JP7447674B2 - Information processing program, information processing method, and information processing device - Google Patents

Description

The present invention relates to an information processing program, an information processing method, and an information processing apparatus.

For manufacturers who ship various products to the market, market quality management after product shipment has become an important management issue. For example, a malfunction that occurs in a field where a product is actually used is reported as a maintenance report (MR). Such a failure report is created, for example, by a field SE that performs field support work, but depending on the target product, it may also be created during call center work that provides support by telephone or email.

Therefore, market quality management uses failure reports of field-operated products on the market to identify occurrences, and determines causes and countermeasures based on past cases. Then, in accordance with the determined countermeasure method, a response to the failure for which the failure report has been submitted is taken.

In addition, in recent years, efforts have been made to improve the quality of field-operated products using the following cycle. Failure reports are reported from various places, and analysis is performed on a large number of reported failure reports. This analysis results in, for example, detection of defects that have increased recently. Such failures that have increased recently are called trend failures and are distinguished from random failures. Trend disorders are likely to continue to increase in the future, so there is a high degree of urgency in formulating countermeasures. For example, if an OS (Operating System) update causes a problem with the firmware used in a field-operated product, the problem will occur in all products using the same firmware in the future. Therefore, it is necessary to take measures as soon as possible.

When such a trend failure is detected, the product manufacturer prevents the occurrence of the failure by documenting the details of the detected failure and countermeasures, and disseminating the information to the market. For example, a product manufacturer detects a recent increase in "malfunctions caused by specific firmware bugs" and informs the field of the nature of the trending faults along with countermeasures for the trending faults. This allows the manufacturer to prevent the occurrence of failures by having users take preventive measures for products equipped with the relevant firmware.

Conventionally, this trend failure detection has been performed manually. On the other hand, due to the rapid increase in the number of failure reports, it has become difficult to properly detect trend failures manually. Therefore, detection of trend failures using information processing devices, such as code information-based detection methods and keyword-based detection methods, has been introduced. The code information-based detection method is a method for detecting trend failures using code information that indicates the nature of the problem when it is attached to an external report. In addition, the keyword-based detection method is a method for detecting trend failures by detecting an increasing number of keywords by regarding the dependency pairs of subjects and predicates in words and sentences written in failure reports as code information. It is also conceivable to use a document-based detection method that handles words in larger units than keyword-based methods.

Note that as a technique for analyzing and searching documents, there is a conventional technique that changes the granularity of clusters to be presented based on geographical location. Furthermore, there is a conventional technique that calculates the degree of similarity between a representative vector at each branch point in the hierarchical clustering result and a vector generated from a search target object, and presents clusters with high calculated degrees of similarity.

Japanese Patent Application Publication No. 2011-118784 Japanese Patent Application Publication No. 2003-316819

However, in the case of a detection method based on code information, the granularity of the defect content represented by the code information is coarse. As a result, disorders that are on the rise are often overlooked. For example, if a fault with an increasing trend and a fault with a decreasing trend are classified into the same code, the fault with an increasing trend may be overlooked.

Furthermore, when performing a keyword-based detection method, it is preferable to perform detection by taking into consideration spelling variations of words, synonyms, and synonymous expressions. For example, the words "manufacturer logo screen," "manufacturer screen," and "BIOS (Basic Input Output System) screen" may all refer to the same screen, in which case these words are synonyms. Furthermore, "to stop" and "to stop" are synonymous, and the negative expression "do not proceed" is also a synonym. However, as the amount of information increases, the cost of creating a dictionary to identify variations in the spelling of words, synonyms, and synonymous expressions increases due to the complexity of creating the dictionary. Therefore, in the current situation where the number of failure reports is rapidly increasing, it is difficult to perform keyword-based detection. Therefore, it is difficult to detect trend failures using these detection methods, and response to trend failures is delayed, making it difficult to improve product quality.

On the other hand, by using a document-based detection method, it is possible to reduce the cost of processing word variations and synonyms. When grouping documents like this, a method using clusters can be considered. When clustering is performed, processing parameters such as the number of clusters are given in advance. However, if the parameters are inappropriate, appropriate clusters may not be generated, and there is a risk of overlooking the increasing trend of each cluster.

For example, if a cluster is too small, the content similarity between elements included in each cluster will be high, but if the number of occurrences of failure reports is aggregated over time, the number will be small, making it difficult to detect a significant trend. For example, if the monthly aggregate results for the past six months are (0.1.0, 2, 2, 3), no significant increasing trend is recognized in the Mann-Kendall trend test with a risk rate of 5%. Here, the numbers in parentheses in the monthly aggregation results represent the number of occurrences in each month from the left to the oldest in the past six months.

In addition, if a cluster is too large and is made up of multiple faults, the increasing trend of the fault that is desired to be detected may be diluted by other faults, and the test may not be significant. For example, a significant increasing trend is observed for failures whose monthly aggregate results for the past six months are (0, 3, 6, 10, 15, 20). On the other hand, for the defects whose monthly aggregate results are (52, 48, 44, 48, 46, 47), neither an increasing trend nor a decreasing trend is observed. The monthly aggregation result for a cluster that includes these two failures is the sum of both, but in that case, no significant increasing trend will be recognized in this cluster.

If clustering is simply performed on a document basis, it is difficult to appropriately determine parameters to be used when performing clustering, and there is a risk of overlooking an increasing trend.

In addition, with conventional technology that changes the granularity of clusters presented based on geographic location, it is difficult to detect cases that are showing an appropriate increasing trend depending on location, and it is difficult to detect cases that are showing an appropriate increasing trend depending on location. Failures are difficult to detect. Further, in the conventional technology that presents clusters based on the similarity of vectors between a target and a result, it is difficult to present clusters without inputting a search query, and it is difficult to detect trend failures. Therefore, it is difficult to improve the quality of products using any of the conventional techniques.

The disclosed technology has been developed in view of the above, and aims to provide an information processing program, an information processing method, and an information processing device that improve the quality of products.

In one aspect of the information processing program, information processing method, and information processing apparatus disclosed in the present application, a computer is caused to execute the following processing. Get multiple document information. Each content of the document information is digitized to calculate digitized information. Hierarchical clustering is performed based on the numerical information to generate dendrogram information indicating a dendrogram. A feature of the document information belonging to a cluster corresponding to a plurality of branch points of the tree diagram indicated by the tree diagram information is identified. A specific cluster included in the cluster corresponding to a plurality of branch points included in the tree diagram indicated by the tree diagram information, the feature having significance, and a hierarchy including the corresponding branch point. The specific cluster having the feature stronger than the cluster corresponding to the upper branching point and the lower branching point in the structure is extracted. Information about the extracted specific cluster is output.

In one aspect, the present invention can improve product quality.

FIG. 1 is a diagram showing a trend failure detection system. FIG. 2 is a diagram illustrating an example of a failure report. FIG. 3 is a block diagram of the server device. FIG. 4 is a block diagram showing details of the hierarchical clustering unit and the trend failure detection unit. FIG. 5 is a diagram showing an example of a dendrogram. FIG. 6 is a diagram showing an example of increasing tendency evaluation values for each branch point. FIG. 7 is a diagram for explaining cluster extraction. FIG. 8 is a flowchart of the trend failure calculation process according to the first embodiment. FIG. 9 is a diagram showing part of a hierarchical cluster in the process of detecting failures that are uniquely common to specific customers. FIG. 10 is a hardware configuration diagram of the server device.

Embodiments of an information processing program, an information processing method, and an information processing apparatus disclosed in the present application will be described in detail below based on the drawings. Note that the information processing program, information processing method, and information processing apparatus disclosed in the present application are not limited to the following embodiments.

FIG. 1 is a diagram showing a trend failure detection system. As shown in FIG. 1, the trend failure detection system 10 includes a server device 1, a failure report input terminal 2, a failure report database 3, a word vector database 4, and a client terminal 5 for outputting search results.

The server device 1 and the failure report input terminal 2 are connected via a network 6. Furthermore, a failure report database 3 and a word vector database 4 are connected to the server device 1 . Further, the search result output client terminal 5 may be connected to the server device 1 or to the network 6.

The failure report input terminal 2 is a device used by a worker who has responded to a failure to input information about the failure when a failure occurs in a product that is used in a company and is in operation. The failure report input terminal 2 acquires failure information input by the worker. Then, the failure report input terminal 2 transmits the acquired failure information to the server device 1 via the network 6 as a failure report. The failure information includes, for example, the publication date when the failure information was registered, the name of the customer where the failure occurred, the type and model name of the product where the failure occurred, the phenomenon of the failure that occurred, and the cause of the failure. , the response to the failure and the date on which the response was completed.

The failure report database 3 receives failure report information collected from each failure report input terminal 2 by the server device 1 . Then, the failure report database 3 compiles the input failure reports and creates a failure report 100 as shown in FIG. 2. FIG. 2 is a diagram illustrating an example of a failure report. The failure report 100 includes a plurality of individual failure reports 101. The failure report 101 is document information, and is each failure report transmitted from the failure report input terminal 2.

Each failure report 101 has a field for writing standard information such as issue date, and a free description field for phenomena, causes, countermeasures, etc. Here, although the issue date is described in this embodiment, the issue date and time including the time may also be used. That is, the failure report 101 is document information associated with time, and time includes date and time. For example, as shown in FIG. 2, the failure report 101 includes standard information such as issue date, customer name, type, model name, and response completion date, as well as free description information such as the phenomenon, cause, and response method. Furthermore, the information in the failure report 100 is not limited to the information listed here, but may also include other information such as part model names of replacement parts, and clusters can be generated from various viewpoints in addition to trend failures. .

The failure report database 3 holds failure reports 100. The failure report 100 held by the failure report database 3 is updated by registering a new failure report 101 when input from the failure report input terminal 2 or periodically.

The word vector database 4 holds word vectors that represent words as vectors called distributed representations. In word vectors expressed using distributed representation, words that have similar usage examples are expressed as similar vectors even if no correct answer is given. In other words, words with similar usages are treated the same way, so words with different spellings and synonyms are treated the same way.

The server device 1 collects failure reports 101 sent from each failure report input terminal 2 via the network 6. Then, the server device 1 stores the failure report 101 in the failure report database 3.

Then, the server device 1 vectorizes each of the failure reports 101 included in the failure report 101 using the word vectors stored in the word vector database 4, and performs clustering using the similarity to generate a plurality of document clusters. Next, the server device 1 totals the number of occurrences of failures in each document cluster and detects trend failures. Thereafter, the server device 1 transmits the detected trend failure to the search result output client terminal 5.

The search result output client terminal 5 is a device that displays the detection results of trend failures and provides them to the user. The search result output client terminal 5 receives information on detected trend failures from the server device 1 . Then, the search result output client terminal 5 notifies the user of the trend failure information by outputting the acquired trend failure information and displaying it on a monitor or the like.

Next, details of the server device 1 will be explained with reference to FIG. 3. FIG. 3 is a block diagram of the server device. As shown in FIG. 3, the server device 1 includes a failure report information acquisition section 11, a hierarchical clustering section 12, a trend failure detection section 13, and an output section 14.

The failure report information acquisition unit 11 collects failure reports sent from the failure report input terminal 2. Then, the failure report information acquisition unit 11 transmits the collected failure report 101 to the failure report database 3 and updates the failure report 100.

The hierarchical clustering unit 12 creates word vectors of words included in the failure report 101, vectorizes the failure report 101 using the created word vectors, and clusters the vectorized failure report 101. FIG. 4 is a block diagram showing details of the hierarchical clustering unit and the trend failure detection unit. As shown in FIG. 4, the hierarchical clustering unit 12 includes a sentence analysis unit 121, a word vector creation unit 122, a document vector creation unit 123, and a dendrogram creation unit 124.

The sentence analysis unit 121 acquires all the failure reports 101 included in the failure reports 100 stored in the failure report database 3. The sentence analysis unit 121 then analyzes the obtained failure report 101 and extracts words. Thereafter, the sentence analysis unit 121 outputs the extracted words to the word vector creation unit 122.

Furthermore, the sentence analysis unit 121 acquires a plurality of failure reports 101 from the failure report 100 to be used for detecting trend failures. For example, when determining trends over the past six months, the sentence analysis unit 121 obtains all failure reports 101 in which failures that have occurred within the past six months are registered. The sentence analysis unit 121 then outputs the obtained failure report 101 to the document vector creation unit 123.

The word vector creation unit 122 digitizes words included in the failure report 100. Specifically, the word vector creation unit 122 receives input of words extracted from the failure report 100 from the sentence analysis unit 121. Then, the word vector generation unit 122 generates a word vector by representing the acquired word using the distributed representation. For example, the word vector generation unit 122 generates a word vector using Word2Vec. The word vector creation unit 122 uses a neural network to learn how words are used in free description sentences, and converts each word into an n-dimensional vector. The number of dimensions, n, is a value specified by the operator. Thereafter, the word vector creation unit 122 stores the generated word vectors in the word vector database 4.

The document vector creation unit 123 digitizes the failure report 101. Specifically, the document vector creation unit 123 receives input of the failure report 101 used for detecting trend failures from the sentence analysis unit 121. Then, the document vector creation unit 123 uses the word vectors stored in the word vector database 4 to vectorize documents such as phenomena, causes, and countermeasures in each failure report 101, and generates document vectors for each failure report 101. do. Thereafter, the document vector creation unit 123 outputs each failure report 101 represented by the document vector to the dendrogram creation unit 124. This vectorization is an example of "digitization," and the document vector is an example of "digitization information." This document vector creation unit 123 is an example of a "digitization unit".

Here, an example of the document vector creation process by the document vector creation unit 123 will be described in detail. Input text information is described in the failure report 101 divided into fields such as phenomenon, cause, and countermeasure. The document vector creation unit 123 calculates a centroid vector from the word vectors of words included in the free description text for each field. Thereby, the document vector creation unit 123 expresses the free description text for each field as an n-dimensional vector.

Alternatively, the document vector creation unit 123 can also employ a weighted centroid vector that takes into consideration the weight of each word included in the free description text. For example, the document vector creation unit 123 may perform weighting by giving more importance to words whose appearance probabilities in the document are biased. In this case, the document vector creation unit 123 can calculate the weight of w=frequency of appearance of w in documents×log (total number of documents/number of documents including w), where w is a word. Further, the document vector creation unit 123 may normalize the calculated gravity center vector. For example, the document vector creation unit 123 can perform normalization by setting the length of each centroid vector to 1.

After calculating the centroid vector of each field, the document vector creation unit 123 calculates the document vector using, for example, one of the following two methods. As one vectorization method, the document vector creation unit 123 uses the center of gravity of the n-dimensional vector obtained from the field as the document vector. In this case, one document is expressed as an n-dimensional vector. Here, the document vector creation unit 123 may normalize and use the obtained document vector.

Further, as another vectorization method, the document vector creation unit 123 may employ a vector obtained by combining n-dimensional vectors of each field as the document vector. For example, when using three fields: phenomenon, cause, and countermeasure, the failure report 101 is expressed as a three-dimensional vector. In this case as well, the document vector creation unit 123 may normalize and use the obtained document vector.

The dendrogram creation unit 124 receives input of each failure report 101 expressed using a document vector from the document vector creation unit 123. Next, the dendrogram creation unit 124 uses the similarity of the document vectors to perform hierarchical cluster analysis in which failure reports 101 with similar degrees of similarity are grouped in order. For example, the dendrogram creation unit 124 generates a large cluster by treating each failure report 101 as one cluster and combining the clusters using the distance between the clusters represented by the document vector. Use.

The dendrogram creation unit 124 creates a dendrogram 200 of the failure report 101, as shown in FIG. FIG. 5 is a diagram showing an example of a dendrogram. The dendrogram 200 is also called a "dendrogram," and information indicating the dendrogram 200 is an example of "dendrogram information."

Each cluster 201 at the lowest level in FIG. 5 corresponds to one failure report 101. One branch point 202 in the dendrogram 200 corresponds to one cluster obtained by integrating two clusters. The dendrogram creation unit 124 outputs the created dendrogram 200 to the trend failure detection unit 13.

Here, since clustering is actually performed using a very large number of failure reports 101, the number of branch points 202 in the dendrogram 200 becomes very large. Further, since hierarchical clustering requires a large calculation cost, it is preferable to reduce the number of calculation processes. Therefore, for example, the dendrogram creation unit 124 may use a method such as the k-means method that has coarse granularity and low calculation cost to broadly divide the failure report 101 and then perform hierarchical clustering. This makes it possible to shorten the clustering processing time and to keep the calculation cost low.

Returning to FIG. 4, the explanation will be continued. The trend failure detection unit 13 includes a branch point evaluation unit 131 and an extraction unit 132.

The branch point evaluation unit 131 receives input of the dendrogram 200 from the dendrogram creation unit 124. Next, the branch point evaluation unit 131 extracts each branch point 202 of the dendrogram 200 shown in FIG. Then, the branch point evaluation unit 131 calculates an increasing tendency evaluation value for each branch point 202.

For example, the branch point evaluation unit 131 calculates an increasing tendency evaluation value for each branch point 202 as shown in FIG. FIG. 6 is a diagram showing an example of increasing tendency evaluation values for each branch point. Clusters 211 to 213 in FIG. 6 correspond to the clusters at the branch point 202 in the dendrogram 200, respectively.

For example, the branch point evaluation unit 131 obtains (0, 1, 0, 2, 2, 3) as the number of occurrences of failures belonging to the cluster 212 in the past six months. Here, the numbers in parentheses represent the number of occurrences in each month in 6-month chronological order starting from the left on the page. Next, the branching point evaluation unit 131 performs a Mann-Kendall trend test on the number of occurrences in the past six months and calculates a tau statistic that is an index of monotony. Here, the branch point evaluation unit 131 calculates tau=0.788 in the cluster 212. Next, the branching point evaluation unit 131 uses tau to calculate the P value, which is an increasing trend evaluation value used in the statistical test. The P value is the probability that a bias greater than the observed result will occur when the null hypothesis that there is no bias holds true. In the case of the Mann-Kendall trend test, the tau statistic follows a standard normal distribution under the null hypothesis, so the branch point evaluation unit 131 can calculate the P value from the tau statistic. For example, the branch point evaluation unit 131 calculates the P value of the cluster 212 as 0.788.

Similarly, the branch point evaluation unit 131 obtains (1, 0, 1, 1, 2, 2) as the number of occurrences of failures belonging to the cluster 213 in the past six months. Then, the branch point evaluation unit 131 calculates the tau statistic of the cluster 213 as 0.701, and calculates the P value as 0.100.

Next, the branching point evaluation unit 131 adds up the number of occurrences in the past six months in the clusters 212 and 213, and calculates the number of occurrences in the past six months of failures belonging to the cluster 211 as (1, 1, 1, 3, 4, 5). Then, the branch point evaluation unit 131 calculates the tau statistic of the cluster 211 as 0.894, and calculates the P value as 0.027.

The branch point evaluation unit 131 thus calculates the P value for the clusters corresponding to all the branch points 202 of the dendrogram 200. Then, the branch point evaluation unit 131 outputs the P value, which is the increasing tendency evaluation value of each cluster, to the extraction unit 132 together with the identification information of each cluster.

Here, in this embodiment, the branching point evaluation unit 131 calculated increasing trend evaluation values for all clusters corresponding to all the branching points 202 of the dendrogram 200. However, the granularity of the clusters on the lower layer side is often too fine to show any significant increasing trend. Therefore, the branching point evaluation unit 131 selects clusters including a predetermined number or more of failure reports 101, calculates an increasing trend evaluation value limited to the selected clusters, and The extraction unit 132 may extract clusters representing trend failures. Thereby, calculation costs can be reduced.

The extraction unit 132 receives from the branch point evaluation unit 131 the identification information of each cluster corresponding to all the branch points 202 in the dendrogram 200 and the P value that is the increasing tendency evaluation value. Then, the branching point evaluation unit 131 compares a predetermined risk rate with the calculated P value as a statistical test to determine whether a significant increasing tendency is observed in each cluster. For example, when the risk rate is set to 5%, if the P value is <0.05, the extraction unit 132 determines that a significant increasing trend is recognized.

For example, since the P value of cluster 212 in FIG. 5 is 0.788, which is greater than the risk rate of 5%, the extraction unit 132 determines that no significant increasing trend is observed in cluster 212. Similarly, since the P value of cluster 213 is 0.701, which is greater than the risk rate of 5%, the extraction unit 132 determines that no significant increasing trend is observed in cluster 213. On the other hand, since the P value of cluster 211 is 0.027, which is a risk rate of 5% or less, the extraction unit 132 determines that a significant increasing tendency is recognized in cluster 211. In this way, by merging clusters 212 and 213 in which no significant increasing trend was observed due to the small number of included failure reports 101, cluster 211 where a significant increasing trend was observed may be generated. be.

Next, when a specific branch point 202 is selected, the extraction unit 132 extracts a cluster of branch points 202 having the maximum increasing tendency in the hierarchical structure that includes that branch point 202. FIG. 7 is a diagram for explaining cluster extraction. For example, when cluster 223 is selected, clusters 221, 222, and 224 to 229 correspond to branch point 202 in a hierarchical structure that includes branch point 202 corresponding to cluster 223. Among them, clusters 221 and 227 to 229 have no significant increasing tendency, and clusters 222 to 226 have a significant increasing tendency. Note that the cluster 221 in this case can be said to have lost its significance due to the integration of lower-order clusters including the cluster 222. In this case, if the increasing tendency of the cluster 223 is larger than the increasing tendency of the clusters 222 and 224 to 226, that is, if the increasing tendency is strong, the extraction unit 132 extracts the cluster 223. The extraction unit 132 extracts such clusters from all branch points 202 in the dendrogram 200. The cluster extracted by this extraction unit 132 is an example of a "specific cluster."

A specific example of the method for extracting clusters by the extraction unit 132 will be described below. The extraction unit 132 creates a selection list in which cluster identifiers corresponding to each branch point 202 are sorted in descending order of P value, which is an increasing tendency evaluation value.

Next, the extraction unit 132 sequentially extracts entries from the top of the list. Then, it is determined whether a significant increasing trend is observed in the clusters of the extracted entries. If a significant increasing trend is recognized, the extraction unit 132 adds the extracted entry to the extraction list. Then, the extraction unit 132 deletes the entry of the cluster corresponding to the branch point 202 below the branch point 202 corresponding to the cluster of the extracted entry from the selection list. Next, the extraction unit 132 re-sorts the entries in the selection list by closing the spaces between the deleted entries. After that, the extraction unit 132 extracts the entry at the next position.

On the other hand, if a significant increasing tendency is not recognized in the cluster of extracted entries, the extraction unit 132 determines that there is no entry in the selection list below which a significant increasing tendency is recognized. Finish creating. Then, the extraction unit 132 outputs information on clusters corresponding to the entries registered in the created extraction list to the output unit 14. Here, the cluster information may be any information that can identify what kind of failure the cluster represents. For example, the cluster information may be information on failure reports 101 belonging to the extracted cluster, or may be words related to failures that have a high degree of similarity among the failure reports 101 included in the cluster.

Further, the extraction unit 132 may determine a lower limit or an upper limit of the number of clusters to be extracted in accordance with the evaluation criteria. In this example, clusters corresponding to a branch point 202 that meet the condition that the increasing tendency evaluation value of the cluster at a certain branch point 202 is the largest increasing tendency evaluation value in the hierarchical structure that includes that branch point 202 are extracted. The number of extracted clusters may not reach the lower limit. In that case, the extraction unit 132 may create the selection list again excluding the extracted cluster, create the above-mentioned extraction list again, and add it to the previous list. Furthermore, when the upper limit is exceeded, the extraction unit 132 may extract the number of clusters that fall within the upper limit in descending order of increasing tendency evaluation value.

The output unit 14 receives input from the extraction unit 132 of information on clusters, which are groups of trend failures in which an increasing trend in incidence is observed. Then, the output unit 14 outputs the information on each acquired cluster to the search result output client terminal 5 as a cluster that is a group of trend failures in which an increasing tendency of occurrence rate is recognized.

The user uses the search result output client terminal 5 to obtain information on clusters, which are groups of trend failures whose incidence rates are increasing. Then, the user uses the information of each cluster to check what kind of trend failures are occurring. This allows the user to quickly take measures against the trending failures that are occurring.

Next, with reference to FIG. 8, the flow of the trend failure detection process according to this embodiment will be described. FIG. 8 is a flowchart of the trend failure calculation process according to the first embodiment.

The failure report information acquisition unit 11 collects failure reports 101 sent from the failure report input terminal 2. Then, the failure report information acquisition unit 11 transmits the collected failure reports to the failure report database 3, and stores the failure report 100 having each failure report 101 as an entry in the failure report database 3 (step S1).

The sentence analysis unit 121 of the hierarchical clustering unit 12 acquires a failure report 100 having data used for trend failure detection processing from the failure report database 3. The sentence analysis unit 121 then analyzes each failure report 101 included in the failure report 100 and extracts words included in the failure report 100. The word vector generation unit 122 generates a word vector by representing the words extracted by the sentence analysis unit 121 using a distributed representation (step S2). Thereafter, the word vector creation unit 122 stores the generated word vectors in the word vector database 4.

The document vector creation unit 123 acquires the target document, which is the failure report 101 used for detecting trend failures, from the sentence analysis unit 121. Then, the document vector creation unit 123 uses the word vectors registered in the word vector database 4 to create a document vector for each failure report 101, which is a target document for the trend failure detection process (step S3). Thereafter, the document vector creation unit 123 outputs the created document vector to the dendrogram creation unit 124.

The dendrogram creation unit 124 acquires the sentence vector of each failure report 101 included in the target document group from the document vector creation unit 123. Next, the dendrogram creation unit 124 executes hierarchical clustering of the target document group using the document vectors of each failure report 101 to create the dendrogram 200 (step S4). Thereafter, the dendrogram creation unit 124 outputs the created dendrogram 200 to the branch point evaluation unit 131 of the trend failure detection unit 13.

The branch point evaluation unit 131 calculates the increasing tendency evaluation value of the cluster corresponding to each branch point 202 of the dendrogram 200 obtained from the dendrogram creation unit 124 (step S5). Then, the branch point evaluation unit 131 outputs the identification information of each cluster together with the calculated increasing tendency evaluation value to the extraction unit 132.

The extraction unit 132 acquires the identification information and increasing tendency evaluation value of each cluster corresponding to each of the branch points 202 of the dendrogram 200 from the branch point evaluation unit 131. Then, the extraction unit 132 generates a selection list in which clusters are sorted and arranged in descending order of increasing tendency evaluation value (step S6).

Next, the extraction unit 132 sets i to 1, which represents the order from the beginning of the entries to be selected in the selection list (step S7).

Next, the extraction unit 132 extracts the i-th entry in the selection list (step S8).

Next, the extraction unit 132 determines whether the i-th entry has been successfully extracted (step S9). If the extraction of the i-th entry fails (step S9: negative), the extraction unit 132 proceeds to step S13.

On the other hand, if the extraction of the i-th entry is successful (step S9: affirmative), the extraction unit 132 uses the increasing tendency evaluation value of the cluster corresponding to that entry to determine whether there is a significant increasing tendency in that cluster. It is determined whether it exists (step S10). If there is no significant increasing trend in that cluster (step S10: negative), the extraction unit 132 proceeds to step S13.

On the other hand, if there is a significant increasing trend in that cluster (step S10: affirmative), the extraction unit 132 adds the extracted entry to the extraction list (step S11).

Next, the extraction unit 132 selects an entry corresponding to a branch point 202 below the branch point 202 corresponding to the extracted entry from the selection list (step S12). Furthermore, the extraction unit 132 fills in the portion of the selection list where entries have been deleted, and renumbers each entry included in the selection list in order from the beginning. After that, the extraction unit 132 returns to step S8.

On the other hand, if entry extraction fails (step S9: negative) or if no significant increasing trend is observed in the cluster represented by the extracted entry (step S10: negative), the extraction unit 132 performs the following processing. I do. The extraction unit 132 outputs cluster information corresponding to the entry registered in the extraction list to the output unit 14. The output unit 14 outputs the clusters acquired from the extraction unit 132 and information on each cluster to the search result output client terminal 5 as clusters representing trend failures (step S13).

As described above, in the trend failure detection process according to the present embodiment, the server device obtains a document vector for each failure report, performs hierarchical clustering using the document vector, and creates a dendrogram. After that, the server device extracts a cluster that has a dominant increasing tendency among the clusters corresponding to the branching points of the dendrogram and has the highest increasing tendency evaluation value in the hierarchical structure that includes the branching point. and notify you of that information.

As a result, it is possible to automatically create clusters with a content granularity that shows the strongest increasing tendency, and it is possible to detect defects that are increasing with high accuracy. Furthermore, since there is no need to maintain synchronized word dictionaries, human costs can be reduced. Therefore, by detecting trend failures with high accuracy, it is possible to quickly and appropriately deal with the occurrence of defects, and it is possible to improve product quality.

Next, Example 2 will be explained. In Example 1, the detection process for a disorder that is on the rise is explained as an example, but the method described in this example can be applied to other detection processes by changing the evaluation value calculated at the branch point 202 of the dendrogram. It is also possible to apply In that case, the evaluation value is a chi-square statistic calculated from the cross-tabulation results of the failure report 101 within the cluster, the failure report 101 outside the cluster, and the customer name in a specific cluster. In this embodiment, a process for discovering failures that are uniquely common to specific customers will be described. The server device 1 according to this embodiment is also shown in FIGS. 3 and 4. In the following description, descriptions of the operations of the same parts as in the first embodiment may be omitted.

FIG. 9 is a diagram showing part of a hierarchical cluster in the process of detecting failures that are uniquely common to specific customers. In Figure 9, for each cluster, there are failures that occur at customer A, failures that occur at customers other than customer A, failures that occur at customer A that is not included in that cluster, and customers other than A that are not included in that cluster. It shows the failures that occurred for customers and the chi-square statistic for that cluster.

For example, cluster 301 is a cluster that includes a total of 10 failures: 8 failures that occurred with customer A and 2 failures that occurred with other customers. The failures not included in the cluster 301 include 17 failures that occurred with customer A and 99973 failures that occurred with other customers. The chi-square statistic of the cluster 301 is 22,493.

Similarly to the first embodiment, the dendrogram creation unit 124 of the hierarchical clustering unit 12 executes hierarchical clustering using the document vector of the failure report 101 used in the detection process, and partially creates a hierarchy as shown in FIG. A dendrogram 200 having the following information is created.

The branch point evaluation unit 131 of the trend failure detection unit 13 calculates a chi-square statistic that is an evaluation value for each cluster corresponding to each branch point 202 of the dendrogram 200.

The larger the value of the chi-square statistic, the stronger the specificity of customer A is. That is, in FIG. 9, cluster 302 exhibits the strongest peculiarity of customer A. The extraction unit 132 extracts a branching point 202 where significant specificity is recognized in the failure that occurred at customer A, and where the evaluation value expressed by the chi-square statistic of that cluster is maximum in the hierarchical structure that includes that cluster. Extract the cluster corresponding to . As a result, clusters representing failures that are uniquely common to customer A are extracted.

By checking the failures represented by the clusters extracted by the server device 1, the user can discover failures that are uniquely common in a particular customer, and can respond quickly and appropriately to the particular customer. It can be performed.

(Hardware configuration)
FIG. 10 is a hardware configuration diagram of the server device. As shown in FIG. 10, the server device 1 includes a CPU 91, a memory 92, a hard disk 93, and a network interface 94. The CPU 91 is connected to a memory 92, a hard disk 93, and a network interface 94 via a bus.

The network interface 94 is an interface for communication with the network 6 and the search result output client terminal 5. The CPU 91 communicates with the failure report input terminal 2 and the search result output client terminal 5 via the network interface 94 .

The hard disk 93 is an auxiliary storage device. The hard disk 93 stores various programs including programs for realizing the functions of the failure report information acquisition section 11, the hierarchical clustering section 12, the trend failure detection section 13, and the output section 14 illustrated in FIG.

Further, in this embodiment, the failure report database 3 and the word vector database 4 are placed outside the server device 1, but the server device 1 may have a configuration in which they are retained. In that case, the hard disk 93 implements the functions of the failure report database 3 and the word vector database 4.

The CPU 91 reads various programs stored in the hard disk 93, expands them into the memory 92, and executes them. Thereby, the CPU 91 and the memory 92 realize the functions of the failure report information acquisition unit 11, the hierarchical clustering unit 12, the trend failure detection unit 13, and the output unit 14 illustrated in FIG.

1 Server device 2 Failure report input terminal 3 Failure report database 4 Word vector database 5 Client terminal for outputting search results 6 Network 10 Trend failure detection system 11 Failure report information acquisition unit 12 Hierarchical clustering unit 13 Trend failure detection unit 14 Output unit 121 Sentences Analysis section 122 Word vector creation section 123 Document vector creation section 124 Dendrogram creation section 131 Branch point evaluation section 132 Extraction section

Claims (7)

Get multiple document information,
digitizing each content of the document information to calculate quantified information;
Performing hierarchical clustering based on the digitized information to generate dendrogram information indicating a dendrogram;
identifying features of the document information belonging to clusters corresponding to a plurality of branch points of the tree diagram indicated by the tree diagram information;
A specific cluster included in the cluster corresponding to a plurality of branch points included in the tree diagram indicated by the tree diagram information, the feature having significance, and a hierarchy including the corresponding branch point. extracting the specific cluster having the feature stronger than the cluster corresponding to the upper branching point and the lower branching point in the structure;
An information processing program that causes a computer to execute a process of outputting the extracted information on the specific cluster. The document information is information associated with time,
Based on the time associated with the document information, identifying an increase/decrease trend in events indicated by the document information as the feature;
Extracting the specific cluster in which the increase/decrease tendency has significance and the increase/decrease tendency is stronger than the cluster corresponding to the upper branch point and the lower branch point in the hierarchical structure including the corresponding branch point. The information processing program according to claim 1, causing a computer to execute the processing. The document information is information indicating an event associated with an occurrence location,
Identifying the occurrence tendency of the event indicated by the document information as the characteristic at the specific occurrence location,
Extracting the specific cluster in which the occurrence tendency has significance and the occurrence tendency is stronger than the cluster corresponding to the upper branch point and the lower branch point in the hierarchical structure including the corresponding branch point. The information processing program according to claim 1, causing a computer to execute the processing. A claim characterized in that a statistical evaluation value indicating the feature is calculated, and a computer is caused to execute a process of determining the prior significance of the feature and the strength of the feature based on the calculated evaluation value. The information processing program described in any one of Items 1 to 3. The information processing according to any one of claims 1 to 4, characterized in that the computer executes a process of performing natural language processing and obtaining each document vector as numerical information based on the content of the document information. program. Get multiple document information,
digitizing each content of the document information to calculate quantified information;
Performing hierarchical clustering based on the digitized information to generate dendrogram information indicating a dendrogram;
identifying features of the document information belonging to clusters corresponding to a plurality of branch points of the tree diagram indicated by the tree diagram information;
A specific cluster included in the cluster corresponding to a plurality of branch points included in the tree diagram indicated by the tree diagram information, the feature having significance, and a hierarchy including the corresponding branch point. extracting the specific cluster having the feature stronger than the cluster corresponding to the upper branching point and the lower branching point in the structure;
outputting information on the extracted specific cluster;
An information processing method characterized by causing a computer to perform processing . a digitization unit that obtains a plurality of pieces of document information and digitizes the content of each of the document information to calculate digitized information;
a dendrogram creation unit that performs hierarchical clustering based on the numerical information to generate dendrogram information indicating a dendrogram;
a branching point evaluation unit that identifies characteristics of the document information belonging to a cluster corresponding to a plurality of branching points of the tree diagram indicated by the tree diagram information;
A specific cluster included in the cluster corresponding to a plurality of branch points included in the tree diagram indicated by the tree diagram information, the feature having significance, and a hierarchy including the corresponding branch point. an extraction unit that extracts the specific cluster having the characteristic stronger than the cluster corresponding to the upper branching point and the lower branching point in the structure;
An information processing device comprising: an output unit that outputs the extracted information on the specific cluster.

Images