JP6723673B2 - Causal relationship extraction system and causal relationship extraction program - Google Patents

Description

The present invention relates to a causal relationship extraction system and a causal relationship extraction program for extracting a causal relationship between economic events and corporate performance.

For example, Patent Literature 1 discloses a stock selection support device capable of searching for a stock of interest using information based on a large amount of text data in order to effectively use a large amount of content such as news articles for investment. ing. Specifically, first, the news article related to the keyword is read from the news article setting means that holds the news article. Next, the brands having a high degree of relevance with the keyword are searched, and those having a high degree of relevance are sequentially extracted. Then, the keywords related to the keyword given by the user and having a high degree of relevance are extracted by searching the news article setting means, and the extracted news articles are provided to the user.

Further, Non-Patent Document 1 discloses a method of extracting a causal relationship chain using a vector expression, which extracts a chain of economic events and financial events as a causal relationship from text data. Specifically, first, a clue expression (causal relationship expression) indicating a causal relationship is extracted with reference to the text of the settlement brief. With regard to this expression extraction, by listing the syntactic patterns in which a causal relationship exists and using the clue expression, the causal expression is extracted regardless of simple sentences, compound sentences, and compound sentences. Then, from the extracted causal relationship nodes, a causal relationship node describing the situation of the market and the performance of the company is manually selected, and is set as the end node of the causal relationship chain. Next, referring to the text of the newspaper, causal relationship nodes that are candidates for addition to the causal relationship chain ending at the terminal node are extracted. The causal relations are extracted using the same method as the text of settlement of accounts, and those that are older than the end node and basically within the search period based on the end node are added to the causal relation chain. Become a candidate. Finally, the similarity between the causal relation nodes is calculated, the causal relation chain is expanded, and additional candidate nodes to the causal relation chain are extracted, thereby constructing the causal relation chain.

JP, 2003-162639, A Nishimura et al., "Extracting Causal Relationship Chains Using Vector Representation", [online], March 20, 2018, 20th AIJ Financial Informatics Research Group (SIG-FIN), Japan, AI Society, [Search February 15, 2019], Internet <URL:http://sigfin.org/?plugin=attach&refer=020-09&openfile=SIG-FIN-020-09.pdf>

Needless to say, economic events that affect economic activities also have a major effect on corporate performance such as stock prices. Therefore, such an economic phenomenon is of great concern to users such as institutional investors, securities companies, and individual investors, and many users wonder how an economic phenomenon affects a company's business performance. We are always predicting. However, since various economic phenomena occur all the time in the world, it becomes a great burden for the user to read a huge number of news and analyze the influence of the economic phenomena on the business performance.

In this regard, Non-Patent Document 1 describes a method of automatically expressing a causal relationship between economic and financial events by expressing an economic/financial event with a numerical vector, but between events in causal relationship building. Similarity determination is performed by a general method such as IDF or cosine similarity, and nothing is described about a device for establishing a causal relationship in economic and financial phenomena.

The present invention has been made in view of such circumstances, and an object thereof is to efficiently automate the extraction of a causal relationship between an economic phenomenon and business performance by computer processing.

In order to solve such a problem, the first invention has a causal relation database, a news analysis unit, a settlement analysis unit, a node registration unit, and a causal relation registration unit, and correlates economic events with corporate performance. A causal relationship extraction system for extracting sex as a causal relationship is provided. The causal relationship database is composed of a plurality of event nodes to which a digest of economic events is added, a plurality of performance nodes to which a digest of corporate performance is added, a connection between nodes related to a specific economic event pair, and a specific economic event. And connection between nodes for a particular pair of corporate performance. The news analysis unit analyzes the content of the news to extract the economic phenomenon and also generates a digest of the economic phenomenon. The financial statement analysis unit analyzes the content of the settlement related material to extract the economic phenomenon and the business performance, and also generates the digest of the economic phenomenon and the digest of the business performance. The node registration unit registers, in the causal relation database, the event node to which the digest of the economic event is added, provided that the event node having the same digest is not registered for the extracted economic event. At the same time, the node registration unit registers, in the causal relationship database, a performance node added with a digest of the corporate performance, provided that no performance node of the same digest is registered for the extracted corporate performance. The causal relationship registration unit uses a first extraction method based on a specific word or phrase indicating a causal relationship, or a machine learning method using learning data that is a set of two known factors that are known to have a causal relationship. When a causal relationship between specific economic events different from news is extracted by the extraction method of No. 2, the connection between nodes related to this economic event pair is registered in the causal relationship database based on the extracted causal relationship. .. At the same time, the causal relationship registration unit compares the morpheme sequence abstracted by the first extraction method, the second extraction method, or the attribute label with the factor pattern, and compares the morpheme sequence with the factor pattern. If a causal relationship between a particular corporate performance and a particular economic event is extracted from the financial statements-related material by the third extraction method that extracts the part specified by the factor pattern as a factor, the extracted causal relationship Based on this, the connection between nodes related to this economic event and this company performance pair is registered in the causal relationship database. Here, the digest of an economic event is structured by dividing the content of the economic event into a plurality of predetermined items, and these items are items that represent the quantitative amount or tendency of the economic event and their fluctuations. There is an item indicating a direction (however, a variation numerical value is not included). The corporate performance digest is structured by dividing the content of the corporate performance into a plurality of predetermined items. These items are an item representing an account item and an item representing the changing direction thereof. Have.

Here, in the first invention, the node registration unit refers to a name identification dictionary in which a plurality of expression patterns regarded as the same digest are registered, and with respect to the extracted economic event and the extracted corporate performance, the same digest event node And it is preferable to judge whether or not the performance nodes of the same digest are registered in the causal relationship database.

In the first aspect of the invention, the causal relationship registering unit copies a series of registered nodes that are registered in a causal relationship database and that have a connection relationship as an event node and an achievement node, after reversing the direction of change. By doing so, it may be registered in the causal relationship database as a predicted event node and a predicted achievement node. In this case, the causal relationship registration unit refers to an inversion dictionary in which a plurality of sets of expressions having an inversion relationship are registered, and on the condition that there is an expression corresponding to the inversion of the fluctuation direction in the registered node, the prediction event It is preferable to register as a node and a forecast performance node.

A second invention provides a causal relationship extraction program for causing a computer to execute the following steps to extract a correlation between an economic phenomenon and a business performance as a causal relationship. In the first step, the content of the news is analyzed to extract the economic phenomenon and a digest of the economic phenomenon is generated. In the second step, the contents of the settlement related material are analyzed to extract the economic phenomenon and the business performance, and at the same time, the economic phenomenon digest and the business performance digest are generated. In the third step, with respect to the extracted economic event, the event node to which the digest of the economic event is added is registered in the causal relationship database, provided that no event node having the same digest is registered. In the fourth step, with respect to the extracted corporate performances, the performance node added with the digest of the corporate performances is registered in the causal relationship database, provided that the performance nodes of the same digest are not registered. In the fifth step, a first extraction method using a specific word or phrase indicating a causal relationship as a clue, or a machine learning method using learning data that is a set of two known factors that have a causal relationship is known. When a causal relationship between specific economic events different from news is extracted by the extraction method of No. 2, the connection between the nodes related to the pair of these economic events is registered in the causal relationship database based on the extracted causal relationship. To do. In the sixth step, the morpheme sequence abstracted by the first extraction method, the second extraction method, or the attribute label is compared with the factor pattern, and the morpheme sequence matching the factor pattern is identified by the factor. When the causal relationship between a specific corporate performance and a specific economic phenomenon is extracted from the financial statements-related material by the third extraction method that extracts the portion specified by the pattern as a factor, based on this extracted causal relationship, Register the connections between nodes for this economic event and this pair of company performances in a causal relationship database. Here, the digest of an economic event is structured by dividing the content of the economic event into a plurality of predetermined items, and these items are items that represent the quantitative amount or tendency of the economic event and their fluctuations. There is an item indicating a direction (however, a variation numerical value is not included) . The corporate performance digest is structured by dividing the content of the corporate performance into a plurality of predetermined items. These items are an item representing an account item and an item representing the changing direction thereof. Have.

Here, in the second invention, the third and fourth steps refer to a name identification dictionary in which a plurality of expression patterns regarded as the same digest are registered, and the extracted economic phenomenon and the extracted corporate performance are It is preferable to have a step of determining whether or not the same digest event node and the same digest performance node are registered in the causal relationship database.

In the second invention, a causal relationship database is created by inverting the direction of change and copying a series of registered nodes that are registered in the causal relationship database and that are connected as event nodes and achievement nodes. May have a seventh step of registering as a predicted event node and a predicted achievement node. In this case, the seventh step refers to the inversion dictionary in which a plurality of sets of expressions having an inversion relation are registered, and the prediction event is provided on condition that there is an expression corresponding to the inversion of the fluctuation direction in the registered node. It is preferable to register as a node and a forecast performance node.

According to the present invention, causal relationship extraction processing is performed on economic events and corporate performance on a digest basis that eliminates redundancy of information, including determination of identity of content when registering in a causal relationship database. The economic event digest is structured in such a way that it briefly characterizes the economic event, taking into account the description characteristics of the news, and an item that indicates the quantitative amount or tendency of the economic event and an item that indicates the direction of change. Have. In addition, the corporate performance digest is structured in a way that briefly characterizes the corporate performance in consideration of the description characteristics of the financial statements-related materials. Have. This makes it possible to efficiently automate the extraction of the causal relationship between economic events and corporate performance by computer processing.

Conceptual diagram of the service model according to the present embodiment The figure which shows an example of the display screen in news analysis Diagram showing an example of correlation graph in news analysis Block diagram of correlation graph display device Conceptual diagram of causality database Diagram showing an example of correlation graph in profit impact factor analysis Block diagram of correlation graph generation system Diagram showing the structure of the news database Figure showing the structure of the settlement related database Block diagram of causal relationship extraction system Explanatory diagram of extraction of causal relationship using text structure analysis Illustration of causal relationship extraction method using machine learning Diagram showing economic events extracted from news Diagram showing event nodes extracted from news Diagram showing causal relationships between event nodes extracted from news Diagram showing the causal relationship between the event node and the achievement node extracted from the financial statements-related materials Explanatory diagram of correlation graph when reversing new news Explanatory diagram of generation process of forecast event node and forecast achievement node

FIG. 1 is a conceptual diagram of a service model according to this embodiment. The analysis server managed by the operator of the service or a person who is entrusted with the service acquires news distributed from outside such as a news company. The analysis server also acquires settlement-related materials distributed from the outside such as stock exchanges and companies, that is, disclosure materials such as financial statements and securities reports, and analyzes these. Then, the analysis server provides users such as institutional investors, securities companies, and individual investors with various services including news analysis and profit influence factor analysis described later. One of the characteristic services provided by the analysis server is to provide a correlation graph that visualizes the correlation between economic events and corporate performance. In the present specification, the term "economic phenomenon" is a concept that broadly encompasses phenomena that can affect economic activities, and not only economic phenomena in a narrow sense, but also natural phenomena that affect the economy, For example, it is used in the concept including cold summer, large earthquake, and large flood.

FIG. 2 is a diagram showing an example of a display screen in news analysis. The display screen 1 is displayed on the display device included in the client operated by the user. The display screen 1 has a search candidate display area 1a, a correlation graph display area 1b, a news text display area 1c, and an affected company display area 1d, which are provided in different areas as a part of the display screen 1. .. In the search candidate display area 1a, a list of search candidate groups selectable by the user is displayed. The displayed search candidate group varies depending on the content of the service, but in the case of news analysis, various news titles are displayed in a list. When the user selects a search target (news) from the displayed search candidate group (news group), the correlation graph display area 1b displays a correlation graph corresponding thereto. The news text selected by the user is displayed in the news text display area 1c. In the affected company display area 1d, a list of companies that can be affected by the economic phenomenon shown in the news selected by the user is displayed in a list.

FIG. 3 is a diagram showing an example of a correlation graph in news analysis. This correlation graph visualizes the correlation between economic events and corporate performance. The correlation graph includes at least one economic event node and at least one performance node. The structure of the correlation graph may be in the form of a tree having branches, or may be in the form of a column. In the following explanation, nodes related to economic events such as "India copper plant closure" and "India copper product production decrease" are referred to as "event nodes", and "○○ Paint business profit decrease" and "△△ automobile business profit decrease". A node related to corporate performance such as "is referred to as "performance node".

In the correlation graph, specific nodes (a pair of nodes) are connected by directed edges. This directed edge indicates that there is a causal relationship between the nodes connected by this, and the origin of the directed edge means “factor” and its end point means “result”, respectively. When a pair of event nodes are connected by a directed edge, it means that the economic event corresponding to the event node on the starting side is the cause of the economic event corresponding to the event node on the starting side. To do. In addition, when the event node and the performance node are connected by a directed edge, an economic event corresponding to the event node on the starting side causes the corporate performance corresponding to the performance node on the ending side. Means that It should be noted that in the present specification, the “directed side” is a concept that broadly includes symbols that the user can distinguish between the factor side and the result side, and typically, an arrow from the factor side to the result side. However, other than that, a triangle with the base on the factor side, the vertex on the result side, the “factor” on the factor side, and the line segment “result” on the result side may be used.

The example of the figure shows a case where the user selects “Indian copper factory closed” as a search target. In this case, the “closed Indian copper factory”, which is the search target, is the highest node, and other economic events or corporate performances that have a direct or indirect causal relationship with this are shown as lower nodes. In other words, with regard to the economic phenomenon of "Indian copper factory closure", this is a direct factor, and the economic phenomenon (result) of "decrease in Indian copper product production" has occurred. In addition, two economic events (results), "decrease in Indian car production" and "increase in Indian copper import", are caused by "decrease in Indian copper product production". These two economic events are directly caused by the latest “decrease in Indian copper product production”, but indirectly due to the closure of the highest Indian copper factory, which has a causal relationship with this. It was caused. Furthermore, with regard to the "decrease in Indian automobile production," this is a direct factor, which results in two corporate performances (results): "○○ Paint operating profit decrease" and "△△ automobile operating profit decrease". .. This is also caused by an indirect factor, which is the closure of the highest-ranking Indian copper plant, which has a causal relationship with the “decrease in Indian automobile production”.

From this, as for the closure of the Indian copper factory, this business phenomenon caused the decrease in operating income of ○○ Paint, the decrease in operating income of △△ automobiles, the increase in operating income of □□ materials, and It can be understood that fluctuations in the business performance of each company such as “XX increase in electric wire operating income” are caused. Thus, when a particular news item is selected as a search target, the correlation graph visualizes a series of economic events having a direct or indirect causal relationship with the economic event indicated by the particular news item. , Corporate performance that has a direct or indirect causal relationship with this is also visualized.

In addition, in the example of the same figure, a plurality of corporate performances are arranged in the lowest node of the correlation graph, but this is an example, and in some cases, the corporate performance may be derived from the highest node or the intermediate node. is there. However, the performance node indicating the business performance is basically the terminal node of each branch in the correlation graph. Further, the achievement node may have a plurality of nodes assigned to each business domain of the company. For example, a company called “□□ material” is subdivided into a plurality of performance nodes such as “cement business”, “metal business”, “aluminum business”, and “electronic material business”. Further, when the number of achievement nodes to be displayed in the correlation graph is equal to or larger than the predetermined number, the companies having a high influence degree may be displayed first. It is preferable to quantitatively determine the degree of this degree of influence by making full use of the strength of expression in a sentence, the appearance frequency of a specific expression, statistical analysis, and the like. The number of display steps of the correlation graph starting from the search target may be set in advance as the system, or may be arbitrarily set by the user. Further, in the search candidate display area 1a, only news relating to a specific company may be displayed by the user's selection or designation.

FIG. 4 is a block diagram of the correlation graph display device according to the present embodiment. In the present embodiment, the correlation graph display device 10 is constructed on the client side including a browser. The correlation graph display device 10 includes a display switching unit 2 in addition to the display device that displays the search candidate display area 1a and the correlation graph display area 1b described above. The display switching unit 2 accepts a user's selection of the search candidate group displayed in the search candidate display area 1a, and causes the correlation graph display area 1b to display a correlation graph corresponding to the search target selected by the user. When constructing the correlation graph, the causal relationship database 3 provided on the analysis server side is referred to. Therefore, the display switching unit 2 transmits a search request regarding the search target to the analysis server and receives the search result (correlation graph) from the analysis server. It should be noted that such a display switching unit 2 can be realized as a function of a web browser installed in the client.

FIG. 5 is a conceptual diagram of the causal relationship database 3. The causal relationship database 3 holds information necessary for constructing a correlation graph, and specifically, a plurality of event nodes, a plurality of achievement nodes, and a connection between nodes related to a specific pair of event nodes. And a connection between nodes for a particular event node and a particular performance node pair. Each node held in the causal relationship database 3 is provided with identification information (node ID) unique to the node, and each node is managed by this node ID. As described above, the node includes the event node corresponding to the economic event and the performance node corresponding to the corporate performance. In order to distinguish the two, the node IDs of the event nodes are “a1” to “a7”. , And the node IDs of the achievement nodes are shown as “b1” to “b3”. In addition, if there is a causal relationship between a pair of economic events, the connection between the pair of event nodes corresponding to these is maintained in a form that can identify the factors and the results (corresponding to the directed side of the correlation graph). There is. Furthermore, when there is a causal relationship between a certain economic phenomenon and a certain business performance, the connection between the event node corresponding to the former and the performance node corresponding to the latter is also maintained. As will be described later, the causal relationship database 3 also holds node information (digest) regarding each node.

The display switching unit 2 switches the correlation graph to be displayed in the correlation graph display area 1b when the search target is changed by the user's operation. That is, in the search candidate display area 1a, when the first news is selected as the search target from the news group displayed as the search candidate group, the first correlation graph is displayed. In the first correlation graph, the first economic event corresponding to the first news to be searched is the top node, and other economic events and company performances that correlate with the first economic event are subordinate nodes. .. Further, when the second news different from the first news is selected as the search target, the second correlation graph different from the first correlation graph is displayed. In the second correlation graph, the second economic event corresponding to the second news item to be searched is the top node, and other economic events and company performances correlated with the second economic event are the subordinate nodes. ..

Next, the profit influence factor analysis will be described. In the case of profit influence factor analysis, unlike the news analysis, in the search candidate display area 1a, a group of company achievements selectable by the user is displayed. When the user selects a search target (corporate achievement) from the search candidate group (corporate achievement group), a correlation graph corresponding to this is displayed in the correlation graph display area 1b.

FIG. 6 is a diagram showing an example of a correlation graph in the profit influence factor analysis. The example in the figure shows the case where “Increase in operating income of □□ material” is selected as the search target. In this case, “Increase in operating income of □□ material” is the lowest node, and the economic phenomenon that is a direct or indirect factor of this corporate performance is displayed as the upper node. From this correlation graph, the corporate performance of "Increase in operating income of □□ material", which is the search target, has "Increase in production and export volume of copper products in Japan" as a direct factor, and "Increase in import volume of copper products in India", It can be understood that "decrease in Indian copper product production" and "India copper factory closure" are indirect factors. In this way, when a specific company performance is selected as a search target, the correlation graph visualizes a series of economic events having a direct or indirect causal relationship with the specific company performance.

In the profit influence factor analysis, as in the news analysis described above, the display switching unit 2 switches the correlation graph to be displayed in the correlation graph display area 1b when the search target is changed by the user's operation. That is, in the search candidate display area 1a, when the first corporate performance is selected as the search target from the corporate performance group displayed as the search candidate group, the third correlation graph is displayed. In the third correlation graph, the first corporate performance is the lowest node, and the economic phenomenon having a correlation with the first corporate performance is the upper node. When a second company performance different from the first company performance is selected, a fourth correlation graph different from the third correlation graph is displayed. In the fourth correlation graph, the second corporate performance is the lowest node, and the economic phenomenon having the correlation with the second corporate performance is the upper node.

Regarding the search of the causal relation database 3, the news analysis and the profit influence factor analysis have opposite connection tracing. In the case of news analysis, the connection relationship is traced from the factor (upper) to the result (lower) according to the direction of the directed side. On the other hand, in the case of profit influence factor analysis, the factor (upper) is traced from the result (lower), contrary to the direction of the directed side.

FIG. 7 is a functional block diagram of a correlation graph generation system that provides a correlation graph to the correlation graph display device 10 described above. In the present embodiment, the correlation graph generation system 20 is constructed on the analysis server side. It should be noted that this generation system 20 can be equivalently realized by installing a computer program (correlation graph generation program) that causes a computer to function and operate as blocks 4 to 6 into the analysis server.

The correlation graph generation system 20 has a correlation graph generation unit 4, a causal relationship extraction system 5, a news database 7, and a settlement related database 8 in addition to the above-described causal relationship database 3.

The correlation graph generation unit 4 searches the causal relationship database 3 based on a search request (including a search target instruction) from the client side using the event/achievement node related to the search target as a search key. As a result, at least one node that is directly and indirectly connected to the search key is extracted, including the connection relationship between the nodes. Then, the correlation graph generation unit 4 generates a correlation graph based on the search result of the causal relationship database 3. This correlation graph includes a node serving as a search key and at least one node extracted as a search result of the causality database 3. In addition, the directed side of the correlation graph is formed based on the connection relation between the nodes extracted as the search result of the causal relation database 3.

In the case of news analysis, the causal relationship database 3 is searched using the event node corresponding to the specific news selected by the user as a search key, and based on the search result, the event node corresponding to the specific news and the A correlation graph is generated that includes other event nodes and achievement nodes that are connected indirectly or indirectly. On the other hand, in the case of profit influence factor analysis, the causal relationship database 3 is searched with the performance node corresponding to the specific corporate performance selected by the user as a search key, and the performance corresponding to the specific corporate performance is based on this search result. A correlation graph including a node and an event node directly or indirectly connected to the node is generated. The correlation graph generated by the correlation graph generation unit 4 is transmitted to the client side as the search result for the search request.

FIG. 8 is a diagram showing the configuration of the news database 7. The news database 7 holds a large number of news acquired from the outside. The database 7 holds a large number of records composed of "news ID", "news information" and "node ID". The "news ID" is unique identification information for identifying each piece of news. This news ID is assigned each time news acquired from outside is registered in the news database 7. The "news information" is the contents of the news itself, and includes the title of the news, the text, and other incidental information (date, etc.). “Node ID” indicates the node ID of the event node corresponding to the economic event extracted from this news. Regarding the node ID, a plurality of node IDs may be described in one news item, for example, news ID=“N002” to “N004”. This means that a plurality of economic events are extracted from one news item and event nodes corresponding to these are registered in the causal relationship database 3. On the contrary, one node ID (=“a1”) may be redundantly described for a plurality of news items such as news ID=“N001” and “N002”. This means that, although economic events are individually extracted from each news item, only one event node a1 is registered in the causal relationship database 3 as a result of determining that they are the same economic event.

FIG. 9 is a diagram showing the structure of the settlement related database 8. The settlement related database 8 holds a large number of settlement related materials acquired from the outside. This database 8 holds a large number of records composed of “account settlement ID”, “account settlement related information” and “node ID”. The “account settlement ID” is unique identification information for identifying each settlement related material. This settlement ID is assigned each time a settlement-related material obtained from the outside is registered in the settlement-related database 8. “Financial statement related information” is the content itself of the financial statement related material. The "node ID" indicates the node ID (for example, a2) of the event node corresponding to the economic event extracted from this settlement related material and the node ID (for example, b1) of the performance node corresponding to the extracted corporate performance. .. For example, for the settlement ID=“K001”, it means that the economic phenomenon extracted from the settlement related material exists as the event node of “a2” and the business achievement exists as the achievement node of “b1”.

FIG. 10 is a functional block diagram of the causal relationship extraction system 5. The causal relationship extraction system 5 analyzes the news read from the news database 7 and the financial statement-related materials read from the financial statement-related database 8 and determines the correlation between economic events and corporate performance as a causal relationship. The information is extracted and necessary information is registered in the causal relationship database 3. The causal relationship extraction system 5 has a news analysis unit 5a, a settlement analysis unit 5b, a node registration unit 5c, and a causal relationship registration unit 5d. In the present embodiment, the causal relationship extraction system 5 is constructed on the analysis server side, but by installing a computer program (causal relationship extraction program) that causes a computer to function and operate as blocks 5a to 5d into the analysis server. It is also possible to realize equivalently.

The news analysis unit 5a reads out the news stored in the news database 7, analyzes the contents of the news, extracts the economic phenomenon, and generates a digest of the economic phenomenon. The economic event digest is structured by dividing the content of the economic event into a plurality of predetermined items, and directly expresses the characteristic (feature amount) of the economic event without redundancy. .. In the present embodiment, in order to characterize the content of the news in consideration of the description characteristics, the plurality of items that define the digest of the economic phenomenon are "area", "name", " "Element" and "predicate" are used. Here, the “area” is an item representing the area name where the economic phenomenon occurs, such as “Southeast Asia” and “Japan”. The “name” is an item representing the name of an economic phenomenon, such as “soft drink” or “crude oil”. "Element" means "price", "sales volume", "demand", "tax rate", "economy", "market conditions", "trade friction", "business condition", "regulation". It is an item that represents the quantitative amount or tendency of economic events. “Fluctuation (predicate)” is an item that indicates the direction of change in economic events (“elements”), such as “fall”, “decrease”, “increase”, “intensify”, “deteriorate”, and “strengthen”. .. For example, for the economic phenomenon of “increase in import quantity of Indian copper products”, a digest is generated in which “region”=India, “name”=copper product, “element”=import quantity, and “fluctuation”=increase. .. However, of these four items, the most important ones that characterize economic events are "element" and "predicate", that is, "what" is "what happened" (eg "price"). Has “fallen”). Therefore, "element" and "predicate" are indispensable, but "area" and "name" may be appropriately adopted as necessary, or Other items other than the above may be added. In addition, depending on the content of the economic phenomenon, not all “area”, “name”, “element”, and “predicate” clearly exist. Therefore, as the digest pattern of the economic phenomenon, it is possible to allow the digest with a smaller number of items than the above four items.

The node registration unit 5c assigns a node ID to the extracted economic event on condition that no economic event having the same content is registered in the causal relationship database 3 (elimination of duplicate registration), and then a new one is created. It is registered in the causal relationship database 3 as an event node. It is necessary to judge whether or not economic phenomena are the same after taking into account notational fluctuations. Therefore, the identity is judged after standardizing the expressions by referring to a name identification dictionary in which many patterns of expressions considered to be the same are registered (data cleansing). For example, economic events such as “Rise in crude oil price”, “Oil price up”, and “Crude oil price” are the same as a result of being standardized to “name”=crude oil, “factor”=price, and “fluctuation”=upward. Is judged. Note that, for example, if the digest including the “consumption tax increase ” and the digest including the “consumption tax” and the “ increased ” are determined to be the same by referring to the name identification dictionary, the digests having different numbers of items are the same. It becomes possible to judge the sex.

In the causality database 3, a digest whose expression is standardized by a name identification dictionary is registered as node information of an event node. The same applies to the digest of corporate performance extracted from the materials related to financial results.

The settlement analysis unit 5b reads out the settlement-related materials stored in the settlement-related database 8 and analyzes the contents of the settlement-related materials to extract economic events and corporate performance. Regarding the extracted economic event and corporate performance, the node registration unit 5c provides a new event node and a new business performance, provided that no causal relationship database 3 has registered the same economic business or corporate performance. It is registered in the causal relationship database 3 as a node. Further, the settlement analysis unit 5b generates a digest of economic events and a digest of corporate performance for the extracted economic phenomena and corporate performance. The digest of economic events is similar to that of news. Further, the corporate performance digest is structured by dividing the content of the corporate performance into a plurality of predetermined items, and directly expresses the characteristic (feature amount) of the corporate performance without redundancy. ing. In the present embodiment, “accounting item” and “fluctuation” are used as a plurality of items that define the digest of the corporate performance in order to characterize the content of the settlement-related material in consideration of the description characteristic. The “account item” is an item representing the account item, and the “fluctuation” is an item representing the variation direction of the account item. As in the case of news, the economic events and corporate performances extracted from the financial statements-related materials are numbered node IDs on condition that the same thing is not registered in the causal relationship database 3, and then the causal relationship database is created. Registered in 3.

The digest of economic events and the digest of corporate performance registered in the causal relationship database 3 are used as node information in the correlation graph. That is, the digests of economic events read from the causal relationship database 3 are individually added as node information of event nodes in the correlation graph. Further, the corporate performance digest read from the causal relationship database 3 is individually added as node information of the performance node in the correlation graph.

The causal relationship registration unit 5d reads the news stored in the news database 7 and extracts the causal relationship between different economic events. Basically, when one news contains a plurality of economic events and it is determined that a pair of economic events has a cause-effect relationship, a causal relationship is recognized. Specifically, first, one economic event is specified based on the structured digest as described above. When two or more digests are extracted from one news, a set in which a causal relationship is recognized between the pair of digests is extracted based on the text expression of the news. For example, if there is news that "... this year in Japan, sales of soft drinks have fallen due to record cold summer...", the phrase "due to" will be followed by "record cool summer" and "soft drinks". It is judged that there is a causal relationship between the pair of economic events of "falling water". Note that temporal information may be taken into consideration when determining whether there is a causal relationship, in order to prohibit the association of economic events with extremely different dates and times.

In addition to the simple method using a specific word as a clue, if the following method is adopted, the causal relationship between economic events can be more accurately extracted. FIG. 11 is an explanatory diagram of a causal relationship extraction method using text structure analysis as the first method. In order to identify the sentences in the sentence, paragraph numbers 1 to 5 are given to the individual sentences for convenience. In the first method, it is determined whether the two factors are located in the cause and effect paragraphs in the structure of the entire sentence. Specifically, the relation between clauses, sentences, and paragraphs is estimated by the text structure analysis, and the "name", "element", and "variation" contained in the relation are the cause and effect. Be a candidate for a causal relationship. This is a more extensive systematization of techniques that use specific words as clues.

FIG. 12 is an explanatory diagram of a causal relationship extraction method using machine learning as the second method. In the second method, each of learning data (a set of two known factors with clear causal relationship) is converted into a factor vector in advance. The factor vector is a numerical sequence, and is obtained by concatenating individual vectorized words of “name”, “element”, and “variation” using a technique called word2vec, for example. Then, the learning data is machine-learned to prepare a learned database. As a machine learning method, a two-class classification method such as support vector machine (SVM) or random forest can be used. Whether or not there is a causal relationship can be determined by calculating a factor vector from the pairs of "region", "name", "element", and "variation", and inputting this factor vector into a trained model.

In the third method, a news summary sentence is generated as a preprocessing by the document summarization technique, and the first technique or the second technique is applied to the summary sentence. There are extraction type and abstract type in the abstract. The extraction type is a type that summarizes using the original words and phrases, and since redundant structure of sentences is eliminated, there is a possibility that the above-mentioned “name”, “element” and “variation” detection performance can be improved. .. On the other hand, the abstract type is a type that generates a phrase that does not exist in the original sentence in the summary sentence, and there is a possibility that "name", "element" and "fluctuation" and causality that cannot be obtained directly from the original sentence are obtained. Can be expected.

Further, the causal relationship registration unit 5d reads out the settlement-related material stored in the settlement-related database 8, and the settlement-related material includes an economic phenomenon that is a factor affecting the corporate performance (account item and its fluctuation). If it is determined that the connection is established, the connection relationship between the company performance node corresponding to the settlement related information and the event node corresponding to the influencing factor is registered in the causal relationship database 3. As a method of extracting the causal relationship between the business performance and the economic phenomenon, the same extraction method as the causal relationship between the economic phenomena described above may be used, but the extraction of Japanese Patent No. 6155409 already obtained by the present applicant. It is preferable to use the method. This patent gazette discloses a method for analyzing factors related to accounting and extracting factors of accounting events. Specifically, first, morphological analysis is performed on each sentence obtained by decomposing the sentence included in the financial statement-related material, and a morpheme string of the sentence is generated. Next, a unique attribute label is assigned to at least each of the attributes for classifying the subject and the amount of money to the morphemes forming the morpheme string or the combination thereof. Then, the morpheme string abstracted by the attribute label is compared with the factor pattern to determine whether or not the morpheme string matches the factor pattern. Finally, for the morpheme string that matches the factor pattern, the portion designated by the factor pattern is extracted as a factor, and the extracted factor is linked to the subject and amount information in the morpheme string.

The causal relationship between the pair of economic events is registered in the causal relationship database 3 in a form showing that the event node corresponding to the factor and the event node corresponding to the result are connected. Further, the causal relationship between the economic event and the corporate performance is registered in the causal relationship database 6 in a form showing that the event node corresponding to the factor is connected to the performance node corresponding to the result.

Hereinafter, the flow of processing in the correlation graph generation system 20 will be specifically described with reference to FIGS. First, the news analysis unit 5a analyzes news individually and extracts economic phenomena in each news. As shown in FIG. 13, economic event 1 is from news “N001”, economic events 2 to 3 are from news “N002”, economic events 4 to 6 are from news “N003”, economic events 8 to 9 are news from “N004”. Economic events 10 to 12 are extracted from “N005”.

Next, under the condition that duplicate registration is prohibited, the node registration unit 5c registers the economic event extracted from the news in the causal relationship database 3 as a new event node. As shown in FIG. 14, since the economic event 1 extracted from the news “N001” has not been registered, it is registered as a new event node “a1”. For the news “N002”, two economic events 2-3 are extracted, but the economic event 2 is the same as the already registered economic event 1. Therefore, the economic event 2 is assigned the event node “a1” corresponding to the economic event 1, and only the economic event 3 is registered as a new event node “a2”. For the news “N003”, three economic events 4 to 6 are extracted, but the economic event 4 is the same as the already registered economic event 1. Therefore, the economic event 4 is assigned the event node “a1” corresponding to the economic event 1, and only the economic events 5 to 6 are registered as new event nodes “a3” to “a4”. For the news “N004”, two economic events 7 to 8 are extracted, but the economic event 7 is the same as the already registered economic event 6. Therefore, the economic event 7 is assigned the event node “a4” corresponding to the economic event 6, and only the economic event 8 is registered as a new event node “a5”. For the news “N005”, three economic events 9 to 11 are extracted, but the economic event 9 is the same as the already registered economic event 5. Therefore, the economic event 9 is assigned the event node “a3” corresponding to the economic event 5, and only the economic events 10 to 11 are registered as new event nodes “a6” to “a7”.

Next, the causal relationship registration unit 5d extracts the causal relationship between the economic events from the news and registers it in the causal relationship database 3 as a connection relationship in which the factor and the result can be identified. As shown in FIG. 15, a causal relationship between economic event 2 (factor) and economic event 3 (result) is extracted from news “N002”. As a result, a connection having the event node “a1” corresponding to economic event 2 as the starting point and the event node “a2” corresponding to economic event 3 as the ending point is registered. From the news “N003”, the causal relationship between economic event 4 (factor) and economic event 5 (result) and the causal relationship between economic event 5 (factor) and economic event 6 (result) are extracted. As a result, for the former, a connection in which the event node “a1” corresponding to economic event 4 is the starting point and the event node “a3” corresponding to economic event 4 is the ending point is registered. For the latter, a connection is registered with the event node “a3” corresponding to economic event 5 as the starting point and the event node “a4” corresponding to economic event 6 as the ending point. A causal relationship between economic event 7 (factor) and economic event 8 (result) is extracted from news “N004”. As a result, the event node “a4” corresponding to the economic event 7 is registered as a starting point, and the event node “a5” corresponding to the economic event 8 is registered as an end point. From the news “N005”, the causal relationship between the economic event 9 (factor) and the economic event 10 (result) and the causal relationship between the economic event 10 (factor) and the economic event 11 (result) are extracted. As a result, for the former, a connection with the event node “a3” corresponding to the economic event 9 as the starting point and the event node “a6” corresponding to the economic event 10 as the ending point is registered. For the latter, a connection is registered with the event node “a6” corresponding to economic event 10 as the starting point and the event node “a7” corresponding to economic event 11 as the ending point.

Then, the causal relationship extracting unit 5d extracts a causal relationship between the event node and the achievement node from the settlement related material, and registers the causal relationship in the causal relationship database 3 as a connection relationship capable of identifying the factor and the result. As shown in FIG. 16, a causal relationship between a certain economic event (factor) and a certain corporate performance (result) is extracted from the settlement related material “K001”, and the former is the same as the economic event of the event node “a2”. And the latter is assumed to be unregistered. As a result, a new achievement node “b1” corresponding to the corporate achievement is registered, and a connection having the event node “a2” as a starting point and the achievement node “b1” as an ending point is registered. A causal relationship between a certain economic event (factor) and a certain corporate performance (result) is extracted from the financial statement-related material "K001", the former is the same as the economic event of the event node "a4", and the latter is unregistered. Suppose As a result, a new achievement node “b2” corresponding to the corporate achievement is registered, and a connection with the event node “a4” as the starting point and the achievement node “b2” as the ending point is registered. A causal relationship between a certain economic event (factor) and a certain corporate performance (result) is extracted from the financial statement-related material "K003", the former is the same as the economic event of the event node "a5", and the latter is not registered. Suppose As a result, a new achievement node “b3” corresponding to the corporate achievement is registered, and a connection having the event node “a5” as a starting point and the achievement node “b3” as an ending point is registered.

Through the series of processes described above, for example, a causal relationship as shown in FIG. 5 is constructed as the content retained in the causal relationship database 3.

Finally, as an extended example of the correlation graph, a predictive correlation graph that predicts the causal relationship between economic events and corporate performance will be described. The upper part of FIG. 17 shows an example of the correlation graph generated by the extraction processing described above. That is, with the event node a11 of “decrease” of “record cold summer” in “Japan” as a route, “decrease” (event node a12) of “soft drink sales volume” in “Japan” and “aluminum” in “Japan”. After "decrease" of "demand" (event node a13) and "decrease" of "aluminum price" in "Japan" (event node a14), performance node b10, that is, a specific company that manufactures products using aluminum as a raw material (○) ○ Material) has reached “increased operating income”.

Here, as a hypothetical case, consider a case where the variation direction of the event node a11 is reversed from “decrease” to “increase”. In this case, from the above correlation graph, it is expected that the changing directions of the event nodes a12 to a14 and the achievement node b10 after the event node a11 are also reversed. That is, with the predictive event node a51 of "increasing" of "record cold summer" in "Japan" as the route, "increasing" (predictive event node a52) of "soft drink sales volume" in "Japan", and in "Japan" After "increasing" of aluminum demand (forecast event node a53) and "increasing" of "aluminum price" in Japan (forecast event node a54), finally, forecast performance node b50, that is, aluminum as a raw material It is reasonably expected that a specific company that manufactures products (○○ Material) will have a "operating loss". In this way, even if it does not actually occur as an economic event or corporate performance, it is useful information for the user if it is predicted as a reversal event or reversal performance based on what actually occurred and visualized as a prediction correlation graph. ..

FIG. 18 is an explanatory diagram of the generation process of the predicted event node and the predicted achievement node. First, the causal relationship registering unit 5d registers a new node ID=a51-a11 with respect to a series of registered nodes a11-a14, b10 that are registered in the causal relationship database 3 and that are connected as event nodes and achievement nodes. After assigning numbers a54 and b50, they are registered in the causal relationship database 3 as prediction nodes (prediction event node and prediction achievement node). The prediction node is obtained by inverting (for example, “decreasing”→“increasing”) the changing direction of the registered node corresponding thereto, and then copying it including the connection relationship. As a result, the predicted correlation graph shown in the lower part of FIG. 17 is generated.

Here, when registering a prediction node, an inversion dictionary in which a plurality of sets of expressions having an inversion relation are registered is referred to, and an expression corresponding to inversion of a fluctuation direction in a registered node (for example, “increase” to “decrease”) Is required to exist. Therefore, for a registered node for which an inverted expression does not exist, the corresponding prediction node is not registered. For example, when it is determined that the inverted expression does not exist for the registered node a13, the prediction node a53 is not registered. As a result, the two predictive nodes a51 and a52 connected to each other and the two predictive nodes a54 and b50 connected to each other are registered in a separated form.

As described above, according to the present embodiment, the correlation between the economic phenomenon and the business performance is visualized through the correlation graph, and the factor of the change in the business performance can be easily grasped, so that the convenience of the user can be enhanced.

Further, according to the present embodiment, the user can easily obtain a desired correlation graph by only selecting the search candidate group displayed in the search candidate display area 1a without requiring a complicated operation. At the same time, the correlation graph can be easily switched by changing the search target. At that time, if different analysis services such as news analysis and profit influence factor analysis are provided, user convenience can be further enhanced.

Further, according to the present embodiment, by displaying the digest of the economic phenomenon and the corporate performance as the node information of the nodes forming the correlation graph, the convenience of the user can be further enhanced.

Further, according to the present embodiment, the causal relationship extraction process is performed on the basis of the digest, which eliminates the redundancy of information, including the judgment of the identity of the contents when registering in the causal relationship database 3 for the economic phenomenon and the business performance. I do. The economic event digest is structured in such a way that it briefly characterizes the economic event, taking into account the description characteristics of the news, and an item that indicates the quantitative amount or tendency of the economic event and an item that indicates the direction of change. Have. In addition, the corporate performance digest is structured in such a way that it briefly characterizes corporate performance in consideration of the description characteristics of the financial statements-related materials. Have. This makes it possible to efficiently automate the extraction of the causal relationship between economic events and corporate performance by computer processing.

1 Display Screen 1a Search Candidate Display Area 1b Correlation Graph Display Area 1c News Body Display Area 1d Affected Company Display Area 2 Display Switching Section 3 Causal Relationship Database 4 Correlation Graph Generation Section 5 Causal Relationship Extraction System 5a News Analysis Section 5b Financial Statement Analysis Section 5c Node registration section 5d Causal relationship registration section 7 News database 8 Financial statements related database 10 Correlation graph display device 20 Correlation graph generation system

Claims (8)

In a causal relationship extraction system that extracts the causal relationship between economic phenomena and corporate performance,
Multiple event nodes with economic event digests, multiple performance nodes with corporate performance digests, connections between nodes for specific economic event pairs, and specific economic events and specific corporate performances A causal database holding connections between nodes for pairs of
A news analysis unit that analyzes the content of news and extracts economic events, and also generates a digest of economic events,
A settlement analysis unit that analyzes the contents of settlement related materials to extract economic events and corporate performance, and also generates a digest of economic events and a corporate performance digest,
Regarding the extracted economic phenomenon, on the condition that the event node of the same digest is not registered, while registering the event node added with the digest of the economic phenomenon in the causal relationship database, regarding the extracted corporate performance, A node registration unit that registers a performance node to which a digest of corporate performance is added to the causal relationship database, provided that no performance node of the same digest is registered.
By the first extraction method using a specific word or phrase indicating a causal relationship as a clue, or the second extraction method by machine learning using learning data that is a set of two known factors that have a causal relationship, When a causal relationship between different specific economic events is extracted from the news, based on the extracted causal relationship, the connection between the nodes related to the economic event pair is registered in the causal relationship database and the first causal relationship is registered. The extraction method, the second extraction method, or the morpheme sequence abstracted by the attribute label is compared with the factor pattern, and the morpheme sequence that matches the factor pattern is caused by the portion specified by the factor pattern as a factor. When a causal relationship between a specific corporate performance and a specific economic event is extracted from the financial statements-related material by the third extraction method to extract, based on the extracted causal relationship, the relationship between the economic event and the corporate performance is extracted. And a causal relationship registration unit that registers a connection between nodes related to the causal relationship database,
The digest of the economic phenomenon is structured by dividing the content of the economic phenomenon into a plurality of predetermined items, and the plurality of items are items that represent the quantitative amount or tendency of the economic phenomenon and the changing direction thereof. (However, the variable value is not included.)
The corporate performance digest is structured by dividing the content of the corporate performance into a plurality of predetermined items, and the plurality of items include an item representing an account item and an item representing the direction of change thereof. A causal relationship extraction system characterized by having. The node registration unit refers to a name identification dictionary in which a plurality of expression patterns regarded as the same digest are registered, and with respect to the extracted economic event and the extracted corporate performance, the same digest event node and the performance node of the same digest are The causal relationship extraction system according to claim 1, wherein it is determined whether or not the causal relationship database is registered. The causal relationship registration unit, the series of registered nodes that are registered in the causal relationship database, and have a connection relationship as an event node and a performance node, by reversing the changing direction, and then copying. The causal relationship extraction system according to claim 1, wherein the causal relationship database is registered as a predicted event node and a predicted achievement node. The causal relationship registration unit refers to an inversion dictionary in which a plurality of sets of expressions having an inversion relationship are registered, and the prediction event node is provided on condition that an expression corresponding to the inversion of the fluctuation direction in the registered node exists. And the causal relationship extraction system according to claim 3, wherein the causal relationship extraction system is registered as the predicted achievement node. In the causal relationship extraction program that extracts the causal relationship between economic phenomena and corporate performance,
A first step of generating a digest of economic events while analyzing the contents of news and extracting economic events;
The second step of analyzing the contents of the settlement related material to extract the economic phenomenon and the business performance, and generating the digest of the economic phenomenon and the business performance.
A third step of registering, in the causal relationship database, an event node to which a digest of an economic event is added, provided that no event node having the same digest is registered for the extracted economic event.
A fourth step of registering, in the causal relationship database, a performance node to which a digest of the corporate performance is added, provided that no performance node having the same digest is registered for the extracted business performance.
By the first extraction method using a specific word or phrase indicating a causal relationship as a clue, or the second extraction method by machine learning using learning data that is a set of two known factors that have a causal relationship, When a causal relationship between different specific economic events is extracted from the news, a fifth step of registering a connection between nodes related to the economic event pair in the causal relationship database based on the extracted causal relationship. ,
The morpheme sequence abstracted by the first extraction method, the second extraction method, or the attribute label is compared with the factor pattern, and the morpheme sequence that matches the factor pattern is specified by the factor pattern. When a causal relationship between a specific corporate performance and a specific economic event is extracted from the financial statements-related material by the third extraction method that extracts as a factor, the economic event and the relevant causal relationship are extracted based on the extracted causal relationship. A sixth step of registering a connection between nodes relating to a pair of corporate achievements in the causal relationship database, causing a computer to execute the processing.
The digest of the economic phenomenon is structured by dividing the content of the economic phenomenon into a plurality of predetermined items, and the plurality of items are items that represent the quantitative amount or tendency of the economic phenomenon and the changing direction thereof. (However, the variable value is not included.)
The corporate performance digest is structured by dividing the content of the corporate performance into a plurality of predetermined items, and the plurality of items include an item representing an account item and an item representing the direction of change thereof. A causal relationship extraction program having: The third and fourth steps refer to a name identification dictionary in which a plurality of expression patterns regarded as the same digest are registered, and with respect to the extracted economic event and the extracted corporate performance, the same digest event node and the same digest 6. The causal relationship extraction program according to claim 5, further comprising a step of determining whether or not the achievement node of is registered in the causal relationship database. For a series of registered nodes that are registered in the causal relationship database and that have a connection relationship as an event node and a performance node, the predicted event is recorded in the causal relationship database by reversing the direction of change and copying. The causal relationship extraction program according to claim 5, further comprising a seventh step of registering as a node and a predicted achievement node. The seventh step refers to an inversion dictionary in which a plurality of sets of expressions having an inversion relation are registered, and the prediction event node is provided on condition that an expression corresponding to the inversion of the variation direction in the registered node exists. And the causal relationship extraction program according to claim 7, which is registered as the predicted achievement node.

Images