Method for expecting correction information in a natural language processing system

The present invention relates to a natural language processing system, as well as to a natural language processing method, for processing entered natural language information and, more particularly, to a natural language processing system and method for recognizing or correcting natural language information.

Natural language recognizing systems have been known such as an optical character reader (OCR), a voice recognition machine, or an on-line or off-line hand-written character recognition machine. Such known natural language recognition system employs a pattern matching technique in which a character is identified as a candidate character which is selected from a group of candidate characters and which exhibit's the best pattern matching with the character to be recognized.

Systems also have been proposed and used for verifying whether a row of recognized characters form a valid word, by making a reference to a dictionary.

It is true that the recognition efficiency can be remarkably improved through restricting the recognition results to valid words. Practically, however, it is often experienced that two or more valid candidate words are presented as a result of the recognition and verification so that the identification of the word cannot be performed.

It is also to be pointed out that the known recognition systems, when attempting to identify a word, do not take into consideration that a row of characters entered is a part of phrase or sentence which is described in a certain language and that each kind of information in a given context is followed by a specific type of information in accordance with meaning and/or grammatical construction of the sentence.

Operations of known natural language systems are based on an assumption that the entered information is correct without doubt. This causes an impediment to the construction of a natural language recognizing/processing system in which the processing system performs processing on information which is output from the recognition system and which may contain an error.

It is considered that execution of processing for expecting correct information as the input to the processing system, based on knowledge of a specific domain or context, contributes to correction of entered natural language information, even when the entry is made through means other than a recognition system, e.g., a keyboard. For instance, any typographical error incurred during entry can be corrected without difficulty by using a simple interface which is designed to restrict the scope or type of the information to be entered.

Accordingly, an object of the present invention is to provide a natural language processing method, as well as a natural language processing system, which, when processing a series of natural language information having a regular construction, expects the natural language information to be processed and executes required processing based on the expected information.

Another object of the present invention is to provide a natural language processing method, as well as a natural language processing system, which, when recognizing natural language information, expects the information to be processed and restricts the candidate information based on the result of the expectation.

Still another object of the present invention is to provide a natural language processing method, as well as a natural language processing system, which, when recognizing natural language information, expects the information to be processed and performs correction of the recognition based on the result of the expectation.

A further object of the present invention is to provide a natural language processing method, as well as a natural language processing system, which, when processing natural language information, expects the information to be input and performs correction of the input information based on the result of the expectation.

According to one aspect, the present invention which achieves these objectives relates to a natural language processing apparatus comprising: recognition means for recognizing natural language information; knowledge base means for storing knowledge concerning the types of said natural language information and knowledge concerning restriction in regard to the sequence of units of a series of natural language information groups of a plurality of types; and expectation means which, when said recognition means progressively and sequentially recognizes said series of natural language information groups of a plurality of types, makes reference to the knowledge stored in said knowledge base means so as to expect the subsequent object of recognition.

According to another aspect, the present invention which achieves these objectives relates to a natural language processing apparatus comprising: input means for inputting natural language information; knowledge base means for storing knowledge concerning the types of said natural language information and knowledge concerning restriction in regard to the sequence of statement of a series of natural language information groups of a plurality of types; and expectation means which expects information regarding a part of the series of natural language information groups of a plurality of types to be inputted through said input means, by making reference to the knowledge stored in said knowledge base means.

According to still another aspect, the present invention which achieves these objectives relates to a natural language processing method comprising: recognizing step in which natural language information is recognized; and expecting step in which, when a series of natural language information groups of a plurality of types is progressively and sequentially recognized in said recognizing step, the subsequent object to be recognized is expected, with a reference to a knowledge base which contains knowledge concerning the types of said natural language information and knowledge concerning restriction in regard to the sequence of units of said series of natural language information groups of a plurality of types.

According to yet another aspect, the present invention which achieves these objectives relates to a natural language processing method comprising: inputting step in which natural language information is inputted; and expecting step in which information regarding part of said series of natural language information groups inputted in said inputting step is expected, with reference to knowledge stored in a knowledge base which stores knowledge concerning the types of said natural language information and knowledge concerning restriction in regard to the sequence of statement of said series of natural language information groups of a plurality of types.

Embodiments of the present invention will now be described with reference to the drawings in which: <ul id="ul0001" list-style="none" compact="compact"><li>Fig. 1 is a block diagram showing the hardware structure of a natural language processing system embodying the present invention;</li><li>Fig. 2 is an illustration of the detail of a knowledge base;</li><li>Fig. 3 is a flow chart showing a process for processing recognition result structure;</li><li>Fig. 4 is a flow chart showing a recognition processing;</li><li>Fig. 5 is an illustration of an example of a census family register;</li><li>Fig. 6 is a chart illustrative of a knowledge structure for an action;</li><li>Fig. 7 is a chart illustrative of a hierarchy of expression of a Japanese address;</li><li>Fig. 8 is an illustration of knowledge contained in an ordinary dictionary;</li><li>Fig. 9 is an illustration of a general rule of description;</li><li>Fig. 10 is an illustration of a rule of description in the domain of census family register;</li><li>Fig. 11 is an illustration of the structure of a prepared census family register;</li><li>Fig. 12 is an illustration of the structure of an output census family register information;</li><li>Fig. 13 is a flow chart showing a census family register recognition processing;</li><li>Fig. 14 is an illustration of an out-door duty form;</li><li>Fig. 15 is an illustration of knowledge contained in an ordinary dictionary; and</li><li>Fig. 16 is an illustration of an example of the result of recognition.</li></ul>

Fig. 1 is a block diagram showing the hardware structure of a natural language processing system as an embodiment of the present invention. The hardware comprises an input unit 1, a CPU 2, an output unit 3, a program memory 4, a data memory 5 and a BUS 6.

The input unit 1 serves as means through which information is entered or inputted in the form of a natural language. In the following description, the input information is referred to as a "text". It is to be understood, however, that the present, invention can be applied to processing of natural language information of various kinds, even when the information is not a sentence which is grammatically complete, provided that the information has a structure which follows a certain rule.

The input unit 1 may comprise, for example, a voice recognition apparatus which recognizes vocal input information, a keyboard having keys for entering characters, a character recognition machine for optically reading and recognizing characters from a document, an on-line/off-line hand-written character recognition device, or a receiver which receives information from another system, e.g., a receiver which receives results of recognition performed by a character recognition system. The arrangement may be such that information generated through a different processing performed in the same system is used as the input information. It is also possible to arrange such that two or more types of input unit 1 mentioned above are prepared for selective use.

The CPU 2 performs computation and logical judgment necessary for various kinds of processing, and controls the system components which are connected to the BUS 6.

The output unit 3 serves as means for outputting data obtained through an analytical unit. The output unit 3 may comprise, for example, a voice synthesizer which synthesizes voice information from character information and outputs the synthetic voice information, a display device such as a CRT or a liquid crystal display device, a printer for printing and outputting characters on a document, or a transmission device which transmits the information to another device or system such as a data base. The arrangement may be such that information output from the output unit 3 is used as information to be input to another processing unit in the same system. The arrangement also may be such that two or more types of the output units mentioned above are prepared and used selectively.

The program memory 4 serves to store programs which are used for various controls performed by the CPU 2 including processings which will be described later with reference to flow charts. The program memory 4 may comprise a ROM or a RAM which is adapted to be externally loaded with programs.

The data memory 5 stores data generated through various processings, as well as various kinds of knowledge of a later-mentioned knowledge base. The data memory 5 may be a RAM. The knowledge of the knowledge base may be loaded from an external non-volatile storage medium prior to the processing to be executed, or referred to each time such knowledge is required.

The BUS 6 provides paths for transmitting various signals such as address signals for appointing the components to be controlled by the CPU 2, control signals for controlling various components, data signals exchanged between different components of the system, and so forth.

Fig. 2 shows the detail of the construction of a knowledge base 22 provided in the data memory 5. The knowledge base 22 includes a world knowledge base 221 which contains general knowledge such as knowledge of action, a domain knowledge base 222 which possesses knowledge peculiar to the domain of the information to be processes, and a linguistic knowledge base 223 which contains linguistic information such as parts of speech, grammar and so forth.

Fig. 3 is a flow chart illustrative of a processing which is executed after the completion of a recognition process for the purpose of correcting the recognition result based on expectation information. Although not exclusive, the processing in this case is executed in word-by-word fashion.

In Step S301, initial expectation is set to expect a subsequent word.

Step S302 determines whether or not any subsequent word exists. The process terminates when there is no subsequent word. When any subsequent word exists, the process advances to Step S303 which checks validity of the word based on the expectation information. Correction is performed when the word is determined as being invalid. Step S304 conducts later-mentioned conceptual analysis on the basis of practical cases. In Step S305, the expectation is updated to expect the next word. The process then returns to Step S302.

Fig. 4 is a flow chart showing a process for incorporating the expectation information in recognition processing. It is assumed here that the object of the processing is sentences which extend over a plurality of pages each containing one or more blocks, and that the processing proceeds from the first block of the first page and advances to the next page after completion of processing of the last block of the first page.

Initial expectation is set in Step S401.

In Step S402, whether or not any subsequent page exists is determined. The process terminates when there is no subsequent page. If a subsequent page exists, the process proceeds to Step S403 which conducts page analysis. In Step S404, determination is made as to whether any subsequent block exists. If no subsequent block exists, the processing on the instant page is finished and the process returns to Step S402. However, if a subsequent block exists, the process advances to Step S405 which conducts block analysis.

Step S406 executes determination as to whether any subsequent word exists. If no subsequent word exists, the process returns to Step S402 after finishing the processing of the block. However, if a subsequent word exists, the process advances to Step S407 which performs conceptual analysis and then advances to Step S408 which updates the expectation. The process then returns to Step S406.

The process performed by the illustrated embodiment of the system in accordance with the invention will be described, on an assumption that the system is used for the purpose of processing Japanese census family register information. Japanese census family register information is not written in the form of complete sentences, but is expressed in the form of consecutive words which describe predetermined items such as birth, marriage and so forth of each individual in a predetermined sequence or order.

Fig. 5 shows an example of such a Japanese census family register form. In the following description of the embodiment, an assumption is made that information such as that shown in Fig. 5 is read by an OCR and then processed.

Fig. 6 is a chart showing the knowledge structure concerning 〈ACTION〉. S1 indicates agent of the action, S2 indicates a direct object of the action, S3 indicates an indirect object of the action, S4 indicates an action which is the reason for the indirect object, S6 indicates an action serving as an instrument, S7 indicates a place serving as a support, and S8 indicates time.

Fig. 7 shows hierarchy of information concerning addresses used in Japan.

Fig. 8 shows the contents of an ordinary dictionary. Concepts of information shown in the left-hand-side column is shown in the right-hand column.

Fig. 9 shows general rule of description of natural language information, the rule being stored in the aforesaid world knowledge base.

Fig. 10 shows general rule of description of natural language information pertaining to a specific domain, the rule being stored in the aforesaid domain knowledge base.

A description will now be given of the process performed by the system of one embodiment of the invention, by employing a practical example of information. More specifically, in the following description, it is assumed that Japanese statement reading<img file="EP0717364A2_D0001.tif" /><img file="EP0717364A2_D0002.tif" /> is recognized by a recognition technique using expectation and the recognized information is analyzed into corresponding conceptual structure.

In this case, based on knowledge that all the statement contained in the portion in question of the census family register information pertains to an action of<img file="EP0717364A2_D0003.tif" /> (notify or declaration), i.e., MTRANS for a certain EVENT which actually took place, empty instances are generated for MTRANS 1 and EVENT 1 shown in Fig. 11.

At the same time, the 〈Time-Phrase〉 of EVENT 1 is set as expectation information, with the owner of the block being determined as the agent of the EVENT.

The above-described input information is progressively and successively processed. Since the expectation is for the Time-Phrase, the sub-expectation is for 〈ERA〉, from the general rules shown in Fig. 9. Namely,<img file="EP0717364A2_D0004.tif" /> (Heisei),<img file="EP0717364A2_D0005.tif" /> (Showa),<img file="EP0717364A2_D0006.tif" /> (Meiji) and so forth are listed as candidate era. This expectation information is given to the recognition section to be used during recognition process. Thus, only few <i>kanji</i> characters are selected and presented for the recognition from among more than several thousands of <i>kanji</i> characters.

As described above, expectation information is given to the recognition section on word-by-word basis. The expectation information given on word-by-word basis is useful and further enhances the accuracy of the recognition, particularly when the character recognition in a character row is conducted in parallel fashion on a plurality of characters. For instance, the expectation information given on word-by-word basis, when used together with parallel processing of a plurality of characters, offers the following advantage. The first character<img file="EP0717364A2_D0007.tif" /> of the word representing an era<img file="EP0717364A2_D0008.tif" /> (Showa) and the first character<img file="EP0717364A2_D0009.tif" /> of the word representing another era<img file="EP0717364A2_D0010.tif" /> (Meiji) are similar in form to each other. Therefore, when one of these two characters is recognized as a correct character, there is a large possibility that the other character also is listed as a candidate, if the recognition is performed on character basis. Consequently, it is impossible to definitely determine the word, even though the expectation information is used. However, this problem is overcome when the result of the recognition of the second character, obtained through recognition processing conducted in parallel with the first character, is taken into consideration based on the expectation information given on word basis.

It is not essential that a plurality of characters are recognized simultaneously. Namely, the recognition may be conducted sequentially on consecutive characters. In such a case, the process may proceed while leaving a plurality of candidate characters undecided for the first character and one of such candidate characters is determined as the correct character in accordance with the result of recognition of the next character. In some cases, erroneous combination of characters can be excluded by effective use of plural candidate characters.

If<img file="EP0717364A2_D0011.tif" /> (Heisei) has been recognized and determined as being the correct word based on the expectation information, the concept of this word is stored as<img file="EP0717364A2_D0012.tif" /> in the slot S1 of TIME 2.

The next expectation is for the 〈year〉 which is expressed in terms of a 〈Number〉. The scope of the number is restricted, in accordance with the above-mentioned concept of 〈ERA〉. Namely, the number is limited to range from 1 to 64 when the recognized 〈ERA〉 is<img file="EP0717364A2_D0013.tif" /> (Showa), and to range from 1 to indefinite when the recognized 〈ERA〉 is<img file="EP0717364A2_D0014.tif" /> (Heisei), as can be seen from Fig. 8. The knowledge shown in Fig. 8 also contains the fact that<img file="EP0717364A2_D0015.tif" /> (first year) is represented as<img file="EP0717364A2_D0016.tif" /> Characters such as<img file="EP0717364A2_D0017.tif" /> (second) and so forth are therefore expected as the number which represents the year.

Written Japanese language does not have "space" between successive words, so that the number of characters which should be determined as the 〈Number〉 is indefinite. A concept of 〈YEAR MARK〉 is used in order to determine the length of the word expressing the year. In this case, the determined word is recognized as<img file="EP0717364A2_D0018.tif" /> based on the expectation information. This result of recognition is treated as 〈NUMBER (=1)〉 which is stored in the slot S1 of TIME 2. The next character<img file="EP0717364A2_D0019.tif" /> (year) is the 〈Year Mark〉 and is treated as a delimiter which is neglected. Thus, a delimiter is used to determine words from successive pieces of input information and also serves as an index which indicates the kind of information which precedes this delimiter, as well as the kind of information which follows the delimiter. Similarly,<img file="EP0717364A2_D0020.tif" /> (first = 1) is stored in the slot S2 of TIME 2, and<img file="EP0717364A2_D0021.tif" /> (month) is neglected as being 〈Month Mark〉.<img file="EP0717364A2_D0022.tif" /> (twenty seven) is stored in the slot S3 of TIME 2, and<img file="EP0717364A2_D0023.tif" /> (day) is neglected as being 〈Day Mark〉.

In the above-described process for recognizing the number<img file="EP0717364A2_D0024.tif" /> (twenty seven), since a number indicating a date has been expected, there is no risk that the number<img file="EP0717364A2_D0025.tif" /> is erroneously recognized through pattern matching as a different character<img file="EP0717364A2_D0026.tif" /> or<img file="EP0717364A2_D0027.tif" /> which is similar in shape to<img file="EP0717364A2_D0028.tif" /> Similarly,<img file="EP0717364A2_D0029.tif" /> is never recognized as<img file="EP0717364A2_D0030.tif" /> The 〈Day Mark〉 has been expected when the character<img file="EP0717364A2_D0031.tif" /> is recognized. This character<img file="EP0717364A2_D0032.tif" /> therefore is not erroneously recognized as being a different similar character<img file="EP0717364A2_D0033.tif" />

A word<img file="EP0717364A2_D0034.tif" /> (Yokohama) is then processed. This word is a 〈City Name〉, so that inference is made based on the rule shown in Fig. 10 to determine that an Address-Phrase has been started. The expectation information is therefore changed to ADDRESS. The name<img file="EP0717364A2_D0035.tif" /> (Yokohama) is stored in the slot S4 of ADDRESS 2. The next character<img file="EP0717364A2_D0036.tif" /> is neglected as being a 〈City Mark〉. The next word<img file="EP0717364A2_D0037.tif" /> (Tsurumi) is stored in the slot S3 of ADDRESS 2, and the subsequent character<img file="EP0717364A2_D0038.tif" /> (ku = ward) is neglected as being a 〈Ku Mark〉.

In the foregoing processing pieces of expectation information are given to the recognition section, in accordance with the number of candidates. When a city name is expected, it is not efficient to send the names of all Japanese cities as the pieces of expectation information. However, if the first character of the city name has been correctly identified, the number of the candidate city names can be restricted. For instance, when the first character is<img file="EP0717364A2_D0039.tif" /> city names such as<img file="EP0717364A2_D0040.tif" /> (Yokohama),<img file="EP0717364A2_D0041.tif" /> (Yokosuka) and other small number of city names starting with<img file="EP0717364A2_D0042.tif" /> are listed as the candidates which are sent as expectation information to the recognition section. When it is expected that the identified city name is followed by the name of a Ku (ward), all the Ku names may be sent to the recognition section so as to improve recognition accuracy, because the number of wards in a city is not so large.

The next word is<img file="EP0717364A2_D0043.tif" /> Since this word is a casemarker as shown in Fig. 8, an inference is made in accordance with the rule shown in Fig. 10 to determine that the processing of the Address-Phrase is over. The kind of expectation information is therefore changed to BIRTH which is a type of declarations.

Then. BIRTH as the concept of 〈BIRTH〉 is obtained, and inference is made in accordance with the rule shown in Fig. 10 so as to determine that the type of the Event 1 as the object of the MTRANS is BIRTH. The kind of expectation information is then changed to Time-Phrase.

The next word<img file="EP0717364A2_D0044.tif" /> (same) is a〈PRONOUN 1〉 and is regarded as information pertaining to the slot S2 (Month) of TIME 1 from the fact that the next to next word is<img file="EP0717364A2_D0045.tif" /> (Month Mark), the contents of slot 12 is copied to slot S2 of time 1. When this word is recognized, the content of the slot S1 of TIME 1 (Era, Year), which shows the preceding time information, is copied to form the information to be stored in the slot S1 of TIME 2. The word<img file="EP0717364A2_D0046.tif" /> (month) is neglected as being a 〈Month Mark〉, and the word<img file="EP0717364A2_D0047.tif" /> (twenty nine), is stored in the slot S3 of TIME 1. The subsequent single-character word<img file="EP0717364A2_D0048.tif" /> (day) is neglected as being a 〈Day Mark〉.

The next single-character word<img file="EP0717364A2_D0049.tif" /> (father) is determined to be to 〈Person 1〉<img file="EP0717364A2_D0050.tif" /><img file="EP0717364A2_D0051.tif" /> so that an inference is made to determine that the Time-Phrase is over and 〈Person〉 has started. The person identified by the word<img file="EP0717364A2_D0052.tif" /> can be identified from the block information.

The next word<img file="EP0717364A2_D0053.tif" /> belongs to the concept of MTRANS, so that an inference is made in accordance with the rule shown in Fig. 10 to determine that the MTRANS 1 has been done by the 〈Person 1〉. The 〈person 1〉 is therefore stored in the slot S1 of MTRANS 1, and the kind of expectation information is changed to the Time-phrase which is at the beginning of the {Diff Birthplace}, in accordance with the rule shown in Fig. 10.

The next word<img file="EP0717364A2_D0054.tif" /> belongs to the concept of 〈Entry in Family Register〉. Since no Time-Phrase could be found, the rule of Diff Birth-place does not apply in accordance with the rule shown in Fig. 10. Thus, an inference is made to conclude that the statement in question has been successfully analyzed. It is therefore inferred that the content of the 〈Address 1〉 of the MTRANS 1 is identical to the permanent domicile recorded in the family register holder, and the domicile information is copied to fill the 〈Address 1〉. Consequently,<img file="EP0717364A2_D0055.tif" /> is filled in 〈Address 1〉.

Fig. 12 shows the construction of the family register information produced from the input information through the processings described above.

The types of expectation information employed in the processings described above are fixed terms such as<img file="EP0717364A2_D0056.tif" /> (year) and<img file="EP0717364A2_D0057.tif" /> (month) or variable words such as numerals, e.g., dates, city names and so forth which can be determined by trailing words.

It is assumed here that an erroneous recognition result has been obtained to read<img file="EP0717364A2_D0058.tif" /><img file="EP0717364A2_D0059.tif" /> when the original statement reads<img file="EP0717364A2_D0060.tif" /><img file="EP0717364A2_D0061.tif" /> A description will now be given as to a process for correcting such an erroneous input.

In this case, based on knowledge that all the statement contained in the portion in question of the census family register information pertains to an action of<img file="EP0717364A2_D0062.tif" /> (notify or declaration), i.e., MTRANS for a certain EVENT which actually took place, empty instances are generated for MTRANS 1 and EVENT 1 shown in Fig. 11.

At the same time, the 〈Time-Phrase〉 of EVENT 1 is set as expectation information, with the owner of the block being determined as the agent of the EVENT.

The above-described input information is progressively and successively processed. Since the expectation is for the Time-Phrase, from the general rules shown in Fig. 9, the sub-expectation is for 〈ERA〉. Namely,<img file="EP0717364A2_D0063.tif" /> (Heisei),<img file="EP0717364A2_D0064.tif" /> (Showa),<img file="EP0717364A2_D0065.tif" /> (Meiji) and so forth are listed as candidate era.

Since the recognition result<img file="EP0717364A2_D0066.tif" /> (Heisei) is found in the expectation information, the concept of this result is stored in the slot S1 of TIME 2 as<img file="EP0717364A2_D0067.tif" />

The next expectation is for the 〈year〉 which is expressed in terms of a 〈Number〉. The scope of the number is restricted in accordance with the above-mentioned concept of 〈ERA〉. Namely, the number is limited to range from 1 to 64 when the recognized 〈ERA〉 is<img file="EP0717364A2_D0068.tif" /> (Showa), and to range from 1 to indefinite when the recognized 〈ERA〉 is<img file="EP0717364A2_D0069.tif" /> (Heisei), as will be seen from Fig. 8. The knowledge shown in Fig. 8 also contains the fact that<img file="EP0717364A2_D0070.tif" /> (first year) is represented as<img file="EP0717364A2_D0071.tif" /> Characters such as<img file="EP0717364A2_D0072.tif" /> (second) and so forth are therefore expected as the number which represents the year.

Written Japanese language does not have "space" between successive words, so that the number of characters which should be determined as the 〈Number〉 is indefinite. A concept of 〈YEAR MARK〉 is used in order to define the length of the word expressing the year. In this case, the recognized word<img file="EP0717364A2_D0073.tif" /> is found in the expectation information And, therefore, is treated as 〈NUMBER (= 1)〉 which is stored in the slot S1 of TIME 2 as 〈Year (= 1)〉.

The next word<img file="EP0717364A2_D0074.tif" /> (year) is the 〈Year Mark〉 and is treated as a delimiter so as to be neglected. Similar processings are performed so that<img file="EP0717364A2_D0075.tif" /> (first = 1) is stored in the slot S2 of TIME 2, and<img file="EP0717364A2_D0076.tif" /> (month) is neglected as being 〈Month Mark〉.

A number indicative of a date is expected for the next word<img file="EP0717364A2_D0077.tif" /> It is therefore determined by inference that<img file="EP0717364A2_D0078.tif" /> has been wrongly recognized for<img file="EP0717364A2_D0079.tif" /> i.e., that<img file="EP0717364A2_D0080.tif" /> should be corrected to<img file="EP0717364A2_D0081.tif" /> Thus,<img file="EP0717364A2_D0082.tif" /> (twenty seven) is stored in the slot S3 of TIME 2, and<img file="EP0717364A2_D0083.tif" /> (day) is neglected as being 〈Day Mark〉.

The word<img file="EP0717364A2_D0084.tif" /> (Yokohama) is then processed. This word is a 〈City Name〉, so that inference is made based on the rule shown in Fig. 10 to determine that an Address-Phrase has been started. The expectation information is therefore changed to ADDRESS. The name<img file="EP0717364A2_D0085.tif" /> (Yokohama) is stored in the slot S3 of ADDRESS 2. The next character<img file="EP0717364A2_D0086.tif" /> is neglected as being a 〈City Mark〉.

As to the next word<img file="EP0717364A2_D0087.tif" /> a ward<img file="EP0717364A2_D0088.tif" /> name has been expected. The word cannot be found in the data base which lists names of all the wards existing in Yokohama city, whereas a similar ward name<img file="EP0717364A2_D0089.tif" /> is found in the data base. An inference is therefore made to determine that<img file="EP0717364A2_D0090.tif" /> is wrong and should be corrected to<img file="EP0717364A2_D0091.tif" /> The corrected ward name<img file="EP0717364A2_D0092.tif" /> is stored in the slot S4 of ADDRESS 2, and the subsequent word<img file="EP0717364A2_D0093.tif" /> (ku = ward) is neglected as being a 〈Ku Mark〉.

The next word is<img file="EP0717364A2_D0094.tif" /> Since this word is a casemarker as shown in Fig. 8, an inference is made in accordance with the rule shown in Fig. 10 to determine that the processing of the Address-Phrase is over. The kind of expectation information is therefore changed to BIRTH which is a type of declarations.

Then. BIRTH as the concept of 〈BIRTH〉 is obtained, and inference is made in accordance with the rule shown in Fig. 10 so as to determine that the type of the Event 1 as the object of the MTRANS is BIRTH. The kind of expectation information is then changed to Time-Phrase.

The next word<img file="EP0717364A2_D0095.tif" /> (same) is a〈PRONOUN 1〉 and is regarded as information pertaining to the slot S2 (Month) of TIME 1 and stored in the same slot of TIME 2. When this word is recognized, the content of the slot S1 of TIME 1 (Era, Year), which shows the preceding time information, is copied to form the information to be stored in the slot S1 of TIME 2. The word<img file="EP0717364A2_D0096.tif" /> (month) is neglected as being a 〈Month Mark〉. Since a 〈Day Mark〉 as a delimiter is expected subsequently, an inference is made that<img file="EP0717364A2_D0097.tif" /> is wrong and should be corrected to<img file="EP0717364A2_D0098.tif" /> Due to the presence of the delimiter, the subsequent word<img file="EP0717364A2_D0099.tif" /> (twenty nine) is cut out and stored in the slot S3 of TIME 1. The subsequent single-character word<img file="EP0717364A2_D0100.tif" /> (day) is neglected as being a 〈Day Mark〉.

The next single-character word<img file="EP0717364A2_D0101.tif" /> (father) is determined to be 〈Person 1〉<img file="EP0717364A2_D0102.tif" /> so that an inference is made to determine that the Time-Phrase is over and 〈Person〉 has started. The person identified by the word<img file="EP0717364A2_D0103.tif" /> can be identified from the block information.

The next word<img file="EP0717364A2_D0104.tif" /> belongs to the concept of MTRANS, so that an inference is made in accordance with the rule shown in Fig. 10 to determine that the MTRANS 1 has been done by the 〈Person 1〉. The 〈person 1〉 is therefore stored in the slot S1 of MTRANS 1, and the kind of expectation information is changed to the Time-phrase which is at the beginning of the (Diff Birthplace) in accordance with the rule shown in Fig. 10.

Next expectation is for Time-Phrase. The next word is<img file="EP0717364A2_D0105.tif" /> This word cannot be found as a part of Time-Phase so that a decision is made not to apply the rule of Diff birthplace, and the kind of expectation information is changed to Entry in Family Register<img file="EP0717364A2_D0106.tif" /> which belongs to the concept of 〈Entry in Family Register〉. An inference is therefore made that<img file="EP0717364A2_D0107.tif" /> is wrong and should be corrected to<img file="EP0717364A2_D0108.tif" /> An inference is therefore made to conclude that the statement in question has been successfully analyzed. It is therefore inferred that the content of the 〈Address 1〉 of the MTRANS 1 is identical to the permanent domicile recorded in the family register holder, and the domicile information is copied to fill the 〈Address 1〉. Consequently,<img file="EP0717364A2_D0109.tif" /> is correctly registered in 〈Address 1〉.

A description will now be given as to a recognition processing relying upon layout information, with specific reference to a flow chart of Fig. 13 showing this processing, on an assumption that the census family register information as shown in Fig. 5 has been input.

Referring to Fig. 13, Step S1301 conducts setting of blocks for a new page and determines the sequence of processing. Step S1302 gives expectation for a new block. Step S1303 determines whether expectation has been made for a new line. The term "line" is used here to mean a small unit which provides a meaning. When expectation cannot be done, the process proceeds to Step S1304 in which types of candidate lines are selected. In Step S1305, character recognition is executed on the selected types of line. The process then returns to Step S1303 which determines whether expectation has been made for the line type which has been found as a result of the character recognition.

If the determination in Step S1303 has proved that the expectation has safely been done, the process proceeds to Step S1306 which selects a list of candidate characters and the character rule which determines the break between successive words, for each of the words in the next line. In Step S1307, character recognition processings are executed on expected candidate characters. In Step S1308, meaning processing is executed to determine the meaning of the word. In Step S1309, meaning processing is executed for the sentence down to the instant ward.

Step S1310 executes determination as to whether or not the processing has been finished for the line. The process returns to Step S1306 if the processing has not been finished. If the processing has been completed for the line, the process proceeds to Step S1311 which executes meaning analysis of the line and determines whether the processing has been finished for the block. If the processing for the block has been finished, the process advances to Step S1312 which executes meaning processing on the block, followed by determination as to whether or not the processing is finished for the page. If the process has been finished for the page, the process proceeds to Step S1313 which determined whether or not the processing has been completed for the census family register. If the process has not been completed, the process returns to Step S1301, otherwise the results are entered in a data base.

A description will now be given of the census family register form shown in Fig. 5.

Each census family register form contains a plurality of pages, each including one or more blocks. Fixed positional or layout information indicative of the relationship between each page and blocks contained therein has been stored in a knowledge base of a specific domain, in order to enable extraction of all the necessary information.

The layout information is effectively used by the system when the system performs various processings such as setting of expectation information, checking of validity of information and identification of person indicated by a common noun such as<img file="EP0717364A2_D0110.tif" /> (father). In the illustrated example, the content of data is determined based on the block position and header, so as to facilitate the setting of the candidate context.

The block headed by<img file="EP0717364A2_D0111.tif" /> (permanent domicile) contains the actual address (Address-Phase) as data indicating a registered place. Setting of the expectation to 〈Address-Phrase〉 therefore helps the system in recognizing the data shown in this block.

The block headed by<img file="EP0717364A2_D0112.tif" /> (name) contains information concerning the register holder, written in a format which is family name-space-given name. In contrast, in the block relating to BIRTH declaration, the name is written without giving a space between the family name and the given name. This information contained in the block headed by<img file="EP0717364A2_D0113.tif" /> (name) can therefore be used effectively to assist analysis of the statement of a person's name in the block pertaining to birth declaration.

The next line contain information as to when the census family register in question was formed. Usually, a new family register is formed at the time of marriage. The marriage date is therefore set as being default.

Each page is divided into blocks allocated for individuals, so that each block contains BIRTH information and declaration of events concerning the individual person. The name (family name and given name) of the individual, name of father, name of mother, date of birth, sex, and other information such as the order of the individual among brothers and sisters, e.g.,<img file="EP0717364A2_D0114.tif" /> (eldest daughter) are written in lower part of the blank.

As will be seen from Fig. 5, each block is headed by a header which indicates the content of information contained in the block. Thus, the header<img file="EP0717364A2_D0115.tif" /> (father) naturally leads to an expectation that the content of the data in the block is the father's name.

Information or data mentioned above can effectively be used when the system confirms data written in declarations. For instance, data concerning the date of birth, which is clearly written in the BIRTH declaration can be checked up with the data concerning the date of birth appearing in a lower block. Similarly, persons indicated by common nouns such as<img file="EP0717364A2_D0116.tif" /> (father) and<img file="EP0717364A2_D0117.tif" /> (mother) can be correctly correlated to their names.

Declarations of event relating to a plurality of persons has a block for each of such persons, and the matter or the fact of declaration of event is written in each of the blocks. For instance, the fact of marriage is written both in the block for the husband and the block for the wife. Such knowledge assists the system in making confirmation and expectation.    a description will now be given of another example of recognition processing which relies upon layout information.

Fig. 14 shows an out-door duty form which is an object of recognition by the recognition system in accordance with one embodiment of the present invention.

In, this case, the system performs recognition and analysis of input information by using knowledge of area concerning the layout of this form. For instance, the heading reading "Employee code" leads to an expectation that a number follows this heading. Once the number is recognized, the system can make access to the data base in order to expect and set the name and the division of the employee. Any ambiguity which may exists in the result of recognition of the number can be removed when reference is made to the employee's name and the division. Some of the blanks have only limited candidates to fill such blanks. For instance, the line headed by "Mode of Transport" can be filled only by one of train, bus, taxi and personal car, or a combination thereof.

A description will now be given of the case where the invention is applied to voice recognition.

As in the case of character recognition, accuracy of voice correction is enhanced when the recognition is conducted with the help of expectation.

It is supposed here that a voice message "I shall meet you at five p.m.." has been input. Once the recognition has been completed till "I shall", it is expected that the subsequent word is a verb. In this case, although both "meet" and "meat" are listed as candidates through a voice pattern matching, the former, i.e., "meet", is selected since "meat" is a noun rather than a verb.

After the analysis has proceeded to "I shall meet you at", it is expected that "at" is followed by a noun indicating a place or time. The word "five", therefore, is never recognized as "fine", although the pronunciation is similar.

When the analysis has proceeded to "pm" while leaving "five" and "nine" as candidates, it might be considered that nine p.m. is inappropriate for a meeting time, from a knowledge base or the context, so that the "nine" is deleted from the group of candidates or put to a lower order of preference of candidate words.

It is assumed here that the above-mentioned message has been input through a keyboard, with the word "meet" wrongly spelled as "meat". In such a case, however, the system can automatically correct the word "meat" into "meet", since it expects a verb as the word which follows "I shall".

A description will now be given of an example in which a first candidate word for a certain part of the recognized result is changed into another candidate part, with a consequence that the first candidate for another part of the recognition result is changed to another candidate correspondingly.

Fig. 15 is an illustration of a common knowledge used in the above-mentioned example of processing, while Fig. 16 is a diagrammatic illustration of the recognition result to be processed in accordance with the method of one embodiment of the present invention. Encircled characters are the first candidates, and characters show below circles are other candidates. It is assumed that the user has changed<img file="EP0717364A2_D0118.tif" /> into<img file="EP0717364A2_D0119.tif" /> The system then reviews the recognition result which already has been acquired. The system recognizes that the word<img file="EP0717364A2_D0120.tif" /> does not make any sense and considers alternative words such as<img file="EP0717364A2_D0121.tif" /> and<img file="EP0717364A2_D0122.tif" /> The system then refers to the knowledge base as shown in Fig. 15 to reach a conclusion that the word<img file="EP0717364A2_D0123.tif" /> is inappropriate as a word which expresses a date. The system therefore selects<img file="EP0717364A2_D0124.tif" /> as the most appropriate candidate and changes<img file="EP0717364A2_D0125.tif" /> into numeral 1.

Obviously, the present invention can be carried out in various forms, provided that they can perform the function of the invention. Thus, the method of the, invention may be realized in the form of a single unit apparatus, a system composed of a plurality of apparatuses or a program which may be loaded on an apparatus or a system.

As will be understood from the foregoing description, according to the described various forms of invention, it is possible to progressively and sequentially process a series of natural language information groups of different types, while expecting, each time a recognition is performed, the object to be recognized subsequently.

It is also possible to input natural language information and to expect information of a part of the inputted series of natural language information groups of different types.

Furthermore, the present invention makes it possible to produce correct analysis result, even when any error is contained in the inputted information.

The present invention as described may be applied to a system which is composed of a plurality of computers or only to a selected computer of a system having a plurality of computers. The invention also can be applied when the computer is achieved by executing a program. Such a program may be supplied from an external memory medium, and such external storage medium also falls within the scope of the present invention.

The natural language which has been recognised by the natural language processing apparatus can form a command for carrying out the predetermined operation such as storage or display of the recognised natural language, or translation of the natural language into another natural language or another form of language e.g. instructions for a machine.

Although the present invention has been described in its preferred form with a certain degree of particularity, many apparently widely different embodiments of the invention can be made without departing from the scope of the claims. It is also to be understood that the invention is not limited to the specific embodiments.