Text classification and device

[0057] In order to make the objectives, technical solutions and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below in conjunction with the accompanying drawings.

[0058] Before explaining the embodiments of the present invention in detail, the application scenarios involved in the embodiments of the present invention will be described first.

[0059] The customer service platform is often the most important service window for telecom operators or Internet operators, such as the mobile 10086 platform and Taobao customer service platform. Take the mobile 10086 platform as an example. The average daily customer service calls in the first half of 2015 were about 5.5 million. There are millions of pieces of customer service data stored on the mobile 10086 platform every day. Conversation record. Since the customer service data is often stored in the form of recordings, in order to facilitate the processing of the customer service data, it is often possible to perform voice analysis on the customer service data to convert the customer service data into customer service text. Figure 1A Shown is a specific example of a customer service text.

[0060] Since massive customer service data often contains a lot of real and valuable information, such as business development, business problem feedback, etc., this information can provide important references for the implementation of measures such as product innovation, marketing improvement, and network optimization. Therefore, in order to quickly and effectively discover valuable information from the massive customer service data, the customer service text corresponding to the customer service data can be classified.

[0061] In related technologies, technicians can manually construct the keyword database and matching rule database of the business information database. After the construction is completed, based on the keyword database and the matching rule database, determine the text to be classified from multiple preset categories. Category. Because when the keyword database and matching rule database of the business information database are completely constructed manually, the construction time of the business information database is relatively long, the construction efficiency is low, and the labor cost for establishing and maintaining the business information database is relatively high. Because when the business information database is completely dependent on the experience summaries of the technical personnel, the accuracy and coverage of text classification based on the business information database will be low. Therefore, the embodiment of the present invention provides a text classification method. On the premise of reducing labor costs, improve the accuracy and coverage of text classification.

[0062] Figure 1B It is a schematic diagram of a system architecture involved in a text classification method provided by an embodiment of the present invention. See Figure 1B The system architecture includes: a server 110, a client 120, and a network 130 that facilitates the establishment of a communication connection between the server 110 and the client 120. Among them, the user can communicate with the server 110 through the client 120. It should be noted that different users can communicate with the server 110 through different types of clients 120. For example, user A can communicate with the server 110 through a personal computer 120-1, and user B can communicate with the server through a mobile terminal 120-2. 110 communicates, which is not specifically limited in the embodiment of the present invention. Communication data is generated when the user communicates with the server 110 through the client 120. The communication data can be expressed in the form of text. For example, the communication data can be expressed as Figure 1A Customer service text shown. The server 110 is used to obtain the text to be classified and classify the text to be classified. The text to be classified may be obtained in real time by the server 110 during the communication with the client 120, or may be stored in advance. The text on the server 110 or other storage devices is not specifically limited in the embodiment of the present invention.

[0063] Figure 1C It is a schematic structural diagram of a server 110 provided by an embodiment of the present invention. See Figure 1C The server 110 includes: a business information database construction module 1101, a preprocessing module 1102, a pattern matching classification module 1103, a semantic classification module 1104, a weighted fusion module 1105, and a category determination module 1106.

[0064] When the server 110 needs to classify a certain text to be classified, it can first input the text to be classified into the preprocessing module 1102; the preprocessing module 1102 can preprocess the text to be classified to remove the After some significant noise interference in the text to be classified, the preprocessing module 1102 can input the preprocessed text to be classified into the pattern matching classification module 1103 and the semantic classification module 1104 respectively; the pattern matching module 1103 can be based on business The business information database constructed by the information database construction module 1101 classifies the text to be classified through a pattern matching classifier, and obtains the first probability that the text to be classified belongs to each of the plurality of preset categories. The first probability is input to the weighted fusion module 1105; the semantic classification module 1104 can classify the text to be classified through a semantic classifier, and obtain that the text to be classified belongs to the first of each of the plurality of preset categories. Two probabilities, and input the second probability into the weighted fusion module 1105; the weighted fusion module 1105 can be based on the first probability that the text to be classified belongs to each of the plurality of preset categories and the text to be classified The second probability belonging to each preset category in the plurality of preset categories, determining the weighted probability of the text to be classified belonging to some preset category in the plurality of preset categories, and inputting the weighted probability to the category determining module 1106: The category determination module 1106 can determine the category to which the text to be classified belongs according to the weighted probability that the text to be classified belongs to certain preset categories in the plurality of preset categories.

[0065] Specifically, the respective functions of the multiple modules included in the server 110 are described as follows:

[0066] The business information database construction module 1101 is used to construct the business information database according to the word vector model. The business information database construction module 1101 may include a seed database construction sub-module 11011 and a keyword expansion sub-module 11012. Wherein, the seed database construction submodule 11011 includes a seed information database constructed in advance by a technician, the seed information database includes a keyword database and a matching rule database, the keyword database includes multiple keywords, and the matching rule database includes Multiple matching rules, each of the multiple matching rules includes at least one keyword in the keyword library. The keyword expansion submodule 11012 is used to determine the potential expansion words of each keyword in the plurality of keywords according to the word vector model. After that, the technician can expand and update the keyword database and matching rule database of the seed information database based on the potential expansion words of the multiple keywords, and determine the expanded and updated seed information database as the business information database.

[0067] The preprocessing module 1102 is used to preprocess the text to be classified. Since the text to be classified includes a large amount of useless information, such as modal particles, auxiliary words, conjunctions, etc., some significant noise interference in the text to be classified can be removed by preprocessing the text to be classified. In a possible design, the preprocessing may include at least one of Chinese word segmentation, part-of-speech filtering, and stop word filtering. Among them, Chinese word segmentation refers to the conversion of text into a collection of Chinese words, part-of-speech filtering refers to removing modal particles, auxiliary words, conjunctions, etc. in the text, and stop word filtering refers to removing words that have no actual meaning in the text, such as removing words in the text. The "of", "then" etc.

[0068] The pattern matching classification module 1103 is used to classify the text to be classified after the preprocessing module 1102 is preprocessed by the pattern matching classifier based on the business information database constructed by the business information database building module 1101, and the text to be classified is multi The first probability of each preset category in the preset categories. The pattern matching classification module 1103 may include a keyword determination sub-module 11031 and a rule classification sub-module 11032. Among them, the keyword determining submodule 11031 is used to determine the keywords of the text to be classified through the pattern matching classifier based on the keyword database of the business information database. In a possible design, the pattern matching classifier may be based on the keyword database , Determine the keywords of the text to be classified through a multi-pattern matching algorithm. Among them, the rule classification sub-module 11032 is used to classify the text to be classified by the pattern matching classifier based on the keywords of the text to be classified and the matching rule library of the business information library, to obtain that the text to be classified belongs to multiple The first probability of each preset category in the preset categories.

[0069] The semantic classification module 1104 is configured to classify the text to be classified after the preprocessing module 1102 is preprocessed by the semantic classifier, and obtain the second probability that the text to be classified belongs to each of the multiple preset categories. When the semantic classification module 1104 classifies the text to be classified, it may determine that the text to be classified belongs to the first of each of the plurality of preset categories based on multiple preset semantic feature vectors included in the semantic classifier. Two probability. Specifically, the semantic feature vector of each word in the text to be classified can be determined by the semantic classifier; based on the semantic feature vector of each word in the text to be classified, the semantic classifier can be used to determine the semantics of the text to be classified Feature vector; for each preset semantic feature vector in the plurality of preset semantic feature vectors, the similarity between the preset semantic feature vector and the semantic feature vector of the text to be classified is calculated by the semantic classifier; The calculated similarity is determined as the second probability that the text to be classified belongs to the preset category corresponding to the preset semantic feature vector.

[0070] The weighted fusion module 1105 is used for the first probability that the text to be classified belongs to each of a plurality of preset categories according to the pattern matching classification module 1103 and the text to be classified obtained by the semantic classification module 1104 belongs to multiple The second probability of each preset category in the preset categories determines the weighted probability that the text to be classified belongs to certain preset categories in the plurality of preset categories. In a possible design, based on the first weight corresponding to the pattern matching classifier and the second weight corresponding to the semantic classifier, the first probability and the second probability corresponding to a certain preset category can be weighted and averaged to obtain the waiting The weighted probability that the classified text belongs to the preset category.

[0071] The category determining module 1106 is configured to determine the weighted probability that the text to be classified belongs to certain preset categories among the plurality of preset categories according to the weighted fusion module 1105 obtained The category to which the text belongs. In a possible design, when the weighted probability of the text to be classified belonging to a certain preset category among the plurality of preset categories is greater than the specified probability, the preset category may be determined as the category to which the text to be classified belongs . The specified probability can be preset according to specific business requirements, which is not specifically limited in the embodiment of the present invention.

[0072] figure 2 Is a schematic structural diagram of a computer device provided by an embodiment of the present invention, Figure 1B or Figure 1C Server 110 in figure 2 The computer equipment shown in to achieve. See figure 2 The computer device includes at least one processor 201, a communication bus 202, a memory 203, and at least one communication interface 204.

[0073] The processor 201 may be a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits for controlling the execution of the program of the present invention.

[0074] The communication bus 202 may include a path for transferring information between the aforementioned components.

[0075] The memory 203 can be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, random access memory (RAM), or other types that can store information and instructions. The type of dynamic storage device can also be Electrically Erasable Programmable Read-Only Memory (EEPROM), CD-ROM (CompactDisc Read-Only Memory, CD-ROM) or other optical disk storage, optical disk storage (Including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or can be used to carry or store desired program codes in the form of instructions or data structures and can be used by a computer Any other media accessed, but not limited to this. The memory 203 may exist independently and is connected to the processor 201 through a communication bus 202. The memory 203 may also be integrated with the processor 201.

[0076] The communication interface 204 uses any device such as a transceiver to communicate with other devices or communication networks, such as Ethernet, wireless access network (RAN), wireless local area network (Wireless Local Area Networks, WLAN), etc.

[0077] In a specific implementation, as an embodiment, the processor 201 may include one or more CPUs, for example figure 2 CPU0 and CPU1 shown in.

[0078] In specific implementation, as an embodiment, the computer device may include multiple processors, for example figure 2 The processor 201 and the processor 208 shown in. Each of these processors can be a single-CPU (single-CPU) processor or a multi-core (multi-CPU) processor. The processor here may refer to one or more devices, circuits, and/or processing cores for processing data (for example, computer program instructions).

[0079] In a specific implementation, as an embodiment, the computer device may further include an output device 205 and an input device 206. The output device 205 communicates with the processor 201, and can display information in a variety of ways. For example, the output device 205 may be a liquid crystal display (LCD), a light emitting diode (LED) display device, a cathode ray tube (CRT) display device, or a projector. Wait. The input device 206 communicates with the processor 201 and can receive user input in a variety of ways. For example, the input device 206 may be a mouse, a keyboard, a touch screen device, or a sensor device.

[0080] The above-mentioned computer equipment may be a general-purpose computer equipment or a special-purpose computer equipment. In a specific implementation, the computer device may be a desktop computer, a portable computer, a network server, a personal digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. The embodiments of the present invention do not limit the type of computer equipment.

[0081] The memory 203 is used to store program codes for executing the solutions of the present invention, and the processor 201 controls the execution. The processor 201 is configured to execute the program code 210 stored in the memory 203. The program code 210 may include one or more software modules (for example, a business information library building module, a preprocessing module, a pattern matching classification module, a semantic classification module, a weighted fusion module, a category determination module, etc.). Figure 1B or Figure 1C The server 110 shown in FIG. 1 can classify the text to be classified through the processor 201 and one or more software modules in the program code 210 in the memory 203.

[0082] image 3 It is a flowchart of a text classification method provided by an embodiment of the present invention, and the method is used in a server. See image 3 , The method includes:

[0083] Step 301: For each of the multiple keywords included in the keyword database of the business information database, determine the word vector corresponding to the keyword according to the word vector model.

[0084] It should be noted that the business information database is used to assist the pattern matching classifier to classify the text to be classified, and the business information database may include a keyword database, and the keyword database may include multiple keywords.

[0085] In addition, the word vector model is used to convert words into word vectors. The word vector model is obtained through training by a word vector tool. The word vector tool can be a Word2vec (Word to vector) tool, etc. This embodiment of the present invention does not do this Specific restrictions.

[0086] Among them, according to the word vector model, the operation of determining the word vector corresponding to the keyword can refer to related technologies, which is not described in detail in the embodiment of the present invention.

[0087] Further, according to the word vector model, before the word vector corresponding to the keyword is determined, a text data set can also be obtained, and multiple texts included in the text data set can be preprocessed to obtain the first training corpus. After that, use the The first training corpus is used as the input of the word vector tool, and the word vector tool is trained to obtain the word vector model.

[0088] It should be noted that the text data set can be stored in the server or in other storage devices, and the server and the other storage devices can communicate via a wired network or a wireless network, which is not specifically limited in the embodiment of the present invention . The text data set includes multiple texts, and the multiple texts are texts related to multiple preset categories. For example, the multiple texts may be explanatory texts, customer service texts, etc., of the multiple preset categories. This is not specifically limited. The explanatory text is a text in which explanatory data of a preset category is recorded, and the customer service text is a text in which dialogue data between a user and a customer service related to the preset category is recorded.

[0089] In addition, the multiple preset categories can be set in advance. For example, the multiple preset categories can be points redemption-point query, international roaming-tariff consultation, credit startup-shutdown reasons, etc. The embodiment of the present invention does not specifically limit this .

[0090] Furthermore, the preprocessing may include at least one of Chinese word segmentation, part-of-speech filtering, and stop word filtering, which is not specifically limited in the embodiment of the present invention. Chinese word segmentation refers to the conversion of text into a collection of Chinese words. Part-of-speech filtering refers to removing modal particles, auxiliary words, conjunctions, etc. in the text. Stop word filtering refers to removing words that have no actual meaning in the text, such as removing the " "Of", "then", etc.

[0091] It should be noted that the operation of preprocessing multiple texts included in the text data set can refer to related technologies, which are not described in detail in the embodiment of the present invention.

[0092] In addition, using the first training corpus as the input of the word vector tool to train the word vector tool to obtain the word vector model can refer to related technologies, which are not described in detail in the embodiment of the present invention.

[0093] Furthermore, the word vector model obtained by the word vector tool after training can convert each word included in the text data set into a K-dimensional word vector, where K is a natural number greater than or equal to 1, and the value of K It can be preset by a technician, which is not specifically limited in the embodiment of the present invention.

[0094] It should be noted that each keyword in the keyword database of the business information database can be defined by a technician based on experience and observation of the text data set, which is not specifically limited in the embodiment of the present invention.

[0095] Step 302: Determine potential expansion words of the keyword based on the word vector corresponding to the keyword.

[0096] Specifically, according to the word vector model, determine the word vector corresponding to each word included in the text data set; calculate the similarity between the word vector corresponding to the keyword and the word vector corresponding to each word included in the text data set; The similar words in the text data set are determined as potential expansion words of the keyword, and the similarity between the word vector corresponding to the similar word and the word vector corresponding to the keyword is greater than the specified similarity.

[0097] It should be noted that the designated similarity can be preset according to specific business needs. For example, the designated similarity can be 0.75, 0.8, etc., which is not specifically limited in the embodiment of the present invention.

[0098] Among them, when calculating the similarity between the word vector corresponding to the keyword and the word vector corresponding to each word included in the text data set, the Euclidean distance between the two can be calculated, and the calculated Euclidean distance is determined as both Or, the cosine similarity between the two can be calculated, and the calculated cosine similarity can be determined as the similarity between the two. Of course, in practical applications, the key can also be calculated by other methods The embodiment of the present invention does not specifically limit the similarity between the word vector corresponding to the word and the word vector corresponding to each word included in the text data set.

[0099] Step 303: When an expansion rule input for the potential expansion word is received, and when an add instruction for the potential expansion word is detected, the potential expansion word is added to the keyword library, and the expansion rule is added to In the matching rule database of the business information database.

[0100] It should be noted that the add instruction is used to instruct to add the potential expansion word to the keyword library, and add the expansion rule input for the potential expansion word to the matching rule library. The addition instruction can be triggered by a technician. The technician can trigger through a designated operation, which can be a single-click operation, a double-click operation, a voice operation, etc., which is not specifically limited in the embodiment of the present invention.

[0101] In addition, the matching rule library may include multiple matching rules, and each matching rule in the multiple matching rules may include at least one keyword in the keyword library. For example, the matching rule may use a logical “&” symbol (&) Associating multiple keywords means that when a text contains all the keywords in the matching rule at the same time, it is determined that the text satisfies the matching rule.

[0102] Furthermore, each of the plurality of preset categories may correspond to at least one matching rule in the matching rule library, which is not specifically limited in the embodiment of the present invention.

[0103] In the embodiment of the present invention, the technician may store multiple keywords in the keyword database of the business information database in advance, and store multiple matching rules in the matching rule database in advance, and each of the multiple matching rules includes For at least one of the multiple keywords, the server can determine the potential expansion word of each keyword in the multiple keywords according to the word vector model. After that, the technician can identify and analyze the potential expansion word. When the technician confirms that the potential expansion word can be used to construct a matching rule for a certain preset category, the technician can input the expansion rule for the potential expansion word, and add the potential expansion word to the keyword library, and the expansion The rule is added to the matching rule library, thereby completing the expansion update of the keyword library and the matching rule library.

[0104] It is easy to find that the keyword database and matching rule database in the embodiments of the present invention are constructed in a semi-manual manner, which not only saves construction time, improves construction efficiency, and reduces the establishment and maintenance of the keyword database and The labor cost of the matching rule base. In addition, since the keyword database and the matching rule database can be expanded and updated according to the potential expansion words determined by the server, the coverage rate of the keyword database and the matching rule database can be improved, and because the potential expansion words and The expansion rule is added to the keyword database and the matching rule database based on the add instruction, that is, the potential expansion word and the expansion rule are added to the keyword database and the matching rule database after confirmation by the technician Therefore, the accuracy of the keyword database and the matching rule database can be improved, and the coverage and accuracy of subsequent text classification based on the keyword database and the matching rule database can be improved.

[0105] Step 304: Based on the keyword library and the matching rule library, a pattern matching classifier is used to determine the first probability that the text to be classified belongs to each of the plurality of preset categories.

[0106] It should be noted that the text to be classified may be one of the above-mentioned text data sets, or it may be obtained in real time by the server during communication with the client, which is not specifically limited in the embodiment of the present invention.

[0107] In addition, based on the keyword library and the matching rule library, before the pattern matching classifier determines the first probability that the text to be classified belongs to each of the plurality of preset categories, the text to be classified can also be Preprocessing, so that based on the preprocessed text to be classified, the pattern matching classifier can quickly determine the first probability that the text to be classified belongs to each of the multiple preset categories, thereby improving the determination efficiency.

[0108] It should be noted that the operation of preprocessing the text to be classified is similar to the related operation in step 301, which will not be repeated in this embodiment of the present invention.

[0109] Specifically, based on the keyword database and the matching rule database, when the pattern matching classifier determines the first probability that the text to be classified belongs to each of the plurality of preset categories, the server may select the text to be classified , Obtain the same words as the keywords in the keyword library through the pattern matching classifier; determine the obtained words as the keywords of the text to be classified; based on the keywords of the text to be classified and the matching rule library, The text to be classified is classified by the pattern matching classifier, and the first probability that the text to be classified belongs to each of the plurality of preset categories is obtained.

[0110] Among them, from the text to be classified, when the pattern matching classifier obtains the same words as the keywords in the keyword library, the pattern matching classifier can obtain the same words from the text to be classified by specifying the matching algorithm. The keywords in the keyword database are the same words. Of course, in practical applications, the pattern matching classifier can also obtain the same words as the keywords in the keyword database from the text to be classified in other ways. The embodiment of the invention does not specifically limit this.

[0111] It should be noted that the designated matching algorithm can be preset. For example, the designated matching algorithm can be a multi-pattern matching algorithm, and the multi-pattern matching algorithm can be a Wu-Manber algorithm. The embodiment of the present invention does not specifically limit this. In addition, because the Wu-Manber algorithm can quickly perform string matching through the jump table, the pattern matching classifier uses the Wu-Manber algorithm to obtain the same keywords as the keywords in the keyword library from the text to be classified It can improve the acquisition efficiency.

[0112] In addition, from the text to be classified, the operation of obtaining the same words as the keywords in the keyword library through the pattern matching classifier may refer to related technologies, which are not described in detail in the embodiment of the present invention.

[0113] Wherein, based on the keywords of the text to be classified and the matching rule library, the text to be classified is classified by the pattern matching classifier, and the text to be classified is obtained in each of the plurality of preset categories. For the first probability of the category, the pattern matching classifier can, for each of the multiple matching rules included in the matching rule library, determine whether the text to be classified satisfies the keyword based on the keywords of the text to be classified Matching rule; when the text to be classified meets the matching rule, the first probability of determining that the text to be classified belongs to the preset category corresponding to the matching rule is 1; when the text to be classified does not meet the matching rule, Determine the keyword vector of the text to be classified based on the keywords of the text to be classified, and determine the keyword vector of the matching rule based on at least one keyword contained in the matching rule, and then calculate the keyword vector of the text to be classified The similarity between the keyword vector of the text and the keyword vector of the matching rule determines the calculated similarity as the first probability that the text to be classified belongs to the preset category corresponding to the matching rule.

[0114] The operation of calculating the similarity between the keyword vector of the text to be classified and the keyword vector of the matching rule is similar to the related operation in step 302, which will not be repeated in this embodiment of the present invention.

[0115] Among them, when judging whether the text to be classified meets the matching rule based on the keywords of the text to be classified, it can be judged whether the text to be classified contains the matching rules based on the keywords of the text to be classified All keywords; when the text to be classified contains all the keywords in the matching rule, it is determined that the text to be classified meets the matching rule; when the text to be classified does not contain all the keywords in the matching rule For keywords, it is determined that the text to be classified does not satisfy the matching rule.

[0116] For example, if the keyword of the text to be classified is Guoman, Cancel, and the matching rule is Guoman&Cancel, that is, if at least one keyword included in the matching rule is Guoman, Cancel, then the to-be-classified text can be determined The text contains all the keywords in the matching rule, and it is determined that the text to be classified meets the matching rule.

[0117] For another example, if the keywords of the text to be classified are national comics, cancel, and the matching rule is roaming&cancel&international, that is, at least one of the keywords included in the matching rule is roaming, cancel, international, it can be determined The text to be classified does not contain all the keywords in the matching rule, and it is determined that the text to be classified does not satisfy the matching rule.

[0118] Wherein, when the text to be classified does not satisfy the matching rule, based on the keywords of the text to be classified, when the keyword vector of the text to be classified is determined, the number of keywords included in the keyword library can be determined Then, all the keywords of the text to be classified are converted into a vector whose dimension is equal to the number, and the vector is determined as the keyword vector of the text to be classified.

[0119] For example, if the number of keywords included in the keyword library is 8, and all the keywords of the text to be classified are Guoman, cancel, then all the keywords of the text to be classified can be converted into a dimension A vector equal to 8, for example, all keywords of the text to be classified can be converted into a vector (0, 1, 1, 0, 0, 0, 0, 0).

[0120] Wherein, based on at least one keyword contained in the matching rule, when determining the keyword vector of the matching rule, the number of multiple keywords included in the keyword library can be determined, and then, at least A keyword is converted into a vector whose dimension is equal to the number, and the vector is determined as the keyword vector of the matching rule.

[0121] For example, the number of keywords included in the keyword library is 8, and the matching rule is roaming&cancellation&international, that is, at least one of the keywords included in the matching rule is roaming, canceling, and international. Convert at least one keyword contained in the matching rule into a vector with a dimension equal to 8. For example, at least one keyword contained in the matching rule can be converted into a vector (1, 0, 1, 1, 0, 0, 0, 0).

[0122] It should be noted that, based on the keyword library and the matching rule library, the operation of determining the first probability that the text to be classified belongs to each of the multiple preset categories through the pattern matching classifier can also refer to related technologies. This embodiment of the present invention will not elaborate on this.

[0123] Step 305: Based on the first probability that the text to be classified belongs to each of the multiple preset categories, determine the category to which the text to be classified belongs from the multiple preset categories.

[0124] In a possible design, step 304 can determine the first probability that the text to be classified belongs to each of the multiple preset categories. When the text to be classified belongs to a certain preset of the multiple preset categories When the first probability of the category is greater than the specified probability, the preset category may be determined as the category to which the text to be classified belongs.

[0125] It should be noted that the designated probability can be preset according to specific business needs. For example, the designated probability can be 0.8, 0.85, etc., which is not specifically limited in the embodiment of the present invention.

[0126] For example, if the specified probability is 0.8, and the first probability that the text to be classified belongs to a preset category among multiple preset categories is 0.85, it can be determined that the first probability of the text to be classified belongs to the preset category is greater than the specified probability To determine the preset category as the category to which the text to be classified belongs.

[0127] In another possible design, the second probability that the text to be classified belongs to each of the plurality of preset categories may be determined based on multiple preset semantic feature vectors included in the semantic classifier. One preset semantic feature vector corresponds to the plurality of preset categories; based on the first probability that the text to be classified belongs to each of the plurality of preset categories and the text to be classified belongs to the plurality of preset categories The second probability of each preset category in the preset categories is to determine the category to which the text to be classified belongs from the multiple preset categories.

[0128] Wherein, based on the plurality of preset semantic feature vectors included in the semantic classifier, when it is determined that the text to be classified belongs to the second probability of each of the plurality of preset categories, the semantic classifier can determine the The semantic feature vector of each word in the classified text; based on the semantic feature vector of each word in the text to be classified, the semantic feature vector of the text to be classified is determined by the semantic classifier; for the plurality of preset semantic features For each preset semantic feature vector in the vector, the similarity between the preset semantic feature vector and the semantic feature vector of the text to be classified is calculated by the semantic classifier; the calculated similarity is determined as the to be classified The second probability that the text belongs to the preset category corresponding to the preset semantic feature vector.

[0129] It should be noted that the semantic classifier is used to classify the text to be classified based on the semantic information of the text to be classified. For example, the semantic classifier may be a recursive convolutional neural network, etc., which is not specifically limited in the embodiment of the present invention. .

[0130] In addition, the semantic feature vector of a word is a feature vector that can reflect the semantic information of the word and the dependency relationship of the word in the context, and the semantic feature vector of the text to be classified is a feature vector that can reflect the semantic information of the text to be classified The embodiment of the present invention does not specifically limit this.

[0131] Furthermore, multiple preset semantic feature vectors may be preset, which is not specifically limited in the embodiment of the present invention.

[0132] Wherein, when the semantic feature vector of each word in the text to be classified is determined by the semantic classifier, the semantic feature vector of each word in the text to be classified can be determined by the word feature extraction layer in the semantic classifier. The present invention The embodiment does not specifically limit this. In addition, the operation of determining the semantic feature vector of each word in the text to be classified by the semantic classifier may refer to related technologies, which are not described in detail in the embodiment of the present invention.

[0133] Wherein, based on the semantic feature vector of each word in the text to be classified, when the semantic feature vector of the text to be classified is determined by the semantic classifier, the semantic feature vector of each word in the text to be classified can be determined by The text feature extraction layer in the semantic classifier determines the semantic feature vector of the text to be classified, which is not specifically limited in the embodiment of the present invention. In addition, based on the semantic feature vector of each word in the text to be classified, the operation of determining the semantic feature vector of the text to be classified by the semantic classifier may refer to related technologies, which are not described in detail in the embodiment of the present invention.

[0134] It should be noted that the word feature extraction layer and the text feature extraction layer can adopt a recursive structure to ensure that the semantic feature vector of a word determined by the word feature extraction layer can not only reflect the semantic information of the word, but also reflect the word The dependency relationship in the context further ensures that based on the semantic feature vector of the word, the semantic feature vector of the text to be classified determined by the text feature extraction layer can completely reflect the semantic information of the text to be classified.

[0135] Wherein, for each preset semantic feature vector of the plurality of preset semantic feature vectors, when calculating the similarity between the preset semantic feature vector and the semantic feature vector of the text to be classified by the semantic classifier, The similarity between the preset semantic feature vector and the semantic feature vector of the text to be classified is calculated by the classification layer in the semantic classifier, which is not specifically limited in the embodiment of the present invention. In addition, the operation of calculating the similarity between the preset semantic feature vector and the semantic feature vector of the text to be classified by the semantic classifier is similar to the related operation in the foregoing step 302, which will not be repeated in this embodiment of the present invention.

[0136] It should be noted that when text classification is performed in related technologies, the term frequency-inverse document frequency (TF-IDF) feature of the text to be classified is often extracted first, and then the support vector machine (English: Support Vector Machine (abbreviation: SVM) The classifier determines the category to which the TF-IDF feature of the text to be classified belongs, and determines the category to which the TF-IDF feature of the text to be classified belongs as the category to which the text to be classified belongs . However, because TF-IDF features are only statistics on the frequency of important words and lack high-level information on text semantic understanding, texts belonging to different categories are likely to have very similar TF-IDF features, resulting in texts in related technologies. The classification accuracy is low. In the embodiment of the present invention, the semantic classifier classifies the text to be classified based on the semantic information of the text to be classified, thereby effectively avoiding misclassification of the text to be classified due to lack of information, and improving the text The accuracy of classification.

[0137] Further, before determining the second probability that the text to be classified belongs to each of the plurality of preset categories based on the plurality of preset semantic feature vectors included in the semantic classifier, a preset text set may also be obtained , And preprocess multiple preset texts included in the preset text set to obtain a second training corpus. After that, use the second training corpus to train the semantic classifier to be trained to obtain the semantic classifier. The text set includes a plurality of preset texts, and each preset category of the plurality of preset categories corresponds to at least one preset text.

[0138] It should be noted that the preset text set can be preset, and the preset text set can be stored in the server or in other storage devices, and the server and the other storage devices can communicate through a wired network or a wireless network The embodiment of the present invention does not specifically limit this.

[0139] In addition, the multiple preset texts may also be preset, and the multiple preset texts may be customer service texts of the multiple preset categories, etc., which is not specifically limited in the embodiment of the present invention. Each of the plurality of preset categories corresponds to at least one preset text in the preset text set, that is, the plurality of preset texts are all texts with a category identifier.

[0140] Furthermore, in the embodiment of the present invention, when the semantic classifier to be trained is trained through a preset text set, the semantic classifier to be trained may be trained by means of supervised learning. The supervised learning means that the semantic classifier is given In the case of input and output, the parameters in the semantic classifier are continuously adjusted by specifying the adjustment algorithm to make the semantic classifier achieve the required performance.

[0141] It should be noted that the designated adjustment algorithm may be preset. For example, the designated adjustment algorithm may be a stochastic gradient descent algorithm, which is not specifically limited in the embodiment of the present invention.

[0142] In addition, the second training corpus is used to train the semantic classifier to be trained, and the operation of obtaining the semantic classifier can refer to related technologies, which will not be described in detail in the embodiment of the present invention.

[0143] Furthermore, the operation of preprocessing the multiple preset texts included in the preset text set is similar to the related operation in the above step 301, which will not be repeated in this embodiment of the present invention.

[0144] Wherein, based on the first probability that the text to be classified belongs to each of the multiple preset categories and the second probability that the text to be classified belongs to each of the multiple preset categories, from Among the multiple preset categories, when determining the category to which the text to be classified belongs, at least one target preset category may be determined from the multiple preset categories; for each target in the at least one target preset category The preset category determines the first probability and the second probability corresponding to the target preset category; based on the first weight corresponding to the pattern matching classifier and the second weight corresponding to the semantic classifier, the first probability corresponding to the target preset category Perform a weighted average with the second probability to obtain a weighted probability; when the weighted probability is greater than the specified probability, it is determined that the target preset category is the category to which the text to be classified belongs.

[0145] Among the multiple preset categories, the operation of determining at least one target preset category may include the following three methods:

[0146] The first method is to determine each of the plurality of preset categories as the target preset category.

[0147] The second way: according to the multiple first probabilities corresponding to the multiple preset categories in descending order, from the multiple first probabilities, obtain N first probabilities, and combine the N first probabilities The preset category corresponding to each of the first probabilities is determined as the target preset category, and the N is a natural number greater than or equal to 1.

[0148] The third way: According to the multiple second probabilities corresponding to the multiple preset categories in descending order, obtain N second probabilities from the multiple second probabilities, and combine the N second probabilities The preset category corresponding to each of the second probabilities in is determined as the target preset category.

[0149] It should be noted that the above second method and the above third method can not only be executed separately to determine at least one target preset category from multiple preset categories, of course, the above second method and the above first The three methods are combined to determine at least one target preset category from multiple preset categories, which is not specifically limited in the embodiment of the present invention.

[0150] In addition, the embodiment of the present invention may directly determine each preset category of the multiple preset categories as the target preset category. In this case, the server does not need to perform other operations, thereby improving the efficiency of determining the target preset category. In addition, the embodiment of the present invention may also determine at least one target preset category based on the first probability and the second probability. In this case, the at least one target preset category is part of the preset categories. Therefore, the server In the subsequent steps, there is no need to calculate the weighted probabilities for the multiple preset categories, but only the weighted probabilities for some of the multiple preset categories, which can save server processing resources and improve text Classification efficiency.

[0151] Among them, based on the first weight corresponding to the pattern matching classifier and the second weight corresponding to the semantic classifier, the first probability and the second probability corresponding to the target preset category are weighted and averaged. When the weighted probability is obtained, the first weight can be Multiply the first probability corresponding to the target preset category to obtain the first value; multiply the second weight and the second probability corresponding to the target preset category to obtain the second value; and multiply the first value with the second value Add them to get the weighted probability.

[0152] It should be noted that the first weight and the second weight can be set in advance, and in actual applications, the first weight and the second weight can be set and adjusted according to the reliability of the pattern matching classifier and the semantic classifier, and can also The settings and adjustments are made according to different business requirements, which are not specifically limited in the embodiment of the present invention.

[0153] It should be noted that in the embodiment of the present invention, the pattern matching classifier and the semantic classifier can be combined to perform text classification, thereby avoiding the inaccuracy caused by relying solely on a certain method for text classification. The first weight and the second weight are set and adjusted according to different business requirements. Therefore, it can be ensured that the weighted probability obtained meets the technical personnel's text classification requirements.

[0154] In addition, the text classification method provided in the embodiment of the present invention can be used to classify customer service text generated in the customer service platform of a telecommunication operator or an Internet operator. Because the embodiment of the present invention can efficiently construct a keyword database and services Rule library, and the pattern matching classifier and semantic classifier can be combined for text classification. Therefore, when the matching rules of the customer service text are cumbersome and the amount of data is large, the customer service text can also be better performed Classification, the accuracy of text classification can be guaranteed. Of course, the text classification method provided in the embodiment of the present invention can also be applied to other fields to perform other types of text classification, which is not specifically limited in the embodiment of the present invention.

[0155] It should be noted that, in the embodiment of the present invention, when the server is a server cluster composed of multiple node servers, the underlying architecture of the server may adopt a Hadoop distributed computing platform, which is not specifically limited in the embodiment of the present invention. In addition, when the underlying architecture of the server adopts the Hadoop distributed computing platform, the Hadoop distributed computing platform can include components such as HDFS, Map-Reduce, and Hive, and the HDFS component is used to store text, keyword libraries, and matching rules Libraries, etc., Map-Reduce components and Hive components are used to support the core operations of text classification. The core operations may include operations such as determining the first probability, the second probability, and the weighted probability.

[0156] In the embodiment of the present invention, for each of the multiple keywords included in the keyword database of the business information database, the word vector corresponding to the keyword is determined according to the word vector model, and the word vector corresponding to the keyword is determined based on the word corresponding to the keyword. The vector determines the potential expansion word of the keyword, and then, when the expansion rule input for the potential expansion word is received, and when the addition instruction for the potential expansion word is detected, the potential expansion word is added to the keyword Database, and add the extended rule to the matching rule database of the business information database, that is, the keyword database and the matching rule database are constructed in a semi-manual manner, which not only saves construction time and improves construction efficiency , And can reduce the labor cost of establishing and maintaining the keyword database and the matching rule database. In addition, since the keyword database and the matching rule database can be expanded and updated according to the potential expansion words determined by the server, the coverage rate of the keyword database and the matching rule database can be improved, and because the adding instruction is made by a technician Therefore, the potential expansion word and the expansion rule are added to the keyword database and the matching rule database after confirmation by the technician, so that the accuracy of the keyword database and the matching rule database can be improved. Furthermore, since the coverage and accuracy of the keyword database and the matching rule database are improved, based on the keyword database and the matching rule database, it is determined that the text to be classified belongs to each of the multiple preset categories. Set the first probability of the category, and based on the first probability that the text to be classified belongs to each of the multiple preset categories, determine the category to which the text to be classified belongs from the multiple preset categories It can improve the coverage and accuracy of text classification.

[0157] Figure 4 It is a schematic structural diagram of a text classification device under the same inventive concept as the foregoing method embodiment provided by an embodiment of the present invention. Such as Figure 4 As shown, the structure of the text classification device is used to perform the above image 3 The function of the server in the illustrated method embodiment includes: a first determining unit 401, a second determining unit 402, an adding unit 403, a third determining unit 404, and a fourth determining unit 405.

[0158] The first determining unit 401 is configured to determine the word vector corresponding to the keyword according to the word vector model for each of the multiple keywords included in the keyword database of the business information database;

[0159] The second determining unit 402 is configured to determine the potential expansion words of the keywords based on the word vectors corresponding to the keywords;

[0160] The adding unit 403 is used to add the potential expansion word to the keyword library and add the expansion rule to the business information when the expansion rule input for the potential expansion word is received, and when the addition instruction for the potential expansion word is detected, The matching rule library of the library;

[0161] The third determining unit 404 is configured to determine the first probability that the text to be classified belongs to each of the plurality of preset categories through the pattern matching classifier based on the keyword library and the matching rule library;

[0162] The fourth determining unit 405 is configured to determine the category to which the text to be classified belongs from the multiple preset categories based on the first probability that the text to be classified belongs to each of the multiple preset categories.

[0163] Optionally, the second determining unit 402 is configured to:

[0164] According to the word vector model, determine the word vector corresponding to each word included in the text data set;

[0165] Calculate the similarity between the word vector corresponding to the keyword and the word vector corresponding to each word included in the text data set;

[0166] The similar words in the text data set are determined as potential expansion words of the keywords, and the similarity between the word vectors corresponding to the similar words and the word vectors corresponding to the keywords is greater than the specified similarity.

[0167] Optionally, the third determining unit 404 is configured to:

[0168] From the text to be classified, the same words as the keywords in the keyword library are obtained through the pattern matching classifier;

[0169] Determine the acquired words as keywords of the text to be classified;

[0170] Based on the keywords of the text to be classified and the matching rule library, the text to be classified is classified by the pattern matching classifier to obtain the first probability that the text to be classified belongs to each of a plurality of preset categories; where, matching The rule library includes multiple matching rules, each of the multiple matching rules includes at least one keyword in the keyword library, and each preset category corresponds to at least one matching rule in the matching rule library.

[0171] Optionally, the fourth determining unit 405 is configured to:

[0172] Based on the multiple preset semantic feature vectors included in the semantic classifier, determine the second probability that the text to be classified belongs to each of the multiple preset categories, multiple preset semantic feature vectors and multiple preset categories One-to-one correspondence

[0173] Based on the first probability that the text to be classified belongs to each of the multiple preset categories and the second probability that the text to be classified belongs to each of the multiple preset categories, from the multiple preset categories To determine the category of the text to be classified.

[0174] Optionally, the fourth determining unit 405 is further configured to:

[0175] Determine the semantic feature vector of each word in the text to be classified through the semantic classifier;

[0176] Based on the semantic feature vector of each word in the text to be classified, the semantic feature vector of the text to be classified is determined through the semantic classifier;

[0177] For each preset semantic feature vector in the plurality of preset semantic feature vectors, the similarity between the preset semantic feature vector and the semantic feature vector of the text to be classified is calculated by the semantic classifier;

[0178] The calculated similarity is determined as the second probability that the text to be classified belongs to the preset category corresponding to the preset semantic feature vector.

[0179] Optionally, the fourth determining unit 405 is further configured to:

[0180] Determine at least one target preset category from multiple preset categories;

[0181] For each target preset category in at least one target preset category, determine the first probability and the second probability corresponding to the target preset category;

[0182] Based on the first weight corresponding to the pattern matching classifier and the second weight corresponding to the semantic classifier, weighted average the first probability and the second probability corresponding to the target preset category to obtain the weighted probability;

[0183] When the weighted probability is greater than the specified probability, it is determined that the target preset category is the category to which the text to be classified belongs.

[0184] Optionally, the fourth determining unit 405 is further configured to:

[0185] Determine each preset category of the plurality of preset categories as the target preset category.

[0186] Optionally, the fourth determining unit is also used in at least one of the following two ways:

[0187] According to the multiple first probabilities corresponding to multiple preset categories in descending order, obtain N first probabilities from the multiple first probabilities, and assign each of the N first probabilities to the corresponding The preset category is determined as the target preset category, and N is a natural number greater than or equal to 1; or,

[0188] According to the multiple second probabilities corresponding to multiple preset categories in descending order, N second probabilities are obtained from the multiple second probabilities, and each of the N second probabilities is corresponding to The preset category is determined as the target preset category.

[0189] In the embodiment of the present invention, the text classification device is presented in the form of a functional unit. The "unit" here may refer to an ASIC, a processor and memory that executes one or more software or firmware programs, an integrated logic circuit, and/or other devices that can provide the above-mentioned functions. In a simple embodiment, those skilled in the art can imagine that the text classification device can adopt figure 2 The form shown. The first determining unit 401, the second determining unit 402, the adding unit 403, the third determining unit 404 and the fourth determining unit 405 can pass figure 2 Specifically, the first determining unit 401, the second determining unit 402, and the adding unit 403 can be implemented by the processor executing the business information database building module, and the third determining unit 404 can be implemented by the processor The pattern matching classification module is executed, and the fourth determining unit 405 may be realized by executing the semantic classification module, the weighted fusion module, and the category determination module by the processor.

[0190] The embodiment of the present invention also provides a computer storage medium for storing and realizing the above Figure 4 The computer software instructions of the text classification device shown include the programs designed to execute the above method embodiments. By executing the stored program, the text to be classified can be classified.

[0191] It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described sequence of actions. Because according to the present invention, some steps can be performed in other order or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the involved actions and modules are not necessarily required by the present invention.

[0192] Although the present invention has been described in conjunction with various embodiments, in the process of implementing the claimed invention, those skilled in the art can understand and understand by looking at the drawings, the disclosure, and the appended claims. Implement other changes of the disclosed embodiment. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "one" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. Certain measures are described in mutually different dependent claims, but this does not mean that these measures cannot be combined to produce good results.

[0193] Those skilled in the art should understand that the embodiments of the present invention can be provided as methods, devices (equipment), or computer program products. Therefore, the present invention may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The computer program is stored/distributed in a suitable medium, provided with other hardware or as a part of the hardware, and can also be distributed in other forms, such as through the Internet or other wired or wireless telecommunication systems.

[0194] The present invention is described with reference to the flowcharts and/or block diagrams of the methods, devices (equipment) and computer program products of the embodiments of the present invention. It should be understood that each process and/or block in the flowchart and/or block diagram, and the combination of processes and/or blocks in the flowchart and/or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing equipment to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing equipment can be generated It is a device that realizes the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.

[0195] These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing equipment to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including the instruction device. The device implements the functions specified in one process or multiple processes in the flowchart and/or one block or multiple blocks in the block diagram.

[0196] These computer program instructions can also be loaded on a computer or other programmable data processing equipment, so that a series of operation steps are executed on the computer or other programmable equipment to produce computer-implemented processing, so as to execute on the computer or other programmable equipment. The instructions provide steps for implementing functions specified in a flow or multiple flows in the flowchart and/or a block or multiple blocks in the block diagram.

[0197] Although the present invention has been described with reference to specific features and embodiments thereof, it is obvious that various modifications and combinations can be made without departing from the spirit and scope of the present invention. Accordingly, this specification and drawings are merely exemplary descriptions of the present invention defined by the appended claims, and are deemed to have covered any and all modifications, changes, combinations or equivalents within the scope of the present invention. Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. In this way, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.