WO2010119615A1 - 学習データ生成装置、及び固有表現抽出システム - Google Patents
学習データ生成装置、及び固有表現抽出システム Download PDFInfo
- Publication number
- WO2010119615A1 WO2010119615A1 PCT/JP2010/001901 JP2010001901W WO2010119615A1 WO 2010119615 A1 WO2010119615 A1 WO 2010119615A1 JP 2010001901 W JP2010001901 W JP 2010001901W WO 2010119615 A1 WO2010119615 A1 WO 2010119615A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- learning data
- specific
- cluster
- learning
- candidate
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
- G06F40/295—Named entity recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Definitions
- the present invention relates to a learning data generation apparatus that generates learning data necessary for creating a specific expression extraction rule, a specific expression extraction system using the same, a learning data generation method, and a computer-readable recording medium.
- correct data (learning data) is created by annotating the expression of the class to be extracted in the text, and then extracted from the correct data by machine learning (specific expression extraction rule)
- machine learning specific expression extraction rule
- Patent Document 1 discloses an example of a system that can create conventional correct answer data at low cost.
- the system described in Patent Literature 1 includes a correct data storage unit that stores correct answer data, a correct answer expansion unit, and a rule learning unit.
- the correct answer extension unit calls up the correct answer data from the correct answer data storage unit, performs word order operation, syntactic expression conversion, specific expression conversion, and the like, and generates extended data obtained by extending the correct answer data.
- the rule learning unit learns the extraction rule using both the correct answer data and the generated extended data as learning data.
- Patent Document 1 can reduce the cost by reducing the number of steps for creating correct data, the cost for creating correct data is still high for the following reasons.
- the object of the present invention is to solve the above-mentioned problems, eliminate the need for accurate annotation work without omission, and reduce the generation cost of learning data for creating a specific expression extraction rule. It is an object to provide a learning data generation method and a computer-readable recording medium.
- a learning data generation device is a learning data generation device that generates learning data for creating a specific expression extraction rule, A learning data candidate clustering unit for clustering a plurality of learning data candidates to which a label indicating an annotation class is assigned based on a feature amount including each context information; Referring to each cluster obtained by the clustering result, the distribution of learning data candidate labels in each cluster is obtained, and based on the obtained distribution, learning data candidates satisfying a set condition are specified, A learning data generating unit that generates learning data using the specified learning data candidates; It is characterized by having.
- a named entity extraction system includes the learning data generating device according to the present invention and a named entity extracting device,
- the specific expression extraction device learns a specific expression extraction rule for extracting a specific expression, using the learning data generated by the learning data device, and uses the learned specific expression extraction rule, Extracting the named entity from a corpus containing documents; It is characterized by that.
- a learning data generation method for generating learning data for creating a specific expression extraction rule, (A) clustering a plurality of learning data candidates to which a label indicating an annotation class is assigned based on a feature amount including each context information; and (B) Referring to each cluster obtained by the clustering result in step (a), the distribution of learning data candidate labels in each cluster is obtained, and a set condition is determined based on the obtained distribution. Identifying learning data candidates to be satisfied, and generating learning data using the identified learning data candidates; and It is characterized by having.
- a computer-readable recording medium is a computer-readable recording medium on which a program for generating learning data for creating a specific expression extraction rule is recorded by a computer. Because In the computer, (A) clustering a plurality of learning data candidates to which a label indicating an annotation class is assigned based on the feature amount including each context information; and (B) Referring to each cluster obtained by the clustering result in step (a), the distribution of learning data candidate labels in each cluster is obtained, and the set condition is determined based on the obtained distribution. Identifying learning data candidates to be satisfied, and generating learning data using the identified learning data candidates; and A program including an instruction for executing is recorded.
- the learning data generation device the specific expression extraction system, the learning data generation method, and the computer-readable recording medium in the present invention, accurate annotation work without omission is unnecessary, thereby The generation cost of learning data for creating extraction rules can be reduced.
- FIG. 1 is a block diagram showing a configuration of a learning data generation apparatus according to Embodiment 1 of the present invention.
- FIG. 2 is a flowchart showing the operation of the learning data generation apparatus according to Embodiment 1 of the present invention.
- 3 is a diagram showing a specific example of data stored in the incomplete correct answer corpus storage unit shown in FIG. 1,
- FIG. 3 (a) shows a text data table, and
- FIG. 3 (b) shows an annotation data table. Is shown.
- FIG. 4 is a diagram illustrating a specific example of learning data candidates generated by the learning data candidate generation unit illustrated in FIG. 1.
- FIG. 5 is a diagram showing a specific example of the processing result of the learning data candidate clustering unit shown in FIG.
- FIG. 1 is a block diagram showing a configuration of a learning data generation apparatus according to Embodiment 1 of the present invention.
- FIG. 2 is a flowchart showing the operation of the learning data generation apparatus according to Embodiment 1 of the present invention.
- FIG. 6 is a block diagram showing a configuration of a learning data generation apparatus according to Embodiment 2 of the present invention.
- FIG. 7 is a diagram showing a specific example of dictionary data stored in the dictionary data storage unit shown in FIG.
- FIG. 8 is a flowchart showing the operation of the learning data generation apparatus according to Embodiment 2 of the present invention.
- FIG. 9 is a block diagram showing a configuration of a learning data generation apparatus according to Embodiment 3 of the present invention.
- FIG. 10 is a flowchart showing the operation of the learning data generation apparatus according to Embodiment 3 of the present invention.
- FIG. 11 is a block diagram showing the configuration of the named entity extraction system in the embodiment of the present invention.
- FIG. 12 is a block diagram showing a computer that can execute the program according to any one of the first to third embodiments of the present invention.
- learning data can be generated from an incomplete correct corpus (hereinafter referred to as “incomplete correct corpus”).
- An incomplete correct answer corpus is data obtained by giving an annotation to a part of a specific expression in a text.
- the incomplete correct corpus may include an incorrect annotation as long as the ratio is small.
- Such an incomplete correct corpus can be created at a very low cost. For example, a character string of a known unique expression is searched for a corpus, and an annotation can be added to text constituting the corpus by reading only the searched portion.
- An incomplete correct corpus may include annotation mistakes, so if it can be expected that most of the assigned annotations are correct, it will be created by mechanically adding annotations to the occurrences of the character strings of the searched unique expressions. It may be.
- the created learning data includes “annotation mistake” and “annotation omission”.
- An example of “annotation mistake” and “annotation omission” will be described below by taking the case where the target specific expression class is “person name” and “organization name”.
- “annotation mistake” means that the annotated specific expression class is not the original class of the specific expression. For example, an example in which “Tsuchida” of “Tsuchida Electric Co., Ltd.” is annotated as a personal name is applicable. The correct name is “Tsuchida Denki” and the class is the organization name. Also, “annotation omission” refers to a case where no annotation is performed on the specific expression. For example, an example where “Yamada” of “Yamada works for NEC” is not annotated is applicable. To be correct, "Yamada" must be annotated with a person name.
- the learning data including the above-mentioned annotation mistakes and annotation omissions is not appropriate for use in the technique of recognition of proper expressions.
- the clustering method is used to detect “annotation miss” and “annotation omission” from the incomplete correct corpus that is a candidate for learning data, thereby automatically generating correct learning data.
- learning data candidates are prepared or generated in units of words in the incomplete correct corpus or in units of word strings compiled by some means (described later). (See “Learning Data Candidate Generation Unit 20” shown in FIG. 1).
- learning data candidates are generated, the annotation of the incomplete correct corpus is referred to, and class information is given to the label of each learning data candidate.
- the expression “UNKNOWN” is assigned to the expression to which no annotation is attached because it is not known whether the expression is a specific expression.
- words or word strings that are candidates for learning data may be narrowed down by the part of speech.
- the learning data candidates to be used and the learning data generated therefrom are each configured by data holding a label and a feature amount.
- the label is given by referring to the annotation information of the incomplete correct answer corpus.
- the feature amount includes at least contextual information (context information) such as a word, a dependency, and a clause of a dependency destination around a learning data candidate or a unit constituting the learning data (see FIG. 4 described later).
- the feature amount may be expressed in a format suitable for a learning algorithm for learning an extraction rule, such as a vector or an attribute-attribute value table.
- clustering is performed on the prepared or generated learning data candidates (see “learning data candidate clustering unit 21” shown in FIG. 1 described later).
- any existing method such as k-means or PLSI (Probabilistic Latent Semantic Indexing) can be used.
- learning data is generated from the learning data candidates using the distribution of learning data candidate labels in each cluster as a clue.
- generation of learning data based on the above-mentioned idea that “many parts with annotations are correct” and “data with similar feature values are likely to be in the same class” Is done.
- learning data candidates having similar feature quantities belong to the same cluster.
- the learning data candidates among them are likely to be class A.
- the learning data candidates among them are not likely to be class A.
- words of the same class are likely to be used in the same way, and if the feature quantity includes at least context information, words of the same class tend to belong to the same cluster. Therefore, in the present invention, the feature amount includes context information.
- the distribution of learning data candidate labels in a cluster is used for selection (generation) of learning data. Therefore, the following four points [A] to [D] are considered as an index of good clustering results.
- Clustering parameters and termination conditions are set using the following [A] to [D] as indices.
- [A] The distribution of labels in each cluster differs between clusters.
- [B] There is one label with the highest ratio in each cluster.
- [C] The number of learning data candidates in the cluster is greater than a specific number.
- the number of divisions (number of clusters) during clustering is less than a specific number.
- annotation mistakes and annotation leakage are detected.
- data of class A included in a cluster having many classes other than class A is detected as an annotation error.
- data of other labels (including “UNNOWN”) in the cluster including many classes A are detected as missing annotations.
- the learning data is generated in consideration of the detected annotation mistake and annotation omission (see the learning data generation unit 22 shown in FIG. 1 described later).
- “Takeda” in (3) sentence is the same as “Tsuchida” in (1) sentence, (2) sentence, and (3) sentence in the clause “Working”. For this reason, “Takeda” in the (3) sentence is the same cluster as “Tsuchida” in the (1) sentence, (2) sentence, and (3) sentence. Therefore, (3) “Takeda” in the sentence belongs to the cluster with many “person names”, but is not annotated, so it can be determined that the annotation is missing.
- annotation mistakes and annotation omissions detected in this way are truly annotation mistakes or annotation omissions can only be determined manually, but in the present invention, all of them are removed from the learning data candidates. Learning data. The reason for this will be described below.
- the class indicated by the annotation is a correct class, but it can be assumed that there are many other learning data of the correct class that are detected incorrectly.
- the mistake that the annotation omission could not be detected corresponds to the fact that the annotation is not given to the specific expression and is treated as a non-specific expression.
- the correct answer data may include one or both of an annotation mistake and an annotation omission. That is, since the correct answer data includes incorrect correct answer data, it is adversely affected when the extraction rule is learned.
- the learning data candidate label may be changed to the correct label, and the learning data candidate after the label change may be added to the learning data.
- the correct label can be estimated by calculating the probability (reliability) that the label of the learning data candidate determined as an annotation mistake or annotation omission becomes a label that occupies a high proportion in the cluster to which it belongs. Then, when the reliability is higher than the set threshold value, the label of the learning data candidate determined as an annotation mistake or annotation omission may be changed to a label that occupies a high ratio in the cluster to which the user belongs.
- the target learning data candidate when the distance from the representative point of the target learning data candidate cluster is small, when the target learning data candidate exists in a high-density space within the cluster, the target learning data candidate When the probability of belonging to the cluster is high, a parameter having a high value may be used. Further, the threshold used for the determination may be set as appropriate according to the recognition result using the learning data.
- only label information may be adopted from learning data candidates, and a necessary feature amount may be newly acquired.
- the contextual information may be used as the feature amount of the learning data candidate, and in-word information such as a word or a word string itself and a surface character string may be further added in the generation of learning data.
- the learning data generation apparatus of the present invention assigns a label by referring to incomplete annotation information in units of either words in incomplete correct answer data or word strings compiled by some means.
- a learning data candidate generating unit for preparing or generating the learning data candidates, a learning data candidate clustering unit for clustering the learning data candidates, and learning data in each cluster with reference to each cluster obtained by the clustering result A learning data generation unit that generates learning data from learning data candidates using a distribution of candidate labels.
- FIG. 1 is a block diagram showing a configuration of a learning data generation apparatus according to Embodiment 1 of the present invention.
- the learning data generation device 2 in the first embodiment shown in FIG. 1 is a device that generates learning data for creating a specific expression extraction rule.
- the specific expression extraction rule is a rule used for recognition of a specific expression in a question answering system, a document document system, a machine translation system, or the like.
- the learning data generation device 2 includes a learning data candidate clustering unit 21 and a learning data generation unit 22.
- the learning data candidate clustering unit 21 performs clustering on a plurality of learning data candidates to which a label indicating an annotation class is assigned based on the feature amount including each context information.
- the learning data generation unit 22 refers to each cluster obtained as a result of the clustering, and obtains a distribution of learning data candidate labels in each cluster. In addition, the learning data generation unit 22 specifies learning data candidates that satisfy the set conditions based on the obtained distribution, and generates learning data using the specified learning data candidates.
- the distribution of labels is obtained by clustering based on the feature amount. From the label distribution result, it is possible to easily guess the learning data candidate accompanied by the annotation mistake or the annotation omission. For this reason, it is possible to select learning candidates that are less likely to be accompanied by annotation mistakes or omissions without human intervention. Therefore, if the learning data generation device 2 is used, it is possible to generate learning data for creating a correct unique expression extraction rule accurately and without an annotation operation without omission, and thus the generation cost of learning data can be reduced.
- the learning data generation device 2 constitutes a learning data generation system 3 together with the storage device 1.
- the storage device 1 stores incomplete correct corpus, learning data candidates generated by the learning data generation device 2 from the incomplete correct corpus, and learning data.
- the learning data generation device 2 can be realized by a computer and operates by program control.
- the storage device 1 can be realized by a storage device included in a computer that implements the data generation device 2 or by a storage device such as another computer connected to the computer via a network.
- the storage device 1 includes an incomplete correct answer corpus storage unit 10, a learning data candidate storage unit 11, and a learning data storage unit 12.
- the incomplete correct answer corpus storage unit 10 stores an incomplete correct answer corpus.
- an incomplete correct corpus is obtained by adding annotations indicating classes to specific expressions that are part of the extraction objective in the text, and these annotations may contain mistakes and omissions. It is data with.
- an incomplete correct corpus is composed of text data and annotation data indicating what class of specific expression exists from what character to what character in the text data.
- the annotation data is “0: 2: personal name (persons from the 0th character to the second character)” and “3: 7: organization Name (the third to seventh characters are organization names) ”.
- the annotation data if the text data is subjected to morphological analysis, the number of words can be used instead of the number of characters.
- the learning data candidate storage unit 11 stores learning data candidates generated by a learning data candidate generation unit 20 described later.
- the learning data candidate is a set of data including a label indicating a class (hereinafter also referred to as “class label”) and a feature amount.
- class label a label indicating a class
- feature amount is preferably expressed in a format suitable for the learning algorithm for learning the extraction rule, such as a vector or an attribute-attribute value table.
- the learning data storage unit 12 stores learning data for creating a specific expression extraction rule generated by the learning data generation unit 22 using the learning data candidates.
- the learning data is also a set of data composed of class labels and feature amounts, like the learning data candidates.
- the feature amount is preferably expressed in a format suitable for the learning algorithm for learning the extraction rule, such as a vector or an attribute-attribute value table.
- the learning data generation device 2 further includes a learning data candidate generation unit 20 in addition to the learning data candidate clustering unit 21 and the learning data generation unit 22.
- the learning data candidate generation unit 20 reads text data from the incomplete correct answer corpus storage unit 10 and generates learning data candidates from the text data in the read corpus. Further, the learning data candidate generation unit 20 stores the generated learning data candidates in the learning data candidate storage unit 11.
- the learning data candidate generation unit 20 divides the text data read from the incomplete correct corpus storage unit 10. Each divided part becomes a unit of one learning data candidate.
- the unit of one learning data candidate is a word or a word string collected by some means.
- a word string put together by some means for example, a word string matched using a technical term string extracted using an existing technical term extraction technique or an item name of a large-scale dictionary such as Wikipedia Etc.
- the learning data candidate generation unit 20 uses the text data and the annotation information given to each word included in the text data for each unit that becomes a learning data candidate. A feature amount including information is detected, a feature amount is added to each unit, and a learning data candidate is generated.
- the learning data candidate generation unit 20 first refers to the annotation of the incomplete correct corpus and assigns class information to each learning data candidate label. If there is no annotation, a class such as “UNKNOWN” Add a label indicating that it is unknown. The feature amount includes contextual information. Moreover, the learning data candidate production
- the learning data candidate clustering unit 21 extracts data from the learning data candidate storage unit 11, performs clustering using the feature amount of the learning data candidate, and passes the clustering result to the learning data generation unit 22. Any method can be used for clustering. Further, the learning data candidate clustering unit 21 considers the above four points [A] to [D] as an index of the goodness of the clustering result, and any one or more, preferably [A] to [D]. Consider 4 points. Clustering parameters and termination conditions are set as appropriate so that a desired clustering result can be obtained.
- the above-mentioned index is taken into consideration because the learning data generation unit 22 that executes the following process detects an annotation error and an annotation omission based on the distribution of labels and generates learning data based on this. That is, if the distribution of labels is uneven among clusters and one class with a high ratio exists in one cluster, it is easy to detect an annotation mistake and an annotation omission.
- an algorithm using the number of clusters as a parameter such as k-means
- an algorithm using a similarity threshold between two data as a parameter it is preferable to set the parameter value so that the number of data in the cluster does not become too small.
- the learning data generation unit 22 receives the result of the learning data candidate clustering unit 21 and detects a learning data candidate that is an annotation error or an annotation omission from the distribution of labels in each cluster. In addition, the learning data generation unit 22 generates learning data from the learning data candidates by removing the detected learning data candidates or by changing the label of the detected learning data candidates. The generated learning data is stored in the learning data storage unit 12.
- the learning data generation unit 22 sets the learning data to which the specific label is given in the cluster including the specific label at a predetermined ratio or more, according to the set condition. It is specified as a learning candidate that satisfies. The identified learning candidate is used as learning data.
- a learning data candidate to which a specific label is not assigned is a learning data candidate in which an annotation error or an annotation is missing.
- the learning data generation unit 22 determines that an annotation error of a certain class is data of that class included in a cluster with many labels of other classes from the clustering result. Further, the learning data generation unit 22 determines that the annotation omission is data of another label in the cluster including many classes from the clustering result. Whether or not there are many classes in the cluster can be determined based on whether or not a certain class exceeds (determines) a predetermined ratio, which will be described later.
- the learning data generation part 22 is the degree which belongs to the said cluster about the learning data candidate which is not provided with the specific label in the cluster which contains a specific label more than a predetermined ratio. (Hereinafter referred to as “confidence”).
- the learning data generation unit 22 can delete learning data candidates whose certainty factor is less than the threshold from the entire learning data candidates, with learning data candidates having the certainty factor equal to or greater than the threshold value as learning data.
- the certainty factor can be regarded as the “reliability factor” described above.
- the learning data generation unit 22 uses learning data candidates with annotation mistakes and annotation omissions with high certainty as learning data of labels determined from the clusters, and removes learning data candidates with annotation errors and annotation omissions other than those. May be.
- the certainty factor includes the distance from the representative point of the learning data candidate cluster, the estimated density of the learning candidate data in the cluster, the ratio of the cluster including a specific label, and the like.
- the certainty factor includes the probability of belonging to a cluster of learning data candidates. Whether or not the certainty factor is high is determined based on a predetermined threshold value as described above. The threshold used for the determination may be set as appropriate according to the recognition result using the learning data. Further, each learning data candidate may be sorted based on the certainty factor, and a certain percentage of learning data candidates may be adopted as learning data.
- the learning data generation unit 22 can also add a new feature amount to the learning data candidate set as the learning data. That is, in the first embodiment, the learning data generation unit 22 extracts only the label information from the feature amount of the learning data candidate, and adds a new feature amount to the feature information to obtain the feature amount of the learning data. it can. For example, when only contextual information is used as the feature amount of the learning data candidate, the learning data generation unit 22 further generates a part of speech or surface layer of the word or the word string itself as the feature amount when generating the learning data. In-word information such as character strings can be added.
- the learning data generation device 2 may not have the learning data candidate generation unit 20. That is, the first embodiment may be an aspect in which processing corresponding to the processing performed by the learning data candidate generation unit 20 is performed in advance, and the learning data candidates are stored in the learning data candidate storage unit 11. .
- FIG. 2 is a flowchart showing the operation of the learning data generation apparatus according to Embodiment 1 of the present invention.
- generation of learning data for creating a specific expression extraction rule for recognizing a person name and a place name is taken as an example.
- the learning data generation method according to the first embodiment can be implemented by operating the learning data generation apparatus 2 shown in FIG. Therefore, the description of the learning data generation method according to the first embodiment is replaced with the following description of the operation of the learning data generation device 2.
- FIG. 1 is taken into consideration as appropriate.
- the learning data candidate generation unit 20 reads data from the incomplete correct corpus storage unit 10 and generates learning data candidates from the text in the corpus (step S1 in FIG. 2).
- the learning data candidates generated in step S1 are stored in the learning data candidate storage unit 11.
- FIG. 3 is a diagram showing a specific example of data stored in the incomplete correct answer corpus storage unit shown in FIG. 1, FIG. 3 (a) shows a text data table, and FIG. 3 (b) shows an annotation data table. Is shown.
- the incomplete correct corpus storage unit 10 stores a text data table that specifies text data and an annotation data table that specifies annotation data.
- the text data table shown in FIG. 3A is composed of text data and an ID that is an identifier thereof.
- the text ID of the annotation table shown in FIG. 3B has a relation with the ID of the text data table.
- “start” represents the number of start characters
- “end” represents the number of end characters
- “class” represents the annotated class information. That is, the data in the first row of the annotation table indicates that the fifth to seventh characters of the text having the ID of 1 in the text data table are annotated as a person name.
- the number of characters is a value counted from 0.
- FIG. 4 is a diagram illustrating a specific example of learning data candidates generated by the learning data candidate generation unit illustrated in FIG. 1. An example of learning data candidates created from the text of ID 2 shown in FIGS. 3A and 3B will be described with reference to FIG.
- the learning data candidate generation unit 20 refers to the annotation table (FIG. 3B), assigns class information to a label of a word with an annotation, and applies to a label of a word without an annotation. “UNKNOWN” indicating that the class is unknown is assigned.
- the feature quantity includes “two previous words”, “one previous word”, “one next word”, “two words after”, and “the main word of the related clause”.
- the learning data candidate is composed of a label representing class information and a feature amount.
- the learning data candidate generation unit 20 generates learning data candidates in this way, and stores the results in the learning data candidate storage unit 11.
- the learning data candidate clustering unit 21 reads the learning candidate data stored in the learning data candidate storage unit 11, and clusters the learning data candidates using the feature amount (step S2 in FIG. 2).
- the learning data candidate clustering unit 21 passes the clustering result to the learning data generation unit 22.
- FIG. 5 is a diagram showing a specific example of the processing result of the learning data candidate clustering unit 21 shown in FIG. Referring to FIG. 5, in FIG. 5, the learning data candidates are divided into n clusters. In the example of FIG. 5, the label ratio of the cluster 1 is 80% personal name, 10% place name, and 10% UNKNOWN.
- the learning data generation unit 22 receives the clustering result of the learning data candidate clustering unit 21, looks at the distribution of labels of each cluster, and generates learning data from clusters including a specific label at a certain ratio or more. (Step S3 in FIG. 2).
- the generated learning data is stored in the learning data storage unit 12.
- the learning data generation unit 22 generates learning data using the clustering result shown in FIG. 5 will be described.
- learning data is generated from a cluster including 70% or more of one class.
- cluster 1 since the personal name exceeds 70%, it is determined that the personal name is reliable.
- the place name and UNKNOWN included in the cluster 1 are determined to be unreliable and are removed from the cluster 1.
- the place name since the place name exceeds 70%, it is determined that the place name is reliable.
- the person name and UNKNOWN are determined to be unreliable and are removed.
- cluster n since UNKNOWN exceeds 70%, it is determined that only non-unique expressions are included. In class n, it is determined that the person name and place name are not reliable, and these are removed.
- the learning data generation unit 22 combines the remaining learning data candidates after the execution of step S3 to obtain learning data.
- “Tsuchida” of cluster 1 that is an annotation omission is deleted.
- “Fukushima” of cluster 2 that is an annotation mistake is also deleted. In this way, correct learning data is generated by executing steps S1 to S3.
- learning data candidates are created based on annotation information from an incomplete correct corpus that can be created at low cost. Then, the learning data candidates are clustered, and an annotation error and an annotation omission that are inappropriate learning data candidates are detected using the class label distribution of the learning data candidates in the cluster. Then, learning data is generated by removing the detected inappropriate learning data candidates. Therefore, according to the first embodiment, it is possible to create correct learning data for extracting a specific expression at low cost.
- FIG. 6 is a block diagram showing a configuration of a learning data generation apparatus according to Embodiment 2 of the present invention.
- FIG. 7 is a diagram showing a specific example of dictionary data stored in the dictionary data storage unit shown in FIG.
- the learning data generation system 6 includes a storage device 4 and a learning data generation device 5 unlike the first embodiment.
- the storage device 4 is different from the storage device 1 shown in FIG. 1 in the first embodiment, instead of the incomplete correct corpus storage unit 10, a corpus storage unit 13 and a dictionary data storage unit 14.
- the corpus storage unit 13 stores a corpus having no annotation information.
- the dictionary data storage unit 14 stores a dictionary in which unique expressions and their classes are registered. Except for the above points, the storage device 4 is configured in the same manner as the storage device 1 shown in FIG.
- the learning data generation device 5 in the second embodiment of the present invention further includes an annotation information generation unit 23, unlike the learning data generation device 2 shown in FIG. 1 in the first embodiment. ing.
- the annotation information generation unit 23 generates annotation information.
- the learning data generation device 5 is configured in the same manner as the learning data generation device 2 shown in FIG. The difference between the second embodiment and the first embodiment will be specifically described below.
- the corpus storage unit 13 stores text data.
- An example of the text data stored in the corpus storage unit 13 is the text data table shown in FIG.
- the corpus storage unit 13 stores only the text data table.
- the dictionary data storage unit 14 stores a dictionary (dictionary data) in which unique expressions and their classes are registered.
- dictionary dictionary data
- “Fukushima” and “Yamada” are registered as personal names in the dictionary data.
- “Osaka” and “Nara” are registered as place names.
- the annotation information generation unit 23 first reads text data from the corpus storage unit 13 and reads dictionary data from the dictionary data storage unit 14 respectively. Then, the annotation information generation unit 23 compares the plurality of unique expressions registered in the dictionary data with the text data stored in the corpus storage unit 13, and determines the character strings of the unique expressions registered in the dictionary. Search from text data.
- annotation information generation unit 23 specifies the position of the appearance position of the specific expression and the class of the specific expression by character string matching, and creates annotation information based on these.
- the annotation information is passed to the learning data candidate generation unit 20. Since the annotation information corresponds to the annotation table shown in FIG. 3B, it can be said that the annotation information generation unit 23 has a function of creating an incomplete correct corpus.
- FIG. 8 is a flowchart showing the operation of the learning data generation apparatus according to Embodiment 2 of the present invention.
- the learning data generation method can also be implemented by operating the learning data generation device 5 shown in FIG. Therefore, the description of the learning data generation method according to the second embodiment is replaced with the following description of the operation of the learning data generation device 5.
- FIG. 6 is taken into consideration as appropriate.
- the annotation information generation unit 23 reads data from the corpus storage unit 13 and the dictionary data storage unit 14 respectively, and for the portion where the specific expression registered in the dictionary data appears in the corpus. To generate annotation information (step S11 in FIG. 8). After execution of step S1, the annotation information generation unit 23 passes the corpus and annotation information stored in the corpus storage unit 13 to the learning data candidate generation unit 20 as incomplete correct corpus data.
- the learning data candidate generation unit 20 generates learning data candidates from the incomplete correct answer corpus data created in step S11 (step S12 in FIG. 8).
- clustering by the learning data candidate clustering unit 21 (step S13 in FIG. 8) and generation of learning data by the learning data generation unit 22 (step S14 in FIG. 8) are performed.
- Steps S12 to S14 are the same as steps S1 to S3 shown in FIG. 2 in the first embodiment.
- an incomplete correct corpus is automatically created by the corpus and the dictionary. Therefore, as long as a dictionary is prepared, it is not necessary to manually create an incomplete correct corpus, and the cost for creating an incomplete correct corpus can be reduced. According to the second embodiment, learning data can be generated at a lower cost than in the first embodiment.
- FIG. 9 is a block diagram showing a configuration of a learning data generation apparatus according to Embodiment 3 of the present invention.
- the learning data generation system 9 includes a storage device 7 and a learning data generation device 8, unlike the first and second embodiments.
- the storage device 7 includes an extraction rule storage unit 15 in addition to the configuration of the storage device 4 shown in FIG. 6 in the second embodiment.
- the learning data generation device 8 includes an extraction rule learning unit 24 and a dictionary multiplication unit 25 in addition to the configuration of the learning data generation device 5 shown in FIG. 6 in the second embodiment. The difference between the third embodiment and the second embodiment will be specifically described below.
- the extraction rule learning unit 24 learns a specific expression extraction rule for extracting a specific expression using the learning data generated by the learning data generation unit 22.
- the extraction rule learning unit 24 reads the learning data from the learning data storage unit 12, learns the specific expression extraction rule using an existing arbitrary extraction rule learning method, and extracts the result as the extraction rule storage unit. 15 is output.
- the extraction rule storage unit 15 stores the unique expression extraction rule learned by the extraction rule learning unit 24.
- the dictionary multiplication unit 25 extracts a specific expression from the corpus including a plurality of documents using the extraction rule learned by the extraction rule learning unit 24. Then, the dictionary multiplication unit 25 additionally registers the unique expression satisfying the set condition among the extracted unique expressions and the class information for specifying the class in the dictionary stored in the dictionary data storage unit 14. To do.
- the dictionary multiplication unit 25 reads data from the corpus storage unit 13 and the extraction rule storage unit 15 respectively, and extracts a specific expression from the corpus text data stored in the corpus storage unit 13. In addition, the dictionary multiplication unit 25 obtains the certainty factor of the extracted unique expression, and determines the unique expression having the certainty factor equal to or higher than the threshold, several cases having the highest certainty factor, It is stored in the dictionary data storage unit 14 together with the class information.
- the dictionary multiplication unit 25 can also calculate the certainty factor using the reliability of the extraction rule.
- the threshold value that serves as a criterion for determining whether or not to store in the dictionary data storage unit 14 may be set as appropriate in accordance with the degree of annotation mistakes and annotation omissions included in the learning data candidates generated by the annotation information generation unit 23. .
- FIG. 10 is a flowchart showing the operation of the learning data generation apparatus according to Embodiment 3 of the present invention.
- the learning data generation method can be implemented by operating the learning data generation device 8 shown in FIG. Therefore, the description of the learning data generation method in the third embodiment is replaced with the following description of the operation of the learning data generation device 8.
- FIG. 9 is referred to as appropriate.
- the annotation information generating unit 23 generates incomplete correct corpus data (step S21 in FIG. 10).
- the learning data candidate generation unit 20 generates learning data candidates from the incomplete correct answer corpus data generated in step S21 (step S22 in FIG. 10).
- clustering by the learning data candidate clustering unit 21 step S23 in FIG. 10
- generation of learning data by the learning data generation unit 22 step S24 in FIG. 10.
- the extraction rule learning unit 24 reads the learning data from the learning data storage unit 12, and learns the specific expression extraction rule according to any known extraction rule learning method (S25 in FIG. 10). .
- the extraction rule learning unit 24 stores the generated extraction rule in the extraction rule storage unit 15.
- the dictionary multiplication unit 25 reads the specific expression extraction rule from the extraction rule storage unit 15 and reads the text data from the corpus. Then, using these, the dictionary multiplication unit 25 extracts a specific expression in the corpus, and among the extracted specific expressions, the specific expression having a certainty level equal to or greater than a threshold is stored together with the class information in the dictionary data storage unit. 14 (step S26 in FIG. 10).
- the dictionary multiplication unit 25 determines whether or not the termination condition is satisfied (S7 in FIG. 10). If the end condition is not satisfied as a result of the determination, step S1 is executed again. On the other hand, if the end condition is satisfied as a result of the determination, the operation of the system ends. Examples of the termination condition include the number of executions of steps S21 to S27 (end if repeated a prescribed number of times) and the number of words added to the dictionary (end if the number of additional words exceeds a prescribed value). Moreover, when no specific expression is added, the process may be ended.
- the “confidence level” used in the third embodiment will be described.
- the following scores can be used as the certainty factor.
- the certainty factor is calculated so as to be high enough to satisfy the two points that the specific expression w is correct as the class c and the specific expression w is not used as a class other than the class c in the corpus.
- the certainty factor (w, c) is calculated by the following (Equation 1) and (Equation 2).
- “ext_cnt (w, c)” indicates the number of specific expressions w extracted as class c.
- D (w) indicates a set of learning data in which the specific expression w appears in the corpus.
- X indicates one piece of learning data including the specific expression w.
- “Fci (x)” indicates the reliability assigned to the extraction rule used to determine whether “x” is a class ci.
- “Fc (x)” indicates the reliability assigned to the extraction rule used when determining whether “x” is class c.
- ext_cnt (w, c) is the number of times that the specific expression w is extracted as the class c, so that the larger the value, the more accurate the specific expression w is as the specific expression of the class c. Further, ( ⁇ _x f (x, c)) /
- class information includes “person name”, “place name”, and “UNKNOWN”.
- the certainty factor is calculated as follows from the above (Equation 1) and (Equation 2).
- the certainty factor is calculated to be high.
- the specific expression “Fukuoka” is often used as a person name or a place name, so the certainty factor is calculated to be low.
- a specific expression is extracted from the corpus using a specific expression extraction rule obtained from the generated learning data, and a reliable specific expression is added to the dictionary. Then, incomplete correct answer data is created from the corpus using the dictionary to which the specific expression is added. Further, the addition of the specific expression to the dictionary and the generation of incomplete correct answer data can be repeated.
- the dictionary for creating the incomplete correct corpus may be initially small in scale, and can be increased later. According to the third embodiment, since learning data can be created even if only a small dictionary is prepared, learning data can be created at a lower cost than in the first and second embodiments. become able to.
- FIG. 11 is a block diagram showing the configuration of the named entity extraction system in the embodiment of the present invention.
- the specific expression extraction system shown in FIG. 11 is a system including the learning data generation apparatus described in the first to third embodiments as a constituent element. As illustrated in FIG. 11, the specific expression extraction system includes a learning data generation system 100, a specific expression extraction device 106, and a storage device 107.
- a learning data generation system 100 is the learning data generation system shown in any one of the first to third embodiments.
- the specific expression extraction device 106 includes a specific expression extraction rule learning unit 101 and a specific expression extraction unit 104.
- the storage device 107 includes an extraction rule storage unit 102, a processing target text data storage unit 103, and an extraction result storage unit 105.
- the specific expression extraction rule learning unit 101 reads the learning data from the learning data storage unit 12 provided in the learning data generation system 100 and learns the specific expression extraction rule for extracting the specific expression.
- the specific expression extraction rule learning unit 101 has the same function as the extraction rule learning unit 24 shown in FIG. 9 in the third embodiment.
- the learned specific expression extraction rule is stored in the extraction rule storage unit 102.
- the specific expression extraction unit 104 extracts a specific expression from the text data stored in the processing target text data storage unit 103 using the specific expression extraction rule read from the extraction rule storage unit 102.
- the specific expression extraction unit 104 stores the extracted specific expression in the extraction result storage unit 105.
- the processing target text data storage unit 103 stores a corpus including a plurality of documents to be processed.
- the named entity extraction system in the present embodiment includes the learning data generation system of the present invention. Therefore, since the specific expression extraction rule can be learned from the learning data created at low cost, it is possible to reduce the cost for constructing the specific expression extraction system.
- FIG. 12 is a block diagram showing a computer that can execute the program according to any one of the first to third embodiments of the present invention.
- the computer shown in FIG. 12 functions as an information processing system that implements the learning data generation apparatus according to the present invention.
- the computer (information processing system) includes a processor 300, a program memory 301, a storage medium 302, an interface 303 connected to an external device, and a video card 304 for outputting display data. It has.
- the storage medium 302 various memories such as a RAM and a magnetic storage medium such as a hard disk can be used.
- an input device 305 such as a keyboard and a display device 306 such as a liquid crystal display device are connected to the computer.
- the program memory 301 stores a program for operating the processor 300.
- the program is provided by being recorded on a computer-readable recording medium, for example.
- the program may be provided via a network or the like.
- Examples of the computer-readable recording medium include an optical disk, a magnetic disk, a magneto-optical disk, a semiconductor memory, and a flexible disk.
- the program stored in the program memory 301 includes instructions for executing the processing steps shown in any of the first to third embodiments.
- the processor 300 functions as the learning data candidate generation unit 20, the learning data candidate clustering unit 21, and the learning data generation unit 22, and the storage medium 302 functions as the storage device 1.
- a program including instructions for causing the processor 300 to execute steps S11 to S14 shown in FIG. 8 is stored in the program memory 301, and when this program is executed, the learning data generating apparatus 5 in the second embodiment is used. Realized.
- the processor 300 functions as the learning data candidate generation unit 20, the learning data candidate clustering unit 21, the learning data generation unit 22, and the annotation information generation unit 23.
- the storage medium 302 functions as the storage device 4.
- a program including instructions for causing the processor 300 to execute steps S21 to S27 shown in FIG. 10 is stored in the program memory 301, and if this program is executed, the learning data generating apparatus 8 in the third embodiment is Realized.
- the processor 300 functions as the learning data candidate generation unit 20, the learning data candidate clustering unit 21, the learning data generation unit 22, the annotation information generation unit 23, the extraction rule learning unit 24, and the dictionary multiplication unit 25.
- the storage medium 302 functions as the storage device 7.
- a storage medium of another external computer may function as the storage devices 1, 4, and 7.
- the learning data generation device, the specific expression extraction system, the learning data generation method, and the computer-readable recording medium in the present invention have the following characteristics.
- a learning data generation device that generates learning data for creating a specific expression extraction rule, A learning data candidate clustering unit for clustering a plurality of learning data candidates to which a label indicating an annotation class is assigned based on a feature amount including each context information; Referring to each cluster obtained by the clustering result, the distribution of learning data candidate labels in each cluster is obtained, and based on the obtained distribution, learning data candidates satisfying a set condition are specified, A learning data generating unit that generates learning data using the specified learning data candidates; A learning data generation device comprising:
- the learning data generation unit sets, as the learning data of the specific label, a learning data candidate to which the specific label is given in the cluster including a specific label at a predetermined ratio or more.
- the learning data generation device according to (1).
- the learning data generation unit deletes, from the entire learning data candidate, a learning data candidate to which the specific label is not given in the cluster including a specific label at a predetermined ratio or more.
- the learning data generation unit obtains a degree of belonging to the cluster with respect to learning data candidates to which the specific label is not assigned in the cluster including a specific label at a predetermined ratio or more.
- the learning data according to (2) wherein the learning data candidate whose degree of determination is greater than or equal to a threshold is the learning data, and the learning data candidate whose degree of determination is less than the threshold is deleted from the entire learning data candidate. Generator.
- the learning data generation unit determines the distance from the representative point of the cluster of the learning data candidate as the degree to which the learning data candidate not assigned the specific label belongs to the cluster, the learning data candidate The learning data generation according to (4), wherein at least one of the probability of belonging to the cluster, the estimated density of the learning data candidate in the cluster, and the ratio of the cluster including the specific label is used. apparatus.
- the learning data candidate clustering unit when the clustering, the label distribution in each cluster is different between the clusters, there is a label having the highest ratio in each cluster, learning in each cluster Learning according to (1) above, wherein the clustering condition is set so that any one or more of the number of data candidates is greater than a specific number and the number of clusters is less than a specific number is satisfied.
- Data generator when the clustering, the label distribution in each cluster is different between the clusters, there is a label having the highest ratio in each cluster, learning in each cluster Learning according to (1) above, wherein the clustering condition is set so that any one or more of the number of data candidates is greater than a specific number and the number of clusters is less than a specific number is satisfied.
- a learning data candidate generation unit uses text data and annotation information given to each word included in the text data, for each of a plurality of words or word strings obtained by dividing the text data, The learning data generation apparatus according to (1), wherein a feature amount including context information is detected, and the feature amount is added to each word or each word string to generate the learning data candidate.
- An annotation information generation unit that creates the annotation information; A dictionary in which a plurality of specific expressions and the class of each of the plurality of specific expressions are registered; The annotation information generation unit compares the plurality of specific expressions registered in the dictionary with the text data, and identifies an appearance location where any of the plurality of specific expressions in the text data appears.
- the learning data generation device according to (8), wherein the annotation information that identifies the position of the appearance location and the class of the specific expression that appears is created.
- the extraction rule learning unit learns a specific expression extraction rule for extracting the specific expression using the learning data generated by the learning data generation unit
- the dictionary multiplication unit extracts a specific expression from a corpus including a plurality of documents by using an extraction rule learned by the extraction rule learning unit, and sets a set condition among the extracted specific expressions.
- the learning data generation device according to (9), wherein a specific expression to be satisfied and class information for specifying the class are additionally registered in the dictionary.
- the learning data generation device includes: A learning data candidate clustering unit for clustering a plurality of learning data candidates to which a label indicating an annotation class is assigned based on a feature amount including each context information; Referring to each cluster obtained by the clustering result, the distribution of learning data candidate labels in each cluster is obtained, and based on the obtained distribution, learning data candidates satisfying a set condition are specified, A learning data generation unit that generates learning data using the specified learning data candidates,
- the specific expression extraction device learns a specific expression extraction rule for extracting a specific expression, using the learning data generated by the learning data device, and uses the learned specific expression extraction rule, Extracting the named entity from a corpus containing documents; A unique expression extraction system characterized by that.
- a learning data generation method for generating learning data for creating a specific expression extraction rule (A) clustering a plurality of learning data candidates to which a label indicating an annotation class is assigned based on a feature amount including each context information; and (B) Referring to each cluster obtained by the clustering result in step (a), the distribution of learning data candidate labels in each cluster is obtained, and a set condition is determined based on the obtained distribution. Identifying learning data candidates to be satisfied, and generating learning data using the identified learning data candidates; and A learning data generation method characterized by comprising:
- step (b) a learning data candidate to which the specific label is assigned in the cluster including a specific label at a predetermined ratio or more is set as the learning data of the specific label.
- step (b) the learning data candidates to which the specific label is not given in the cluster including the specific label at a predetermined ratio or more are deleted from the entire learning data candidate.
- step (b) for the learning data candidates to which the specific label is not given in the cluster including the specific label at a predetermined ratio or more, the degree of belonging to the cluster is obtained,
- Data generation method for the learning data candidates to which the specific label is not given in the cluster including the specific label at a predetermined ratio or more, the degree of belonging to the cluster is obtained.
- step (b) as a degree to which the learning data candidate not assigned with the specific label belongs to the cluster, the distance from the representative point of the cluster to the learning data candidate, the learning data candidate
- step (a) at the time of clustering, the distribution of labels in each cluster differs between clusters, the presence of the highest percentage of labels in each cluster, and learning in each cluster Learning according to (12) above, wherein the clustering condition is set so that at least one of the number of data candidates larger than a specific number and the number of clusters smaller than a specific number is satisfied Data generation method.
- (c) Context information for each of a plurality of words or word strings obtained by dividing the text data using text data and annotation information given to each word included in the text data.
- a computer-readable recording medium recording a program for generating learning data for creating a specific expression extraction rule by a computer, In the computer, (A) clustering a plurality of learning data candidates to which a label indicating an annotation class is assigned based on the feature amount including each context information; and (B) Referring to each cluster obtained by the clustering result in step (a), the distribution of learning data candidate labels in each cluster is obtained, and the set condition is determined based on the obtained distribution. Identifying learning data candidates to be satisfied, and generating learning data using the identified learning data candidates; and A computer-readable recording medium on which is recorded a program including instructions for executing
- step (b) a learning data candidate to which the specific label is assigned in the cluster including a specific label at a predetermined ratio or more is set as the learning data of the specific label.
- the degree of belonging to the cluster is obtained.
- step (b) as a degree to which the learning data candidate not assigned with the specific label belongs to the cluster, the distance from the representative point of the learning data candidate to the cluster, the learning data candidate The computer reading according to (25), wherein at least one of a probability of belonging to the cluster, an estimated density of the learning data candidates in the cluster, and a ratio of the cluster including the specific label is used. Possible recording media.
- step (a) at the time of clustering, the distribution of labels in each cluster differs between clusters, the presence of the highest percentage of labels in each cluster, learning in each cluster
- the program is stored in the computer.
- (C) A feature including context information for each of a plurality of words or word strings obtained by dividing the text data using text data and annotation information given to each word included in the text data.
- the program is stored in the computer.
- the computer-readable recording medium according to (29) further including instructions.
- the program is stored in the computer.
- a specific expression extraction rule can be created at a low cost. For this reason, this invention is useful for the system aiming at the effective utilization of text information, such as an information search system and a text mining system.
- the present invention has industrial applicability.
- Storage device (Embodiment 1) 2 Learning data generation device (Embodiment 1) 3 Learning data generation system (Embodiment 1) 4. Storage device (Embodiment 2) 5. Learning data generation device (Embodiment 2) 6 Learning data generation system (Embodiment 2) 7.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Machine Translation (AREA)
Abstract
Description
アノテーションのクラスを示すラベルが付与された複数の学習データ候補を、それぞれの文脈情報を含む特徴量に基づいて、クラスタリングする、学習データ候補クラスタリング部と、
前記クラスタリング結果で得られた各クラスタを参照して、前記各クラスタ内の学習データ候補のラベルの分布を求め、求めた前記分布に基づいて、設定された条件を満たす学習データ候補を特定し、特定した前記学習データ候補を用いて学習データを生成する、学習データ生成部と、
を備えていることを特徴とする。
前記固有表現抽出装置は、前記学習データ装置により生成された前記学習データを用いて、固有表現を抽出するための固有表現抽出ルールを学習し、学習した前記固有表現抽出ルールを用いて、複数の文書を含むコーパスから、前記固有表現を抽出する、
ことを特徴とする。
(a)アノテーションのクラスを示すラベルが付与された複数の学習データ候補を、それぞれの文脈情報を含む特徴量に基づいて、クラスタリングする、ステップと、
(b)前記ステップ(a)のクラスタリング結果で得られた各クラスタを参照して、前記各クラスタ内の学習データ候補のラベルの分布を求め、求めた前記分布に基づいて、設定された条件を満たす学習データ候補を特定し、特定した前記学習データ候補を用いて学習データを生成する、ステップと、
を有することを特徴とする。
前記コンピュータに、
(a)アノテーションのクラスを示すラベルが付与された複数の学習データ候補を、それぞれの文脈情報を含む特徴量に基づいて、クラスタリングする、ステップと、
(b)前記ステップ(a)のクラスタリング結果で得られた各クラスタを参照して、前記各クラスタ内の学習データ候補のラベルの分布を求め、求めた前記分布に基づいて、設定された条件を満たす学習データ候補を特定し、特定した前記学習データ候補を用いて学習データを生成する、ステップと、
を実行させる、命令を含むプログラムを記録していることを特徴とする。
本発明では、例えば、不完全な正解コーパス(以下「不完全正解コーパス」という。)から学習データを生成することができる。不完全正解コーパスとは、テキスト中の一部の固有表現に対してアノテーションを付与して得られたデータである。また、不完全正解コーパスには、割合が少なければ、間違ったアノテーションが含まれていても良い。
[A]各クラスタ内のラベルの分布がクラスタ間で異なる。
[B]各クラスタ内で割合が最も多いラベルが1つある。
[C]クラスタ内の学習データ候補の数が特定の数より多くなっている。
[D]クラスタリング時の分割数(クラスタ数)が特定の数より少なくなっている。
以下、本発明の実施の形態1における、学習データ生成装置及び学習データ生成方法について、図1~図5を参照しながら説明する。最初に、本実施の形態1における学習データ生成装置の構成について図1を用いて説明する。図1は、本発明の実施の形態1における学習データ生成装置の構成を示すブロック図である。
次に本発明の実施の形態2における、学習データ生成装置及び学習データ生成方法について、図6及び図7を参照しながら説明する。最初に、本実施の形態2における学習データ生成装置の構成について図6及び図7を用いて説明する。図6は、本発明の実施の形態2における学習データ生成装置の構成を示すブロック図である。図7は、図6に示す辞書データ記憶部に記憶されている辞書データの具体例を示す図である。
次に本発明の実施の形態3における、学習データ生成装置及び学習データ生成方法について、図9及び図10を参照しながら説明する。最初に、本実施の形態3における学習データ生成装置の構成について図9を用いて説明する。図9は、本発明の実施の形態3における学習データ生成装置の構成を示すブロック図である。
確信度(w,c)=ext_cnt(w,c)×(Σ_x f(x,c))/|D(w)|
f(x,c)=fc(x)/Σ_ci fci(x)
一か所目/「人名」ルール:0.6、「地名」ルール:0
二か所目/「人名」ルール:0.6、「地名」ルール:0
三か所目/「人名」ルール:0.5、「地名」ルール:0.8
=ext_cnt(斎藤,人名)×(Σ_x f(x,人名))/|D(斎藤)|
=3 × (0.6 / (0.6+0) + 0.6 /
(0.6+0) + 0.5 / (0.5+0.8)) / 3
=2.38
一か所目/「人名」ルール:0.6、「地名」ルール:0.6、
二か所目/「人名」ルール:0.6、「地名」ルール:0.6、
三か所目/「人名」ルール:0.5、「地名」ルール:0.8
=ext_cnt(福岡,人名)×(Σ_x f(x,人名))/|D(福岡)|
=3 × (0.6 / (0.6+0.6) + 0.6 /
(0.6+0.6) + 0.5 / (0.5+0.8)) / 3
=1.38
次に本発明の実施の形態における固有表現抽出システムについて、図11を参照しながら説明する。図11は、本発明の実施の形態における固有表現抽出システムの構成を示すブロック図である。
アノテーションのクラスを示すラベルが付与された複数の学習データ候補を、それぞれの文脈情報を含む特徴量に基づいて、クラスタリングする、学習データ候補クラスタリング部と、
前記クラスタリング結果で得られた各クラスタを参照して、前記各クラスタ内の学習データ候補のラベルの分布を求め、求めた前記分布に基づいて、設定された条件を満たす学習データ候補を特定し、特定した前記学習データ候補を用いて学習データを生成する、学習データ生成部と、
を備えていることを特徴とする学習データ生成装置。
前記学習データ候補生成部は、テキストデータと、テキストデータに含まれる各単語に付与されたアノテーション情報とを用いて、前記テキストデータを分割して得られた複数個の単語又は単語列毎に、文脈情報を含む特徴量を検出し、各単語又は各単語列に前記特徴量を付加して、前記学習データ候補を生成する、上記(1)に記載の学習データ生成装置。
複数の固有表現と、前記複数の固有表現それぞれの前記クラスとが登録されている辞書とを更に備え、
前記アノテーション情報生成部は、前記辞書に登録されている前記複数の固有表現と前記テキストデータとを比較し、前記テキストデータ中の前記複数の固有表現のいずれかが出現している出現箇所を特定し、前記出現箇所の位置と前記出現している固有表現の前記クラスとを特定するアノテーション情報を作成する、上記(8)に記載の学習データ生成装置。
前記抽出ルール学習部は、前記学習データ生成部により生成された前記学習データを用いて、前記固有表現を抽出するための固有表現抽出ルールを学習し、
前記辞書増殖部は、複数の文書を含むコーパスから、前記抽出ルール学習部によって学習された抽出ルールを用いて固有表現を抽出し、そして、抽出された前記固有表現のうちの設定された条件を満たす固有表現と、そのクラスを特定するクラス情報とを、前記辞書に追加登録する、上記(9)に記載の学習データ生成装置。
前記学習データ生成装置は、
アノテーションのクラスを示すラベルが付与された複数の学習データ候補を、それぞれの文脈情報を含む特徴量に基づいて、クラスタリングする、学習データ候補クラスタリング部と、
前記クラスタリング結果で得られた各クラスタを参照して、前記各クラスタ内の学習データ候補のラベルの分布を求め、求めた前記分布に基づいて、設定された条件を満たす学習データ候補を特定し、特定した前記学習データ候補を用いて学習データを生成する、学習データ生成部とを備え、
前記固有表現抽出装置は、前記学習データ装置により生成された前記学習データを用いて、固有表現を抽出するための固有表現抽出ルールを学習し、学習した前記固有表現抽出ルールを用いて、複数の文書を含むコーパスから、前記固有表現を抽出する、
ことを特徴とする固有表現抽出システム。
(a)アノテーションのクラスを示すラベルが付与された複数の学習データ候補を、それぞれの文脈情報を含む特徴量に基づいて、クラスタリングする、ステップと、
(b)前記ステップ(a)のクラスタリング結果で得られた各クラスタを参照して、前記各クラスタ内の学習データ候補のラベルの分布を求め、求めた前記分布に基づいて、設定された条件を満たす学習データ候補を特定し、特定した前記学習データ候補を用いて学習データを生成する、ステップと、
を有することを特徴とする学習データ生成方法。
(d)複数の文書を含むコーパスから、前記(e)のステップで学習された抽出ルールを用いて固有表現を抽出し、そして、抽出された前記固有表現のうちの設定された条件を満たす固有表現と、そのクラスを特定するクラス情報とを、前記辞書に追加登録するステップとを、更に有する、上記(20)に記載の学習データ生成方法。
前記コンピュータに、
(a)アノテーションのクラスを示すラベルが付与された複数の学習データ候補を、それぞれの文脈情報を含む特徴量に基づいて、クラスタリングする、ステップと、
(b)前記ステップ(a)のクラスタリング結果で得られた各クラスタを参照して、前記各クラスタ内の学習データ候補のラベルの分布を求め、求めた前記分布に基づいて、設定された条件を満たす学習データ候補を特定し、特定した前記学習データ候補を用いて学習データを生成する、ステップと、
を実行させる、命令を含むプログラムを記録していることを特徴とするコンピュータ読み取り可能な記録媒体。
(c)テキストデータと、テキストデータに含まれる各単語に付与されたアノテーション情報とを用いて、前記テキストデータを分割して得られた複数個の単語又は単語列毎に、文脈情報を含む特徴量を検出し、各単語又は各単語列に前記特徴量を付加して、前記学習データ候補を生成するステップを実行させる、命令を更に含む上記(22)に記載のコンピュータ読み取り可能な記録媒体。
(d)複数の固有表現と前記複数の固有表現それぞれの前記クラスとが登録されている辞書を用い、前記辞書に登録されている前記複数の固有表現と前記テキストデータとを比較し、前記テキストデータ中の前記複数の固有表現のいずれかが出現している出現箇所を特定し、前記出現箇所の位置と前記出現している固有表現の前記クラスとを特定するアノテーション情報を作成するステップを実行させる、命令を更に含む上記(29)に記載のコンピュータ読み取り可能な記録媒体。
(e)前記(b)のステップで生成された前記学習データを用いて、前記固有表現を抽出するための固有表現抽出ルールを学習するステップと、
(d)複数の文書を含むコーパスから、前記(e)のステップで学習された抽出ルールを用いて固有表現を抽出し、そして、抽出された前記固有表現のうちの設定された条件を満たす固有表現と、そのクラスを特定するクラス情報とを、前記辞書に追加登録するステップとを、実行させる、命令を更に含む上記(30)に記載のコンピュータ読み取り可能な記録媒体。
2 学習データ生成装置(実施の形態1)
3 学習データ生成システム(実施の形態1)
4 記憶装置(実施の形態2)
5 学習データ生成装置(実施の形態2)
6 学習データ生成システム(実施の形態2)
7 記憶装置(実施の形態3)
8 学習データ生成装置(実施の形態3)
9 学習データ生成システム(実施の形態3)
10 不完全正解コーパス記憶部
11 学習データ候補記憶部
12 学習データ記憶部
13 コーパス記憶部
14 辞書データ記憶部
15 抽出ルール記憶部
20 学習データ候補生成部
21 学習データ候補クラスタリング部
22 学習データ生成部
23 アノテーション情報生成部
24 抽出ルール学習部
25 辞書増殖部
100 学習データ生成システム
101 固有表現抽出ルール学習部
102 抽出ルール記憶部
103 処理対象テキストデータ記憶部
104 固有表現抽出部
105 抽出結果記憶部
106 固有表現抽出装置
107 記憶装置
300 プロセッサ
301 プログラムメモリ
302 記憶媒体
303 インターフェイス
304 ビデオカード
305 入力装置
306 表示装置
Claims (31)
- 固有表現抽出ルールの作成用の学習データを生成する、学習データ生成装置であって、
アノテーションのクラスを示すラベルが付与された複数の学習データ候補を、それぞれの文脈情報を含む特徴量に基づいて、クラスタリングする、学習データ候補クラスタリング部と、
前記クラスタリング結果で得られた各クラスタを参照して、前記各クラスタ内の学習データ候補のラベルの分布を求め、求めた前記分布に基づいて、設定された条件を満たす学習データ候補を特定し、特定した前記学習データ候補を用いて学習データを生成する、学習データ生成部と、
を備えていることを特徴とする学習データ生成装置。 - 前記学習データ生成部が、特定のラベルを予め定められた割合以上含む前記クラスタ内における、前記特定のラベルが付与された学習データ候補を、前記特定ラベルの前記学習データとする、請求項1に記載の学習データ生成装置。
- 前記学習データ生成部が、特定のラベルを予め定められた割合以上含む前記クラスタ内における、前記特定のラベルが付与されていない学習データ候補を、前記学習データ候補全体から削除する、請求項2に記載の学習データ生成装置。
- 前記学習データ生成部が、特定のラベルを予め定められた割合以上含む前記クラスタ内における、前記特定のラベルが付与されていない学習データ候補について、当該クラスタに所属する度合いを求め、求められた度合いが閾値以上となる前記学習データ候補を前記学習データとし、求められた度合いが閾値未満となる前記学習データ候補を前記学習データ候補全体から削除する、請求項2に記載の学習データ生成装置。
- 前記学習データ生成部が、前記特定のラベルが付与されていない学習データ候補が当該クラスタに所属する度合いとして、前記学習データ候補の当該クラスタの代表点からの距離、前記学習データ候補の当該クラスタへの所属確率、前記学習データ候補の当該クラスタ内での推定密度、及び当該クラスタが前記特定のラベルを含む割合のうち、1つ以上を用いる、請求項4に記載の学習データ生成装置。
- 前記学習データ候補クラスタリング部が、前記クラスタリング時に、前記各クラスタ内のラベルの分布がクラスタ間で異なる事、前記各クラスタ内で最も割合が多いラベルが存在する事、各クラスタ内の学習データ候補数が特定の数より多い事、及びクラスタ数が特定の数より少ない事のうちいずれか1つ以上が満たされるように、前記クラスタリングの条件を設定する、請求項1から5のいずれかに記載の学習データ生成装置。
- 前記学習データ生成部が、前記学習データとされた前記学習データ候補に、新たな特徴量を付加する、請求項1から6のいずれかに記載の学習データ生成装置。
- 学習データ候補生成部を更に備え、
前記学習データ候補生成部は、テキストデータと、テキストデータに含まれる各単語に付与されたアノテーション情報とを用いて、前記テキストデータを分割して得られた複数個の単語又は単語列毎に、文脈情報を含む特徴量を検出し、各単語又は各単語列に前記特徴量を付加して、前記学習データ候補を生成する、請求項1から7のいずれかに記載の学習データ生成装置。 - 前記アノテーション情報を作成するアノテーション情報生成部と、
複数の固有表現と、前記複数の固有表現それぞれの前記クラスとが登録されている辞書とを更に備え、
前記アノテーション情報生成部は、前記辞書に登録されている前記複数の固有表現と前記テキストデータとを比較し、前記テキストデータ中の前記複数の固有表現のいずれかが出現している出現箇所を特定し、前記出現箇所の位置と前記出現している固有表現の前記クラスとを特定するアノテーション情報を作成する、請求項8に記載の学習データ生成装置。 - 抽出ルール学習部と、辞書増殖部とを更に備え、
前記抽出ルール学習部は、前記学習データ生成部により生成された前記学習データを用いて、前記固有表現を抽出するための固有表現抽出ルールを学習し、
前記辞書増殖部は、複数の文書を含むコーパスから、前記抽出ルール学習部によって学習された抽出ルールを用いて固有表現を抽出し、そして、抽出された前記固有表現のうちの設定された条件を満たす固有表現と、そのクラスを特定するクラス情報とを、前記辞書に追加登録する、請求項9に記載の学習データ生成装置。 - 上記請求項1から上記請求項10のいずれかに記載の学習データ生成装置と、固有表現抽出装置とを備え、
前記固有表現抽出装置は、前記学習データ装置により生成された前記学習データを用いて、固有表現を抽出するための固有表現抽出ルールを学習し、学習した前記固有表現抽出ルールを用いて、複数の文書を含むコーパスから、前記固有表現を抽出する、
ことを特徴とする固有表現抽出システム。 - 固有表現抽出ルールの作成用の学習データを生成するための学習データ生成方法であって、
(a)アノテーションのクラスを示すラベルが付与された複数の学習データ候補を、それぞれの文脈情報を含む特徴量に基づいて、クラスタリングする、ステップと、
(b)前記ステップ(a)のクラスタリング結果で得られた各クラスタを参照して、前記各クラスタ内の学習データ候補のラベルの分布を求め、求めた前記分布に基づいて、設定された条件を満たす学習データ候補を特定し、特定した前記学習データ候補を用いて学習データを生成する、ステップと、
を有することを特徴とする学習データ生成方法。 - 前記(b)のステップにおいて、特定のラベルを予め定められた割合以上含む前記クラスタ内における、前記特定のラベルが付与された学習データ候補を、前記特定のラベルの前記学習データとする、請求項12に記載の学習データ生成方法。
- 前記(b)のステップにおいて、特定のラベルを予め定められた割合以上含む前記クラスタ内における、前記特定のラベルが付与されていない学習データ候補を、前記学習データ候補全体から削除する、請求項13に記載の学習データ生成方法。
- 前記(b)のステップにおいて、特定のラベルを予め定められた割合以上含む前記クラスタ内における、前記特定のラベルが付与されていない学習データ候補について、当該クラスタに所属する度合いを求め、求められた度合いが閾値以上となる前記学習データ候補を前記学習データとし、求められた度合いが閾値未満となる前記学習データ候補を前記学習データ候補全体から削除する、請求項13に記載の学習データ生成方法。
- 前記(b)のステップにおいて、前記特定のラベルが付与されていない学習データ候補が当該クラスタに所属する度合いとして、前記学習データ候補の当該クラスタの代表点からの距離、前記学習データ候補の当該クラスタへの所属確率、前記学習データ候補の当該クラスタ内での推定密度、及び当該クラスタが前記特定のラベルを含む割合のうち、1つ以上を用いる、請求項15に記載の学習データ生成方法。
- 前記(a)のステップにおいて、前記クラスタリング時に、前記各クラスタ内のラベルの分布がクラスタ間で異なる事、前記各クラスタ内で最も割合が多いラベルが存在する事、各クラスタ内の学習データ候補数が特定の数より多い事、及びクラスタ数が特定の数より少ない事のうちいずれか1つ以上が満たされるように、前記クラスタリングの条件を設定する、請求項12から16のいずれかに記載の学習データ生成方法。
- 前記(b)のステップにおいて、前記学習データとされた前記学習データ候補に、新たな特徴量を付加する、請求項12から17のいずれかに記載の学習データ生成方法。
- (c)テキストデータと、テキストデータに含まれる各単語に付与されたアノテーション情報とを用いて、前記テキストデータを分割して得られた複数個の単語又は単語列毎に、文脈情報を含む特徴量を検出し、各単語又は各単語列に前記特徴量を付加して、前記学習データ候補を生成するステップを、更に有する、請求項12から18のいずれかに記載の学習データ生成方法。
- (d)複数の固有表現と前記複数の固有表現それぞれの前記クラスとが登録されている辞書を用い、前記辞書に登録されている前記複数の固有表現と前記テキストデータとを比較し、前記テキストデータ中の前記複数の固有表現のいずれかが出現している出現箇所を特定し、前記出現箇所の位置と前記出現している固有表現の前記クラスとを特定するアノテーション情報を作成するステップを、更に有する、請求項19に記載の学習データ生成方法。
- (e)前記(b)のステップで生成された前記学習データを用いて、前記固有表現を抽出するための固有表現抽出ルールを学習するステップと、
(d)複数の文書を含むコーパスから、前記(e)のステップで学習された抽出ルールを用いて固有表現を抽出し、そして、抽出された前記固有表現のうちの設定された条件を満たす固有表現と、そのクラスを特定するクラス情報とを、前記辞書に追加登録するステップとを、更に有する、請求項20に記載の学習データ生成方法。 - コンピュータによって、固有表現抽出ルールの作成用の学習データを生成させるための、プログラムを記録した、コンピュータ読み取り可能な記録媒体であって、
前記コンピュータに、
(a)アノテーションのクラスを示すラベルが付与された複数の学習データ候補を、それぞれの文脈情報を含む特徴量に基づいて、クラスタリングする、ステップと、
(b)前記ステップ(a)のクラスタリング結果で得られた各クラスタを参照して、前記各クラスタ内の学習データ候補のラベルの分布を求め、求めた前記分布に基づいて、設定された条件を満たす学習データ候補を特定し、特定した前記学習データ候補を用いて学習データを生成する、ステップと、
を実行させる、命令を含むプログラムを記録していることを特徴とするコンピュータ読み取り可能な記録媒体。 - 前記(b)のステップにおいて、特定のラベルを予め定められた割合以上含む前記クラスタ内における、前記特定のラベルが付与された学習データ候補を、前記特定のラベルの前記学習データとする、請求項22に記載のコンピュータ読み取り可能な記録媒体。
- 前記(b)のステップにおいて、特定のラベルを予め定められた割合以上含む前記クラスタ内における、前記特定のラベルが付与されていない学習データ候補を、前記学習データ候補全体から削除する、請求項23に記載のコンピュータ読み取り可能な記録媒体。
- 前記(b)のステップにおいて、特定のラベルを予め定められた割合以上含む前記クラスタ内における、前記特定のラベルが付与されていない学習データ候補について、当該クラスタに所属する度合いを求め、求められた度合いが閾値以上となる前記学習データ候補を前記学習データとし、求められた度合いが閾値未満となる前記学習データ候補を前記学習データ候補全体から削除する、請求項23に記載のコンピュータ読み取り可能な記録媒体。
- 前記(b)のステップにおいて、前記特定のラベルが付与されていない学習データ候補が当該クラスタに所属する度合いとして、前記学習データ候補の当該クラスタの代表点からの距離、前記学習データ候補の当該クラスタへの所属確率、前記学習データ候補の当該クラスタ内での推定密度、及び当該クラスタが前記特定のラベルを含む割合のうち、1つ以上を用いる、請求項25に記載のコンピュータ読み取り可能な記録媒体。
- 前記(a)のステップにおいて、前記クラスタリング時に、前記各クラスタ内のラベルの分布がクラスタ間で異なる事、前記各クラスタ内で最も割合が多いラベルが存在する事、各クラスタ内の学習データ候補数が特定の数より多い事、及びクラスタ数が特定の数より少ない事のうちいずれか1つ以上が満たされるように、前記クラスタリングの条件を設定する、請求項22から26のいずれかに記載のコンピュータ読み取り可能な記録媒体。
- 前記(b)のステップにおいて、前記学習データとされた前記学習データ候補に、新たな特徴量を付加する、請求項22から27のいずれかに記載のコンピュータ読み取り可能な記録媒体。
- 前記プログラムが、前記コンピュータに、
(c)テキストデータと、テキストデータに含まれる各単語に付与されたアノテーション情報とを用いて、前記テキストデータを分割して得られた複数個の単語又は単語列毎に、文脈情報を含む特徴量を検出し、各単語又は各単語列に前記特徴量を付加して、前記学習データ候補を生成するステップを実行させる、命令を更に含む請求項22から28のいずれかに記載のコンピュータ読み取り可能な記録媒体。 - 前記プログラムが、前記コンピュータに、
(d)複数の固有表現と前記複数の固有表現それぞれの前記クラスとが登録されている辞書を用い、前記辞書に登録されている前記複数の固有表現と前記テキストデータとを比較し、前記テキストデータ中の前記複数の固有表現のいずれかが出現している出現箇所を特定し、前記出現箇所の位置と前記出現している固有表現の前記クラスとを特定するアノテーション情報を作成するステップを実行させる、命令を更に含む請求項29に記載のコンピュータ読み取り可能な記録媒体。 - 前記プログラムが、前記コンピュータに、
(e)前記(b)のステップで生成された前記学習データを用いて、前記固有表現を抽出するための固有表現抽出ルールを学習するステップと、
(d)複数の文書を含むコーパスから、前記(e)のステップで学習された抽出ルールを用いて固有表現を抽出し、そして、抽出された前記固有表現のうちの設定された条件を満たす固有表現と、そのクラスを特定するクラス情報とを、前記辞書に追加登録するステップとを、実行させる、命令を更に含む請求項30に記載のコンピュータ読み取り可能な記録媒体。
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2011509187A JP5424001B2 (ja) | 2009-04-15 | 2010-03-17 | 学習データ生成装置、固有表現抽出システム、学習データ生成方法、及びプログラム |
US13/263,280 US9195646B2 (en) | 2009-04-15 | 2010-03-17 | Training data generation apparatus, characteristic expression extraction system, training data generation method, and computer-readable storage medium |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2009099183 | 2009-04-15 | ||
JP2009-099183 | 2009-04-15 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2010119615A1 true WO2010119615A1 (ja) | 2010-10-21 |
Family
ID=42982290
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2010/001901 WO2010119615A1 (ja) | 2009-04-15 | 2010-03-17 | 学習データ生成装置、及び固有表現抽出システム |
Country Status (3)
Country | Link |
---|---|
US (1) | US9195646B2 (ja) |
JP (1) | JP5424001B2 (ja) |
WO (1) | WO2010119615A1 (ja) |
Cited By (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2015129988A (ja) * | 2014-01-06 | 2015-07-16 | 日本電気株式会社 | データ処理装置 |
JP2017058816A (ja) * | 2015-09-15 | 2017-03-23 | 株式会社東芝 | 情報抽出装置、方法およびプログラム |
JP2017151707A (ja) * | 2016-02-24 | 2017-08-31 | ヤフー株式会社 | 判定装置、判定方法、及び判定プログラム |
JP2017167854A (ja) * | 2016-03-16 | 2017-09-21 | 株式会社東芝 | 学習装置、方法、及びプログラム |
US9858266B2 (en) | 2014-03-27 | 2018-01-02 | International Business Machines Corporation | Calculating correlations between annotations |
JP2018013857A (ja) * | 2016-07-19 | 2018-01-25 | 富士通株式会社 | センサデータ学習方法、センサデータ学習プログラム、及びセンサデータ学習装置 |
US9898464B2 (en) | 2014-11-19 | 2018-02-20 | Kabushiki Kaisha Toshiba | Information extraction supporting apparatus and method |
JP2018081569A (ja) * | 2016-11-17 | 2018-05-24 | 株式会社東芝 | 情報抽出支援装置、情報抽出支援方法およびプログラム |
JP2019148847A (ja) * | 2018-02-26 | 2019-09-05 | 株式会社Kddi総合研究所 | 学習データ作成装置並びに分類モデル学習装置及びカテゴリ付与装置 |
US10444742B2 (en) | 2016-02-09 | 2019-10-15 | Kabushiki Kaisha Toshiba | Material recommendation apparatus |
JP2020042797A (ja) * | 2018-09-07 | 2020-03-19 | バイドゥ オンライン ネットワーク テクノロジー (ベイジン) カンパニー リミテッド | サンプル処理方法、装置、機器および記憶媒体 |
JP2020144755A (ja) * | 2019-03-08 | 2020-09-10 | 日立オートモティブシステムズ株式会社 | 演算装置 |
JP2020190935A (ja) * | 2019-05-22 | 2020-11-26 | 富士通株式会社 | 機械学習プログラム、機械学習方法および機械学習装置 |
JP2021015533A (ja) * | 2019-07-16 | 2021-02-12 | 株式会社ジャパン・カレント | 画像分類システム及び画像分類方法 |
US10936806B2 (en) | 2015-11-04 | 2021-03-02 | Kabushiki Kaisha Toshiba | Document processing apparatus, method, and program |
WO2021098397A1 (zh) * | 2019-11-21 | 2021-05-27 | 腾讯科技(深圳)有限公司 | 数据处理方法、设备及存储介质 |
WO2022201256A1 (ja) * | 2021-03-22 | 2022-09-29 | 日本電気株式会社 | 情報処理装置、情報処理システム、情報処理方法、及びプログラム |
JP7456289B2 (ja) | 2020-05-28 | 2024-03-27 | 富士通株式会社 | 判定プログラム、判定方法および情報処理装置 |
Families Citing this family (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP5751251B2 (ja) * | 2010-03-26 | 2015-07-22 | 日本電気株式会社 | 意味抽出装置、意味抽出方法、および、プログラム |
JP5640773B2 (ja) * | 2011-01-28 | 2014-12-17 | 富士通株式会社 | 情報照合装置、情報照合方法および情報照合プログラム |
WO2014189521A1 (en) * | 2013-05-24 | 2014-11-27 | Nokia Siemens Networks Oy | Automatic correlation and prevention of improper user plane traffic wireless networks |
US9230009B2 (en) * | 2013-06-04 | 2016-01-05 | International Business Machines Corporation | Routing of questions to appropriately trained question and answer system pipelines using clustering |
US9146987B2 (en) * | 2013-06-04 | 2015-09-29 | International Business Machines Corporation | Clustering based question set generation for training and testing of a question and answer system |
US9348815B1 (en) | 2013-06-28 | 2016-05-24 | Digital Reasoning Systems, Inc. | Systems and methods for construction, maintenance, and improvement of knowledge representations |
US10191893B2 (en) | 2013-07-22 | 2019-01-29 | Open Text Holdings, Inc. | Information extraction and annotation systems and methods for documents |
US8856642B1 (en) * | 2013-07-22 | 2014-10-07 | Recommind, Inc. | Information extraction and annotation systems and methods for documents |
US9348900B2 (en) | 2013-12-11 | 2016-05-24 | International Business Machines Corporation | Generating an answer from multiple pipelines using clustering |
US10885013B2 (en) * | 2014-06-20 | 2021-01-05 | Jpmorgan Chase Bank, N.A. | Automated application lifecycle tracking using batch processing |
US9860200B1 (en) * | 2014-08-27 | 2018-01-02 | Google Llc | Message suggestions |
JP6544131B2 (ja) * | 2015-08-14 | 2019-07-17 | 富士ゼロックス株式会社 | 情報処理装置及び情報処理プログラム |
US10530721B2 (en) * | 2015-11-24 | 2020-01-07 | International Business Machines Corporation | Customized timing for access to shared media files |
US11544579B2 (en) * | 2016-11-23 | 2023-01-03 | Primal Fusion Inc. | System and method for generating training data for machine learning classifier |
US11205103B2 (en) | 2016-12-09 | 2021-12-21 | The Research Foundation for the State University | Semisupervised autoencoder for sentiment analysis |
US20180225590A1 (en) * | 2017-02-07 | 2018-08-09 | International Business Machines Corporation | Automatic ground truth seeder |
US11048762B2 (en) | 2018-03-16 | 2021-06-29 | Open Text Holdings, Inc. | User-defined automated document feature modeling, extraction and optimization |
US10762142B2 (en) | 2018-03-16 | 2020-09-01 | Open Text Holdings, Inc. | User-defined automated document feature extraction and optimization |
KR102129843B1 (ko) * | 2018-12-17 | 2020-07-03 | 주식회사 크라우드웍스 | 검증용 주석 처리 작업을 이용한 실전용 주석 처리 작업의 검증 방법 및 장치 |
US11610277B2 (en) | 2019-01-25 | 2023-03-21 | Open Text Holdings, Inc. | Seamless electronic discovery system with an enterprise data portal |
US11514691B2 (en) * | 2019-06-12 | 2022-11-29 | International Business Machines Corporation | Generating training sets to train machine learning models |
US11270061B2 (en) * | 2020-02-25 | 2022-03-08 | International Business Machines Corporation | Automatic generation of training data for scientific paper summarization using videos |
US11328117B2 (en) | 2020-05-17 | 2022-05-10 | International Business Machines Corporation | Automated content modification based on a user-specified context |
US11886820B2 (en) * | 2020-10-06 | 2024-01-30 | Genpact Luxembourg S.à r.l. II | System and method for machine-learning based extraction of information from documents |
US11664010B2 (en) | 2020-11-03 | 2023-05-30 | Florida Power & Light Company | Natural language domain corpus data set creation based on enhanced root utterances |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2004005648A (ja) * | 2002-05-10 | 2004-01-08 | Microsoft Corp | 自然言語理解システムに関するトレーニングデータの自動注釈付けのための方法およびユーザインターフェース |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7366705B2 (en) * | 2004-04-15 | 2008-04-29 | Microsoft Corporation | Clustering based text classification |
JP4347226B2 (ja) | 2004-05-20 | 2009-10-21 | 富士通株式会社 | 情報抽出プログラムおよびその記録媒体、情報抽出装置ならびに情報抽出規則作成方法 |
-
2010
- 2010-03-17 JP JP2011509187A patent/JP5424001B2/ja active Active
- 2010-03-17 US US13/263,280 patent/US9195646B2/en active Active
- 2010-03-17 WO PCT/JP2010/001901 patent/WO2010119615A1/ja active Application Filing
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2004005648A (ja) * | 2002-05-10 | 2004-01-08 | Microsoft Corp | 自然言語理解システムに関するトレーニングデータの自動注釈付けのための方法およびユーザインターフェース |
Non-Patent Citations (1)
Title |
---|
TETSUJI NAKAGAWA ET AL.: "Shusei Gakushuho ni yoru Keitai Sokaiseki", INFORMATION PROCESSING SOCIETY OF JAPAN KENKYU HOKOKU, 20 November 2001 (2001-11-20) * |
Cited By (24)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2015129988A (ja) * | 2014-01-06 | 2015-07-16 | 日本電気株式会社 | データ処理装置 |
US9858266B2 (en) | 2014-03-27 | 2018-01-02 | International Business Machines Corporation | Calculating correlations between annotations |
US9858267B2 (en) | 2014-03-27 | 2018-01-02 | International Business Machines Corporation | Calculating correlations between annotations |
US9898464B2 (en) | 2014-11-19 | 2018-02-20 | Kabushiki Kaisha Toshiba | Information extraction supporting apparatus and method |
JP2017058816A (ja) * | 2015-09-15 | 2017-03-23 | 株式会社東芝 | 情報抽出装置、方法およびプログラム |
US10936806B2 (en) | 2015-11-04 | 2021-03-02 | Kabushiki Kaisha Toshiba | Document processing apparatus, method, and program |
US10444742B2 (en) | 2016-02-09 | 2019-10-15 | Kabushiki Kaisha Toshiba | Material recommendation apparatus |
JP2017151707A (ja) * | 2016-02-24 | 2017-08-31 | ヤフー株式会社 | 判定装置、判定方法、及び判定プログラム |
JP2017167854A (ja) * | 2016-03-16 | 2017-09-21 | 株式会社東芝 | 学習装置、方法、及びプログラム |
US11037062B2 (en) | 2016-03-16 | 2021-06-15 | Kabushiki Kaisha Toshiba | Learning apparatus, learning method, and learning program |
JP2018013857A (ja) * | 2016-07-19 | 2018-01-25 | 富士通株式会社 | センサデータ学習方法、センサデータ学習プログラム、及びセンサデータ学習装置 |
US11481663B2 (en) | 2016-11-17 | 2022-10-25 | Kabushiki Kaisha Toshiba | Information extraction support device, information extraction support method and computer program product |
JP2018081569A (ja) * | 2016-11-17 | 2018-05-24 | 株式会社東芝 | 情報抽出支援装置、情報抽出支援方法およびプログラム |
JP2019148847A (ja) * | 2018-02-26 | 2019-09-05 | 株式会社Kddi総合研究所 | 学習データ作成装置並びに分類モデル学習装置及びカテゴリ付与装置 |
JP2020042797A (ja) * | 2018-09-07 | 2020-03-19 | バイドゥ オンライン ネットワーク テクノロジー (ベイジン) カンパニー リミテッド | サンプル処理方法、装置、機器および記憶媒体 |
JP2020144755A (ja) * | 2019-03-08 | 2020-09-10 | 日立オートモティブシステムズ株式会社 | 演算装置 |
WO2020183776A1 (ja) * | 2019-03-08 | 2020-09-17 | 日立オートモティブシステムズ株式会社 | 演算装置 |
JP7079745B2 (ja) | 2019-03-08 | 2022-06-02 | 日立Astemo株式会社 | 演算装置 |
JP2020190935A (ja) * | 2019-05-22 | 2020-11-26 | 富士通株式会社 | 機械学習プログラム、機械学習方法および機械学習装置 |
JP7197795B2 (ja) | 2019-05-22 | 2022-12-28 | 富士通株式会社 | 機械学習プログラム、機械学習方法および機械学習装置 |
JP2021015533A (ja) * | 2019-07-16 | 2021-02-12 | 株式会社ジャパン・カレント | 画像分類システム及び画像分類方法 |
WO2021098397A1 (zh) * | 2019-11-21 | 2021-05-27 | 腾讯科技(深圳)有限公司 | 数据处理方法、设备及存储介质 |
JP7456289B2 (ja) | 2020-05-28 | 2024-03-27 | 富士通株式会社 | 判定プログラム、判定方法および情報処理装置 |
WO2022201256A1 (ja) * | 2021-03-22 | 2022-09-29 | 日本電気株式会社 | 情報処理装置、情報処理システム、情報処理方法、及びプログラム |
Also Published As
Publication number | Publication date |
---|---|
US9195646B2 (en) | 2015-11-24 |
US20120030157A1 (en) | 2012-02-02 |
JPWO2010119615A1 (ja) | 2012-10-22 |
JP5424001B2 (ja) | 2014-02-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP5424001B2 (ja) | 学習データ生成装置、固有表現抽出システム、学習データ生成方法、及びプログラム | |
CN109933785B (zh) | 用于实体关联的方法、装置、设备和介质 | |
US20240028837A1 (en) | Device and method for machine reading comprehension question and answer | |
Cohen et al. | Exploiting dictionaries in named entity extraction: combining semi-markov extraction processes and data integration methods | |
EP2812883B1 (en) | System and method for semantically annotating images | |
CN117076653B (zh) | 基于思维链及可视化提升上下文学习知识库问答方法 | |
JP2020126493A (ja) | 対訳処理方法および対訳処理プログラム | |
US20180075017A1 (en) | Computer-readable recording medium, retrieval device, and retrieval method | |
Kou et al. | High-recall protein entity recognition using a dictionary | |
US20230342400A1 (en) | Document search device, document search program, and document search method | |
Hay et al. | Representation learning of writing style | |
Zong et al. | Information extraction | |
JP2024091709A (ja) | 文作成装置、文作成方法および文作成プログラム | |
Schaback et al. | Multi-level feature extraction for spelling correction | |
CN111738009B (zh) | 实体词标签生成方法、装置、计算机设备和可读存储介质 | |
CN111133429A (zh) | 提取表达以供自然语言处理 | |
Gero et al. | Word centrality constrained representation for keyphrase extraction | |
Gao et al. | Active entity recognition in low resource settings | |
Pandi et al. | A novel similarity measure for sequence data | |
CN113449063B (zh) | 一种构建文档结构信息检索库的方法及装置 | |
Banisakher et al. | Improving the identification of the discourse function of news article paragraphs | |
CN113553410B (zh) | 长文档处理方法、处理装置、电子设备和存储介质 | |
JP2022002034A (ja) | 抽出方法、抽出プログラム、及び、抽出装置 | |
Zhu et al. | Semantic document distance measures and unsupervised document revision detection | |
CN111552780A (zh) | 医用场景的搜索处理方法、装置、存储介质及电子设备 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 10764202 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2011509187 Country of ref document: JP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 13263280 Country of ref document: US |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 10764202 Country of ref document: EP Kind code of ref document: A1 |