CN1452098A - File classing system and program for carrying out same - Google Patents

Abstract

The invention provides a document classification system and its realization program. It is possible to efficiently classify according to the content of letters. The frequency of occurrence of words stored in the important word dictionary is measured by the character recognition device, and the document type is estimated by the document type recognition unit.

Description

Document Classification System and Its Realization Program

technical field

The present invention relates to a document classification technology for automatically classifying inquiry letters, e-mails, etc. by computers at customer windows of companies, institutions, etc., and a support system for answering inquiries.

Background technique

In manufacturing, insurance, telecommunications, institutions, etc., the business of directly receiving inquiries in the form of e-mails, letters, FAX, etc. from customers has become increasingly important in recent years. In many cases, it is difficult to ask a person to efficiently answer various inquiries. Usually, there are more inquiries than one person can handle. In addition, a lot of content involves multiple aspects. For example, in the manufacturing industry, it is necessary to deal with documents for inquiries about a product, a purchase method, an operation method, and the like. It takes a wide range of knowledge for one to deal with all these issues. Often, it is difficult to secure staff with such extensive knowledge.

Therefore, there is a need for a system that recognizes the types of inquiry documents, classifies them, distributes them to specialist staff according to each content, and allows these staff to answer.

The present invention relates to the technology of using computer to classify the above query documents, and the technology of using computer to answer the query support system.

Techniques for sorting e-mail inquiries using computers are known. As a typical method, there is a method of using a multivariate pattern recognition technique characterized by the frequency of occurrence of a specific word (important word) group in a document. The body text and file name of the e-mail are text data, and the occurrence frequency of words can be obtained through simple word comparison or morpheme analysis. Once the type of inquiry e-mail is recognized, a technology is known in which an e-mail or the like is automatically answered according to the type.

Also, as a method for identifying the type of inquiry letter, there is a method of converting the contents of the letter into text by a character recognition device and using the same technique as above.

However, there is a problem of recognition accuracy in conventional letter type recognition using character recognition. Generally, in order to increase the accuracy of character recognition, it is necessary to store possible words in advance as a dictionary. In particular, in order to recognize handwritten characters, the number of words must be reduced to about hundreds. However, in conventional character recognition devices used for letter type recognition, it is difficult to narrow down the range of words that can appear in advance. Therefore, it is difficult to identify the type of letter with sufficiently high accuracy.

In addition, in general character recognition, the result of character segmentation will leave ambiguity in the result of character recognition. For example, a plurality of candidate characters for character recognition can be obtained for each partial image divided by normal character recognition. There are also times when text segmentation itself leaves ambiguity. Estimating the frequency of appearance of a specific word from such a result of character recognition is not an obvious process, and the method of calculating the frequency of appearance of a word from text data cannot be used as it is. Also, if such ambiguity is not allowed and the character recognition result is treated as text, many words in the document will be missed.

In addition, in character recognition, recognition errors cannot be avoided. Due to these errors, it often happens that the type of the document cannot be recognized and an error occurs. In the existing method, when such inability to identify and misidentify occur, the work efficiency will be greatly reduced.

Summary of the invention

The first problem to be solved by the present invention is to realize high-speed and high-accuracy recognition of the type of letters using character recognition in such an inquiry-response support system.

The second problem to be solved by the present invention is to improve maintainability of a dictionary for character recognition.

The third problem to be solved by the present invention is to improve the interface of character recognition processing with ambiguous results and document classification processing of input text data, and improve the affinity in the system.

The fourth problem to be solved by the present invention is to provide an environment in which an answering task can be continued efficiently when the document type cannot be recognized or the recognition error occurs.

A collection of important words in document type recognition is used as a word dictionary for text recognition. In addition, instead of reading all the characters as in the past, word positioning technology is used to measure only the occurrence frequency of important words. The output format of the character recognition process is different from the conventional one in that a vector indicating the frequency of occurrence of a word is created. The obtained frequency of occurrence is input into the existing document type identification to identify the document type.

Not only the recognition result of the document type but also two or less candidates for the document type recognition are transmitted from the device performing character recognition to the device performing the answering task together with the word recognition result. The answering work device supports the answering work by highlighting important words on the image of the question letter. Also, an environment is provided in which a letter image can be forwarded to an appropriate respondent by using a candidate document type of 2 digits or less when an error occurs in the document type.

Description of drawings

Fig. 1 is the hardware structure of the embodiment.

FIG. 2 is a data flow diagram showing the processing flow of the embodiment.

FIG. 3 is a data flow diagram showing the flow of learning processing.

FIG. 4 is a data flow diagram showing the flow of e-mail classification processing.

FIG. 5 is a data flow diagram showing the flow of letter sorting processing.

FIG. 6 is a data flow diagram showing the flow of answer processing.

Fig. 7 is a display screen for answering tasks.

FIG. 8 is an operation sequence diagram showing a learning process procedure.

Fig. 9 is an operation sequence diagram showing a letter processing procedure.

Fig. 10 is the data form of the word occurrence frequency.

Fig. 11 shows the data format of data for answering tasks.

Figure 12 is an example of an input image.

Figure 13 is a split virtual network.

Figure 14 shows the detected important words.

Detailed ways

FIG. 1 shows the configuration of an inquiry answering system according to an embodiment of the present invention, that is, a support system for automatically classifying and answering inquiries from customers. The input to this system is an inquiry by e-mail, letter, telephone or the like. The output is an answer to an inquiry by e-mail, letter, telephone, or the like. In order to communicate with the outside, the system contacts the outside 1 through an interface and a telephone line. In addition, the computers of this system exchange information via LAN.

Before receiving an inquiry using this system, it is necessary to calculate necessary information for identifying the type of the inquiry document and generate a dictionary, that is, learning is required. 101 is a learning computer for managing learning. The computer for learning 101 refers to learning data collected in advance in the data file system 102 for learning, calculates information necessary for document type recognition, and stores it in the dictionary file system 103 as a dictionary for classification. The so-called learning data is a collection of text data that textualizes the question content and its question type identifier pair. The text data in the learning data uses examples of past queries. The corresponding document type is specified manually. The generated classification dictionary is copied to the dictionary file system 107 of the electronic mail classification computer 106, the dictionary file system 109 of the letter classification computer 108, and the dictionary file system 115 of the voice query classification computer 114 via LAN at any time.

Inquiries by electronic mail are received by the electronic mail sorting computer 106 from the Internet 104 outside the system via the gateway 105 . The e-mail sorting computer 106 recognizes the type of e-mail based on the content of the inquiry, and transfers the recognition result of the document type and the positions of important words described later to the automatic answering computer 116 in association with the e-mail.

The inquiry letter is photoelectrically converted by a scanner 110 with a sorter connected to the letter sorting computer 108, and input as an image. The letter classification computer 108 recognizes characters in the image by using character positioning technology described later, and recognizes the type of letter from the content of the inquiry. The recognition result is transferred to the computer 116 for automatic answering in association with the position of the image and the important words described later. Inquiries sent by facsimile are also sent from the telephone line 111 to the letter sorting computer 108 via the line control device 113, and the same processing is carried out. After the above processing, the letters are sorted and stored by the classifier of the scanner according to the identification results of the document types.

Inquiry by telephone is sent from the telephone line 111 through the line control device 113 to the computer 114 for voice inquiry classification. The voice inquiry sorting computer 114 recognizes the voice and converts it into text, sorts it according to the content of the inquiry, and transfers it to the answering phone 112 . An expert worker corresponding to the content makes an answer using the telephone 112 for answering the call.

The computer 116 for automatic answering, when the type of the transferred document can be answered automatically, retrieves a suitable answering sentence example from the answering sentence example file system 117, and replies by e-mail, or prints it in the form of a letter by using the automatic sealing printer 118 answer. Also, if the automatic answer is not possible, it is forwarded to the answering operation device (121, 125, 126) where an appropriate expert is on standby according to the document type corresponding to the inquiry document.

As indicated by 121, the composition of the answering operation device includes a computer 122, an input device 123 composed of a keyboard and a mouse, and an image display device 124. Using these devices, each worker generates an answer document by referring to the inquiry document, and transfers it to the computer 116 for automatic answering. The computer 116 for automatic answering sends out an answering e-mail or prints out an answering letter in the same manner as above.

The following uses the data flow diagram in Figure 2 to describe the processing flow of this system. In this figure, according to the notation of ゲ-ン・サ-ソン (J. マ-チン "Software Structured Technology" Modern Science Society, ISBN4-7649-0124-2 C3050 P5562E), the solid arrow indicates the information flow, and the hollow Arrows indicate logistics. In addition, the rounded rectangle represents processing, and the empty rectangle on the right represents stored information.

First, learning data 201 is used to perform learning 202 , that is, a process of calculating information necessary for identifying the type of query document, and word statistics dictionary 203 and important word dictionary 204 are generated. The important word dictionary 204 stores a collection of words that are important features in identifying document types. The word statistics dictionary 203 stores statistics necessary to identify the type of document based on the frequency of occurrence of important words. This processing is realized by the learning computer 101 .

When receiving an inquiry e-mail, the e-mail classification 205 refers to the word statistics dictionary 203 and the important word dictionary 204 to identify the type of the inquiry document. The type of the query document obtained is output as data for answering work corresponding to the electronic mail itself and the position information of important words obtained during processing. This process is realized by the computer 106 for e-mail sorting.

When there is a query in the letter, the letter classification 206 refers to the word statistics dictionary 203 and the important word dictionary 204 to identify the type of the query document. The type of the obtained question document is outputted as data for answering tasks in correspondence with the appearance positions of important words obtained in the process of document image processing. On the other hand, for the convenience of storage, etc., the letters themselves are classified according to the types of the obtained query documents. This process is realized by the computer 108 for sorting letters and the scanner 110 with a sorter.

Using the answer data, an answer message is generated from the answer 208, and an answer letter is printed or an answer e-mail is sent. At the same time, in order to update the word statistics dictionary 203 , the frequency of occurrence of important words and the document type information identified at the time of answering are output to the dictionary update 210 . When the input is e-mail, the text data to which the type of the document and the content of the e-mail are added is learning data. By performing relearning using such additional learning data, it is possible to adapt the word statistics dictionary 203 and the important word dictionary 204 to actual conditions of use. This processing is realized by the computer 116 for automatic answering, the automatic sealing printer 118 and the answering operation devices 121 , 125 , and 126 .

In dictionary update 210, the classification dictionary is updated based on the frequency of occurrence of important words in the query document and the type of document obtained during operation. The updating of the statistics used for this recognition is accomplished using standard pattern recognition methods. This processing is realized by the learning computer 101 .

In the case of an inquiry by telephone, the type of the inquiry content is recognized by voice inquiry classification 207, and the answering call 208 is made by an expert worker corresponding to the content. This processing is realized by the computer 114 for voice inquiry classification and the telephone 112 for answering calls.

Next, the processing flow of learning 202 is shown in the data flow diagram of FIG. 3 . First, important words for recognizing document types are extracted from the learning data 201 by the important word extraction 301 and stored in the important word dictionary 204 . This processing is realized by using morpheme analysis technology of natural language and feature selection technology of pattern recognition. Next, the statistics necessary for identifying the document type are calculated from the word statistics calculation 302 and stored in the word statistics dictionary 203 . The recognition method here uses standard methods in pattern recognition, such as quadratic recognition functions and neural networks. When the secondary recognition function is used, word statistics mean the frequency of occurrence of each word and its covariance coefficient. In the case of neural networks, word statistics are the connection weights of the network.

Next, the processing flow of the e-mail classification 205 is shown in the data flow diagram of FIG. 4 . First, the word extraction 401 refers to the important word dictionary 204 to calculate the frequency of appearance of each important word in the electronic mail. To detect important words, general language processing techniques such as morphological analysis are used. Next, the document type identification 402 uses word statistics in the word statistics dictionary 203 to identify document types. Recognition utilizes standard pattern recognition methods such as quadratic recognition functions and neural networks. Finally, the data for answering task generating 403 generates and outputs the data for obtaining the correspondence relationship between the occurrence positions of important words and the recognition results of document types and e-mails, that is, answering task data. Usually, a plurality of candidates with different reliability levels can be listed for the document category. All are stored in the data for answering assignments.

Next, the processing flow of letter classification 206 is shown in the data flow diagram of FIG. 5 . First, the area where the article is written in the letter is input as an image. Next, the word location recognition 502 refers to the important word dictionary 204 to identify important words from the image, and outputs the frequency of occurrence of each important word. Next, the document type identification 503 refers to the word statistics stored in the word statistics dictionary 203 to identify the document type based on the frequency of occurrence of important words. Next, from the data generation 505 for answering tasks, the frequency of appearance of images and important words and the recognition results of document types are output in association with each other. The letter itself is classified according to the document type identification result.

Next, the processing flow of the answer 208 will be described using the data flow in FIG. 6 . First, by sending destination determination 601, answering task data is transferred to answering tasks 1 to 3 (602, 603, 604) or automatic answering 605 according to the document type to which the answering task data is added. This process is realized by the computer 116 for automatic answering. In answering tasks 1 to 3 (602, 603, 604), expert staff for each document category discusses the content of the inquiry and generates an answer text. These are realized by the workers using the answering work devices 1 to 3 (121, 125, 126). In the automatic answer 605, an answer sentence example corresponding to the document type is retrieved from the answer example text collection 606 and output. This processing is realized by the computer 116 for automatic answering. An answer e-mail sending 607 sends an e-mail example of an answer sentence obtained by answering assignments 1 to 3 (602, 603, 604) or the automatic answer 605. FIG. This process is realized by the computer 116 for automatic answering. In addition, the answer text printing 608 prints the answer text on paper to generate an answer letter. This process is realized by the automatic seal printer 118 .

In fact, computer document type recognition is not necessarily correct. In addition, in the recognition process, sometimes the recognition cannot be recognized or the recognition is rejected. Therefore, in this system, errors and rejections of document type identification are dealt with as follows. In answering task 1, 2, or 3 (602, 603, 604), if the assigned query document is not in the field that the employee is in charge of a lot, the worker performs a forwarding operation on an operation screen described later. As described above, a plurality of candidates with different degrees of reliability are usually available as a document recognition result. Utilizing this point, when a forwarding operation is performed, the answering work data is automatically forwarded to the forwarding destination corresponding to the document type candidates of 2 or less digits. The forwarding destination can also be specified by the worker himself. In addition, when the recognition is rejected, important words detected in the text of the e-mail or in the image are highlighted to support the answering task.

FIG. 7 shows an example of a display screen on the image display device in the answering operation devices 1 to 3 ( 121 , 125 , 126 ). An inquiry document is displayed in an inquiry document window 708 on the screen 701 . Text is displayed in the case of an e-mail inquiry document, and a letter image is displayed in the case of a letter form inquiry. In addition, by using the important words appearing positions in the answering task data, the important words are highlighted in the same window, so that answering tasks can be easily generated. The worker edits the answer text in the answer text editing window 709 . A normal word processor can be used here. Also, when the document displayed in the query document window 708 is not the document under his/her responsibility, the worker can click the automatic transfer button 703 using a mouse provided as an input device. Accordingly, the data for answering work is transferred to the answering work device or the computer for automatic answering corresponding to the two or less candidates of the recognition result. In window 702 , candidates of document recognition results including candidates of 2 or less digits are displayed. When the operator specifies to forward the data for answering work to the destination, he clicks the forward button 704 after specifying the type of document using the radio buttons provided in the window 702 . When it is desired to search past sentence examples corresponding to the document type, the sentence example search button 705 is clicked. Then, the answer sentence example is transferred from the answer sentence example file system 117 via the LAN, and displayed on the answer text editing window 709 . If the send button 706 is clicked, the edited answer text will be sent back to the person who sent the email inquiry. In addition, when the print button 707 is clicked, the edited reply text is printed by the automatic sealing printer 118 .

FIG. 8 shows the process steps of learning 202 . In extracting important words, first, words are extracted from each learning text data Ti (1≤i≤N, N: number of data) by morpheme analysis, and the frequency of appearance is stored as a vector ui=(ui1, ui2, ... uiM) (1≤i≤N, .uij: number of occurrences of word j in Ti, M: total number of words). From the set {(ui\ci)} of the vector ui and the type ci (manually assigned) of each text data Ti, use known methods such as feature selection of the Branch and Bound (branch and bound) algorithm to select important words in the classification M' (M'<<M). Important words can also be manually selected when needed. Next, in calculating word statistics, the frequency of occurrence of each important word is calculated and stored as a vector vi=(vi1, vi2, ... viM'). In addition, statistics necessary for document type identification are calculated. For example, when the identification method is a quadratic identification function, statistics such as the average and correlation coefficient of the variables vi1, vi2, ... viM' are calculated.

FIG. 9 shows the processing steps of letter classification 206 . This processing includes steps of image input, character location recognition, document type recognition, classification, and generation of data for work.

In normal character recognition, all characters in an image are recognized. On the other hand, in word positioning recognition, a word to be read from outside is specified before execution. In the recognition process, the type of characters to be recognized is limited to characters that may appear in the designated word, and a seemingly correct character string is detected as the designated word. In this embodiment, words are recognized from an image by the method described in JP-A-11-85909, and the frequency of appearance is calculated. This processing includes a virtual generation step of character segmentation, a step of exploring and identifying important words, and a step of calculating the frequency of occurrence of important words. The frequency of occurrence of important words is represented by a vector w=(w1, w2, . . . wM'). By using this method of exploratory recognition for a given set of words, the accuracy and speed of recognition can be greatly improved. In addition, although the recognition and accuracy are inferior to this method, it is also possible to recognize all the characters in the image as in the existing method, and use the existing word comparison technology to obtain the vector w.

In document type recognition, general pattern recognition methods such as quadratic recognition functions and neural networks are used to calculate the reliability of each document type from the vector w and word statistics, and assign them in order according to the reliability of the document type. In addition, following a general method, when the difference between the reliability of the 1-digit and 2-digit document type candidates is smaller than a certain value, and when the reliability of the 1-digit document type candidate is smaller than a certain value, it is determined that recognition is rejected.

When categorizing, according to the recognition result of the document type, the scanner 110 with a sorting and sorting device is controlled to classify letters into specified document piles.

In generating answer data, data corresponding to an appearance position of an important word, a recognition result of a document type, and an image, ie, answer work data, is generated and output.

Fig. 10 shows the data form of the frequency of occurrence of words recognized as a word location. This is an array consisting of M' records. The frequency of occurrence of the M'th important word is stored in each record. The stored frequency of occurrence can be an integer value or a real value corresponding to the reliability of recognition.

Fig. 11 shows the data format of data for answering tasks. Stored in the variable kindOfMessage 1101 is a flag for indicating whether the query is expressed in text such as an e-mail or in an image such as a letter or a facsimile. The variable sizeOfMsg 1102 indicates the size of the document stored in the answer job data. Next, the inquiry letter entity is stored in the area 1103 of sizeOfMsg bytes. Store text in e-mail, etc., and store image data in letters and faxes. The variable numberOfCandidate 1104 stores the number of document type candidates obtained as a result of document type identification. Next, in the area 1105, document type candidate records of the numberOfCandidate number are stored. The configuration of each record includes a pair of an integer identifier indicating the type of document and its reliability value. The number of important words detected is stored in the variable numberOfWords 1106 . Next, the numberOfWords number of important word detection results records are stored in the area 1107 . The configuration of each record includes a pair of an identifier wordID of an important word and a record location indicating a detected position. As the detection position, in the case of text data, the number of bytes in which the first character of an important word in the text data appears. In the case of image data, the coordinates of the upper end, lower end, left end, and right end of the region where important words are recognized are stored.

The outline of character position recognition will be described below. Fig. 12 schematically shows an example of a letter of inquiry. Usually, there is no specific format in an inquiry letter. Therefore, the position of the character line and the size of the character cannot be known in advance. In addition, many occasions do not know whether to write horizontally or vertically. In addition, the interval between lines of characters like this example is very small, and it is sometimes difficult to judge whether the dot above "え" in the second line belongs to the previous line or the next line.

In order to solve such a problem, the method of Japanese Patent Laid-Open No. 11-85909 is adopted in the character positioning recognition of the present invention. In this method, character pattern candidates are extracted from an input image, these relationships are expressed in a segmented virtual network, and then pre-specified words are tentatively recognized in the segmented virtual network. This is a method by which a word can be recognized with high precision and high speed by utilizing the information of the word in a predictive manner. As a method of extracting the candidate of the character pattern, it is possible to use the combination of any number of connected components in the character line, and then select and combine them so that the height and width of the figure obtained by combining them are at the upper limit and lower limit specified in advance. method between values. As an exploratory method, a common breadth-first search is used, and the expansion of the search tree is determined based on the result of character recognition.

In JP-A-11-85909, in order to solve the difficulty of cutting out characters in a character line, a divided virtual network is introduced. In the example of FIG. 12 , it is also difficult to extract character lines before character extraction. Therefore, in the character positioning recognition of the present invention, while extracting the candidates of the character pattern from the whole image, the virtual network is divided to indicate whether the candidates of the character pattern can be connected in any direction in the vertical and horizontal directions, that is to say, whether the characters can be connected together as Words are read together. FIG. 13 shows an example of the divided virtual network thus obtained. Ellipses in the figure indicate candidates for the character mode. For example, 1301 shows that two connection components are combined to generate one character pattern candidate. In this case, character pattern candidate 1301 corresponds to "ら". In addition, the candidate pattern 1302 corresponds to the dot above "ら". A side 1303 indicates that the candidate pattern 1302 and the candidate pattern 1304 can be connected, that is, it indicates that they can be connected as characters in a word. In addition, here, when the side 1303 goes out from the inside, it means that the candidate pattern 1301 and the candidate pattern 1304 may be connected. Whether it can be connected depends on the distance between the alternate modes. Connection is possible if the distance is below a predetermined threshold.

Using the segmented virtual network obtained in this way as input, it is possible to search for important words using character recognition, and it is possible to detect places where important words appear. For example, when "ちらし" and "price" are important words, as shown in 1401 and 1402 in FIG. 14, the position of the important words can be detected. By performing character recognition using a word dictionary composed of the minimum words required for document classification, it is possible to perform automatic classification of letters with high precision and speed.

The word dictionary for the character recognition device can be easily created by sharing it with the important word dictionary for e-mail classification. In addition, the word dictionary can be automatically updated based on examples obtained by e-mail during operation.

The output of the character recognition processing is the frequency of occurrence of words, and the affinity improvement in the document type recognition processing and system based on the frequency of occurrence of words. Therefore, most existing document type identification devices can be used, and the document type identification based on text recognition and the document type identification based on text can easily coexist in the system.

When classification is not possible, it is also possible to support the answering worker's work by pointing out important words on the image to the answering worker. In addition, even in the event of a wrong classification, an environment can be provided in which the answering work can be continued efficiently.

Claims (10)

1. document classification system is characterized in that comprising:

Be used for importing the input media of the view data of document,

The memory storage of the information of the relevant important words of in the kind identification of above-mentioned document, using of storage and the frequency of occurrences thereof, and

Handle the treating apparatus of above-mentioned view data,

Above-mentioned treating apparatus, the word location technology of the above-mentioned important words of utilization from the view data of above-mentioned input media input, count its occurrence number, according to the information and the above-mentioned counting that are stored in the above-mentioned memory storage, discern the kind of above-mentioned document, above-mentioned document kind recognition result and above-mentioned view data are exported accordingly.

2. the document classification system that puts down in writing as claim 1 is characterized in that

Said system also has display device,

Above-mentioned display device shows above-mentioned document kind recognition result and above-mentioned view data accordingly.

3. the document classification system that puts down in writing as claim 2 is characterized in that

Said system will be recorded in the pen recorder accordingly to the answer literary composition example and the mentioned kind of above-mentioned document,

Above-mentioned display device shows above-mentioned view data and the corresponding above-mentioned answer literary composition example of above-mentioned document kind recognition result.

4. the document classification system that puts down in writing as claim 2 to 3 is characterized in that

Above-mentioned treating apparatus is also exported the positional information of important words in above-mentioned view data of above-mentioned counting,

On above-mentioned display device, emphasize to show this important words in the above-mentioned view data according to above-mentioned positional information.

5. as any one document classification system that puts down in writing in the claim 1 to 3, it is characterized in that

Said system has the classification and ordination device,

Above-mentioned classification and ordination device will be sorted out discharge by kind through the above-mentioned document of identification.

6. as any one document classification system that puts down in writing in the claim 1 to 3, it is characterized in that

Above-mentioned document classification system is connected to communication network,

Above-mentioned treating apparatus also carries out mentioned kind identification for the Email that receives through this communication network.

7. as any one document classification system that puts down in writing in the claim 1 to 3, it is characterized in that

Above-mentioned treating apparatus utilizes the occurrence number of above-mentioned recognition result and above-mentioned important words that the information that is stored in the above-mentioned memory storage is upgraded.

8. a document classification system is characterized in that

Have document kind note identification apparatus and a plurality of document processing device, document processings that are connected through network,

Above-mentioned document kind note identification apparatus comprises:

Obtain the view data of document or the device of text data,

The pen recorder that the information of the important words of using in will discerning with this kind about the kind of above-mentioned document writes down accordingly,

Handle the treating apparatus of above-mentioned view data or text data,

Above-mentioned treating apparatus is discerned above-mentioned document according to above-mentioned information, each confidence level of this recognition result and mentioned kind is together exported, according to the above-mentioned document processing device, document processing of this confidence level decision output.

9. the document classification system that puts down in writing as claim 8 is characterized in that

Above-mentioned treating apparatus, in the occasion of accepting the wrong input of this identification,

According to above-mentioned confidence level, above-mentioned document is transferred to other above-mentioned document processing device, document processings.

10. program is a kind of by carrying out the program of document recognition method with the computing machine that the view data input media is connected, has data storage device and control device,

It is characterized in that

The document recognition methods comprises:

Obtain the step of document data by above-mentioned view data input media,

Utilize the word location technology from above-mentioned view data, to discern the important words of depositing in advance in the above-mentioned memory storage, count each the step of occurrence number of this important words,

According to the step of above-mentioned count value identification document kind, and

The step that above-mentioned document data and above-mentioned document kind recognition result are exported accordingly.

Images