Future event prediction system and future event prediction method

Abstract

To predict possible future events based on events described in documents.SOLUTION: A future event prediction system predicts possible future events based on documents describing past events and the dates on which the past events occurred in natural language. The system includes: a past event tree generation unit configured to read past events from documents and generate a past event tree for documents whose vector distances are below a predetermined threshold; a language model generation unit configured to generate a future event prediction model from the past event tree; a prediction unit configured to accept event specifications and predict future events from the future event prediction model based on the accepted events; and an output unit configured to output the future events determined by the prediction unit.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a future event prediction system and a future event prediction method. [Background technology]

[0002] Future event predictions are made using technologies such as natural language processing, simulation, and machine learning. When using natural language processing, information is extracted from text data, and future trends are predicted based on that content. News articles and social media posts are often used.

[0003] Simulation is a method of modeling phenomena or systems and predicting their future behavior based on that model. Physical models and economic models are used. Machine learning is a technology that learns patterns and relationships from large amounts of data and uses them to predict future events.

[0004] Patent document 1 discloses a scenario generation device that includes a causal phrase pair DB that stores causal phrase pairs, a synonym generation unit that searches for causal phrase pairs that have cause phrases that are causally consistent with the result phrase for each causal phrase pair and generates linking information for the causal phrase pairs, and a causal linking unit that uses the linking information and the linking DB that stores the linking information to link causal phrase pairs that have cause phrases that are causally consistent with the result phrase of the causal phrase pair, thereby linking the causal relationships.

[0005] Patent document 2 also discloses an information processing device that has a prediction unit and an output unit, in which the prediction unit predicts events that may occur in the future, and the output unit outputs information based on the assumption that a predicted event, which is an event predicted by the prediction unit, has occurred. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Japanese Patent Application Laid-Open No. 2015-121897 [Patent Document 2] Patent Publication No. 2021-33320 Summary of the Invention [Problem to be solved by the invention]

[0007] The scenario generation device and computer program disclosed in Patent Document 1 extract phrases that are causally related based on the results of morphological analysis and dependency analysis. However, there is a problem in that when a phrase appears multiple times in a document, it is not possible to uniquely identify them.

[0008] In other words, no consideration is given to predicting the possibility of multiple events occurring in time series when one event occurs.

[0009] In addition, the information processing device, information processing method, and information processing program disclosed in Patent Document 2 generate a predictive model from historical information such as player history information, team history information, progress logs, venue information, and weather information, but do not generate a predictive model from a document.

[0010] After identifying the target of prediction, it predicts new events that will occur when a possible future event occurs. [Means for solving the problem]

[0011] The above problem is solved by a future event prediction system that predicts future events that may occur in the future based on documents in which past events that occurred in the past and the dates on which the past events occurred are written in natural language, the future event prediction system comprising: a past event tree generation unit that reads past events from the documents and generates a past event tree for documents whose vector distance is equal to or less than a predetermined threshold; a language model generation unit that generates a future event prediction model from the past event tree; a prediction unit that receives event designation and predicts future events from the future event prediction model based on the received events; and an output unit that outputs the future events determined by the prediction unit. [Effects of the Invention]

[0012] The present invention makes it possible to predict events that may occur in the future based on events described in documents. [Brief explanation of the drawings]

[0013] [Figure 1] 1 is a block diagram illustrating an example of a configuration of a future event prediction system according to an embodiment of the present invention. [Figure 2] 1 is an example of a flowchart illustrating processing of a future event prediction system according to an embodiment of the present invention. [Figure 3] 10 is an example of a flowchart illustrating a past event tree generation process according to an embodiment of the present invention. [Figure 4] 10 is an example of a flowchart illustrating a related event search process according to an embodiment of the present invention. [Figure 5] 10 is an example of antecedent events to be searched for in an embodiment of the present invention. [Figure 6] 10 is an example of a retrieved document chunk in an embodiment of the present invention. [Figure 7] 10 is an example of a flowchart illustrating a subsequent event extraction process in an embodiment of the present invention. [Figure 8] 10 is an example of antecedent events to be searched for in an embodiment of the present invention. [Figure 9] 10 is an example of extracted subsequent events in an embodiment of the present invention. [Figure 10]10 is an example of a flowchart illustrating a language model generation process according to an embodiment of the present invention. [Figure 11] 1 is a diagram illustrating an example of an event and an input event according to an embodiment of the present invention. [Figure 12] 10 is an example of a flowchart illustrating a future event prediction process according to an embodiment of the present invention. [Figure 13] 10 is an example of an event tree output screen in an embodiment of the present invention. [Figure 14] 10 is an example of a date-specified prediction result output screen in the embodiment of the present invention. [Figure 15] 10 is an example of an event specification prediction result output screen in the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, embodiments of the present invention will be described with reference to the drawings. In each drawing for explaining the embodiments, the same components are given the same names and reference numerals as much as possible, and repeated description thereof will be omitted.

[0015] The present invention is not limited to the following examples, and includes various modifications and equivalent configurations within the spirit and scope of the appended claims. For example, the above examples have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those including all of the described configurations.

[0016] Furthermore, the processing units and processing modules described in the embodiments may be realized in part or in whole in hardware, for example by designing them as integrated circuits, or may be realized in software by having a processor interpret and execute a program that realizes each function.

[0017] The information explained in the embodiment may be a table, a database (DB), or data stored in the main memory.

[0018] FIG. 1 is a block diagram showing an example of the configuration of a future event prediction system according to an embodiment of the present invention.

[0019] In this embodiment, an example will be described in which a future event prediction system 100 is implemented using a stand-alone computer. The future event prediction system 100 is implemented by a computer including a CPU (Central Processing Unit) 101, a main memory device 102, an external memory device 103, and an input / output unit 104.

[0020] The CPU 101 may use a GPU (Graphics Processing Unit) for high-speed operation.

[0021] The main storage device 102 is composed of a ROM (Read Only Memory), a RAM (Random Access Memory), etc., and the external storage device is composed of devices such as an HDD (Hard Disk Drive), an SSD (Solid State Drive), etc.

[0022] The input / output unit is a NIC (Network Interface Card) or the like that performs input / output of an input / output device or communication with other devices via a network.

[0023] The main memory device 102 contains a control unit 110 that controls the entire system, and a future event prediction model construction unit 140 that creates a future event prediction model 117 from documents. The future event prediction model construction unit 140 stores, as software modules, a root node extraction unit 111 that extracts root nodes, a past event tree generation unit 113 that generates a tree from documents in which past events are described, and a language model generation unit 114 that adjusts the future event prediction model 117.

[0024] The past event tree generation unit 113 includes a related event search unit 118 that searches for related events from the document corpus 115, and a subsequent event extraction unit 119 that extracts subsequent events.

[0025] The external storage device 103 stores a document corpus 115, a learning source model 116 for a future event prediction model, and a future event prediction model 117. The learning source model 116 for a future event prediction model may be a stored learning model that has been created once. When creating a learning model again, the stored learning source model 116 for a future event prediction model can be used to learn only the differences, thereby creating a learning model in a short time.

[0026] Each software module in the main storage device references data stored in the external storage device 103 and is executed by the CPU 101 .

[0027] The future event prediction model 117 is created by inputting multiple pairs of dates and events from an input / output terminal 130 such as a personal computer, and then outputs a tree-structured representation of future events generated in a chain reaction based on the input date and event pairs.

[0028] FIG. 2 is an example of a flowchart showing the processing of a future event prediction system according to an embodiment of the present invention.

[0029] The future event prediction model 117 is a generalized model that uses a language model to learn event sequences, which are multiple pairs of dates and events, and can predict events that may occur in the future and their dates. To do this, it is necessary to acquire a variety of event sequences that have occurred in the past.

[0030] First, the document corpus is read (step 200), and the root node is extracted (step 201).

[0031] A past event tree is generated from the document corpus (step 202), a language model is generated (step 203), and finally a future event prediction model is output (step 204).

[0032] FIG. 3 is an example of a flowchart showing a past event tree generation process in an embodiment of the present invention.

[0033] The depth d of the past event tree to be created is accepted (step 300), and the event to be the root node is accepted (step 301).

[0034] The event that will become the accepted root node is set to the root of the tree together with the date (step 302).The tree of the future event prediction model 117 is searched in depth-first search order, and the selected node U is selected (step 303).

[0035] It is determined whether the depth of node U in the tree is d (step 304), and if it is d, the past event tree is output (step 308).

[0036] If the depth does not reach d, the related event search unit 118 performs related event search processing (step 405), and the subsequent event extraction unit 119 performs subsequent event extraction processing (step 306).

[0037] For the multiple subsequent events obtained, multiple nodes having information on the dates of the subsequent events, the subsequent events, and related documents are added to the past event tree as child nodes of node U.

[0038] Returning to step 303, the child node added to node u is set as node u and the tree is searched in depth-first search order (step 307).

[0039] FIG. 4 is an example of a flowchart showing a related event search process in an embodiment of the present invention.

[0040] The related event search unit 118 receives the antecedent event and the date of the antecedent event, and reads the document corpus 115 (step 400). The document corpus is divided into chunks, which are short documents containing the date and the event (step 401).

[0041] Examples of document corpora include Wikipedia (registered trademark) and corpora that compile news articles.

[0042] Documents vary in length, which reduces search accuracy, so split the document into as small chunks as possible.

[0043] As shown in Figure 6, the divided chunks include the document chunk ID, the date included in the document chunk, the document chunk vector, and named entities such as place names and disaster types included in the document chunk.

[0044] Each document in the document corpus is divided into smaller chunks, and for each chunk, the results of these three judgments are checked against (predecessor event, date of predecessor event). Then, documents that contain document chunks that also satisfy the relevance judgment criteria are retrieved from the document corpus. If there are many documents to be retrieved, an upper limit can be set on the number of documents to be retrieved.

[0045] The first chunk is set as the target for determination (step 402).

[0046] As shown in Figure 5, if the antecedent event to be searched for has a date of December 10, 2010, a vector of the antecedent event of (0.1, -0.4, ..., 0.6), and the named entity contained in the antecedent event is "City A," it is determined whether the vector distance, such as the Euclidean distance, between the antecedent event and the set chunk is below a predetermined threshold (step 403).

[0047] In this case, the determination may be made based on the cosine similarity between the vectors. In this case, based on a set threshold, if the distance between the events in a document chunk is smaller than the threshold or the cosine similarity is larger than the threshold, the distance between the events in the document chunk is deemed to be less than the threshold.

[0048] If it is not below the threshold, it is determined whether all chunks have been searched (step 407), and if there is an unsearched chunk, the next chunk is set as the object of determination (step 408), and the process returns to step 403.

[0049] If it is equal to or less than the threshold, it is determined whether the chunk contains a named entity of the antecedent event (step 404), and if not, the process from step 403 is carried out for the next chunk.

[0050] Named entities include organization names, personal names, and place names, and named entity extraction refers to the mechanical extraction of named entities from natural language text, including proper nouns such as place names, personal names, and organization names, as well as dates, times, quantities, etc.

[0051] Named entities, excluding the date and time, are extracted from the antecedent events, and a document chunk that contains all of these named entities is determined to contain the named entity.

[0052] If a named entity is included, it is determined whether the date of the preceding event is included in the chunk (step 405). In this case, the three chunks in Figure 6 whose vector distance is below the threshold and which contain the named entity are added to the related event table (step 406).

[0053] If there is a chunk that has not been searched, the process from step 403 is repeated for the next chunk, and if there is no chunk that has not been searched, the process ends.

[0054] If there are many chunks that satisfy the conditions, an upper limit may be set on the chunks to be acquired.

[0055] FIG. 7 is an example of a flowchart showing a subsequent event extraction process in an embodiment of the present invention.

[0056] The subsequent event extraction unit 119 receives the antecedent event and the date of the antecedent event, and reads the related event table (step 701).The first chunk of the related event table is set as the search target (step 702).

[0057] The related event table is searched for a subsequent event of the chunk to be searched for that is a preceding event of the search target as shown in FIG. 8 (step 703), and it is determined whether or not there is a subsequent event (step 704).

[0058] If there is a subsequent event as shown in FIG. 9, the subsequent event information is added to the related event table (step 705); if not, the next chunk is set as the search target (step 706) (step 707), and the process returns to step 703.

[0059] If the search has been completed for all chunks, the related event table is output (step 707) and the process ends.

[0060] FIG. 10 is an example of a flowchart showing a language model generation process in an embodiment of the present invention.

[0061] The related event table is read (step 710), and the related event table is sorted so that the dates of the root events are arranged from the past to the present (step 711).

[0062] Search information is added so that the related event table can be searched by both date and event (step 712).

[0063] For example, as shown in the example 716 in Figure 11, by making approximately 50% of the node sequence the event date and the remaining approximately 50% the text sequence in the order of date and event, it becomes possible to search by either date or event.

[0064] The ratio of event dates to date events may be within a predetermined ratio range, and is preferably close to 50%.

[0065] 11 shows an example of events and input events in an embodiment of the present invention. In the event 715, the event date and the date event are listed alternately, so whether searching by date or by event, bias in the search results can be prevented.

[0066] Next, a constraint is imposed on the text to be learned so that sibling nodes are not identical, and a text to be learned is created (step 713).

[0067] Finally, the adjusted related event table is trained to generate a future event prediction model (step 714).

[0068] The constraint given at this time is for generating an event that is different from the sibling node, such as "Please make this an event that is different from the next event," as shown in input event 716 in FIG.

[0069] When there are multiple sibling nodes of a leaf, the event part is left blank for the first event sequence obtained. Next, for the second leaf event sequence obtained, the event of the first leaf is added as a constraint.

[0070] Similarly, in the third and subsequent cases, by adding previously acquired leaf events as constraints, a model is generated that includes a variety of events from past series.

[0071] FIG. 12 is an example of a flowchart showing a future event prediction process in an embodiment of the present invention.

[0072] A future event prediction model is read (step 720), and a current event and a prediction method are accepted (step 702).

[0073] It is determined whether a date-specified prediction is specified (step 722), and if a date is specified, events that will occur on the specified date are predicted and output (step 726).

[0074] If a date-specified prediction is not specified, it is determined whether an event-specified prediction is specified (step 723).If an event-specified prediction is specified, the date on which the specified event will occur is predicted and output (step 727).

[0075] If no prediction of the event specification is specified, the depth and width of the event tree are accepted (step 724), and a future event tree is output (step 725).

[0076] FIG. 13 is an example of an event tree output screen in an embodiment of the present invention.

[0077] The prediction unit 120 receives an input event 751 from the user, which is composed of input events and dates in the order of date on a screen 750. The input event may be a single event.

[0078] The input events are events within a predetermined period from the date used by the user. For example, the period can be events within the past month or the past year, but the period is determined by the field of events to be predicted and the model to be generated.

[0079] The future event tree 754 is a tree generated by the future event prediction model 117 based on the input events that have been input, and outputs events that may occur in the future and dates that are time-series information.

[0080] The depth and number of branches of the output tree are the depth and number of branches specified by the output tree depth 752 and the output tree branch number 753 .

[0081] It is assumed that the use case is to set the depth of the output tree to 1 when a date is specified, and to set the depth and branching of the output tree to 1 when an event is specified, but the depth and branching are arbitrary. If neither a date nor an event is specified, both are output as the predicted future event tree 203.

[0082] In this example, it has already been predicted that "December 13, 2023: Large-scale power outage in X city" will occur, and in the second branch, it is specified in the input event 202 column to predict events other than "December 13, 2023: Large-scale power outage in X city", and an instruction is given to output the "December 13, 2023" event to the part of the future event tree 203 indicated by the dashed line.

[0083] FIG. 14 shows an example of a date-specified prediction result output screen in the embodiment of the present invention.

[0084] Specify "December 13, 2023 (future date):" in the date-specified prediction condition field 760. By specifying the date on which you want to predict what events will occur, the events that will occur on the specified date are output to the date-specified prediction result field 761.

[0085] In this example, a single date is specified, but a period such as "December 13th - December 20th" can also be specified.

[0086] FIG. 15 shows an example of an event specification prediction result output screen in the embodiment of the present invention.

[0087] A date-specified prediction condition field 762 is used to specify a condition such as "Landslide disaster in the □□ region (future event)." The result of predicting the date on which the event will occur is output to a specified event prediction result field 763. The future event prediction model has the structure of a language model. The language model generates sentences by performing next word prediction autoregressively. The future event prediction model applies a generation method using a language model to a series of consecutive events and dates. [Explanation of symbols]

[0088] 100: Future Event Prediction System 101:CPU 102: Main memory 103: External storage device 104: Input / output section 110: Control unit 111: Root node extraction unit 113: Past event tree generation unit 114: Language model generation unit 115: Document Corpus 116: Model to be learned from for future event prediction model 117: Future Event Prediction Model 118: Related Event Search Unit 119: Subsequent event extraction unit 120: Prediction Department 121: Related Event Table 130: Input / output terminal 750: Output screen 751: Input event 754: Future Event Tree 752: Depth of output tree 753: Output tree division coefficient

Claims

1. A future event prediction system that predicts future events that may occur in the future based on documents that describe past events that occurred in the past and the dates on which the past events occurred in a natural language, a past event tree generation unit that reads past events from documents and generates a past event tree for documents whose vector distance is equal to or less than a predetermined threshold; a language model generation unit that generates a future event prediction model from the past event tree; a prediction unit that receives a designation of an event and predicts a future event from the future event prediction model based on the received event; A future event prediction system comprising an output unit that outputs the future event determined by the prediction unit.

2. 2. The future event prediction system according to claim 1, The language model generation unit rearranges the events included in the past event tree in the order of the date on which the events occurred, modifies sibling nodes so that they are different events, and adjusts the ratio of nodes listed in date-event order to nodes listed in event date order so that it is within a predetermined range.

3. 3. The future event prediction system according to claim 2, the prediction unit receives an input of an event sequence; generating future events including time series information that may occur in the future using the received event sequence and the adjusted future event prediction model; A future event prediction system in which the output unit outputs future events including the generated time-series information.

4. 3. The future event prediction system according to claim 2, A future event prediction system in which the document is a chunk containing a date and a named entity divided from a document written in natural language.

5. 3. The future event prediction system according to claim 2, The prediction unit receives a specified date and predicts future events within the range of the received date from a future event prediction model.

6. 3. The future event prediction system according to claim 2, The prediction unit receives a designation of an event and predicts the date on which the received event will occur from a future event prediction model.

7. 3. The future event prediction system according to claim 2, The prediction unit receives a specification of the depth of the future event tree to be output, and outputs a future event tree that predicts the occurrence of an event from a future event prediction model at the received depth.

8. 3. The future event prediction system according to claim 2, The prediction unit receives a designation of the number of branches of the future event tree to be output, and outputs a future event tree that predicts the occurrence of an event from a future event prediction model with the designated number of branches.

9. A future event prediction method for predicting future events that may occur in the future based on documents describing past events that occurred in the past and the dates on which the past events occurred in a natural language, comprising: A past event tree generation unit reads past events from documents and generates a past event tree for documents whose vector distance is equal to or less than a predetermined threshold value; A language model generation unit generates a future event prediction model from the past event tree, a prediction unit that receives a designation of an event and predicts a future event from the future event prediction model based on the received event; A future event prediction method in which an output unit outputs the future event determined by the prediction unit.

Images