentence type selection means for selecting by search sentence sentence type,
The integrated contents of the related information selected by the sentence similarity calculation and the sentence type, and the search results scored by the search rank determining means are output, and the suitability of the search results and / or the parameter weighting Output means for outputting an input form for inputting the degree of
Correction means for correcting the feature vector and the question vector based on information input to the input form;
The search order determination means determines again the search order of the related information based on the feature vector and the question vector corrected by the correction means, and the output means is scored by the search order determination means. The information acquisition apparatus is characterized in that the search result is output including the temporal relationship between the search information . In the information acquisition device according to claim 1,
The feature vector creating means includes:
Morphological analysis means for performing morphological analysis of the sentence of the search information;
TF calculating means for calculating a word and the number of appearances TF of the word for each information;
IDF calculation means for calculating the document frequency DF and its IDF for all words;
A word weight calculation means for calculating a word weight for each word of each information using the TF-IDF method;
An information acquisition apparatus comprising: a feature vector creating unit that creates a feature vector of each document from the word weighting. In an information acquisition program that causes a computer to function as an information acquisition device that acquires at least one search information,
Feature vector creation means for creating feature vectors of weighted search information;
Information similarity calculation means for calculating the similarity of combinations of feature vectors of all search information;
Similarity matrix calculation means for calculating a numerical similarity matrix obtained by the similarity calculation between the information,
A numerical analysis of the similarity calculation result to obtain an eigenvector of a maximum eigenvalue of a feature vector;
A question vector creating means for creating a question vector of the inquiry content of the search information;
A search order determination means for determining a search order of related information obtained by multiplying the calculated value of the cosine of the feature vector and the query vector by a numerical value of an eigenvector;
Dependency analysis means for analyzing the dependency relationship of parts of speech included in the sentence of the retrieved information,
A noun extraction means for extracting a noun in a clause related to a clause including a verb in each sentence;
Means for calculating similarity between noun sets for calculating similarity S ₁ between nouns of the extracted nouns and similarity S ₂ of noun sets;
A noun display coincidence ratio calculating means for calculating the similarity S ₃ of the ratio of matching of the extracted noun displays;
Sentence similarity calculating means for calculating sentence similarity S by adding similarity S ₃ to similarity S ₂ ;
Sentence type selection means for selecting by search sentence sentence type,
The integrated contents of the related information selected by the sentence similarity calculation and the sentence type, and the search results scored by the search rank determining means are output, and the suitability of the search results and / or the parameter weighting Output means for outputting an input form for inputting the degree of
Based on the information entered in the input form, causing the computer to function as correction means for correcting the feature vector and the question vector,
The search order determination means determines again the search order of the related information based on the feature vector and the question vector corrected by the correction means, and the output means is scored by the search order determination means. An information acquisition program that outputs a search result including a temporal relationship between the search information . A computer of an information acquisition device acquiring at least one search information,
A feature vector creating step for creating a feature vector of weighted search information;
A similarity calculation step between information for calculating the similarity of a combination of feature vectors of all search information;
A similarity matrix calculation step of calculating a numerical similarity matrix obtained by the similarity calculation between the information;
A numerical analysis of the similarity calculation result to obtain an eigenvector of the maximum eigenvalue of the feature vector;
A question vector creating step of creating a question vector of the inquiry content of the search information;
A search order determination step for determining a search order of related information obtained by multiplying the calculated value of the cosine of the feature vector and the query vector by a numerical value of an eigenvector;
A dependency analysis step for analyzing the dependency relationship of parts of speech included in the sentence of the retrieved information;
A noun extraction step for extracting a noun in a clause related to a clause including a verb in each sentence;
A noun-set similarity comparison calculation step of calculating the similarity S ₁ between single nouns of the extracted nouns and the similarity S ₂ of noun sets;
A noun display coincidence ratio calculation step of calculating a similarity S ₃ of the ratio of coincidence of the extracted noun displays;
A sentence similarity calculation step of calculating the sentence similarity S by adding the similarity S ₃ to the similarity S ₂ ;
A sentence type selection step for selecting by search sentence sentence type;
The integrated contents of the related information selected by the sentence similarity calculation and the sentence type, and the search results scored by the search rank determining means are output, and the suitability of the search results and / or the parameter weighting An output step for outputting an input form for inputting the degree of
Executing a correction step of correcting the feature vector and the question vector based on information input to the input form;
The search order determining step determines again the search order of the related information based on the feature vector and the question vector corrected by the correcting means, and the output step is scored by the search order determining means. And a search result including a temporal relationship between the search information is output .

Images