JP4458517B2 - Information extraction apparatus and method - Google Patents

Description

  The present invention relates to an information extracting apparatus and method for extracting information including numerical information from given text sentence data.

  A method of extracting numerical data from document data and extracting noun data corresponding to the numerical data is known (for example, Patent Document 1).

  The above method will be briefly described. For example, a case will be described in which one sentence in the document data is “Today I bought 200 yen pork and 2000 yen cup”.

  First, document data is referred to one character at a time, and a character string with consecutive numbers is extracted as numerical data. For example, “200” and “2000” are extracted as numerical data.

  Next, a sentence in the range up to the punctuation mark including the numerical data is extracted as numerical character string data. For example, “Today I bought 200 yen pork” and “I bought a 2000 yen cup” are extracted as numeric character string data.

  Furthermore, morphological analysis is performed on the numeric character string data to extract nouns, and only nouns included in predetermined item classification character strings are extracted from these nouns as noun data. For example, “pork” and “cup” are extracted as noun data.

  As described above, “200” and “2000” can be obtained as numerical data, and “pork” and “cup” can be obtained as noun data corresponding thereto.

JP-A-8-329165

  However, when numerical data and noun data are extracted using the method described above, it is assumed that the noun data is in the vicinity of the numerical data. In other words, if the noun data corresponding to the numerical data in the sentence is described beyond the numerical data and the punctuation marks, the corresponding noun data cannot be extracted accurately.

  For example, in the sentence “Buy bought today was 200 yen”, “pork” and “200” are described beyond punctuation, so the numerical data and the corresponding noun data are accurately extracted. It is not possible.

  In the method as described above, only nouns included in the item classification character strings provided in advance are extracted as noun data. For this reason, any noun that is not included in the item classification character string cannot be extracted in association with numerical data.

  Further, since only nouns are extracted in association with numerical data, accurate extraction cannot be performed when the numerical data points to sentence elements other than nouns.

  The present invention has been made to solve the above-described problems, and can extract noun data accurately even when noun data corresponding to numerical data is not in the vicinity. It is another object of the present invention to provide an information extraction apparatus that can extract sentence elements other than nouns in correspondence with numerical data.

  (1) (2) (11) An information extraction apparatus or program according to the present invention is an information extraction apparatus for extracting information data including numerical information data from given text sentence data, and the given text sentence data Is divided into multiple sentence element data, the type of each sentence element is determined and recorded in the recording unit, and it is determined whether or not numerical information data is included in the given text sentence data. In the extraction / analysis means for identifying the numerical information data and recording it in the recording unit, with reference to the recording unit, the sentence element data including the numerical information data is identified, and based on the type of the sentence element data, At least one sentence element data having a dependency relationship with the numerical information data is determined, and dependency information extraction means for extracting the dependency information data from the determined sentence element data, and the numerical information extraction means extract And output means for outputting dependency information data extracted by the dependency information extraction means.

  Therefore, even when noun data corresponding to the numerical information data does not exist in the vicinity of the numerical information data, the sentence element data having a dependency relationship with the numerical information data is extracted, and the noun corresponding to the numerical information data is extracted. Data can be extracted accurately.

  (3) In the information extraction apparatus or program according to the present invention, the extraction / analysis means determines whether or not numerical information data is included in the given text sentence data, and if included, the numerical information data Numerical information extraction means for identifying and extracting and recording in the recording section, text sentence analysis for dividing the text sentence data into a plurality of sentence element data, determining the type of each sentence element and recording in the recording section And the dependency information extracting means determines which sentence element data contains the numerical information data recorded in the recording unit, thereby determining the sentence element including the numerical information data. It is characterized by specifying data.

  Therefore, even when noun data corresponding to the numerical information data does not exist in the vicinity of the numerical information data, the sentence element data having a dependency relationship with the numerical information data is extracted, and the noun corresponding to the numerical information data is extracted. Data can be extracted accurately.

  (4) In the information extraction device or program according to the present invention, the extraction / analysis means divides the text sentence data into a plurality of sentence element data, determines the type of each sentence element, and records it in the recording unit. Analyzing means, determining whether or not numerical information data is included in the sentence element data, and if included, numerical information extracting means for extracting the sentence element data and the numerical information data in association with each other; It is characterized by having.

  Therefore, even when noun data corresponding to the numerical information data does not exist in the vicinity of the numerical information data, the sentence element data having a dependency relationship with the numerical information data is extracted, and the noun corresponding to the numerical information data is extracted. Data can be extracted accurately.

  (5) In the information extraction apparatus or program according to the present invention, the text sentence analysis means divides the text sentence data into a plurality of sentence element data, and each sentence element data includes a subject, and a predicate The type of sentence element is determined as a predicate when it contains and as a subordinate phrase when neither subject nor predicate is included.

  Therefore, information can be accurately extracted based on the type of sentence element without being limited to the sentence structure in the text sentence data.

  (6) The information extracting device or program according to the present invention is a method for extracting dependency information when the type of sentence element data including numerical information data or the sentence element data extracted in association with the numerical information data is a subordinate phrase. The means is characterized by determining whether the word to be modified by the sentence element data including the numerical information data is a body word or a predicate, and determining sentence element data having a dependency relationship.

  Therefore, the information indicated by the numerical information can be extracted more accurately.

  (7) In the information extraction apparatus or program according to the present invention, when the word to be modified of the sentence element data including the numerical information data is a body word, the dependency information extracting means includes the sentence element data constituting the text sentence data. If the sentence element data continuous from the sentence element data including the numerical information data is determined as the sentence element data having dependency relation, and the word to be modified by the sentence element from which the numerical information is extracted is a predicate, the relation The received information extracting means is characterized in that sentence element data other than the sentence element data including the numerical information data is determined as a sentence element having a dependency relationship from among the sentence element data constituting the text sentence data. .

  Therefore, the information indicated by the numerical information data can be extracted more accurately. Thereby, semantically sophisticated information data can be extracted.

  (8) In the information extraction apparatus or program according to the present invention, a predetermined numerical unit is recorded in the recording unit in advance, and the extraction / analysis unit has the same character string as the numerical unit read from the recording unit, It is characterized in that whether or not numerical information data is included is determined based on whether or not the text sentence data or sentence element data exists.

  Therefore, only the target numerical information data can be extracted from the text sentence data. For example, when it is desired to obtain numerical information data for a monetary amount, numbers with “yen” and “¥” added before and after are extracted as numerical information data, and dependency information data corresponding to this is extracted. Can be extracted.

  (9) An information extraction apparatus or program according to the present invention determines text sentence data to be extracted based on sentence element data obtained by a text sentence analyzing means and numerical information data extracted by a numerical information extracting means. It is further characterized by further comprising an extraction target sentence determining means.

  Therefore, information can be accurately extracted not only when the text sentence data is a single sentence but also when it is a heavy sentence, a compound sentence or a mixed sentence having two or more sentences.

  (10) The information extraction device or program according to the present invention provides sentence element data obtained by excluding sentence element data including numerical information data and sentence element data including dependency information data from text sentence data as additional information. An additional information extracting means for extracting as data is provided.

  Therefore, it is possible to extract the numerical information data and the dependency information data corresponding thereto, and at the same time extract the additional information data related thereto.

  Embodiments of the present invention will be described below with reference to the drawings.

1. First embodiment 1-1. Functional Block Diagram FIG. 1 shows a functional block diagram of the information extraction apparatus according to this embodiment. In this figure, the information extracting apparatus according to the present invention comprises a cutting means 101, a numerical information extracting means 103, a text sentence analyzing means 105, a dependency information extracting means 107, and an output means 109.

  The cutout unit 101 cuts out text sentence data one sentence at a time from the input text document and records it in the recording unit. For example, a process of cutting out text sentence data is performed using a delimiter “.” Appearing in a text document.

  The numerical information extraction unit 103 refers to the text sentence data read from the recording unit in order from the top, and determines whether or not predetermined numerical information data is included. When predetermined numerical information data is included, the numerical information data is extracted and the text sentence is given to the text sentence analyzing means.

  The text sentence analysis unit 105 performs morphological analysis on the text sentence data read from the recording unit, and then analyzes each morpheme based on predetermined grammar definition information, and divides the text sentence data into predetermined sentence element data.

  The dependency information extracting means 107 is based on the sentence element data including the extracted numerical information data, and other sentence element data having a dependency relation with the numerical information data from the text sentence data divided into the sentence element data. To extract.

  The output unit 109 outputs the numerical information data extracted by the numerical information extraction unit 103 and the dependency information data extracted by the dependency information extraction unit 107. For example, the data is output to a display or a database.

1-2. Hardware Configuration FIG. 2 shows an example of a hardware configuration in which the information extraction apparatus shown in FIG. 1 is realized using a CPU. The information extraction apparatus includes a display 201, a CPU 203, a memory 205, a keyboard / mouse 207, a hard disk 209, a CD-ROM drive 211, and a communication circuit 215.

  The hard disk 209 includes an information extraction program 2091 for performing information extraction processing according to the present invention, a morphological analysis dictionary 2093 for performing morphological analysis processing, grammar definition information 2094 for performing syntax analysis processing, and predetermined numerical information data. A numerical unit master 2095 for extracting data, an extraction result database 2097 for recording extraction results by information extraction processing, and the like are recorded, and these read out data recorded on the CD-ROM 212 via the CD-ROM drive 211. Installed.

  The installation may be performed using data downloaded from the Internet 216 or the like using the communication circuit 215.

1-3. Flowchart Processing based on the information extraction program 2091 will be described using the flowcharts of FIGS. In the following, an example will be described in which a text document including the text sentence “Operating income is 32,763 million yen, which is the same level as the previous year” is input to the clipping unit 101.

  In the flowchart of the information extraction process shown in FIG. 3, the CPU 203 reads a text document and inputs it into the memory 205 (step S301).

  Text sentence data is cut out from the head of the input text document (step S303). In the present embodiment, the punctuation mark “.” Is used as a text sentence delimiter. For example, when the text sentence data is read, the first sentence, “Operating income is 32,976 million yen, which is the same level as the previous year.” Is extracted.

  The CPU 203 performs numerical information extraction processing based on the cut-out text sentence data (step S305).

1-3-1. Numerical Information Extraction Processing In the flowchart of the numerical information extraction processing shown in FIG. 4, the CPU 203 extracts unit characters of numerical information data based on the numerical unit master 2095 (step S401). For example, in this embodiment, unit characters relating to amounts such as “million yen” and “100 million yen” are recorded in the numerical unit master.

  In particular, the numerical unit master may be set according to the contents of a text document from which numerical information data is extracted. For example, in the case of securities reports and financial statements, “Yen”, “Billions of yen”, “Millions of yen”, “Thousands of yen”, etc., which are unit characters relating to monetary amounts, are set, such as historical chronology and newspaper articles In this case, “year”, “month”, “day”, etc., which are unit characters relating to the date, may be set. Thereby, only necessary information can be extracted according to the contents of the document.

  If the numerical unit master is not used, a predetermined unit character for the numerical information data may be set in the information extraction program 2091.

  The CPU 203 searches whether or not the extracted unit character exists from the head of the text sentence data (step S403). For example, it is searched from the head of the text sentence “Operating income was 32,763 million yen, which is the same level as the previous year”, whether or not the unit characters “million yen” or “100 million yen” exist.

  If a predetermined unit character exists in the text sentence data (step S405, YES), numbers consecutive in front of the unit character are extracted as constituent elements of the numerical information data (step S407). In this case, if a plurality of numbers are continuous, a character string of numbers obtained by concatenating all the continuous numbers from the head direction is extracted as a constituent element of the numerical information data. Further, the CPU 203 concatenates the constituent elements of the extracted numerical information data and the unit characters to obtain numerical information data.

  For example, in the case of the above text sentence data, “32776” is extracted as a constituent element of the numerical information data, concatenated with the unit character “million yen”, and “32776 million yen” is extracted as the numerical information data. 205 or a predetermined area of the hard disk 209.

  If there is another unit character in the text sentence data (step S409, NO), the CPU 203 repeats step S407. If there is no other unit character, the process is terminated (step S409). In addition, if no unit character exists in step S405, the process is ended (NO in step S405).

1-3-2. Text sentence analysis processing If numerical information data is extracted in the numerical information extraction process (FIG. 3, step S305) (step S307, YES), the CPU 203 performs a text sentence analysis process (step S309).

  In the flowchart of the text sentence analysis process shown in FIG. 5, the CPU 203 performs a morphological analysis of the text sentence data based on the morphological analysis dictionary 2093 (step S501). For example, for the morphological analysis process, a “tea bowl” of Matsumoto Laboratory of Nara Institute of Science and Technology may be used.

  FIG. 7, FIG. 7a and FIG. 7b show an example in which the information extraction processing is performed on the text sentence data “Operating profit was 32,976 million yen, which is the same level as the previous year” 701.

  The CPU 203 divides the text sentence data 701 (FIG. 7) into morphemes, which are the smallest meaningful language units, through the morphological analysis processing in step S501.

  For example, in FIG. 7, the text sentence data 701 includes “operating profit (noun) / ha (particle) / previous year (noun) / and (particle) / same level (noun) / ) / 32776 million yen (noun) / and (particle) / na (verb) /. (End of sentence). In the text sentence data 703 shown in FIG. 7 and FIG. 7, the text sentence data is displayed by inserting “/” for each morpheme for the sake of explanation, but actually, as shown in A of FIG. It is recorded in a predetermined area of the memory 205 or the hard disk 209 as morphological analysis data.

Based on the grammar definition information 2094, the CPU 203 performs a syntax analysis of the text sentence data subjected to the morphological analysis in the above (step S503). For example, the syntax analysis process may use YACC, which is an LR grammar parser for ascending syntax analysis. The parser refers to a syntax analysis program, which is a program that analyzes an input character string and creates a syntax analysis tree.
The LR grammar parser performs syntax analysis based on preset grammar definition information 2094. FIG. 8 shows an example of grammar definition information 2094 set in this embodiment. In the grammar definition information 2094 shown in FIG. 8, grammar information is defined by BNF notation.

  In step S503, the CPU 203 inputs the text sentence data subjected to morphological analysis to YACC, which is an LR grammar parser, and performs syntax analysis. For example, if the text sentence data 701 shown in FIG. 7 is parsed using the YACC and the grammar definition information 2094, the result is as shown in 705.

  In 705, based on the grammar definition information 2094, syntax analysis is performed step by step as in levels 1-8. Thereby, the whole text sentence data can be divided | segmented into predetermined sentence element data in each level. For example, at level 2, the text sentence data 701 is “operating profit is (subordinate phrase) / previous year (subordinate phrase) / same level (subordinate phrase) / 32976 million yen (subordinate phrase) / Can be divided into each sentence element data of “verb phrase)”.

  As shown in B to D of FIG. 7a and E to I of FIG. 7b, the CPU 203 records the parsed data in a predetermined area of the memory 205 or the hard disk 209 for each level.

  The CPU 203 acquires the text sentence data divided into the sentence element data at the “˜phrase” level from the text sentence data that has been parsed (step S505).

  For example, the text sentence data 707 divided into level 2 of the text sentence data 705 in FIG. 7, specifically, the syntax analysis data (level 2) shown in C of FIG. 7a is acquired.

  The CPU 203 performs a process for determining a main phrase based on the acquired syntax analysis data (level 2) (step S507). Here, the main phrase is sentence element data including the subject, and it is determined whether or not it is a main phrase by the particle at the end of the sentence element data. For example, if the end of the sentence element data is any one of “to”, “to is”, and “to”, the sentence element data is determined to be a main phrase.

  FIG. 9 shows an example in which a main phrase is determined. In the text sentence data 901 shown in FIG. 9, the particle at the end of the sentence element data “Operating profit is (subordinate phrase)” is “ha” from the text sentence data 703 immediately after the morphological analysis in FIG. As a result, it is possible to determine that the sentence element data “operating profit is (subordinate phrase)” ends with “˜ha” and is a main phrase.

  The CPU 203 performs processing for determining a predicate in the acquired text sentence data 707 (step S509). Here, the predicate is statement element data including a predicate. The CPU 203 determines whether the predicate is based on the part of speech at the end of the sentence element data.

  For example, when the part of speech at the end of the sentence element data is “noun”, “adjective”, or “verb”, the sentence element data is determined to be a predicate or a component of the predicate. In particular, in the case of “auxiliary verbs” and the above “nouns” and “adjectives” whose meaning is low even if they are “verbs”, the predicates are connected with the immediately preceding sentence element data. Shall be composed.

  In the present embodiment, when the end of sentence element data is any one of “to”, “becomes”, “was to”, and “becomes to”, The part of speech is determined to be an “auxiliary verb”. Note that the determination as to whether the part of speech at the end of the sentence element data is an “auxiliary verb” may be made based on morphological analysis. In this case, it is only necessary to analyze whether the verb is a main verb (independent) or an auxiliary verb (non-independent) by morphological analysis.

  FIG. 9 shows an example of determining a predicate. In the text sentence data 903 shown in FIG. 9, the sentence element data “Nata (verb phrase)” is a “verb” based on the result of the morphological analysis shown in FIG. Thereby, the sentence element data “Nana (verb phrase)” can be determined to be a predicate or a component of the predicate. Further, since the sentence element data “Nana (verb phrase)” is classified as an “auxiliary verb”, it is concatenated with the immediately preceding sentence element data to form a predicate.

  As a result, the sentence element data “Nata (verb phrase)” and the immediately preceding sentence element data “32776 million yen and (subordinate phrase)” are concatenated to become “32976 million yen” as a predicate. Is done.

  That is, as shown in the text sentence data 905 of FIG. 9, the text sentence data is “Operating profit was (main phrase) / previous year and (subordinate phrase) / same level (subordinate phrase) / 32976 million yen. (Predicate) "is divided into each sentence element data. In practice, the CPU 203 records analysis result data 910 as shown in FIG. 9 a in a predetermined area of the memory 205 or the hard disk 209.

1-3-3. Dependency information extraction processing
In the text sentence analysis process (FIG. 3, step S309), when the text sentence data is divided into predetermined types of sentence element data, the CPU 203 further performs a dependency information extraction process (step S313).

  In the flowchart of the text sentence analysis process shown in FIG. 6, the CPU 203 acquires sentence element data to which the numerical information data belongs (step S601). For example, since the numerical information data extracted in the numerical information extraction process is “32776”, the text element data 905 shown in FIG. Predicate) ”.

  If the CPU 203 determines that the sentence element data including the numerical information data is a main phrase (step S603, YES), the CPU 203 proceeds to step S609 described later, and if it is determined that the sentence element data is not a main phrase (step S603, NO), further It is determined whether or not it is a predicate (step S605).

  If the CPU 203 determines that the sentence element data to which the numerical information data belongs is a predicate (step S605, YES), the noun included in the main phrase of the original text sentence data is extracted as dependency information data of the numerical information data. If it is determined that it is not a predicate (step S605, NO), it is further determined that the sentence element data to which the numerical information data belongs is a subordinate phrase, and whether the sentence element data is a combination modification. It is determined whether or not (step S607).

  If the CPU 203 determines that the sentence element data including the numerical information data is a linkage modification (YES in step S607), the CPU 203 determines the noun included in the sentence element data immediately after the sentence element data as a dependency of the numerical information data. If it is extracted as information data (step S615) and it is determined that it is not a combination modification (NO in step S607), sentence element data other than the sentence element data to which the numerical information data belongs are extracted as dependency information data of the numerical information data. (Step S609).

  In the following description, the sentence element data to which the numerical information data belongs are divided into “predicate”, “main phrase”, and “subordinate phrase”.

1-3-3-1. In the case of “predicate” For example, since the sentence element data “numerical information data that has become 32767 million yen (predicate)” is determined to be a predicate, the text sentence data 905 shown in FIG. The dependency information data is extracted on the basis of “operating profit (main phrase)”. That is, “business profit (noun)” is extracted as dependency information data with reference to the text sentence data 703 shown in FIG.

  As a result, the numerical information data “32776 million yen” and the dependency information data “operating profit” are extracted.

1-3-3-2. In the case of “main phrase” For example, let us consider a case where “Ieyasu Tokugawa was entrusted to the great conqueror in 1603” as text data.

  According to steps S301 to S309 shown in FIG. 3, the above text sentence data is “in the year 1603 (main phrase) / Ieyasu Tokugawa (main phrase) / conquered general (subordinate phrase) / assigned (predicate) ) "Sentence element data. In this case, “year” is used as a unit character, and “1603” is extracted as numerical information data.

  In this case, since the sentence element data to which the numerical information data belongs is “1603 (prominent phrase)”, the CPU 203 relates sentence element data other than “1603 (prominent phrase)” from the text sentence data. Extracted as received information data. That is, “Ieyasu Tokugawa (main phrase) / subordinate phrase (subordinate phrase) / assigned (predicate)” is extracted as dependency information data.

  As a result, the numerical information data “1603” and the dependency information data “Ieyasu Tokugawa was entrusted to the great conqueror” are the extraction results.

1-3-3-3. In the case of “subordinate phrase” When the sentence element data including the numerical information data is a subordinate phrase, the CPU 203 determines whether or not the sentence element data is a combination modification (step S607). That is, it is determined whether or not a statement is included in the subsequent sentence element data.

  For example, if the end of sentence element data including numerical information data is “to” (step S607, YES), it is determined that the word to be modified is a body word, and the sentence element data is determined to be a combined modification. To do.

  For example, consider a case where “I bought a note of 100 yen yesterday” is input as text data. In accordance with steps S301 to S309 shown in FIG. 3, the text data is “I am (main phrase) / yesterday (subordinate phrase) / 100 yen (subordinate phrase) / note (subordinate phrase) / buy (description). Part) "sentence element data. Note that “yen” is used as the unit character in this case, and “100 yen” is extracted as numerical information data.

  In this case, since the sentence element data to which the numerical information data belongs is “100 yen (subordinate phrase)” and the end of the sentence element data is “to”, the CPU 203 determines that the subordinate phrase is a combination modification. judge. That is, it is determined that the word modified by the subordinate phrase is a body word, and the noun of the sentence element data immediately after the sentence element data in the text sentence data is extracted as dependency information data.

  For example, “note (noun)”, which is sentence element data immediately after, is extracted as dependency information data from “note (subordinate phrase)”.

  As a result, the numerical information data “100 yen” and the dependency information data “note” are extracted.

  On the other hand, when the CPU 203 determines that the modification is not a combination modification (NO in step S607), that is, it is determined that the word to be modified is a predicate, and sentence element data other than the sentence element data to which the numerical information data belongs is determined as dependency information data. (Step S609).

  For example, consider a case in which “Yori Yoromo opened the Kamakura Shogunate in 1192” is input as text data.

  According to steps S301 to S309 shown in FIG. 3, the text sentence data is “Gen Yoromo (primary phrase) / 1192 (subordinate phrase) / Kamakura shogunate (subordinate phrase) / opened (predicate)”. Divided into sentence element data. In this case, “year” is used as a unit character, and “1192” is extracted as numerical information data.

  In this case, since the sentence element data to which the numerical information data belongs is “in 1192 (subordinate phrase)” and the end of the sentence element data is other than “to”, the CPU 203 determines that the subordinate phrase is not a combination modification. judge. That is, it is determined that the word modified by the subordinate phrase is a predicate, and sentence element data other than the sentence element data to which the numerical information data belongs in the text sentence data is extracted as dependency information data.

  For example, sentence element data other than the sentence element data “1192 (subordinate phrase)” is “source Yoromo (main phrase) / kamakura shogunate (subordinate phrase) / opened (predicate)” as dependency information data Extracted.

  As a result, the numerical information data “1192” and the dependency information data “Gen Yoromo opened the Kamakura Shogunate” are extracted results.

  The CPU 203 repeats the processing steps S601 to S609 by the number of numerical information data extracted by the numerical information extraction processing (FIG. 4) (step S611, YES), and ends the processing if there is no other numerical information data. (Step S611, NO).

1-3-4. Conclusion When the dependency information extraction process is completed, the CPU 203 records the extraction result in the extraction result database 2097 (FIG. 3, step S315). For example, when the text sentence data “Operating income was 32,763 million yen, the same level as the previous year” shown in FIG. 7, the numerical information data “32776 million yen is entered in the extraction result database 2097. ”, The dependency information data“ operating profit ”is recorded.

  In this embodiment, the extraction result extracted by the information extraction apparatus is recorded in the extraction result database 2097. However, the present invention is not limited to this, and the extraction result is transferred to another application. Also good.

  Further, in order to perform processing for extracting numerical information data and dependency information data from text sentence data, it may be used by being incorporated in another application. Further, the extraction result may be displayed on the display 201 of the information extraction device for presentation to the user.

  The CPU 203 repeats the above process (FIG. 3, steps S303 to S315 (YES in step S317) as many times as the number of text sentence data cut out from the text document (step S317, YES). If there is no text sentence data to be processed, the information extraction process is performed. The process ends (step S317, NO).

  As described above, according to the present invention, nouns having a dependency relationship can be accurately extracted even when nouns corresponding to numerical information data are not in the vicinity. In addition, information extracted as information corresponding to numerical information data is not limited to a specific noun, and a noun or sentence having a dependency relationship can be accurately extracted.

2. Second Embodiment In the above embodiment, the description has been made on the assumption that the input text sentence data is a single sentence. In the present embodiment, a case where the text sentence data is a compound sentence or a double sentence including two sentences will be described.

  Note that a compound sentence is a sentence in which a subject / predicate relationship is established, and a subject / predicate relationship is further found in its constituent parts. A heavy sentence is a sentence formed by combining two or more independent sentences with equal qualifications.

2-1. Functional Block Diagram FIG. 10 shows a functional block diagram of the information extraction apparatus according to this embodiment. In this figure, the information extraction apparatus according to the present invention comprises a cutting means 101, numerical information extraction means 103, text sentence analysis means 105, extraction target sentence determination means 106, dependency information extraction means 107, and output means 109. Yes.

  The extraction target sentence determination unit 106 determines a sentence from which numerical information data is to be extracted from the text sentence data divided into sentence element data by the text sentence analysis unit, and supplies the sentence to the dependency information extraction unit 107. For example, in the case of a compound sentence, after dividing into a main sentence and a sub sentence, it is determined whether or not each sentence is an extraction target.

  Note that the extraction unit 101, the numerical information extraction unit 103, the text sentence analysis unit 105, the dependency information extraction unit 107, and the output unit 109 are the same as those in the first embodiment.

2-2. Hardware configuration The hardware configuration is the same as in the first embodiment.

2-3. Flowchart Processing based on the information extraction program 2091 will be described using the flowchart of FIG. In the following, an example will be described in which a text document including text sentence data “This month's food expense is 45400 yen because my father bought 7000 yen of wine” is input to the clipping unit 101.

  In the flowchart of the information extraction process shown in FIG. 11, the processes from step S301 to S309 performed by the CPU 203 are basically the same as those in the first embodiment.

  For example, when the text sentence data is read, the first sentence, “This month's food expense is 45400 yen because my father bought 7000 yen of wine” is cut out as text sentence data.

  The CPU 203 performs numerical information extraction processing based on the cut-out text sentence data (step S305).

2-3-1. Numerical Information Extraction Processing For example, in this embodiment, “Yen” is recorded as a unit character in the numerical value unit master, and the CPU 203 performs “7000 yen” and “45400 yen” by the numerical information extraction processing shown in FIG. Are respectively extracted as numerical information data.

2-3-2. Text sentence analysis processing If numerical information data is extracted in the numerical information extraction process (FIG. 11, step S305) (step S307, YES), the CPU 203 performs a text sentence analysis process (step S309). As shown in the first embodiment, the CPU 203 divides the text sentence data into predetermined sentence element data according to the flowchart of the text sentence analysis process shown in FIG.

  FIG. 13 shows the text sentence data after the text sentence analysis processing is performed. As shown in the text sentence data 1301, the text sentence data is: “This month's food cost is (main phrase) / father's (main phrase) / 7000 yen (subordinate phrase) / wine (subordinate phrase) / buy. (Predicate) / 45400 Yen (Predicate) ”is divided into each sentence element data.

2-3-3. Extraction target sentence decision processing
When the text sentence analysis process is completed, the CPU 203 performs an extraction target sentence determination process (step S310). FIG. 12 shows a flowchart of the extraction target sentence determination process.

  The CPU 203 determines the type of sentence based on the text sentence data subjected to the syntax analysis process. That is, the CPU 203 determines whether the input text sentence data is a compound sentence, a double sentence, or a single sentence based on the number and arrangement of combinations of the main phrase and predicate (FIG. 12, step S1201).

  If the combination of the main phrase and the predicate is one, the CPU 203 determines that the text sentence data is a single sentence (step S1201, simple sentence). If the combination of the main phrase and predicate is two and the arrangement is “main phrase-main phrase-predicate-predicate”, it is determined that the text sentence data is a compound sentence (step S1201, compound sentence). Further, if there are two combinations of the main phrase and the predicate and the arrangement is “main phrase-predicate-main phrase-predicate”, it is determined that the text sentence data is a heavy sentence (step S1201, heavy sentence).

  If the CPU 203 determines that the input text sentence data is a compound sentence, the CPU 203 divides the text sentence data into a main sentence and a sub sentence (step S1203). For example, in the case of the text sentence data 1301 shown in FIG. 13, it is determined that the sentence structure is “main case phrase-main case phrase-predicate-predicate”, so that it is a compound sentence. Text sentence data 1303 “This month's food cost was (main phrase) / 45400 yen (predicate)” and text sentence data 1305 as a sub sentence (father was (main phrase) / 7000 yen (subordinate phrase) / Wine (subordinate phrase) / Bought (predicate) ".

  On the other hand, if the CPU 203 determines that the input text sentence data is a heavy sentence, the CPU 203 divides the text sentence data into a first half sentence and a second half sentence (step S1205). For example, in the case of the text sentence data 1401 shown in FIG. 14, it is determined that the sentence structure is “major phrase-predicate-prominent phrase-predicate”, so that it is a heavy sentence. Text sentence data 1403 “Food expenses for this month are (primary phrases) / 45400 yen (predicate)” and text sentence data 1305 as sub-sentence is “Father is (primary phrases) / 7000 yen (subordinate phrase) ) / Wine (subordinate clause) / bought (predicate) ”.

  The CPU 203 acquires the divided single sentence or the input single sentence (step S1207), and determines whether the single sentence includes numerical information data (step S1209).

If the acquired single sentence includes numerical information data (step S1209, YES), the CPU 203 determines the single sentence as an extraction target sentence and stores it in the memory 205 (step S1211). If numerical information data is not included, step S1211 is skipped.

  If there is another simple sentence (YES in step S1213), the CPU 203 returns to step S1207, acquires the next simple sentence, and performs the same processing as in steps S1209 to S1213. If there is no single sentence to be processed, the process ends (NO in step S1213).

  For example, when the text sentence data 1303 as the main sentence in FIG. 13 is acquired as a single sentence, the numerical information data “45400 yen” is included, so the text sentence data 1303 is determined as an extraction target sentence. Further, the text sentence data 1305 as the sub sentence is acquired as a single sentence, and the numerical information data “7000 yen” is included, so that the text sentence data 1305 is determined as the extraction target sentence. In the case of the heavy text text data 1401 shown in FIG.

2-3-4. Dependency information extraction processing
In the extraction target sentence determination process (FIG. 11, step S310), when the input text sentence data is divided into predetermined types of sentence element data and divided into simple sentences, the CPU 203 further performs dependency information extraction processing (step S313). I do. The dependency information extraction process is the same as the flowchart of FIG. 6 shown in the first embodiment.

  The CPU 203 repeats the processing steps S313 to S315 as many as the number of simple sentences determined by the extraction target sentence determination process (FIG. 12) (step S316, YES), and if there is no other single sentence, the process ends (step S316). , NO).

  For example, numerical information data “45400 yen” and dependency information data “food expenses” are extracted from the text sentence data 1303 shown in FIG. 13 and recorded in the extraction result database 2097. Similarly, numerical information data “7000 yen” and dependency information data “wine” are extracted from the text sentence data 1305 shown in FIG. 13 and recorded in the extraction result database 2097.

2-3-5. Summary As described above, according to the present invention, it is possible to accurately extract a noun or sentence that is dependent on numerical information data regardless of whether the text sentence data is a single sentence, a compound sentence, or a heavy sentence. it can.

  In the above embodiment, the text sentence data has been described on the premise of a compound sentence or a duplicate sentence in which the combination of the main phrase and the predicate is two, but a combination of two or more may be used. In this case, the arrangement of the main phrase and the predicate may be pre-patterned, and it may be configured to determine whether it is a compound sentence or a heavy sentence based on which pattern it belongs to, and to divide it into simple sentences. .

3. Third Embodiment In the above embodiment, the input text sentence data is divided into sentence element data, and based on the type of sentence element data to which the numerical information data belongs, the numerical information data and the numerical information data are related. Only the dependency information data in the receiving relationship is extracted.

  In the present embodiment, a case will be described in which numerical information data and dependency information data are extracted as main information based on predetermined sentence element data, and additional information is extracted based on other sentence element data.

3-1. Functional Block Diagram FIG. 15 shows a functional block diagram of the information extraction device according to this embodiment. In this figure, an information extracting apparatus according to the present invention includes a cutting means 101, a numerical information extracting means 103, a text sentence analyzing means 105, an extraction target sentence determining means 106, a dependency information extracting means 107, an additional information extracting means 108, Output means 109 is provided.

  The additional information extraction unit 108 extracts sentence element data other than the sentence element data including the numerical information data and the dependency information data based on the text sentence data divided into the sentence element data, and the numerical information data and the dependency information are extracted. This is given to the output means 109 as additional information regarding the information data.

  Note that the extraction unit 101, the numerical information extraction unit 103, the text sentence analysis unit 105, the dependency information extraction unit 107, and the output unit 109 are the same as those in the first embodiment.

3-2. Hardware configuration The hardware configuration is the same as in the first embodiment.

3-3. Flowchart Processing based on the information extraction program 2091 will be described using the flowchart of FIG. In the following, as in the first embodiment, an example in which a text document including text data “operating profit is 32,976 million yen, the same level as the previous year” is input to the clipping unit 101 is described as an example. Will be described.

  In the flowchart of the information extraction process shown in FIG. 11, the processes from step S301 to S315 performed by the CPU 203 are basically the same as those in the first embodiment.

  For example, when the text sentence data is read, the first sentence “Operating profit was 32,976 million yen, the same level as the previous year” is extracted as the text sentence data.

  The CPU 203 performs numerical information extraction processing based on the cut-out text sentence data (step S305).

  If numerical information data is extracted in the numerical information extraction process (FIG. 16, step S305) (step S307, YES), the CPU 203 performs a text sentence analysis process (step S309). As shown in the first embodiment, the CPU 203 divides the text sentence data into predetermined sentence element data according to the flowchart of the text sentence analysis process shown in FIG.

  FIG. 18 shows the text sentence data after the text sentence analysis processing is performed. As shown in the text sentence data 1801, the text sentence data is “Operating profit was (main phrase) / previous year (subordinate phrase) / same level (subordinate phrase) / 32776 million yen (predicate)” Is divided into each sentence element data.

  In the text sentence analysis process (FIG. 16, step S309), when the input text sentence data is divided into predetermined types of sentence element data, the CPU 203 further performs a dependency information extraction process (step S313). The dependency information extraction process is the same as the flowchart of FIG. 6 shown in the first embodiment.

  Therefore, as in the first embodiment, numerical information data “32776 million yen” and dependency information data “operating profit” are extracted results.

3-3-1. Additional information extraction process
When the dependency information extraction processing is completed, the CPU 203 performs additional information extraction processing (step S331). FIG. 17 shows a flowchart in the additional information extraction process.

  The CPU 203 acquires numerical element data extracted in the numerical information extraction process and sentence element data to which the dependency information data extracted in the dependency information extraction process belongs (FIG. 17, step S1701).

  Next, the CPU 203 reads the text sentence data 1801 divided into the sentence element data in the text sentence analysis process, and adds the sentence element data other than the sentence element data to which the numerical information data and the dependency information data acquired above belong to the additional information. (Step S1703).

  For example, in the text sentence data 1801 in FIG. 18, the sentence element data to which the numerical information data belongs is “32776 million yen (predicate)”, and the sentence element data to which the dependency information data belongs is “operating profit is (major Phrase). Therefore, “same year (subordinate phrase)” and “same level (subordinate phrase)” are extracted as sentence element data other than the sentence element data to which the numerical information data and the dependency information data belong.

  The CPU 203 repeats the processing steps S1701 to S1703 by the number of numerical information data extracted by the numerical information extraction processing (FIG. 4) (step S1705, YES), and ends the processing if there is no other numerical information data. (Step S1705, NO).

  When the additional information extraction process is finished, the CPU 203 records the extraction result in the extraction result database 2097 (FIG. 16, step S333).

  For example, as shown by 1803 in FIG. 18, in the extraction result database 2097, in addition to the numerical information data “32776 million yen” and the dependency information data “operating profit”, the additional information “same level as in the previous year”. Is recorded.

  The CPU 203 repeats the above process (FIG. 11, steps S303 to S333 (YES in step S317) as many times as the number of text sentence data cut out from the text document (step S317, YES). If there is no text sentence data to be processed, the information extraction process is performed. The process ends (step S317, NO).

3-4. Summary As described above, according to the present invention, in addition to numerical information data and dependency information data that are main information of text sentence data, additional information related to the main information can be extracted.

  As a result, not only the dependency information directly related to the numerical information data but also the additional information indirectly related to the numerical information data or the dependency information data can be extracted.

  For example, the additional information “same level as last year” regarding the numerical information data “32776 million yen” and the dependency information data “operating profit” is related to the situation where the “operating profit” is “32976 million yen”. -It can be used as additional information as a situation.

  In the above embodiment, the case where the text sentence data is a single sentence has been described. However, the text sentence data may include a compound sentence or a duplicate sentence. In this case, as shown in the second embodiment, text sentence analysis processing (step S309). And an extraction target sentence determination process may be inserted between the dependency information extraction process (step S313).

4). Other Embodiments In the above embodiment, the division process is performed after the numerical information extraction process. However, the numerical information extraction process may be performed during the text sentence analysis process. Specifically, after the text sentence data is divided into sentence element data in the text sentence analysis process, it is determined whether or not numerical information data is included for each sentence element data, and if included, the syntax of the text sentence data What is necessary is just to comprise so that an analysis may be performed.

  In the above embodiment, when the numerical information data is extracted, it is configured to extract, as the numerical information data, numbers consecutive in front of a predetermined unit character existing in the text sentence data. Alternatively, numbers that are continuous in the back may be extracted as numerical information data. For example, this is the case when a currency symbol such as “¥” or “$” is used as a unit character.

  In the above embodiment, YACC, which is an LR grammar parser for ascending syntax analysis, is exemplified in the syntax analysis processing, but other types of parsers may be used. For example, an LL grammar parser for descending parsing corresponds to this.

  In the above embodiment, a text document described in Japanese has been described. However, numerical information data can be extracted and can be divided into sentence element data in morpheme analysis processing and syntax analysis processing. It may be a written text document. For example, in the case of English, numerical information data may be extracted based on English numeric units and configured using an English morphological analysis program and a syntax analysis program.

  In the above embodiment, the CPU 203 is used to realize the function shown in FIG. 1, and this is realized by software. However, some or all of them may be realized by hardware such as a logic circuit. In addition, you may make it make an operating system (OS) process a part of program further.

It is a figure which shows the example of the functional block diagram of the information extraction apparatus in embodiment of this invention. It is an example which shows the example of the hardware block diagram of the information extraction apparatus of this invention. It is a figure which shows the example of the flowchart in the "information extraction process" of this invention. It is a figure which shows the example of the flowchart in the "numerical information extraction process" of this invention. It is a figure which shows the example of the flowchart in the "text sentence analysis process" of this invention. It is a figure which shows the example of the flowchart in the "dependency information extraction process" of this invention. It is a figure which shows the example of "morphological analysis" or "syntax analysis" of this invention. It is a figure which shows the example of "morpheme analysis data" or "syntax analysis data" of this invention. It is a figure which shows the example of the "syntactic analysis data" of this invention. It is a figure which shows the example of the "grammar definition information" of this invention. It is a figure which shows the example which determines the "main phrase" or "predicate" of this invention. It is a figure which shows the example of the "analysis result data" of this invention. It is a figure which shows the example of the functional block diagram of the information extraction apparatus in embodiment of this invention. It is a figure which shows the example of the flowchart in the "information extraction process" of this invention. It is a figure which shows the example of the flowchart in the "extraction target sentence determination process" of this invention. It is a figure which shows the example in the case of "dividing a compound sentence into a main sentence and a sub sentence" of this invention. It is a figure which shows the example in the case of "dividing a heavy sentence into the first half sentence and the latter half sentence" of this invention. It is a figure which shows the example of the functional block diagram of the information extraction apparatus in embodiment of this invention. It is a figure which shows the example of the flowchart in the "information extraction process" of this invention. It is a figure which shows the example of the flowchart in the "additional information extraction process" of this invention. It is a figure which shows the example in the case of extracting additional information from the text sentence data of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 101 Extraction means 103 Numerical information extraction means 105 Text sentence analysis means 106 Extraction target sentence determination means 107 Dependency information extraction means 108 Additional information extraction means 109 Output means

Claims (8)

An information extraction device for extracting information data including numerical information data from given text sentence data,
The given text sentence data is divided into sentence element data as phrases, and when the end of each sentence element data is one of the particles “ha”, “ga”, and “mo”, the sentence element data is “ If the part of speech at the end of each sentence element data is one of "noun", "adjective", or "verb", the sentence element data is "predicate"; otherwise, the sentence element The data is recorded as a “subordinate phrase” in the recording unit, and it is determined whether or not a character indicating the numerical unit is included in the given text sentence data. If included, the numerical unit and the numerical unit are included. Extraction / analysis means for identifying numerical information data as a numerical information data and recording it in a recording section,
Referring to the recording unit, the sentence element data including the numerical information data is identified, and at least one sentence element data having a dependency relation with the numerical information data is determined based on the type of the sentence element data. Dependency information extracting means for extracting dependency information data from the sentence element data,
An information extraction apparatus comprising: output means for outputting numerical information data extracted by the numerical information extraction means and dependency information data extracted by the dependency information extraction means;
The extraction / analysis means determines that the sentence element data is “main phrase” when the sentence element data ends in any one of “to”, “to”, and “to”, and the sentence element data If the part of speech is one of "noun", "adjective", or "verb", the sentence element data is "predicate", and if it does not correspond to any of "main phrase" or "predicate", the sentence element Let the data be a "subordinate phrase"
The dependency information extraction means, when the numerical information data is included in the sentence element data of the “predicate”, sets the noun of “main phrase” in the text sentence data as dependency information data, and the numerical information When the data is included in the sentence element data of the “subordinate phrase” and the word to be modified by the sentence element data is a body word, the noun of the sentence element data next to the sentence element data is used as dependency information data An information extraction apparatus characterized by that. A program for realizing, using a computer, an information extraction device that extracts information data including numerical information data from given text sentence data,
The given text sentence data is divided into sentence element data as phrases, and when the end of each sentence element data is one of the particles “ha”, “ga”, and “mo”, the sentence element data is “ If the part of speech at the end of each sentence element data is "noun", "adjective", or "verb", the sentence element data is "predicate"; otherwise, the sentence element The data is recorded as a “subordinate phrase” in the recording unit, and it is determined whether or not a character indicating the numerical unit is included in the given text sentence data. If included, the numerical unit and the numerical unit are included. Extraction / analysis means for identifying numerical information data as a numerical information data and recording it in a recording section,
Referring to the recording unit, the sentence element data including the numerical information data is identified, and at least one sentence element data having a dependency relation with the numerical information data is determined based on the type of the sentence element data. Dependency information extracting means for extracting dependency information data from the sentence element data,
Numerical information data extracted by the numerical information extraction means, output means for outputting the dependency information data extracted by the dependency information extraction means,
A program to make it function,
The extracting / analyzing means divides the text sentence data into a plurality of sentence element data, and each sentence element data includes a subject phrase including a subject and a predicate when including a predicate. If it doesn't contain, the type of sentence element is determined as a subordinate phrase,
The dependency information extraction means, when the numerical information data is included in the sentence element data of the “predicate”, sets the noun of “main phrase” in the text sentence data as dependency information data, and the numerical information When the data is included in the sentence element data of the “subordinate phrase” and the word to be modified by the sentence element data is a body word, the noun of the sentence element data next to the sentence element data is used as dependency information data A program characterized by that. In the program of Claim 2,
The text sentence analysis means divides the text sentence data into a plurality of sentence element data, and each sentence element data includes a subject phrase including a subject, a predicate when including a predicate, and a subject or predicate. A program characterized by determining the type of sentence element as a subordinate phrase when none of them is included. In the program of Claim 2 or 3,
The dependency information extraction unit includes the numerical value in the text sentence data when the numerical information data is included in the sentence element data of the “subordinate phrase” and the word to be modified by the sentence element data is a predicate. A program characterized in that a combination of all sentence element data other than a “subordinate phrase” including information data is defined as dependency information data. In the program in any one of Claims 2-4,
The dependency information extracting means, when the numerical information data is included in the sentence element data of “main phrase”, the sentence element other than “main phrase” including the numerical information data in the text sentence data A program characterized in that a combination of all data becomes dependency information data. In the program in any one of Claims 2-5,
The computer determines whether the text sentence data has two or more pairs of main clauses and predicates. If there are two or more, the computer decomposes them into a single sentence containing only one pair of main clauses and predicates. A program for further functioning as an extraction target sentence determination means for generating text sentence data. In the program in any one of Claims 2-6,
Additional information extraction for extracting sentence element data as additional information data from the text sentence data, excluding sentence element data including the numerical information data and sentence element data including the dependency information data A program for further functioning as a means. An information extraction method for extracting information data including numerical information data from given text sentence data by a computer,
The computer divides the given text sentence data into sentence element data, which are phrases, and classifies each sentence element data into at least the types of “main phrase”, “subordinate phrase”, and “predicate” and stores them in the recording unit. In addition to recording, it is judged whether or not the character that indicates the numerical unit is included in the given text sentence data, and if it is included, the numerical unit and the consecutive numbers before or after the numerical unit are numerically An extraction / analysis step that is specified as information data and recorded in the recording unit;
The computer refers to the recording unit to identify sentence element data including numerical information data, and based on the type of the sentence element data, at least one sentence element data having a dependency relationship with the numerical information data. A dependency information extraction step for determining and extracting dependency information data from the determined sentence element data;
The computer outputs the numerical information data extracted by the numerical information extraction means, the dependency information data extracted by the dependency information extraction means,
In the information extraction method characterized by executing
In the extraction / analysis step, when the end of the sentence element data is any one of “to”, “to”, and “to”, the computer assumes that the sentence element data is “main phrase”, and the sentence element data If the part-of-speech at the end is either "noun", "adjective", or "verb", the sentence element data is assumed to be "predicate", and if it does not fall under any of "main phrase" or "predicate" Let the sentence element data be a subordinate phrase,
In the dependency information extracting step, when the numerical information data is included in the sentence element data of the “predicate”, the computer sets the noun of “main phrase” in the text sentence data as dependency information data. When the numerical information data is included in the sentence element data of the “subordinate phrase” and the word to be modified by the sentence element data is a formal word, the noun of the sentence element data next to the sentence element data is changed to the dependency information An information extraction method characterized by using data.

Images