CN100511232C - Document retrieving device and method thereof - Google Patents

Abstract

The present invention provides a document retrieval device and method for retrieving document data of a retrieval object according to a retrieval query word and an extended query word of the retrieval query word, and extracting characters consistent with the retrieval query word and the extended query word String, to determine whether the extracted character string contains an unknown word region, when it is judged that it does not contain an unknown word region, adjust, so that the similarity of the extracted character string decreases, and in the order corresponding to the adjusted similarity, the A character string is output as a search result.

Description

Document retrieval device and method

technical field

The present invention relates to a document retrieval device, a retrieval method and a storage medium for retrieving document data according to a retrieval query word (query).

Background technique

With the spread of personal computers (PCs), documents are generally created using application software on PCs such as document creation software. Specifically, operations such as creating and editing various files on a PC screen, and copying and searching files are widely performed. In addition, along with the development and popularization of the network, the electronic document data (electronic document data) thus generated on the PC is usually not printed as a paper document using a printer or the like, but is accessed by other PCs or the like, and sent and distributed by e-mail. , the paperless document generation environment is expanding.

A document management system is constructed to systematically manage such paperless electronic document data by a computer. The electronic document data is extremely convenient in terms of efficient reduction of the amount of information based on document sharing, association between documents, and the like. With the popularity of electronic document data, retrieval operations such as full-text retrieval and keyword retrieval of document data have become popular, and the effectiveness of retrieval has gradually become known.

On the other hand, a paper document in which document data is printed on paper has advantages such as easy reading, general handling, easy handling, and easy grasp of the whole picture compared to electronic document data. For example, when materials need to be distributed, the distribution is still done in the form of a paper document formed after the electronic data is printed by the printing device. However, since paper documents cannot be retrieved in their original form, it is not easy to retrieve paper documents in which desired information is printed. Therefore, in the past, paper documents were scanned and OCR (Optical Character Recognition) was used to process electronic textualized electronic document data for retrieval. However, if misrecognition occurs during the OCR process, document data desired by the user cannot be retrieved correctly.

In order to solve such a problem, various proposals have been made conventionally. As an example, there is a method of performing a search by comparing a search query word (query) and a character string of document data on the assumption of character omission, character mixing, or character distortion in a character string of document data. In addition, a method of retrieving document data by expanding each character of a search query word into an assumed misrecognized character and comparing it with the expanded search query word has also been proposed. Here, these control methods are collectively referred to as fuzzy control.

Through such fuzzy comparison, it is possible to hit character strings missed due to misrecognition, but there are also many disadvantages. For example, if "イラク" is input as a search query word, even misrecognized characters (for example, "イテク") caused by misrecognition of the character "イラク" are hit as search objects. In this case, for example, "イテク" in "プイテク" (not misrecognized) is also hit. Accordingly, whenever the word "ハイテク" is used in the document data, an irrelevant hit will be generated. When such unwanted hits occur frequently, an operation for selecting meaningful hits is required, which increases the workload for the user and makes the search device difficult to use.

As a technique related to this, there is Japanese Patent Laid-Open No. 2004-334334.

者”这一字符串时，尽管该字符串未被误识别，但是该字符串中的“人関”也命中了。这时，“人関”、“告人関”、“人関

”等并不形成字典单词的一部分，在以往技术中无法抑制该命中。However, in the technique disclosed in Japanese Patent Application Laid-Open No. 2004-334334, irrelevant hits still occur. For example, when searching for a character string "Renjian" (or a character string generated by misrecognizing "Renjian"), specify "Renjian" as a query term. Since the Chinese characters "between" and "guan" are similar, "renguan" is also set as a search term for characters that are misrecognized. Therefore, when there is a "defendant related

When the character string ", although the character string has not been misrecognized, the "person off" in the string is also hit. At this time, the

" etc. do not form part of a dictionary word, and this hit cannot be suppressed in previous techniques.

Contents of the invention

The present invention aims to solve the disadvantages of the prior art described above.

Furthermore, the present invention is characterized in that it provides a document search device and method thereof capable of efficiently searching for misrecognized character strings.

The present invention provides a document retrieval device, comprising: a comparison device, based on a retrieval query word and an extended query word similar to the retrieval query word, retrieves document data as a retrieval object, and extracts a document data corresponding to the retrieval query word and the described retrieval query word. Expand the character string that query word is consistent; Analyzer, analyze the document data as described retrieval object, identify unknown word area; Judgment means, judge whether at least a part of the character string that is extracted by described comparison means is contained in described unknown word area and the retrieval result output device, according to the processing results of the comparison device and the judgment device, if it is judged that the character string extracted by the comparison device is not included in the unknown word region, reduce the The similarity of the character strings extracted by the comparison device is calculated, and the character strings are output as the search results.

Moreover, the present invention provides a document retrieval method, including: a comparison step, based on the retrieval query word and the extended query word of the retrieval query word, retrieve the document data as the retrieval object, and extract the document data related to the retrieval query word and the described retrieval query word. Expand the character string that query word is consistent; Analysis step, analyze the document data of described retrieval object, identify unknown word region; Judgment step, judge whether at least a part of the character string extracted by described comparison step is contained in described unknown word region and the retrieval result output step, according to the processing results of the comparison step and the judgment step, if it is judged that the character string extracted by the comparison step is not included in the unknown word region, reduce the character The string similarity is output in the string as a retrieval result.

The summary of the present invention is not an exhaustive list of essential features, therefore sub-combinations of these feature groups can also be inventions.

Other features, objects and advantages of the present invention will become apparent from the following description with reference to the accompanying drawings, in which like reference characters indicate the same or similar parts.

Description of drawings

The accompanying drawings constitute a part of this specification, illustrate embodiments of the invention, and together with the description serve to explain the principles of the invention.

FIG. 1 is a block diagram showing the structure of a document retrieval device according to an embodiment of the present invention.

FIG. 2A and FIG. 2B are diagrams illustrating an operation example of search in this embodiment.

FIG. 3 is a diagram illustrating an example of unknown word analysis processing in this embodiment.

FIG. 4 is a diagram illustrating the data structure of an unknown word area table storing unknown word areas in this embodiment.

FIG. 5 is a diagram illustrating the data structure of a query word expansion table for expanding each character (character string) of a search query word into a similar character string in this embodiment.

FIG. 6 is a diagram illustrating an expanded query word grid that expands the search query word of this embodiment into a character string that may be expanded by misrecognition.

FIG. 7 is a diagram showing the storage format of the rules for identifying which part of the morpheme analysis results should be the unknown word area, that is, the unknown word area identification rule in this embodiment.

FIG. 8 is a diagram showing the structure of a search result table storing candidates of search results in this embodiment.

FIG. 9 is a diagram illustrating a calculation formula of a score that defines the output order of search results according to the present embodiment.

FIG. 10 is a flowchart illustrating the processing of the document retrieval device of this embodiment.

FIG. 11 is a flowchart illustrating retrieval processing as part of the event handling processing in step S4 of FIG. 10 .

FIG. 12 is a flowchart illustrating unknown word analysis processing in step S11 of FIG. 11 .

FIG. 13 is a flowchart illustrating the extraction process of the unknown word region in step S23 of FIG. 12 .

FIG. 14 is a flowchart illustrating the fuzzy matching process of step S13 in FIG. 11 .

FIG. 15 is a flowchart illustrating the point adjustment processing in step S14 of FIG. 11 .

Detailed ways

Preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The following examples do not limit the invention related to the claims, and all combinations of features described in the examples are not necessarily essential to the solution means of the invention.

FIG. 1 is a block diagram showing the structure of a document retrieval device according to an embodiment of the present invention.

In the figure, CPU101 is a microprocessor, according to the program stored in ROM102 or RAM103, is used for image processing (image processing), character processing (character processing), character recognition processing (character recognition processing), retrieval processing (search processing) ) operations, logical judgments, etc., to control each component connected via the bus 120. The bus 120 is a system bus, and transmits address signals, data, and control signals indicating each component to be controlled by the CPU 101 . The ROM 102 is a read-only nonvolatile memory, and stores a boot program executed by the CPU 101 and various data. The boot program loads the control program stored in the hard disk (HD) 108 into the RAM 103 when the system is started, and causes the CPU 101 to execute it. This control program will be described in detail later with reference to a flowchart. RAM 103 is a readable and writable random access memory, stores various programs loaded from HD 108 and executed by CPU 101 , and is used as a work area during operation of CPU 101 to temporarily store various data from each component.

The input unit (input unit) 104 includes a keyboard, a mouse, a touch panel, and the like, and performs selection of menu items, input of various data, and the like through user operations. The display unit 105 has a display such as liquid crystal, CRT, or plasma, and is used to present various menus, processing results, errors, warnings, search results, and the like to the user. A scanner (scanner) 106 performs processes such as optically reading a paper document as an original and digitizing it. A printer (printer) 107 is used when printing documents and images. In this document retrieval apparatus, electronic document data in a PDL (Print Control Language) grid format received by a communication unit 110 can also be printed.

The HD (hard disk) 108 stores a control program 111 executed by the CPU 101, a morpheme analyzing dictionary 112 for natural language analysis, and a document describing rules for identifying unknown-word areas. Unknown word area recognition rule (unknown-word area recognition rule) 113, etc. And, according to need, also store the unknown word area table (unknown-word area table) 114 that is used for managing unknown word area, the search result table (searchresult table) 115 that keeps retrieval result, will search query words expand (develop) and keep Operation data such as query word expansion table (query development table) 116. These various data are loaded into RAM as needed, referred to, changed as needed, and written back to HD 108 . The morpheme analysis dictionary 112 stores necessary information proposed in general natural language analysis, such as word writing, part-of-speech information, conjugation information, word collocation information, and the like.

The removable external storage device 109 is a pluggable storage device (storage device) such as a USB storage device or an IC card. They can also be drives for accessing external storage such as floppy disks, CDs, and DVDs, just like ordinary PCs. This external storage device 109 can be used similarly to HD 108, and can exchange data with other devices via these storage media. All or part of the control program 111 stored in the hard disk 108 can be copied (installed) from the external storage device 109 to the HD 108 as needed. The communication unit 110 is a network controller, and can exchange data with the outside through a communication line.

The document retrieval device of this embodiment having the above-mentioned structure operates according to various events from the input unit 104 and the like. When there is an interrupt from the input unit 104, the interrupt signal is sent to the CPU 101, and an event is generated accordingly. According to this event, CPU 101 reads various commands stored in ROM 102 or RAM 103 and executes the commands to perform various controls according to the control program.

FIG. 2A and FIG. 2B are diagrams illustrating an operation example of search in this embodiment.

In the example 1 shown in FIG. 2A , the text to be retrieved originally is "イラクヘのハイテク战争出の问题を问ぅ" (201). During the OCR process, it is recognized as "イテクヘのハイテク战争出の问题を问ぅ" and registered (202). Here, the character string originally "イラク" is changed to "イテク" because of misrecognition.

Next, the operator issues a search query word "iraku" in order to find the misrecognized character string "iraku" (203). In the retrieval process of this embodiment, similar characters (expanded query word "イテク") are regarded as identical and searched by similar word expansion processing. Accordingly, the misrecognized character string "イテク" is found as a search result (210). Here, the same character string ("イテク") (211) also exists in the character string "プイテク", but they are hits when analyzing the phrase "ハイテク", so they are output as the search results with lowered hit ranks.

者に必要”(205)。与例1同样字符识别该原文时，将“間”误识别为“関”，并被登录(206)。这里，为了检索“人間”，发出检索查询词“人間”的命令(207)。据此，包含相似字符(“人関”)地进行检索。命中字符串“人関”212的“人関”。另外，此时“被告人関

者”中的“人関”213成为可分析为“被告人”“関

者”的短语，所以降低命中位次地作为检索结果输出。In Example 2 shown in FIG. 2B, the retrieved text is "law ni detailed し ぃ human が defendant related

人に必要” (205). When recognizing the original text with the same characters as in Example 1, “between” is misidentified as “off” and is registered (206). Here, in order to retrieve “human”, the search query word “human” is issued. " command (207). Accordingly, search is carried out containing similar characters ("Renguan"). Hit the "Renguan" of the character string "Renguan" 212. In addition, at this time, "Defendant Guan"

"Ren Guan" 213 in "Person" can be analyzed as "Defendant" and "Guan Guan".

"Phrase", so the hit rank is lowered and output as the search result.

FIG. 3 is a diagram illustrating an example of unknown word analysis processing in this embodiment.

A page image 301 represents document data generated by scanning of a paper document or rasterization of an electronic document, which is consistent with the original text. Text 302 represents the result of OCR processing on it. In this text 302, due to misrecognition, "Renjian" becomes "Renguan" (310), "Wang まれる" becomes "Wang Shen れる" (311), and "イラク" becomes "イテク" (312) (misrecognized characters are underlined).

The text 303 represents the analysis result of the morphological analysis of the text 302 subjected to character recognition, and the text is divided into phrase units. "/" indicates division of phrases. Here, a place where morphological analysis cannot be performed (character string 311 that cannot be analyzed) is displayed with a frame 313 . At this time, "人关" (310) and "イテク" (312) that were misrecognized can be analyzed if they are divided into morphemes, so it is not judged that the character string cannot be analyzed. The text 304 represents the result of further performing the extraction process of unknown word regions described later on the analysis results shown in 303 (character strings with frames 314 to 315 represent unknown word regions). Accordingly, in addition to the unanalyzable character string ("Wang Shen れる" (311)) in 303, according to the unknown word area identification rule 113 described later, the misrecognized "person off" (310), "イテク" The part of "(312) is also set to the unknown word region.

FIG. 4 is a diagram illustrating the data structure of the unknown word area table 114 storing unknown word areas in this embodiment.

For each unknown word area, a start position (start position) 401 and an end position (end position) 402 are stored. These start positions and end positions store values indicating the start position and end position of the unknown word region on the text (information indicating the page, line number, and character number of the line). The character code 403 is a character code corresponding to the character string 316 identified as the unknown word area in the text 304 of FIG. 3 . In FIG. 4, the start position and end position of "イテク" 316 designated as the unknown word area are stored.

FIG. 5 is a diagram illustrating the data structure of the query word expansion table 116 for expanding each character (character string) of a search query word into a similar character string in this embodiment.

Pairs of similar characters (strings) that appear to have been misrecognized are stored here. For example, the katakana characters "n" and "ソ" are similar, so "n" is stored in the original character string (original character string) 501, and "ソ" is stored in the corresponding developed character string (developed character string) 502. In addition, since the katakana "デ" may be misrecognized as "テリ" with two characters of katakana, "テリ" is stored as the extended character string 502 corresponding to "デ" of the original character string 501 . In addition, the katakana "ク" may be misrecognized as "ワ" and registered. In addition, since the 2-character katakana "イン" may be misrecognized as the 1-character kanji "仁", "仁" is stored as the extended character string 502 corresponding to the original character string 501 "イン". In addition, characters and character strings with a high possibility of being misrecognized are also registered, but they are omitted here.

FIG. 6 is a diagram illustrating a lattice of an expanded query word that expands the search query word of this embodiment into a character string that may be expanded by misrecognition.

Here, the connection status of each extended character string forms a grid. When selecting a path from the start node to the end node 602, the original search query appears as a query expanded into a similar character string that may be misrecognized. For example, as a search query word, "Index" is changed to "Isoteriwas" according to the rule shown in 610, and is expanded to "Index" according to the rule shown in 611. According to other rules, it is expanded to "Interlyx", "Isodex", "Rendemus" and so on. In this way, the search query word is stored as a character string expanded into the above-mentioned similar character string according to whether the original text of the search query word remains unchanged (here, "Intex") or the expanded query word. In FIG. 6 , the characters inside the ellipses represent the characters that match the characters of the original search query word, and the characters inside the rectangles represent the characters when they were misrecognized.

FIG. 7 is a diagram showing a storage format of an unknown word area identification rule which is a rule for identifying which part of the morphological analysis result should be an unknown word area.

With regard to each rule, conditions satisfied by each of the first phrase 701 and the second phrase 702 are described. For example, in rule 1, when the phrase length (number of characters) of the first phrase 701 is "1", and the length of the independent word of the second phrase 702 (that is, the number of characters excluding dependent words) is "1", rule 1 is satisfied (i.e. identified as an unknown word area). Accordingly, in the example shown at 304 in FIG. 3 , two phrases 314 consisting of "person" and "guanga" are recognized as unknown word regions.

Similarly, the rule described in rule 2, when the phrase length of the first phrase 701 is less than or equal to 2, written in katakana, and the independent word length of the second phrase 702 is less than or equal to 2, written in katakana, it is considered an unknown word . Accordingly, in the example shown at 304 in FIG. 3 , two phrases 316 composed of "i" and "TEK" are recognized as unknown word regions.

FIG. 8 is a diagram showing the structure of a search result table 115 storing candidates of search results in this embodiment.

Here, for each search result candidate, the position on the text of the start character of the search result (hit character string) is stored in the start position 801 . In the end position 802, the position on the text of the end character of the retrieval result (hit character string) is stored. In the score 803, a value specifying the display order of the search result (described later) is stored. The search results are finally sorted in the order of the value of the score 803 and output as a search result.

FIG. 9 is a diagram illustrating a calculation formula of a score defining an output ranking of a search result according to the present embodiment.

First, the similarity is calculated by the expression of (number of fully consistent characters)×2+(number of fuzzy consistent characters). Here, the "number of identical characters" is the number of identical characters between the search query term and the hit character string (without similar character string expansion). In addition, the "number of fuzzy matching characters" is the number of characters matching the hit character string which is the result of expanding the search query term into similar character strings. Here, when a part of the hit character string is in the unknown word region, the calculated "similarity" becomes the "score" as it is. In addition, when the hit string is not in the unknown word region, that is, when the entire region of the hit string has been analyzed into phrases, the "score" is subtracted from the "similarity" by the "maximum analyzed phrase length". Here, the "maximum analyzed phrase length" is the longest phrase length (the number of characters) among the phrases of the morphological analysis result on the hit character string. Details are given below.

Example 1 shows an example when the hit character string is in the unknown word region. In this case, it is defined as "score" = "similarity". In other words, "maximum analysis phrase length"=0.

Examples 2 to 4 show examples when the hit character string is not in the unknown word area. In Example 2, the analysis phrase 1 and the analysis phrase 2 are among the hit character strings, and the number n of characters of the phrase with the largest number of characters becomes the "maximum analysis phrase length".

In Example 3, the hit string covers only one phrase, so the number of characters n of the phrase is the "maximum parsed phrase length".

In Example 4, the hit string covers 3 analysis phrases, and the character number k of the longest phrase among them is the "maximum analysis phrase length".

者”317。此外，检索查询词为“イラク”时，作为与扩展查询词“イテク”一致的字符串，检索“イテク”316和“ハイテク”318。这时，首先“人関が”314的相似度与(例1)对应。这时，根据图9的表达式，完全一致的字符(“人”)的数(1)×2+(模糊一致的字符(“関”)的数(1)＝3，此外，字符串“人関”的一部分在未知词区域，所以分数也变为“3”。Using the example shown at 304 in FIG. 3 , the method of calculating the score described above will be described. When the search query term is "renjian", as a character string consistent with the extended query term "renguan", retrieve "renguan ga" 314 and "defendant Guan

"Those" 317. In addition, when the search query word is "イラク", as a character string consistent with the expanded query word "イテク", search for "イテク" 316 and "ハイテク" 318. At this time, first of all, "人关が" 314 The degree of similarity corresponds to (example 1). At this moment, according to the expression of Fig. 9, the number (1) * 2+(the number (1) of the number (1) * 2+(fuzzy consistent character (" close ") of completely consistent character (" close ") according to the expression of Fig. 9 )=3, in addition, part of the character string "Renguan" is in the unknown word area, so the score also becomes "3".

者に”317的情况，与(例2)对应。这时，相似度也是与上述的计算相同的“3”。可是，这时任何部分都不在未知词区域中，所以分数是相似度“3”-最大分析短语长度(4)(3-4＝)-1。同样，检索查询词为“イラク”时，与(例1)对应。因此，“イテク”316的相似度，根据图9的表达式，成为完全一致的字符(“イ”、“ク”的数(2)×2+(模糊一致的字符“テ”的数(1)＝“5”，此外，由于在未知词区域中，所以分数也为“5”。而在“ハイテク”318的情况下，“ハイテク”318收容在一个分析短语内，所以与(例3)对应。通过上述的计算，相似度为“5”，但由于不在未知词区域，所以分数为相似度“5”-最大分析短语长度(4)(5-4＝)1。And "the accused

者に" 317 corresponds to (Example 2). At this time, the similarity is also "3" which is the same as the above-mentioned calculation. However, at this time, any part is not in the unknown word region, so the score is the similarity "3". "-maximum analysis phrase length (4) (3-4=)-1. Similarly, when the search query word is "イラク", it corresponds to (Example 1). Therefore, the similarity of "イテク" 316, according to Fig. 9 The expression becomes the number (2)×2+(the number (1) of the character “テ” of fuzzy coincidence of characters ("イ" and "ク") that are completely consistent = "5". In addition, because in the unknown word region , so the score is also "5". In the case of "ハイテク" 318, "ハイテク" 318 is contained in an analysis phrase, so it corresponds to (Example 3). Through the above calculation, the similarity is "5", But because it is not in the unknown word area, the score is similarity "5" - maximum analysis phrase length (4) (5-4=)1.

The above-mentioned operation will be described in accordance with the flowchart.

FIG. 10 is a flowchart illustrating the processing of the document retrieval device of this embodiment. A program for executing this processing is stored in RAM 103 and executed under the control of CPU 101 when executed.

First, in step S1, the initialization process of the system is executed, here, initialization of various parameters, display of an initial screen, and the like are performed. Next, in step S2, an arbitrary event such as a request from the input unit 104 or a device connected via a network or the like is waited for. Here, after an event occurs, it proceeds to step S3, and the event that occurred is discriminated, and various processes are branched according to the discriminated event type. Here, a plurality of processes of branch targets corresponding to various events are comprehensively expressed in step S4. As an example of processing of branch destinations corresponding to various events, there is search processing shown in FIG. 11 . As other processing, which is not described in detail, there are processing of specifying retrieval conditions, processing of scanning a document to generate a document image, processing of specifying a document, and other processing of a general retrieval device. Then, it progresses to step S5, and the processing result of each process of step S4 is displayed. The processing here is generally widely performed processing such as display processing of search results, error display when there is an error, and display processing when it ends normally.

FIG. 11 is a flowchart illustrating search processing as part of the event handling processing in step S4 of FIG. 10 .

First, in step S11, unknown word analysis processing described in detail in the flowchart of FIG. 12 is performed. Here, character recognition is performed based on the image of the specified document to generate an OCR text, and an unknown word area table 114 ( FIG. 4 ) is generated through morphological analysis and unknown word analysis. Next, in step S12, the input search query is expanded into similar character strings that may be misrecognized to generate an expanded query grid (see FIG. 6 ). Next, in step S13 , based on the generated expanded query word grid, referring to the flowchart of FIG. 14 , a fuzzy comparison process described later is executed to generate a search result table 115 . Next, in step S14, referring to the flowchart of FIG. 15, the score of the search result table 115 is obtained from the unknown word area table 114 and the expression in FIG. 9 as described later. Next, in step S15 , the search results are sorted in the order of scores in accordance with the order of the obtained scores. Then in step S16, display and output the retrieval results sorted in order of scores.

FIG. 12 is a flowchart illustrating unknown word analysis processing in step S11 of FIG. 11 .

First, in step S21, character recognition is performed on the designated document image to obtain text information (302 in FIG. 3). Next in step S22, morpheme analysis is performed on the text information, and it is divided into phrases (303 in FIG. 3). Next, in step S23, referring to the flowchart of FIG. 13, an unknown word region is extracted as will be described later (304 in FIG. 3). Next, in step S24 , the extracted unknown word region is output as the unknown word region table 114 .

FIG. 13 is a flowchart illustrating the extraction process of the unknown word region in step S23 of FIG. 12 .

First, in step S31, variables and the like are initialized, and the pointer indicating the phrase is initialized to indicate the beginning of the text. Next, in step S32, the information of the phrase indicated by the pointer is obtained. Next, in step S33, referring to the morphological analysis dictionary 112, it is judged whether or not the phrase acquired in step S32 is an unanalyzable phrase. When it is judged to be an unanalyzable phrase, it is regarded as an unknown word region, and branched to step S35, and when it is judged not to be an unanalyzable phrase, it enters step S34, and with reference to the unknown word region determination rule 113, it is judged whether the phrase belongs to an unknown word area. Here, if it is judged not to belong to the unknown word area, it will branch to step S36, and if it is judged to belong to the unknown word area, it will go to step S35, and the necessary information is collected about the extracted unknown word area, and it is set as the unknown word area. Then, enter step S36, update the pointer representing the phrase to indicate the next phrase. Then in step S37, it is judged whether there is a next phrase, if it is judged to exist, it returns to step S32 to execute the described processing. On the other hand, if it is judged that the next phrase does not exist, the extraction process of the unknown word region is terminated.

FIG. 14 is a flowchart illustrating the fuzzy matching process of step S13 in FIG. 11 .

First, in step S41, initial setting is performed so that the pointer indicating the position of the character points to the beginning of the text. Then in step S42, the expanded query word grid is compared with the character on the text indicated by the pointer. Then in step S43, it is judged whether the expanded query word grid is consistent with the character, and if not, skip to step S46 and move to the next character. In step S43, when it is judged that they match, the process proceeds to step S44, and the degree of match is calculated as the degree of similarity. The calculation process of this similarity is performed according to the expression shown in FIG. 9 mentioned above. Next, the process proceeds to step S45, and the matching character position is registered in the search result table 115 (FIG. 8). Here, the degree of similarity obtained in step S44 is set as a score as it is. Then enter step S46, update the pointer indicating the character position, and advance the character position to the next one. Then, in step S47, it is judged whether the character position has reached the end of the text, and when it has not reached the end, it returns to step S42 to execute the described processing, and when it reaches the end, it ends the fuzzy comparison process.

FIG. 15 is a flowchart illustrating the point adjustment processing in step S14 of FIG. 11 .

First, in step S51, initialization is set so that the pointer indicating the search result indicates the start of the search result table 115 (FIG. 8). Next, in step S52, the information (position and score) of the search result indicated by the pointer is acquired. Then in step S53, check whether the hit character string represented by the retrieval result contains unknown word regions in the text according to the unknown word region table 114 (FIG. 4). Then in step S54, if it is judged to include the unknown word region, it will branch to step S58, and the score is determined to be a value equal to the similarity, and if it is judged to not include, it will enter step S55, as shown in Figure 9, to find The length of the longest parsed phrase involved in outputting a hit string. Next, in step S56, the calculated longest analysis phrase length is subtracted from the score to correct the score. Next, in step S57, the corrected score is reflected in the search result table 115 (FIG. 8). Next, in step S58, the pointer indicating the search result is updated to indicate the next search result. Then in step S59, it is judged whether it is the last of the retrieval results, if not, return to step S52, and execute the processing described above, and when it is judged to be the end, the score adjustment processing ends.

(other embodiments)

The present invention is not limited to the above-mentioned embodiments, and appropriate changes can be made as long as they do not deviate from the gist of the present invention. In the above-described embodiments, morphological analysis is used as a method of language analysis, but other implementations are also conceivable. For example, a method based on a technique of dividing only into words is also considered. At this time, the portion of the dependent words is not analyzed at all, and is treated as an unknown word area. According to this, there is a disadvantage that the accuracy of the analysis is lowered, but the analysis process can be easily completed compared with the morphological analysis, and a system with a lighter load can be constructed.

In addition, in the above-mentioned embodiment, as a method of fuzzy comparison, the characters of the query words are expanded into similar character strings for retrieval. However, it is also considered to collect similar character groups without expansion and standardize them as representatives to be represented. The way characters are compared. With such a configuration, the processing load can be reduced, and it can be applied to a smaller-scale device.

Furthermore, completely different implementations of fuzzy controls are also conceivable. For example, it is also possible to adopt a method of judging that the comparison is successful even if there is an inconsistency like a wildcard search. In this case, the calculation method of the similarity is slightly changed, but if the part of the blurred comparison is removed, the configuration can be exactly the same, and the same effect can be obtained. In addition, other than the above, the structure of this embodiment can be appropriately changed as long as it does not deviate from the gist of the present invention.

As described above, according to the present embodiment, it is possible to search for a character string that allows misrecognition for text in which misrecognition exists. In addition, even if there is a hit character string that allows misrecognition, if it is included in an analyzable character string, the score evaluation will be low, so that unwanted false hits can be suppressed. Accordingly, hits of character strings that were actually misrecognized are displayed with relative priority, and a document retrieval device with high operability can be provided.

The present invention is not limited to the above embodiments, and various changes and modifications can be made within the spirit and scope of the present invention. Therefore, to apprise the public of the scope of the present invention, the following claims have been made.

Claims (12)

1. A document retrieval device, characterized in that, comprising: The comparison device retrieves the document data as the retrieval object based on the retrieval query word and an extended query word similar to the retrieval query word, and extracts a character string consistent with the retrieval query word and the expanded query word; An analyzing device, analyzing the document data as the retrieval object, and identifying the unknown word region; Judging means, judging whether at least a part of the character string extracted by the matching means is included in the unknown word region; and The search result output means reduces the number of the character string when it is determined that the character string extracted by the matching means is not included in the unknown word region based on the processing results of the matching means and the judging means. Similarity, output the character string as the retrieval result. 2. The document retrieval device according to claim 1, characterized in that: The retrieval result output device has a score correcting device, which reduces the similarity of the character string extracted by the comparing device when it is judged by the judging device that the unknown word region is not included; The sequence corresponding to the similarity obtained by the score correction device is used to output the character string as the retrieval result. 3. The document retrieval device according to claim 1, characterized in that: The extended query word is a character string formed by replacing a character with a high probability of being misrecognized as another character when character recognition is performed on the characters constituting the search query word with the other character. 4. The document retrieval device according to claim 2, characterized in that: The comparison device, according to the number of characters contained in the consistent character string that is consistent with the characters of the search query word and the number of characters that are consistent with the characters of the extended query word contained in the consistent character string , to obtain the similarity of the consistent character strings. 5. The document retrieval device according to claim 1, characterized in that: The analysis device performs morphological analysis of the document data, and identifies the unknown word region based on whether words included in the phrase obtained by the morphological analysis are included in a word dictionary. 6. The document retrieval device according to claim 2, characterized in that: the analysis means analyzes the document data as the retrieval object, divides the document data in units of phrases, The score correcting means reduces the similarity by subtracting the maximum number of characters of a phrase including at least a part of the character string extracted by the matching means from the similarity of the extracted character string. 7. A document retrieval method comprising: In the comparison step, based on the search query word and the extended query word of the search query word, the document data as the retrieval object is retrieved, and a character string consistent with the search query word and the described extended query word is extracted; An analysis step, analyzing the document data of the retrieval object, and identifying the unknown word region; A judging step, judging whether at least a part of the character string extracted by the matching step is included in the unknown word region; and The retrieval result output step is to reduce the similarity of the character strings when it is judged that the character strings extracted in the comparing step are not included in the unknown word region based on the processing results of the comparing step and the judging step. degree, output the string as the search result. 8. The document retrieval method according to claim 7, characterized in that: In the retrieval result output step, there is a score correction step, which reduces the similarity of the character strings extracted by the comparing step when it is judged by the judging step that the unknown word region is not included; The character strings are output as retrieval results in an order corresponding to the degrees of similarity obtained in the score correction step. 9. The document retrieval method according to claim 7, characterized in that: The extended query word is a character string formed by replacing a character with a high probability of being misrecognized as another character when character recognition is performed on the characters constituting the search query word with the other character. 10. The document retrieval method according to claim 8, characterized in that: The comparison step is based on the number of characters consistent with the characters of the search query word contained in the consistent character string and the number of characters consistent with the characters of the extended query word contained in the consistent character string , to obtain the similarity of the consistent character strings. 11. The document retrieval method according to claim 7, characterized in that: In the analyzing step, the morphological analysis of the document data is performed, and the unknown word region is identified according to whether the words contained in the phrase obtained by the morphological analysis are included in the word dictionary. 12. The document retrieval method according to claim 8, characterized in that: the analysis step analyzes the document data as the retrieval object, divides the document data in units of phrases, The score correction step reduces the similarity by subtracting a maximum number of characters of a phrase including at least a part of the character strings extracted in the comparing step from the similarity of the extracted character strings.

Images