CN107273500A - Text classifier generation method, file classification method, device and computer equipment - Google Patents

Abstract

The invention discloses a text classifier generation method, text classification method, device and computer equipment, which are used to solve the problem of poor classification effect caused by crossover of samples in a training sample set. The method for generating a text classifier includes: performing a merging operation on at least two original sample categories with sample intersections in the training sample set to obtain a new sample category; the training sample set includes a plurality of original sample categories, and each training sample belongs to the One of a plurality of original sample categories; training according to the training sample set after the merging operation to obtain a first classifier; according to the training samples belonging to the new sample category and the original samples to which the training samples belonging to the new sample category belong The category is trained to obtain a second classifier, and the second classifier is used to reclassify the text to be classified whose classification category is the new sample category in the classification result of the first classifier, so as to be divided into corresponding original samples category.

Description

Text classifier generating method, text classifying method, device and computer equipment

technical field

The invention relates to the field of communication technology, in particular to a method for generating a text classifier, a text classification method, a device and computer equipment.

Background technique

In text classification, the quality of training samples largely determines the performance of classifiers.

For example, when inter-class sample crossover occurs in the training samples, there are two or more classes with sample crossover, which will definitely affect the effect of the overall classifier, and the classification accuracy of the two or more classes is low, and the classification effect Difference.

Contents of the invention

The technical problem to be solved by the present invention is to provide a text classifier generation method, text classification method, device and computer equipment to solve the problem of poor classification effect caused by the crossing of samples in the training sample set in the prior art.

On the one hand, the present invention provides a method for generating a text classifier, comprising: performing a merging operation on at least two original sample categories with sample intersections in the training sample set to obtain a new sample category; the training sample set includes a plurality of original sample categories, Each training sample belongs to one of the plurality of original sample categories; the first classifier is obtained according to the training sample set after the merging operation; according to the training samples belonging to the new sample category and the belonging to the new sample category The original sample category to which the training sample belongs is trained to obtain a second classifier, and the second classifier is used to reclassify the text to be classified whose classification category is the new sample category in the classification result of the first classifier, to be classified into the corresponding original sample category.

Optionally, the training to obtain the first classifier according to the training sample set after the merging operation includes: using at least one of the following classification algorithms to train the training sample set after the merging operation to obtain the first classifier: naive Bayesian NB classification algorithm, support vector machine SVM classification algorithm, K nearest neighbor KNN classification algorithm and random forest classification algorithm.

Optionally, performing training according to the training samples belonging to the new sample category and the original sample category to which the training samples belonging to the new sample category belong, to obtain the second classifier includes: according to the training samples belonging to the new sample category and the original sample category to which the training samples belonging to the new sample category belong, to obtain a second classifier for each of the new sample categories respectively.

Optionally, at least one of the following classification algorithms is used to train the training samples belonging to the new sample category and the original sample category to which the training samples belonging to the new sample category belong, to obtain the second classifier: Naive Bayesian NB classification algorithm, support vector machine SVM classification algorithm, K nearest neighbor KNN classification algorithm and random forest classification algorithm; the classification category included in the second classifier is the same as the original class to which the training samples in the new sample category belong corresponding to the sample category.

Optionally, before the merging of at least two original sample categories with sample intersections in the training sample set to obtain a new sample category, the method further includes: preprocessing the training corpus to filter the training corpus and /or the format is unified; the training sample set is obtained by performing word segmentation on the preprocessed training corpus according to the word segmentation dictionary.

Optionally, the training corpus includes sentences and/or text fragments.

Optionally, the method further includes: performing a new word discovery operation on the training corpus and adding the discovered new words to the word segmentation dictionary.

Optionally, the new word discovery operation is realized by at least one of the following methods: mutual information, co-occurrence probability and information entropy.

Optionally, the method further includes: testing the classification accuracy rate of each classification category in the first classifier; testing the classification accuracy rate of each classification category in the second classifier; wherein, the first The classification accuracy rates of the classifiers are respectively P1j, wherein j is an integer greater than or equal to 1 and less than or equal to m, and m is the number of sample categories in the training sample set after the merging operation; the classification accuracy rates of the second classifier are respectively P1h*P2k, wherein k is an integer greater than or equal to 1 and less than or equal to n, n is the number of original sample categories to which the training samples in the new sample category belong; P1h is the number of the new sample category in the first classifier Classification accuracy, h is an integer greater than or equal to 1 and less than or equal to g, g is the number of new sample categories; detecting whether the classification accuracy of each classification category in the first classifier is greater than the first probability threshold, and Whether the classification accuracy rate of each classification category in the second classifier is greater than a second probability threshold; if yes, it is determined that the training of the first classifier and the second classifier is successful.

Optionally, it is detected whether the classification accuracy rate of each classification category in the first classifier is greater than the first probability threshold, and whether the classification accuracy rate of each classification category in the second classifier is greater than the second probability Threshold, then also includes: if no, at least two original sample categories that have sample intersections in the training sample set are re-merged and classifier training until the classification accuracy of each classification category in the first classifier are greater than the first probability threshold, and the classification accuracy of each classification category in the second classifier is greater than the second probability threshold.

Optionally, the method further includes: adjusting the first probability threshold and the second probability threshold to filter out the first classifier and the second classifier with different classification accuracies.

Optionally, the classification accuracy of the first classifier is tested by cross-validation based on the overall sample; the classification accuracy of the second classifier is tested by cross-validation based on the overall sample.

Optionally, the cross-validation method based on the sample population includes: using 60% to 90% of the samples in the sample population as the training sample set, and using the remaining samples as the text to be classified for testing, and the sample population includes all said plurality of original sample classes, each of said samples belonging to one of said plurality of original sample classes.

Optionally, the merging at least two original sample categories with sample intersections into a new sample category includes: merging all original sample categories with sample intersections into a new sample category.

Optionally, before the merging operation of at least two original sample categories with crossed samples in the training sample set to obtain a new sample category, the method further includes: training according to the training sample set to obtain a classification category of all The third classifier of the plurality of original sample categories, testing the classification accuracy rate of each classification category of the third classifier, and initially screening out the original sample categories whose classification accuracy rate is lower than the third threshold; Identify the original sample category for which there is a sample intersection among the above original sample categories.

On the other hand, the present invention also provides a text classification method, using the classifier generated by the text classifier generation method provided by the present invention to classify, the classification method includes: inputting the text set to be classified into the first classifier, Obtaining a first classification result; inputting the text to be classified in the first classification result as the new sample class to a second classifier corresponding to the new sample class to obtain a second classification result.

Optionally, the method further includes: taking the classification result of the unmerged original sample category in the first classification result together with the second classification result as the final classification of the text to be classified result.

On the other hand, the present invention also provides a device for generating a text classifier, including: a merging unit for merging at least two original sample categories with sample intersections in the training sample set to obtain a new sample category; the training sample set Including a plurality of original sample categories, each training sample belongs to one of the plurality of original sample categories; the first training unit is used to train the first classifier according to the training sample set after the merging operation; the second training unit, It is used to obtain a second classifier by training according to the training samples belonging to the new sample category and the original sample category to which the training samples belonging to the new sample category belong, and the second classifier is used to classify the first classifier The text to be classified whose classification category is the new sample category in the classification result of the detector is classified again, so as to be classified into the corresponding original sample category.

Optionally, the first training unit is specifically configured to: use at least one of the following classification algorithms to train the training sample set after the merging operation to obtain the first classifier: Naive Bayesian NB classification algorithm, Support vector machine SVM classification algorithm, K nearest neighbor KNN classification algorithm and random forest classification algorithm.

Optionally, the second training unit is specifically configured to: perform training according to the training samples belonging to the new sample category and the original sample category to which the training samples belonging to the new sample category belong, and obtain each Describe the second classifier for the new sample category.

Optionally, the second classification unit is specifically configured to adopt at least one of the following classification algorithms for training to obtain the second classifier: Naive Bayesian NB classification algorithm, support vector machine SVM classification algorithm, K nearest neighbor KNN classification algorithm and a random forest classification algorithm; the classification category included in the second classifier corresponds to the original sample category to which the training samples in the new sample category belong.

Optionally, the device further includes: a preprocessing unit, configured to perform preprocessing on the training corpus before combining at least two original sample categories with sample intersections in the training sample set to obtain a new sample category. The training corpus is filtered and/or the format is unified; the word segmentation unit is used to perform word segmentation on the preprocessed training corpus according to the word segmentation dictionary to obtain the training sample set.

Optionally, the training corpus includes sentences or text fragments.

Optionally, the device further includes a new word discovery unit, configured to perform a new word discovery operation on the training corpus and add the discovered new words to the word segmentation dictionary.

Optionally, the new word discovery operation is realized by at least one of the following methods: mutual information, co-occurrence probability and information entropy.

Optionally, the device further includes: a first testing unit, used to test the classification accuracy of each classification category in the first classifier; a second testing unit, used to test the classification accuracy of each classification category in the second classifier. The classification accuracy rate of each classification category; wherein, the classification accuracy rate of the first classifier is P1j respectively, wherein j is an integer greater than or equal to 1 and less than or equal to m, and m is the sample category in the training sample set after the merging operation number; the classification accuracy rates of the second classifier are respectively P1h*P2k, wherein k is an integer greater than or equal to 1 and less than or equal to n, and n is the number of original sample categories to which the training samples in the new sample category belong; P1h is the classification accuracy rate of the new sample category in the first classifier, h is an integer greater than or equal to 1 and less than or equal to g, and g is the number of the new sample category; a detection unit is used to detect the first classification Whether the classification accuracy rate of each classification category in the classifier is greater than the first probability threshold, and whether the classification accuracy rate of each classification category in the second classifier is greater than the second probability threshold; the determination unit is used to determine if the The detection result of the detection unit is yes, and it is determined that the training of the first classifier and the second classifier is successful.

Optionally, the device further includes: a return unit, configured to re-merge at least two original sample categories with sample crossover in the training sample set and classifier training if the detection result of the detection unit is negative , until the classification accuracy rate of each classification category in the first classifier is greater than the first probability threshold, and the classification accuracy rate of each classification category in the second classifier is greater than the second probability threshold.

Optionally, the device further includes: an adjustment unit, configured to adjust the first probability threshold and the second probability threshold to select the first classifier and the second classifier with different classification accuracies.

Optionally, the first test unit is specifically used to test the classification accuracy of the first classifier by using a cross-validation method based on the overall sample; the second test unit is specifically used to use a cross-validation method based on the overall sample Testing the classification accuracy of the second classifier.

Optionally, the cross-validation method based on the sample population includes: using 60% to 90% of the samples in the sample population as the training sample set, and using the remaining samples as the text to be classified for testing, and the sample population includes all said plurality of original sample classes, each of said samples belonging to one of said plurality of original sample classes.

Optionally, the merging unit is specifically configured to merge all original sample categories with sample intersections into a new sample category.

Optionally, the device further includes a screening unit, configured to obtain a classification category according to training of the training sample set before performing a merging operation on at least two original sample categories with crossed samples in the training sample set to obtain a new sample category For the third classifier of the plurality of original sample categories, test the classification accuracy of each classification category of the third classifier, and initially screen out the original sample categories whose classification accuracy is lower than the third threshold; Among the original sample classes of , an original sample class with sample intersection is identified.

On the other hand, the present invention also provides a text classification device, which uses any classifier generated by the text classifier generation device provided by the present invention to classify, and the classification device includes: a first input unit for classifying The text set is input into the first classifier to obtain the first classification result; the second input unit is used to input the text to be classified into the new sample category in the first classification result and the new sample category The corresponding second classifier obtains the second classification result.

Optionally, the device further includes: a result output unit, configured to use the classification result of the original sample category in which the classification category is not combined in the first classification result and the second classification result as the The final classification result of the text to be classified.

On the other hand, the present invention also provides a computer device, including a processor and a memory; the memory is used to store computer instructions, and the processor is used to run the computer instructions stored in the memory to execute any text classifier provided by the present invention generate method.

On the other hand, the present invention also provides a computer device, including a processor and a memory; the memory is used to store computer instructions, and the processor is used to run the computer instructions stored in the memory to perform any text classification method provided by the present invention .

On the other hand, the present invention also provides a computer-readable storage medium, and instructions are stored in the storage medium, and when the instructions are executed, any method for generating a text classifier provided by the present invention is executed.

On the other hand, the present invention also provides a computer-readable storage medium, where instructions are stored in the storage medium, and any text classification method provided by the present invention is executed when the instructions are run.

The text classifier generation method, text classification method, device, and computer equipment provided by the embodiments of the present invention can accurately classify the original sample categories with sample intersections and the original sample categories without sample intersections through the training of the first classifier Differentiate, through the training of the second classifier, the original sample category with sample crossover can be separated separately, and more detailed classification training can be carried out in a more specific range, thus greatly improving the classification accuracy of the text classifier.

Description of drawings

Fig. 1 is a kind of flowchart of the text classifier generation method that the embodiment of the present invention provides;

Fig. 2 is a kind of detailed flowchart of the text classifier generating method provided by the embodiment of the present invention;

Fig. 3 is a kind of flowchart of the text classification method provided by the embodiment of the present invention;

FIG. 4 is a schematic structural diagram of a text classifier generation device provided by an embodiment of the present invention;

Fig. 5 is a schematic structural diagram of a text classification device provided by an embodiment of the present invention.

detailed description

The present invention will be described in detail below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

As shown in Figure 1, an embodiment of the present invention provides a method for generating a text classifier, including:

S11, performing a merging operation on at least two original sample categories with sample intersections in the training sample set to obtain a new sample category; the training sample set includes multiple original sample categories, and each training sample belongs to one of the multiple original sample categories one;

S12, training according to the training sample set after the merging operation to obtain the first classifier;

S13. Perform training according to the training samples belonging to the new sample category and the original sample category to which the training samples belonging to the new sample category belong to obtain a second classifier, and the second classifier is used to classify the first classifier The text to be classified whose classification category is the new sample category in the classification result of the detector is classified again, so as to be classified into the corresponding original sample category.

The text classifier generation method provided by the embodiment of the present invention adopts a hierarchical classifier generation method when training the text classifier according to the training samples, and first merges two or more original sample categories with sample intersections in the training sample set into a new sample category, and obtain the first classifier according to the training sample set after the merging operation, and then conduct more detailed classification training on the new sample category to obtain the second classifier. In this way, through the training of the first classifier, the original sample category with sample crossover can be accurately distinguished from the original sample category without sample crossover, and through the training of the second classifier, the original sample category with sample crossover can be distinguished It is separated separately, and more detailed classification training is carried out in a more specific range, thus greatly improving the classification accuracy of the text classifier.

Specifically, the sample crossover mentioned in the embodiments of the present invention refers to that in the provided training sample set, the classification of the sample data is not very clear and accurate, for example, there are sample data that should originally belong to the A category but are classified in the B category. In the case of a class, it is considered that there is a sample intersection between class A and class B. Sample intersection is also known as class overlap or dataset overlap. Since the text classifier is trained by using these training sample sets, the crossover of the training sample sets will inevitably affect the classification accuracy of the text classifier trained. The method for generating a text classifier provided by the embodiment of the present invention can effectively improve the crossover between such sample classes. A detailed description will be given below.

In step S11 , the training sample set includes a plurality of original sample categories, and these original sample categories correspond to target categories expected by the user, and each training sample belongs to one of the multiple original sample categories. In one embodiment of the present invention, the training sample set includes four original sample categories A, B, C, and D, where there is a sample intersection between the original sample category A and the original sample category C, then A and C can be merged The operation generates a new sample category G, and the combined training sample set includes the original sample categories B, D and the new sample category G.

Correspondingly, in step S12, the first classifier is obtained through training according to the training sample set after the merging operation, that is, all elements in the training sample set are divided into sample categories B, D, and G respectively. Through such classification, a first classifier can be trained.

Of course, there may be more than two original sample categories with sample intersection, and more than one new sample category after merging, as long as it is beneficial to correct sample intersection and improve classification accuracy, which is not limited by the embodiment of the present invention.

For example, in the above-mentioned embodiment, it may also be that there is an inter-class sample intersection between the four original sample categories A, B, C, and D, or there is a sample intersection between A and B while there is a sample intersection between C and D, Correspondingly, the merging operation can be to merge the four original sample categories A, B, C, and D into a new sample category G1, that is, to merge all the original sample categories with sample intersections into a new sample category, or Merge A and B into a new sample category G2, and merge C and D into a new sample category G3, that is, merge the original sample categories with sample intersections into multiple new sample categories.

Optionally, one or more of the following classification algorithms can be used to train the training sample set after the merging operation to obtain the first classifier: Naive Bayesian NB classification algorithm, support vector machine SVM classification algorithm, K nearest neighbor KNN Classification algorithms and random forest classification algorithms, etc.

It should be noted that the above-mentioned original sample category with sample intersection may be known or may be yet to be identified. When the number of samples in the training sample set and the data volume of the original sample category are huge, the original sample category with sample crossover can be identified in the following way:

Before performing the merging operation on at least two original sample categories with intersecting samples in the training sample set to obtain a new sample category, training according to the training sample set to obtain a third classifier whose classification category is the plurality of original sample categories, Test the classification accuracy of each classification category of the third classifier, and initially screen out the original sample categories whose classification accuracy is lower than the third threshold;

Among the initially screened original sample categories, an original sample category with sample crossover is identified.

The principle of the above identification method is that if there is inter-class sample intersection in the training sample set, it will inevitably affect the classification accuracy of the trained classifier. Therefore, the sample intersection of the training sample set can be preliminarily screened through the classification accuracy, and then Then, from the screened original sample categories, the original sample categories with sample intersections are identified.

Among them, the original sample category with sample intersection can be identified by manually checking or machine data matching on the original sample category lower than the third threshold.

For example, the higher the third threshold is set, the more sensitive the detection of sample crossing is. In the original sample category of the training sample set, there are sample data that should belong to category A but are classified into category B, and there are no samples that should belong to other categories in category A. At this time, the third classifier’s The classification accuracy rate of class B will be lower than the third threshold. Although the classification accuracy rate of class A is not affected, according to manual check or machine data matching, it can be concluded that there is a sample intersection between class A and class B in the training sample set. In the original sample category of the training sample set, there are sample data that should belong to category A but are classified into category B, and the sample data that originally belonged to category B are classified into category A, then the third classifier at this time Both the classification accuracy rate of class B and the classification accuracy rate of class A are lower than the third threshold, and it can be determined whether the class A and class B are intersected with other class samples through manual checking or machine data matching.

Testing the classification accuracy of each classification category of the third classifier may specifically be: testing the classification accuracy of the third classifier in a cross-validation manner based on the overall sample. After the first classifier is trained, the second classifier can be trained in step S13. Specifically, training may be performed according to the training samples belonging to the new sample category and the original sample category to which the training samples belonging to the new sample category belong, to obtain the second classifiers for each of the new sample categories respectively.

Still taking the above-mentioned embodiment as an example, if there is a sample intersection between the original sample category A and the original sample category C, A and C are merged to generate a new sample category G, and the combined training sample set is used for classification training, and the training Each element in the sample set is classified into the sample category B, D or G to obtain the first classifier. After obtaining the first classifier, use the new sample category G for classification training, and subdivide the elements in G into A and C, where A and C are the original sample categories to which the training samples belonging to the new sample category G belong.

In this way, through the training of the second classifier, the original sample category with sample crossover can be separated separately, and more detailed classification training can be carried out in a more specific range, thereby greatly improving the classification accuracy of the text classifier.

Optionally, the classification algorithm for obtaining the second classifier may include one or more of the following: Naive Bayesian NB classification algorithm, support vector machine SVM classification algorithm, K nearest neighbor KNN classification algorithm and random forest classification algorithm; wherein, The classification category included in the second classifier corresponds to the original sample category to which the training samples in the new sample category belong.

Further, the training sample set in the above embodiment can be obtained from the training corpus after certain processing. In order to obtain the above-mentioned training sample set, in one embodiment of the present invention, before the at least two original sample categories with crossed samples in the training sample set are combined to obtain a new sample category, the text classifier provided by the embodiment of the present invention Generation methods can also include:

Preprocessing the training corpus to filter and/or unify the format of the training corpus;

The preprocessed training corpus is subjected to word segmentation processing according to the word segmentation dictionary to obtain the training sample set.

Specifically, the collected training corpus includes sentences and/or text fragments, and the specific forms can be voice, text, images, etc. First, the format of the acquired training corpus must be unified into a text format through preprocessing, and invalid formats are filtered out. Save for later use. Then, the preprocessed training corpus is subjected to word segmentation processing according to the word segmentation dictionary to obtain a training sample set.

Further, the word segmentation dictionary can be expanded. For example, the new word discovery operation can be performed on the training corpus and the new words found can be added to the word segmentation dictionary. In this way, new words can be obtained by using the new word discovery method. According to the obtained new words Words can be updated in the word segmentation dictionary, then when the word segmentation process is performed, the word segmentation can be performed according to the updated word segmentation dictionary, so that the word segmentation dictionary can be continuously improved, and the accuracy of word segmentation processing can be effectively improved.

Optionally, the new word discovery operation may be realized by one or more of the following methods: mutual information, co-occurrence probability and information entropy.

In order to determine the classification effect of the generated first classifier and the second classifier, further, the text classifier generation method provided by the embodiment of the present invention may further include:

Test the classification accuracy rate of each classification category in the first classifier;

Test the classification accuracy rate of each classification category in the second classifier;

Wherein, the classification accuracy rates of the first classifier are respectively P1j, wherein j is an integer greater than or equal to 1 and less than or equal to m, and m is the number of sample categories in the training sample set after the merging operation;

The classification accuracy rates of the second classifier are respectively P1h*P2k, wherein k is an integer greater than or equal to 1 and less than or equal to n, and n is the number of original sample categories to which the training samples in the new sample category belong; The classification accuracy rate of the new sample category in the first classifier, h is an integer greater than or equal to 1 and less than or equal to g, and g is the number of new sample categories;

Detecting whether the classification accuracy rate of each classification category in the first classifier is greater than a first probability threshold, and whether the classification accuracy rate of each classification category in the second classifier is greater than a second probability threshold;

If yes, it is determined that the training of the first classifier and the second classifier is successful.

For example, if the first probability threshold is 0.98 and the second probability threshold is 0.95, during the test, if the classification accuracy of each classification category in the classification results of the first classifier is greater than 0.98, and the classification of the second classifier In the results, the classification accuracy rate of each classification category is greater than 0.95, indicating that the classification accuracy rate of the text classifier generated by the text classifier generation method provided by the embodiment of the present invention meets the user's requirement.

Optionally, when testing the classification accuracy, data with the same or similar attributes as the classification categories of the training sample set can be used for testing, and these data are marked with relevant classification categories. Among them, the data with the same or similar classification category attributes as the training sample set can be constructed by the algorithm, or can be obtained according to the cross-validation method.

Specifically, the classification accuracy of the first classifier may be tested by cross-validation based on the overall sample; the classification accuracy of the second classifier may be tested by cross-validation based on the overall sample.

Among them, the sample population refers to all sample data related to this classification task. The cross-validation method based on the sample population can be a part of the sample population as a training sample set, and another part as a test sample set, and 60% to 90% (for example, 80%) of the samples in the sample population can be used as a training sample set, Using the remaining samples as the text to be classified for testing, the sample population includes the multiple original sample categories, and each of the samples belongs to one of the multiple original sample categories.

Optionally, when detecting whether the classification accuracy rate of each classification category in the first classifier is greater than the first probability threshold, and whether the classification accuracy rate of each classification category in the second classifier is greater than the second After the probability threshold, the method for generating a text classifier provided by an embodiment of the present invention may further include: if not, re-merging and classifying the at least two original sample categories for which there is a sample cross in the training sample set until the The classification accuracy rate of each classification category in the first classifier is greater than the first probability threshold, and the classification accuracy rate of each classification category in the second classifier is greater than the second probability threshold.

That is to say, after the first classifier and the second classifier are used, if the classification accuracy of any classification category is found to be less than the corresponding probability threshold, it means that the first classifier and the second classifier have not yet met the user's requirements. Therefore, it is necessary to re-merge the original sample categories with sample crossover and classifier training until the accuracy rate reaches the above threshold requirements.

Optionally, in order to meet different requirements on the classification accuracy of the classifier, in an embodiment of the present invention, the first probability threshold and the second probability threshold may be adjusted to filter out different classification accuracy The first classifier and the second classifier.

The method for generating a text classifier provided by the embodiment of the present invention will be described in detail below through specific embodiments.

As shown in Figure 2, the text classifier generation method provided in this embodiment may specifically include the following steps:

S201, preprocessing: unify the acquired training corpus format into a text format, filter invalid formats, and save for later use;

S202, new word discovery: use the existing new word discovery tool to find out the new word candidate words of the training corpus, add the word segmentation dictionary after manual filtering;

S203, performing word segmentation on the preprocessed training corpus according to the word segmentation dictionary;

S204, sample screening: construct a classifier for the original sample category to achieve its classification, and test the accuracy rate P0 of the classifier based on the sample population using cross-validation (the accuracy rates of each class are P01, P02, ..., P0i, ... ). According to the results of the classification (categories with sample intersections have low accuracy rates), select (artificially selected or simple matching methods) category sets with sample intersections (categories that intersect with each other are a category set, so there may be one or multiple category sets);

S205, sample reorganization: merge two or more classes with sample crossover, and keep other classes unchanged;

S206, train and generate the first classifier: classify the training sample set after the merging operation, and test the accuracy rate P1 of the classification based on the overall sample by cross-validation (the accuracy rate of each class is P11, P12, ... , P1j, ...).

S207, training and generating a second classifier: constructing classifiers (one or more) for the new sample categories generated by merging, and testing the classification accuracy P2 based on the sample population using cross-validation (the accuracy of each class is P21, P22, ..., P2k, ...).

S208, accuracy rate test: detecting whether the classification accuracy rate of each classification category in the first classifier is greater than the first probability threshold, and whether the classification accuracy rate of each classification category in the second classifier is greater than the first probability threshold Two probability thresholds;

If yes, determining that the first classifier and the second classifier have been successfully trained;

If not, at least two original sample categories with sample intersections in the training sample set are re-merged and classifier trained until the classification accuracy of each classification category in the first classifier is greater than the first probability threshold , and the classification accuracy rate of each classification category in the second classifier is greater than the second probability threshold.

Correspondingly, as shown in FIG. 3 , an embodiment of the present invention also provides a text classification method. The text classification method utilizes a classifier generated by any of the text classifier generation methods provided in the foregoing embodiments for classification, and the classification Methods include:

S31. Input the text set to be classified into the first classifier to obtain a first classification result;

S32. Input the to-be-classified text whose classification category is the new sample category in the first classification result to a second classifier corresponding to the new sample category, to obtain a second classification result.

The text classification method provided by the embodiment of the present invention applies the text classifier generated by any one of the text classifier generation methods provided in the foregoing embodiments. In this way, through the training of the first classifier, the original sample category with sample crossover can be accurately distinguished from the original sample category without sample crossover, and through the training of the second classifier, the original sample category with sample crossover can be distinguished It is separated separately, and more detailed classification training is carried out in a more specific range, thus greatly improving the accuracy of text classification.

Optionally, the text classification method provided in the embodiment of the present invention may further include: using the classification result of the original sample category in which the classification category is not merged in the first classification result and the second classification result as the Describe the final classification result of the text to be classified.

The text classification method provided by the embodiment of the present invention applies the text classifier generated by any of the text classifier generation methods provided in the foregoing embodiments. The specific classification process and principle have been described in detail above, and will not be repeated here. repeat.

Correspondingly, as shown in FIG. 4, an embodiment of the present invention also provides a text classifier generating device, including:

The merging unit 41 is configured to perform a merging operation on at least two original sample categories with sample intersections in the training sample set to obtain a new sample category; the training sample set includes a plurality of original sample categories, and each training sample belongs to the plurality of original sample categories one of the sample categories;

The first training unit 42 is used to train the first classifier according to the training sample set after the merging operation;

The second training unit 43 is configured to perform training to obtain a second classifier according to the training samples belonging to the new sample category and the original sample category to which the training samples belonging to the new sample category belong, and the second classifier is used for Reclassify the text to be classified whose classification category is the new sample category in the classification result of the first classifier, so as to be classified into the corresponding original sample category.

The text classifier generation device provided by the embodiment of the present invention adopts a hierarchical classifier generation method when training the text classifier according to the training samples, and first merges two or more original sample categories with sample intersections in the training sample set into a new sample category, train the first classifier according to the training sample set after the merge operation, and then conduct more detailed classification training on the new sample category to obtain the second classifier. In this way, through the training of the first classifier, the original sample category with sample crossover can be accurately distinguished from the original sample category without sample crossover, and through the training of the second classifier, the original sample category with sample crossover can be distinguished It is separated separately, and more detailed classification training is carried out in a more specific range, thus greatly improving the classification accuracy of the text classifier.

Specifically, the sample crossover mentioned in the embodiments of the present invention refers to that in the provided training sample set, the classification of the sample data is not very clear and accurate, for example, there are sample data that should originally belong to the A category but are classified in the B category. In the case of a class, it is considered that there is a sample intersection between class A and class B. Since the text classifier is trained by using these training sample sets, the crossover of the training sample sets will inevitably affect the classification accuracy of the text classifier trained. The device for generating a text classifier provided by the embodiment of the present invention can effectively improve the crossover between such sample classes. A detailed description will be given below.

Optionally, the merging unit may be specifically configured to merge all original sample categories with sample intersections into a new sample category.

Optionally, when the merging unit 41 performs the merging operation, the training sample set includes a plurality of original sample categories, and these original sample categories correspond to the target classification expected by the user, and each training sample belongs to one of the multiple original sample categories . In one embodiment of the present invention, the training sample set includes four original sample categories A, B, C, and D, where there is a sample intersection between the original sample category A and the original sample category C, then A and C can be merged The operation generates a new sample category G, and the combined training sample set includes the original sample categories B, D and the new sample category G.

Correspondingly, the first training unit 42 trains the first classifier according to the training sample set after the merging operation, that is, divides all elements in the training sample set into sample categories B, D, and G respectively. Through such classification, a first classifier can be trained.

Of course, there may be more than two original sample categories with sample intersection, and more than one new sample category after merging, as long as it is beneficial to correct sample intersection and improve classification accuracy, which is not limited by the embodiment of the present invention.

For example, in the above-mentioned embodiment, it may also be that there is an inter-class sample intersection between the four original sample categories A, B, C, and D, or there is a sample intersection between A and B while there is a sample intersection between C and D, Correspondingly, the merging operation can be to merge the four original sample categories A, B, C, and D into a new sample category G1, that is, to merge all the original sample categories with sample intersections into a new sample category, or Merge A and B into a new sample category G2, and merge C and D into a new sample category G3, that is, merge the original sample categories with sample intersections into multiple new sample categories.

Optionally, the first training unit 42 can specifically be used for:

The training sample set after the merging operation is trained by at least one of the following classification algorithms to obtain the first classifier: Naive Bayesian NB classification algorithm, support vector machine SVM classification algorithm, K nearest neighbor KNN classification algorithm and random Forest classification algorithm.

It should be noted that the above-mentioned original sample category with sample intersection may be known or may be yet to be identified. When the sample size of the training sample set and the data volume of the original sample category are huge, the text classifier generation device may also include a screening unit for:

Before performing the merging operation on at least two original sample categories with intersecting samples in the training sample set to obtain a new sample category, training according to the training sample set to obtain a third classifier whose classification category is the plurality of original sample categories, Test the classification accuracy of each classification category of the third classifier, and initially screen out the original sample categories whose classification accuracy is lower than the third threshold;

Among the initially screened original sample categories, an original sample category with sample crossover is identified.

The principle of the above identification method is that if there is inter-class sample intersection in the training sample set, it will inevitably affect the classification accuracy of the trained classifier. Therefore, the sample intersection of the training sample set can be preliminarily screened through the classification accuracy, and then Then, from the screened original sample categories, the original sample categories with sample intersections are identified.

Among them, the original sample category with sample intersection can be identified by manually checking or machine data matching on the original sample category lower than the third threshold.

For example, the higher the third threshold is set, the more sensitive the detection of sample crossing is. In the original sample category of the training sample set, there are sample data that should belong to category A but are classified into category B, and there are no samples that should belong to other categories in category A. At this time, the third classifier’s The classification accuracy rate of class B will be lower than the third threshold. Although the classification accuracy rate of class A is not affected, according to manual check or machine data matching, it can be concluded that there is a sample intersection between class A and class B in the training sample set. In the original sample category of the training sample set, there are sample data that should belong to category A but are classified into category B, and the sample data that originally belonged to category B are classified into category A, then the third classifier at this time Both the classification accuracy rate of class B and the classification accuracy rate of class A are lower than the third threshold, and it can be determined whether the class A and class B are intersected with other class samples through manual checking or machine data matching.

Testing the classification accuracy of each classification category of the third classifier may specifically be: testing the classification accuracy of the third classifier in a cross-validation manner based on the overall sample.

After the first classifier is trained, the second training unit 43 can train to form the second classifier. Specifically, training may be performed according to the training samples belonging to the new sample category and the original sample category to which the training samples belonging to the new sample category belong, to obtain the second classifiers for each of the new sample categories respectively.

Still taking the above-mentioned embodiment as an example, if there is a sample intersection between the original sample category A and the original sample category C, A and C are merged to generate a new sample category G, and the combined training sample set is used for classification training, and the training Each element in the sample set is classified into the sample category B, D or G to obtain the first classifier. After obtaining the first classifier, use the new sample category G for classification training, and subdivide the elements in G into A and C, where A and C are the original sample categories to which the training samples belonging to the new sample category G belong.

In this way, through the training of the second classifier, the original sample category with sample crossover can be separated separately, and more detailed classification training can be carried out in a more specific range, thereby greatly improving the classification accuracy of the text classifier.

Optionally, the second classification unit 43 can be specifically configured to adopt at least one of the following classification algorithms for training to obtain the second classifier: Naive Bayesian NB classification algorithm, support vector machine SVM classification algorithm, K nearest neighbor KNN classification algorithm and a random forest classification algorithm; the classification category included in the second classifier corresponds to the original sample category to which the training samples in the new sample category belong.

Further, the training sample set in the above embodiment can be obtained from the training corpus after certain processing. In order to obtain the above-mentioned training sample set, the text classifier generation device provided by the embodiment of the present invention may further include:

A preprocessing unit, configured to perform preprocessing on the training corpus to filter and/or unify the format of the training corpus before merging at least two original sample categories with crossed samples in the training sample set to obtain a new sample category ;

The word segmentation unit is configured to perform word segmentation processing on the preprocessed training corpus according to the word segmentation dictionary to obtain the training sample set.

Specifically, the collected training corpus includes sentences and/or text fragments, and the specific forms can be voice, text, images, etc. First, the format of the acquired training corpus must be unified into a text format through preprocessing, and invalid formats are filtered out. Save for later use. Then, the preprocessed training corpus is subjected to word segmentation processing according to the word segmentation dictionary to obtain a training sample set.

Optionally, the device may further include a new word discovery unit, configured to perform a new word discovery operation on the training corpus and add the discovered new words to the word segmentation dictionary. In this way, new words can be obtained by using the new word discovery method, and the word segmentation dictionary can be updated according to the obtained new words. Then, when performing word segmentation processing, word segmentation can be performed according to the updated word segmentation dictionary, so that the word segmentation dictionary can be continuously improved. Effectively improve the accuracy of word segmentation processing.

Optionally, the new word discovery operation may be realized by one or more of the following methods: mutual information, co-occurrence probability and information entropy.

In order to determine the classification effect of the generated first classifier and the second classifier, further, the device may further include:

The first test unit is used to test the classification accuracy of each classification category in the first classifier;

The second test unit is used to test the classification accuracy of each classification category in the second classifier;

Wherein, the classification accuracy rates of the first classifier are respectively P1j, wherein j is an integer greater than or equal to 1 and less than or equal to m, and m is the number of sample categories in the training sample set after the merging operation;

The classification accuracy rates of the second classifier are respectively P1h*P2k, wherein k is an integer greater than or equal to 1 and less than or equal to n, and n is the number of original sample categories to which the training samples in the new sample category belong; The classification accuracy rate of the new sample category in the first classifier, h is an integer greater than or equal to 1 and less than or equal to g, and g is the number of new sample categories;

A detection unit, configured to detect whether the classification accuracy rate of each classification category in the first classifier is greater than the first probability threshold, and whether the classification accuracy rate of each classification category in the second classifier is greater than the second probability threshold;

A determination unit, configured to determine that the training of the first classifier and the second classifier is successful if the detection result of the detection unit is yes.

For example, if the first probability threshold is 0.98 and the second probability threshold is 0.95, during the test, if the classification accuracy of each classification category in the classification results of the first classifier is greater than 0.98, and the classification of the second classifier In the results, the classification accuracy rate of each classification category is greater than 0.95, indicating that the classification accuracy rate of the text classifier generated by the text classifier generation method provided by the embodiment of the present invention meets the user's requirements.

Optionally, when testing the classification accuracy, data with the same or similar attributes as the classification categories of the training sample set can be used for testing, and these data are marked with relevant classification categories. Among them, the data with the same or similar classification category attributes as the training sample set can be constructed by the algorithm, or can be obtained according to the cross-validation method.

Optionally, the first test unit is specifically configured to test the classification accuracy of the first classifier by using a cross-validation method based on a sample population; the second test unit is specifically used to use a cross-validation method based on a sample population Testing the classification accuracy of the second classifier.

Among them, the sample population refers to all sample data related to this classification task. The cross-validation method based on the sample population can be a part of the sample population as a training sample set, and another part as a test sample set, and 60% to 90% (for example, 80%) of the samples in the sample population can be used as a training sample set, Using the remaining samples as the text to be classified for testing, the sample population includes the multiple original sample categories, and each of the samples belongs to one of the multiple original sample categories.

Optionally, the device may further include: a return unit, configured to re-merge at least two original sample categories with sample intersections in the training sample set and a classifier if the detection result of the detection unit is negative Training until the classification accuracy rate of each classification category in the first classifier is greater than the first probability threshold, and the classification accuracy rate of each classification category in the second classifier is greater than the second probability threshold.

That is to say, after the first classifier and the second classifier are used, if the classification accuracy of any classification category is found to be less than the corresponding probability threshold, it means that the first classifier and the second classifier have not yet met the user's requirements. Therefore, it is necessary to re-merge the original sample categories with sample crossover and classifier training until the accuracy rate reaches the above threshold requirements.

In order to meet different requirements on the classification accuracy of classifiers, in an embodiment of the present invention, optionally, the device may further include: an adjustment unit, configured to adjust the first probability threshold and the second probability threshold Two probability thresholds to filter out the first classifier and the second classifier with different classification accuracy.

Correspondingly, as shown in FIG. 5 , an embodiment of the present invention also provides a text classification device, which classifies using a classifier generated by any text classifier generation device provided in the foregoing embodiments, and the classification device includes:

The first input unit 51 is configured to input the text set to be classified into the first classifier to obtain a first classification result;

The second input unit 52 is configured to input the to-be-classified text classified as the new sample category in the first classification result into a second classifier corresponding to the new sample category to obtain a second classification result.

The text classification apparatus provided by the embodiments of the present invention applies the text classifier generated by any text classifier generation apparatus provided in the foregoing embodiments. In this way, through the training of the first classifier, the original sample category with sample crossover can be accurately distinguished from the original sample category without sample crossover, and through the training of the second classifier, the original sample category with sample crossover can be distinguished It is separated separately, and more detailed classification training is carried out in a more specific range, thus greatly improving the accuracy of text classification.

Further, the text classification device also includes:

The result output unit is configured to use the classification result of the original sample category whose classification category is not combined in the first classification result and the second classification result as the final classification result of the text to be classified.

Correspondingly, an embodiment of the present invention also provides a computer device, including a processor and a memory; the memory is used to store computer instructions, and the processor is used to run the computer instructions stored in the memory to perform any one of the above-mentioned embodiments. The text classifier generation method, therefore, can also achieve corresponding technical effects, which have been described in detail above, and will not be repeated here.

Correspondingly, an embodiment of the present invention also provides a computer device, including a processor and a memory; the memory is used to store computer instructions, and the processor is used to run the computer instructions stored in the memory to perform any of the functions provided in the preceding embodiments. A text classification method, so it can also achieve the corresponding technical effect, which has been described in detail above, and will not be repeated here.

Correspondingly, an embodiment of the present invention also provides a computer-readable storage medium, and instructions are stored in the storage medium, and when the instructions are executed, any method for generating a text classifier provided in the foregoing embodiments is executed, so it can also The realization of the corresponding technical effects has been described in detail above, and will not be repeated here.

Correspondingly, an embodiment of the present invention also provides a computer-readable storage medium, and instructions are stored in the storage medium, and when the instructions are run, any text classification method provided in the foregoing embodiments is executed, so that the corresponding The technical effect of , which has been described in detail above, will not be repeated here.

It should be noted that, in this document, the term "comprising", "comprising" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article or apparatus comprising a set of elements includes not only those elements, It also includes other elements not expressly listed, or elements inherent in the process, method, article, or device. Without further limitations, an element defined by the phrase "comprising a ..." does not preclude the presence of additional identical elements in the process, method, article, or apparatus comprising that element.

Through the description of the above embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus a necessary general-purpose hardware platform, and of course also by hardware, but in many cases the former is better implementation. Based on such an understanding, the essence of the technical solution of the present invention or the part that contributes to the prior art can be embodied in the form of software products, and the computer software products are stored in a storage medium (such as ROM/RAM, disk, CD) contains several instructions to make a terminal device (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the methods described in various embodiments of the present invention.

The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the description of the present invention and the contents of the accompanying drawings, or directly or indirectly used in other related technical fields , are all included in the scope of patent protection of the present invention in the same way.

Claims (38)

1. a kind of text classifier generation method, it is characterised in that including：

Training sample is concentrated at least two original sample classifications that there is sample cross merge operation and obtain new samples class Not；The training sample set includes multiple original sample classifications, and each training sample belongs in the multiple original sample classification One；

First grader is obtained according to the training sample set training after union operation；

According to belonging to the training sample and the training sample for belonging to the new samples classification that belong to the new samples classification Original sample classification be trained and obtain the second grader, second grader is used for the classification to first grader As a result middle class categories carry out subseries again for the text to be sorted of the new samples classification, to be divided into corresponding original sample class Not in.

2. according to the method described in claim 1, it is characterised in that the training sample set according to after union operation is trained Include to the first grader：

Following at least one classification algorithm training is used to the training sample set after the union operation, first classification is obtained Device：The closest KNN sorting algorithms of naive Bayesian NB sorting algorithms, support vector machines sorting algorithm, K and random forest point Class algorithm.

3. according to the method described in claim 1, it is characterised in that according to the training sample for belonging to the new samples classification and Original sample classification belonging to the training sample for belonging to the new samples classification, which is trained, to be obtained the second grader and includes：

According to belonging to the training sample and the training sample for belonging to the new samples classification that belong to the new samples classification Original sample classification be trained, respectively obtain the second grader of each new samples classification.

4. according to the method described in claim 1, it is characterised in that

To the training sample for belonging to the new samples classification and the training sample institute for belonging to the new samples classification The original sample classification of category uses following at least one classification algorithm training, obtains second grader：Naive Bayesian NB The closest KNN sorting algorithms of sorting algorithm, support vector machines sorting algorithm, K and random forest sorting algorithm；Described second The class categories that grader includes are corresponding with the original sample classification belonging to the training sample in the new samples classification.

5. according to the method described in claim 1, it is characterised in that there is sample cross extremely in described concentrate training sample Few two original sample classifications are merged before obtaining new samples classification, and methods described also includes：

Training corpus is pre-processed to filter and/or uniform format the training corpus；

Pretreated training corpus is subjected to word segmentation processing according to dictionary for word segmentation and obtains the training sample set.

6. method according to claim 5, it is characterised in that the training corpus includes sentence and/or text fragments.

7. method according to claim 5, it is characterised in that also include：New word discovery behaviour is performed to the training corpus The neologisms made and will be seen that add the dictionary for word segmentation.

8. method according to claim 7, it is characterised in that the new word discovery operation passes through following at least one mode Realize：Mutual information, co-occurrence probabilities and comentropy.

9. according to the method described in claim 1, it is characterised in that also include：

Test the classification accuracy of each class categories in first grader；

Test the classification accuracy of each class categories in second grader；

Wherein, the classification accuracy of first grader is respectively P1j, and wherein j is whole more than or equal to 1 and less than or equal to m Number, m is to merge the sample class number that the training sample after operation is concentrated；

The classification accuracy of second grader is respectively P1h*P2k, and wherein k is whole more than or equal to 1 and less than or equal to n Number, n is the original sample classification number belonging to the training sample in the new samples classification；P1h is institute in first grader The classification accuracy of new samples classification is stated, h is the integer more than or equal to 1 and less than or equal to g, and g is the new samples classification number；

Detect that whether the classification accuracy of each class categories in first grader is both greater than the first probability threshold value, and described Whether the classification accuracy of each class categories is both greater than the second probability threshold value in second grader；

If it is, determining first grader and second classifier training success.

10. method according to claim 9, it is characterised in that each class categories in detection first grader The classification accuracy whether classification accuracy is both greater than each class categories in the first probability threshold value, and second grader is No both greater than the second probability threshold value, also includes afterwards：

If not, concentrating at least two original sample classifications that there is sample cross to re-start merging behaviour the training sample Make and classifier training, until the classification accuracy of each class categories is both greater than the first probability threshold in first grader Value, and untill the classification accuracy of each class categories is both greater than the second probability threshold value in second grader.

11. method according to claim 10, it is characterised in that also include：

First probability threshold value and second probability threshold value is adjusted to filter out the first grader of different classifications accuracy rate With the second grader.

12. method according to claim 9, it is characterised in that

The classification accuracy of first grader is totally tested using cross validation mode based on sample；

The classification accuracy of second grader is totally tested using cross validation mode based on sample.

13. method according to claim 12, it is characterised in that the cross validation mode bag overall based on sample Include：Using in the sample totality 60% to 90% sample as training sample set, using remaining sample as to be sorted Text is tested, and the sample generally includes the multiple original sample classification, and each sample belongs to the multiple original One in beginning sample class.

14. according to the method described in claim 1, it is characterised in that at least two original samples that there will be sample cross This categories combination includes into new samples classification：

There will be all original sample categories combinations of sample cross into a new samples classification.

15. the method according to any one of claim 1 to 14, it is characterised in that it is described described by training sample set The middle at least two original sample classifications that there is sample cross are merged before operation obtains new samples classification, and methods described is also Including：

The 3rd grader for obtaining that class categories are the multiple original sample classification, test are trained according to the training sample set The classification accuracy of each class categories of 3rd grader, it is original less than the 3rd threshold value that preliminary screening goes out classification accuracy Sample class；

The original sample classification that there is sample cross is identified in the original sample classification that preliminary screening goes out.

16. a kind of file classification method, it is characterised in that utilize the text classifier any one of claim 1 to 15 The grader of generation method generation is classified, and the sorting technique includes：

Text set to be sorted is inputted into first grader, the first classification results are obtained；

By to be sorted text input and the new samples of the class categories in first classification results for the new samples classification Corresponding second grader of classification, obtains the second classification results.

17. method according to claim 16, it is characterised in that also include：

By class categories in first classification results for the original sample classification without merging classification results with it is described Final classification result of second classification results collectively as the text to be sorted.

18. a kind of text classifier generating means, it is characterised in that including：

Combining unit, for concentrating at least two original sample classifications that there is sample cross to merge operation training sample Obtain new samples classification；The training sample set includes multiple original sample classifications, and each training sample belongs to the multiple original One in beginning sample class；

First training unit, for obtaining the first grader according to the training sample set training after union operation；

Second training unit, for according to the training sample for belonging to the new samples classification and described belonging to the new samples class Original sample classification belonging to other training sample, which is trained, obtains the second grader, and second grader is used for described Class categories carry out subseries again for the text to be sorted of the new samples classification in the classification results of first grader, to be divided into In corresponding original sample classification.

19. device according to claim 18, it is characterised in that first training unit, specifically for：

Following at least one classification algorithm training is used to the training sample set after the union operation, first classification is obtained Device：The closest KNN sorting algorithms of naive Bayesian NB sorting algorithms, support vector machines sorting algorithm, K and random forest point Class algorithm.

20. device according to claim 18, it is characterised in that second training unit, specifically for：

According to belonging to the training sample and the training sample for belonging to the new samples classification that belong to the new samples classification Original sample classification be trained, respectively obtain the second grader of each new samples classification.

21. device according to claim 18, it is characterised in that

Second taxon obtains second grader specifically for using following at least one classification algorithm training： The closest KNN sorting algorithms of naive Bayesian NB sorting algorithms, support vector machines sorting algorithm, K and random forest classification are calculated Method；The class categories that second grader includes and the original sample classification belonging to the training sample in the new samples classification It is corresponding.

22. device according to claim 18, it is characterised in that also include：

Pretreatment unit, for training sample to be concentrated at least two original sample classifications progress that there is sample cross described Merging is obtained before new samples classification, and training corpus is pre-processed to filter and/or form the training corpus It is unified；

Participle unit, the training sample is obtained for pretreated training corpus to be carried out into word segmentation processing according to dictionary for word segmentation Collection.

23. device according to claim 22, it is characterised in that the training corpus includes sentence or text fragments.

24. device according to claim 22, it is characterised in that also including new word discovery unit, for the training Language material performs the neologisms addition dictionary for word segmentation that new word discovery is operated and will be seen that.

25. device according to claim 24, it is characterised in that the new word discovery operation passes through following at least one side Formula is realized：Mutual information, co-occurrence probabilities and comentropy.

26. device according to claim 18, it is characterised in that also include：

First test cell, the classification accuracy for testing each class categories in first grader；

Second test cell, the classification accuracy for testing each class categories in second grader；

Wherein, the classification accuracy of first grader is respectively P1j, and wherein j is whole more than or equal to 1 and less than or equal to m Number, m is to merge the sample class number that the training sample after operation is concentrated；

The classification accuracy of second grader is respectively P1h*P2k, and wherein k is whole more than or equal to 1 and less than or equal to n Number, n is the original sample classification number belonging to the training sample in the new samples classification；P1h is institute in first grader The classification accuracy of new samples classification is stated, h is the integer more than or equal to 1 and less than or equal to g, and g is the new samples classification number；

Detection unit, for detecting it is general whether the classification accuracy of each class categories in first grader is both greater than first Whether the classification accuracy of each class categories is both greater than the second probability threshold value in rate threshold value, and second grader；

Determining unit, if the testing result for the detection unit is yes, determines first grader and described second Classifier training success.

27. device according to claim 26, it is characterised in that also include：

Returning unit, if the testing result for the detection unit is no, the training sample is concentrated and there is sample friendship At least two original sample classifications of fork re-start union operation and classifier training, until each in first grader The classification that the classification accuracy of class categories is both greater than each class categories in the first probability threshold value, and second grader is accurate Untill true rate is both greater than the second probability threshold value.

28. device according to claim 27, it is characterised in that also include：

Adjustment unit, for adjusting first probability threshold value and second probability threshold value to filter out different classifications accuracy rate The first grader and the second grader.

29. device according to claim 27, it is characterised in that

First test cell, specifically for totally testing first grader using cross validation mode based on sample Classification accuracy；

Second test cell, specifically for totally testing second grader using cross validation mode based on sample Classification accuracy.

30. device according to claim 29, it is characterised in that the cross validation mode bag overall based on sample Include：Using in the sample totality 60% to 90% sample as training sample set, using remaining sample as to be sorted Text is tested, and the sample generally includes the multiple original sample classification, and each sample belongs to the multiple original One in beginning sample class.

31. device according to claim 18, it is characterised in that the combining unit, specifically for there will be sample friendship All original sample categories combinations of fork are into a new samples classification.

32. the device according to any one of claim 18 to 31, it is characterised in that also including screening unit, for It is described to concentrate at least two original sample classifications that there is sample cross to merge operation and obtain new samples class training sample Before not, the 3rd grader for obtaining that class categories are the multiple original sample classification is trained according to the training sample set, The classification accuracy of each class categories of the 3rd grader is tested, preliminary screening goes out classification accuracy less than the 3rd threshold value Original sample classification；

The original sample classification that there is sample cross is identified in the original sample classification that preliminary screening goes out.

33. a kind of document sorting apparatus, it is characterised in that utilize the text classifier any one of claim 18 to 32 The grader of generating means generation is classified, and the sorter includes：

First input block, for text set to be sorted to be inputted into first grader, obtains the first classification results；

Second input block, for by class categories in first classification results be the new samples classification text to be sorted Input the second grader corresponding with the new samples classification, obtains the second classification results.

34. device according to claim 33, it is characterised in that also include：

As a result output unit, for being the original sample classification without merging by class categories in first classification results Classification results and second classification results collectively as the text to be sorted final classification result.

35. a kind of computer equipment, it is characterised in that including processor and memory；Memory is used to store computer instruction, Processor is used for the computer instruction for running the memory storage, with the text any one of perform claim requirement 1 to 15 This classifier generation method.

36. a kind of computer equipment, it is characterised in that including processor and memory；Memory is used to store computer instruction, Processor is used for the computer instruction for running the memory storage, with the text classification side described in perform claim requirement 16 or 17 Method.

37. a kind of computer-readable recording medium, it is characterised in that be stored with instruction in the storage medium, the instruction fortune Text classifier generation method during row any one of perform claim requirement 1 to 15.

38. a kind of computer-readable recording medium, it is characterised in that be stored with instruction in the storage medium, the instruction fortune File classification method during row described in perform claim requirement 16 or 17.