JP6889038B2 - Financial results analysis system and financial results analysis program - Google Patents

Description

The present invention relates to a settlement analysis system and a settlement analysis program that analyze settlement-related information and extract factors of accounting events.

In general, financial statements such as financial statements disclosed by stock exchanges, revisions of earnings forecasts announced by companies, and financial results analysis reports released by securities companies provide investors and institutional investors with investment transactions such as stocks and bonds. It is an important source of information in doing so. Therefore, various methods have been conventionally proposed for automatically processing settlement-related information using a computer.

For example, Patent Document 1 discloses an article creation system that creates an article based on the analysis result of financial statements. This article creation system includes a template containing conditions related to information obtained from financial statement data and a character string used when the information obtained from financial statement data meets the conditions. When the financial statement data described in XBRL (eXtensible Business Reporting Language) is accepted, it is determined whether or not the information acquired from the financial statement data meets the conditions described in the template. Then, if this condition is met, article data representing the article including the character string described in the template is automatically created. This will reduce the time required to create an article based on the analysis results of financial statements.

Further, Patent Document 2 discloses an apparatus for analyzing a real estate portfolio. In this device, attribute data of a plurality of real estate properties, settlement data of real estate property management, and standardization rule information for standardizing settlement data used for analysis are stored. The settlement data is converted into predetermined standardized data based on the standardization rule information. Then, an index such as a rate of return is calculated based on this standardized data and the attribute data of the real estate property. This makes it possible to easily perform portfolio analysis in real estate investment.

Japanese Unexamined Patent Publication No. 2011-008527 Japanese Unexamined Patent Publication No. 2008-14294

By the way, with the spread of the Internet and the like, it has become easier to obtain various financial statements-related information, and it is useful for investors and institutional investors to collect, analyze and report on this information. At that time, it is particularly useful to extract the factors (for example, "due to the influence of exchange rates") for accounting events (for example, "operating income is XX billion yen") mainly based on the subject and amount.・ To identify. However, conventionally, there is no method for automatically extracting such factors from settlement-related information.

The present invention has been made in view of such circumstances, and its object is had efficiently and flexibly row factors extracted from financial-related information, to provide information useful to investors and institutional investors is there.

In order to solve such a problem, the first invention provides a settlement analysis system that has a pattern storage section, a document analysis section, and a report creation section, and analyzes settlement-related information. The factor pattern is stored in the pattern storage unit. This factor pattern has at least a representation of an accounting event that includes subject and monetary information and a representation of the factors of this event. The document analysis unit compares the morpheme sequence of the sentence to be processed with the factor pattern stored in the pattern storage unit. The morpheme string of the sentence to be processed is abstracted by assigning a unique attribute label to at least each attribute that classifies the subject and the amount of money for the morpheme or a combination thereof constituting the morpheme string. In addition, the document analysis department extracts the part specified by this factor pattern as a factor for the morpheme string that matches the factor pattern, and links the extracted factor with the subject and amount information in the morpheme string to form a set. Store as data. The report creation unit outputs a financial results analysis report based on a set of data.

Here, in the first invention, the document analysis unit is an indefinite repeating portion of the morpheme sequence that matches the factor pattern, which is continuous before or after the expression representing the factor and is not given an attribute label. Is preferable as a factor.

In the first invention, when any element of subject, amount information, and factor is missing in one sentence, the document analysis unit analyzes the other sentences according to the order of the sentences, and the missing sentence is found. Once the elements are obtained, the subject, amount information, and factors may be linked.

In the first invention, a morphological analysis dictionary for storing a predefined character string may be further provided. In this case, the document analysis unit treats the character string stored in the morphological analysis dictionary as one morpheme. Further, a labeling dictionary for storing the combination of the above morphemes in association with the attribute label may be further provided. In this case, the document analysis unit assigns an attribute label specified by the labeling dictionary to the combination of morphemes treated as one morpheme.

Further, the second invention analyzes settlement-related information using a computer in which a factor pattern having at least an expression of an accounting event including subject and amount information and an expression of a factor of this event is stored in advance. Provide a financial results analysis program. This program causes a computer to perform a process having the following first to third steps. In the first step, the morpheme sequence of the sentence to be processed is compared with the factor pattern. The morpheme string of the sentence to be processed is abstracted by assigning a unique attribute label to at least each attribute that classifies the subject and the amount of money for the morpheme or a combination thereof constituting the morpheme string. In the second step, for the morpheme string that matches the factor pattern, the part specified by this factor pattern is extracted as a factor, and the extracted factor is linked with the subject and amount information in the morpheme string to form a set of data. Remember as. In the third step, a settlement analysis report is output based on a set of data.

Here, in the second invention, the second step is an indefinite repetition in which the morpheme sequence matching the factor pattern is continuous in front of or behind the expression representing the factor and is not given an attribute label. It is preferable that the step is to extract the portion as a factor.

In the second invention, if any element of subject, amount information, and factor is missing in one sentence, the second step analyzes the other sentences according to the order of the sentences, and this missing sentence. It may include the step of associating the subject, the amount information, and the factor with the obtained element.

In the second invention, the computer may store in advance a morphological analysis dictionary that stores a combination of morphemes related to a predefined character string. In this case, the first step includes a step of treating the character string stored in the morphological analysis dictionary as one morpheme. Further, the computer may store in advance a labeling dictionary that stores the combination of the above morphemes in association with the attribute label. In this case, the first step includes a step of assigning an attribute label specified by a labeling dictionary to a combination of morphemes treated as one morpheme.

Further, in the first and second inventions, the amount information is preferably information relating to an increase or decrease in the amount.

According to the present invention, not the sentence itself, but the morphological sequence abstracted by the attribute label after morphological analysis of the sentence is compared with the predetermined factor pattern. Then, when both match, the portion specified by the factor pattern in this morpheme sequence is extracted as a factor. By performing comparison with factor patterns on the basis of morpheme sequences abstracted by attribute labels, the number of factor patterns to be defined can be effectively suppressed, and factor extraction can be performed efficiently and flexibly. At the same time, it is possible to provide useful information for investors and institutional investors by storing the subject, amount information and factors as a set of data and outputting a settlement analysis report based on this.

Block configuration diagram of financial results analysis system Diagram showing a display example of a payment analysis report Document analysis routine flowchart Diagram showing an example of the sentence to be processed A list summarizing the names of the labeled morpheme sequences Diagram showing an example of factor patterns Explanatory diagram of factor extraction example in dependency analysis

FIG. 1 is a block configuration diagram of a settlement analysis system according to the present embodiment. The settlement analysis system 1 analyzes the input settlement-related information and creates and outputs a report as the analysis result. This system 1 can be typically realized by a computer, and has a data preprocessing unit 2, a document analysis unit 3, and a report creation unit 4 as functional blocks thereof.

The data preprocessing unit 2 acquires financial statements-related information published on the Internet, for example, financial statements disclosed by stock exchanges, revisions of earnings forecasts announced by companies, financial results analysis reports announced by securities companies, and the like. Information related to settlement of accounts can be obtained by crawling the homepage (IR page) of each company. In addition, the data preprocessing unit 2 analyzes the acquired XBRL data to acquire financial information, and also analyzes the acquired PDF data to acquire financial information by segment, region, and the like. Here, XBRL (eXtensible Business Reporting Language) is an extensible business reporting language, and by electronically documenting business reports such as financial statements, it is possible to improve the efficiency of their creation and to compare and analyze them. It is a language created based on the XML (Extensible Markup Language) standard for the purpose of next use. Further, the data preprocessing unit 2 analyzes the PDF data as the payment-related information and acquires the data of the text described in the PDF.

The document analysis unit 3 extracts and acquires the cause of the accounting event, which is the settlement result, from the text acquired by the data preprocessing unit 2. Here, the accounting event is a description mainly composed of subject and amount information, for example, "operating profit is XX billion yen". In the present specification, "amount information" refers to information on the amount of money that appears in the text, and specifically, the following patterns are assumed, but in particular, attention is paid to important information regarding the increase or decrease of the amount of money. Should be.

[Pattern of amount information]
1. 1. Subject + amount (example: sales 100 million yen)
2. Subject + amount increase / decrease (example: sales increase by 100 million yen)
3. 3. Subject + amount increase / decrease (example: increase in sales)

Further, the factor is a description indicating the cause or factor of the accounting event, for example, "due to the influence of exchange rate or the like". The document analysis unit 3 associates such an event with a factor and stores and holds them as a set of data. The association between such an event and a factor is performed for the entire sentence in sentence units, which will be described later.

The report creation unit 4 analyzes the contents of the settlement of accounts based on the data acquired by the document analysis unit 3, and outputs the analysis result as a settlement analysis report. FIG. 2 is a diagram showing a display example of a payment analysis report. In the example of the figure, the factors of each event such as "sales", "operating profit", and "profit" are described as "commentary" for each business segment. Each event and a set of "commentary" are generated based on the association of information by the factor extraction unit 8 described later. When creating the settlement analysis report, not only the analysis result by the document analysis unit 3 but also the numerical data extracted by the above-mentioned XBRL analysis and PDF analysis is appropriately used.

The document analysis unit 3 has a morphological analysis unit 5, a labeling unit 6, a pattern comparison unit 7, and a factor extraction unit 8 as functional sub-blocks constituting the document analysis unit 3. Further, the document analysis unit 3 includes a morphological analysis dictionary 9, a labeling dictionary 10, and a pattern storage unit 11 as storage units for storing predefined information required for document analysis. The morphological analysis dictionary 9 stores a large number of predefined character strings (for example, subjects conforming to "month", "previous year", "sales", etc.) as one that should be treated as one morpheme in morphological analysis. There is. The labeling dictionary 10 is used in the processing of the labeling unit 6, and stores the combinations of morphemes in association with the attribute labels. Then, the attribute label specified by the labeling dictionary 10 is attached to the combination of morphemes treated as one morpheme. Further, the pattern storage unit 11 stores a large number of predetermined factor patterns. Each factor pattern is a pattern having at least an expression of an accounting event including subject and amount information and an expression of the factor of this event. The stored contents of the morphological analysis dictionary 9, the labeling dictionary 10, and the pattern storage unit 11 can be added or changed as appropriate.

The morphological analysis unit 5 decomposes sentences included in the settlement-related information into sentences, performs morphological analysis on each sentence, and generates a morphological sequence for each sentence. Here, since the PDF does not have the concept of a line, the character string is formed so as to be a line at a break (such as a punctuation mark or an indent) of one sentence prior to the morphological analysis. Morphological analysis is to divide a sentence to be processed into the smallest grammatical units called morphological elements and analyze it. In Japanese analysis, words are cut out from sentences, and verbs, adjectives, nouns, adverbs, etc. The part of speech and utilization of the word are presumed, such as adverbs, conjunctions, adverbs, and adverbs. At that time, the character string defined in advance by the morphological analysis dictionary 9 is treated as one morpheme in the morphological analysis, and predetermined metadata is added.

The labeling unit 6 assigns a unique attribute label for each attribute to the morpheme or a combination thereof constituting the morpheme string of the sentence (labeling). For example, an attribute label of "amount" is given to a combination of a number and "yen". At a minimum, the attribute label is required to be able to classify "subject" and "amount", but attributes other than these may be set as appropriate. Further, for the combination of morphemes defined in advance in the labeling dictionary 10 (those that should be treated as one morpheme), the attribute label specified by the term dictionary 9 is given.

The pattern comparison unit 7 compares the morpheme string of the sentence abstracted by the attribute label with the factor pattern stored in the pattern storage unit 11, and determines whether or not the morpheme string matches the factor pattern.

The factor extraction unit 8 extracts the portion specified by the factor pattern as a factor for the morpheme string that matches the factor pattern, and associates the extracted factor with the subject and amount information in the morpheme string. For example, for a morpheme sequence that matches the factor pattern of "+ amount difference" by "country name + particle + subject + particle + indefinite repetition +", it is continuous in front of "by" and the attribute label. It seems that the part of "indefinite repetition" that is not given is extracted as a factor. Here, "indefinite repetition" means, for example, "." (Any character without labels such as "subject" and "amount information") and "+" (immediately preceding pattern) in regular expressions. It can be expressed as a combination with "one or more repetitions of") ". In addition, there are various expressions that express factors, and depending on the expression, the part of continuous" indefinite repetition "in the rear is considered as a factor. There is also. In this embodiment, the association of the three elements of subject, amount, and factor as a set is basically performed in sentence units, but when these elements are separate sentences. Is extracted from a plurality of sentences as a set while observing the excess or deficiency of the three elements.

FIG. 3 is a flowchart of a document analysis routine executed by the document analysis unit 3. This document analysis process is executed by installing a computer program that causes the computer to execute the process shown in FIG. Hereinafter, the details of the document analysis will be described using the sentence shown in FIG. 4 as an example.

First, in step 1, the morphological analysis unit 5 decomposes the document acquired from the PDF in sentence units, and assigns sentence numbers to each sentence in ascending order according to the order of the sentences. In the following step 2, the circular variable n that specifies the sentence number to be processed is set to 1, and the processing of the first sentence in the sentence is started.

In step 3, the morphological analysis unit 5 performs morphological analysis of the sentence to be processed. As described above, the combination of morphemes defined in the term dictionary is treated as one morpheme, and there is no difference from general morphological analysis.

In step 4, the labeling unit 6 labels the morphemes or combinations thereof that form the morpheme sequence. There are three types of labeling: (1) labeling for simple morpheme sequences, (2) labeling for defined morpheme sequences, and (3) labeling for expressions of amount increase / decrease.

(1) Labeling for simple morpheme sequences Attribute labels are assigned to combinations of simple morpheme sequences such as punctuation marks and numerical values. As for the comma, if the metadata of the morpheme is a "symbol" and a "comma" morpheme and matches ",", the attribute labels of "," and "," are given. As for the punctuation mark, if the metadata of the morpheme is a morpheme of "symbol" and "punctuation mark" and matches ".", It is regarded as ".". As for numerical values, A is defined as having one or more morpheme metadata of "noun" and "number", and morpheme metadata of "noun" and "number" following reading points and punctuation marks. If one or more of are B, the attribute label of "numerical value" is given to the one that matches A or AB. Further, regarding the amount of money, the attribute label of "amount of money" is given to the above-mentioned "numerical value" followed by "yen". In the example sentence of FIG. 4, the attribute label of "amount" is given to each morpheme string of "161.6 billion yen", "362.1 billion yen", "612.8 billion yen", "94.4 billion yen", and "5.3 billion yen". It will be.

(2) Labeling for defined morpheme strings Attribute labels are assigned to morpheme strings defined in the labeling dictionary 10, such as months, previous years, and subjects based on sales. For example, as a subject similar to sales, the expression "consolidated" is A, and the expressions such as "sales revenue", "sales", "sales", and "operating revenue" are B, and the morpheme string enclosed in parentheses. When is C, attribute labels such as "sales" and "sales" are given to those matching B, AB, BC, and ABC.

(3) Labeling for expressions of increase / decrease in amount of money Attribute labels are given by dividing into patterns by simple expressions, parenthesized expressions, expressions in proportions, etc. For example, "the highest ever" is A, the expression immediately before the result is B, the expression of the previous period, the expression of the previous period + the comma is C, the repetition of the amount or ratio is D, and the morpheme enclosed in parentheses is E. , When the reading point + amount or ratio, the reading point + amount or ratio + E is F, the expression of increase / decrease is G, and the reading point is H, the attribute of "amount increase / decrease" is defined as a combination such as ABCBDEFGH, BCBDEFGH, and CBDEFGH. A label is given. However, if all the combinations to be defined are listed, the amount of description becomes enormous. Therefore, in reality, a method such as a regular expression is used instead of listing all the combinations. In the example sentence in Fig. 4, "Sales increase of 362.1 billion yen", "Decrease of 612.8 billion yen", "Increase of sales of 94.4 billion yen", "Decrease of sales of 5.3 billion yen", "Decrease of 161.6 billion yen" are " The attribute label of "Amount increase / decrease" is given.

In step 5, the pattern comparison unit 7 compares the morpheme sequence of the sentence abstracted by the attribute label with the factor pattern stored in the pattern storage unit 11, and determines whether or not they match ( matching). Here, consider the case where the names shown in the list of FIG. 5 are used as the names of the labeled morpheme strings. In this case, as the pattern (factor pattern) of the morpheme sequence to be matched, as shown in FIG. 6, A system (A1 to A5 ...), B system (B1 to B3 ...), C system (C1 to C1 to). C3 ...) etc. can be considered. For example, the factor pattern A is [factor prefix] + "segment prefix" + "segment expression" + "factor (reverse) candidate" + "factor prefix (including: factor)" + "factor (reverse) candidate" + "Operating profit margin is maintained from the previous year," means that it is composed of "modification (quantity)" + "subject expression" + "price expression" + "end of line". A large number of such factor patterns are prepared by investigating a large number of financial statements-related information as a sample and preventing omission of extraction of factors.

If the two sequences match, the affirmative judgment in step 6 proceeds to the factor extraction in step 7, and the factor extraction unit 8 extracts the portion specified by the factor pattern as the result (factor / subject / amount increase / decrease). .. For example, regarding the factor patterns of A and C systems, "CAUSE_THOUGH" is extracted as a factor (reverse), "CAUSE" is a factor (in order), "ACCOUNT_PHRASE" is a subject, and "PRICE_SET" is extracted as an increase or decrease in amount. is there. Regarding the B-type personnel pattern, "CAUSE_THOUGH" is a factor (reverse), "CAUSE" is a factor (order), and "PRICE_SET_WITH_ACCOUNT" is extracted as a subject and amount increase / decrease.

The factors extracted in step 7 are combined with the subject and amount information in the morpheme string to form a set. In the example sentence in Fig. 4, the factor of "increased sales of 362.1 billion yen" was caused by "increased sales volume of crude oil and natural gas", and the event of "decreased sales of 612.8 billion yen" was "decrease in average unit price". The factor "due to" and the factor "due to the depreciation of the yen on the average exchange rate of sales" will be extracted for the event "increased sales of 94.4 billion yen". In addition, notations such as "increased sales" and "decreased sales" represent sales as a subject.

On the other hand, when the two do not match, the process proceeds from the negative determination in step 6 to step 8, and the factor extraction unit 8 performs a combination determination across a plurality of sentences. That is, if any element of subject, amount information, or factor is missing in one sentence, the other sentences are analyzed according to the order of the sentences, and when this missing element is obtained, the subject, Amount information and factors are linked.

Then, in step 9, it is determined whether the circular variable n is the last, in other words, whether the processing of the last sentence in the sentence is completed. If the circular variable n is not the last, the circular variable is incremented in step 10 and then returned to step 3 to execute a new sentence process. On the other hand, when the circular variable is the last, a series of processes is completed.

As described above, according to the present embodiment, not the sentence itself, but the morpheme sequence abstracted by the attribute label after morphological analysis of the sentence is compared with the predetermined factor pattern. Then, when both match, the portion specified by the factor pattern in this morpheme sequence is extracted as a factor. In general, financial statements-related information is often described in sentences in a certain format. In view of this tendency, variations of the description are extracted from a large number of sentences that describe the relationship between the event and the factor, and each is defined as a factor pattern. Then, by performing the comparison with the factor pattern on the basis of the morpheme string abstracted by the attribute label, it is possible to efficiently perform factor extraction while effectively suppressing the number of factor patterns to be defined. At the same time, when a new variation is found, it is only necessary to add the newly defined factor pattern to the pattern storage unit 11, which is excellent in flexibility.

Further, according to the present embodiment, when any element of subject, amount information, and factor is missing in one sentence to be processed, other sentences are analyzed according to the order of the sentences, and this missing sentence is performed. When the above elements are obtained, the event (subject, amount information) and the factor are linked. This makes it possible to perform factor extraction more effectively.

Further, according to the present embodiment, a predetermined character string is registered and stored in the morphological analysis dictionary 9, and this character string is treated as one morpheme in the morphological analysis. As a result, the terms described in the payment-related information can be recognized more accurately, and as a result, the accuracy of factor extraction can be improved.

In the above-described embodiment, an example of matching the arrangement of morphemes in the morpheme sequence and the arrangement of elements in the factor pattern has been described as a part of factor extraction, but the present invention is not limited to this. , The same processing may be performed based on the pattern of dependency between morphemes. In this case, the factor pattern is defined as an event and a pattern of dependency between elements including the factor.

FIG. 7 is an explanatory diagram of an example of factor extraction in the dependency analysis. The results of morphological analysis of the example sentences to be analyzed are "domestic", "no", "sales", "ga", "good", "ni", "transition", "done", "koto", "kara", ",", "increased sales", and "next". In the case of "better" and "ta", the morpheme is first labeled. For example, "domestic" is a segment, "increased sales" is a subject + increase / decrease, and "," "better" "ta" are ignored. Next, it is divided into groups extending from the starting point "next". For example, (1) "domestic sales" is included in the segment, and (2) "because it has performed well" can be a factor because it is composed of morphemes without any label. Highly likely, (3) "," is only ignored, (4) "increased sales" includes subject + increase / decrease, (5) "mashi" is only ignored, (6) "ta" is only ignored, And so on. Finally, by arranging the items excluding neglect, "increased sales" is extracted as the subject + increase / decrease, and "because it has been favorable" as the factor.

1 Payment analysis system 2 Data preprocessing unit 3 Document analysis unit 4 Report creation unit 5 Morphological analysis unit 6 Labeling unit 7 Pattern comparison unit 8 Factor extraction unit 9 Morphological analysis dictionary 10 Labeling dictionary 11 Pattern storage unit

Claims (2)

In the settlement analysis system that analyzes settlement-related information,
A pattern storage unit that stores a factor pattern having at least an expression of an accounting event including subject and amount information and an expression of the factor of the event.
A morpheme string of a sentence to be processed, which is abstracted by assigning a unique attribute label to at least each attribute that classifies a subject and an amount for the morpheme or a combination thereof that constitutes this morpheme string. The column is compared with the factor pattern stored in the pattern storage unit, and the portion specified by the factor pattern is extracted as a factor for the morpheme string that matches the factor pattern, and the extracted factor is extracted. With the document analysis unit that stores as a set of data by associating with the subject and amount information in the morpheme string.
A settlement analysis system characterized by having a report creation unit that outputs a settlement analysis report based on the one set of data. In a settlement analysis program that analyzes settlement-related information using a computer in which a factor pattern having at least an expression of an accounting event including subject and amount information and an expression of the factor of the event is stored in advance.
A morpheme string of a sentence to be processed, which is abstracted by assigning a unique attribute label to at least each attribute that classifies the subject and the amount of money for the morpheme or a combination thereof that constitutes this morpheme string. The first step of comparing the column with the factor pattern,
For the morpheme string that matches the factor pattern, the portion specified by the factor pattern is extracted as a factor, and the extracted factor is associated with the subject and amount information in the morpheme string and stored as a set of data. The second step to do and
A settlement analysis program characterized by causing the computer to execute a process having a third step of outputting a settlement analysis report based on the set of data.

Images