TW201812615A - Sentiment orientation recognition method, object classification method and data processing system - Google Patents

Abstract

Provided by the present application are a sentiment orientation recognition method, object classification method and data processing system. A sentiment degree estimation model, constructed by the present invention in the sentiment ornamentation recognition method, fully considers the category to which a short text belongs. Therefore, sentiment orientation is determined more accurately on the basis of the sentiment degree estimation model. In addition, since the object classification method provided by the present application use text feature information, image feature information and other feature information of an object as the basis of object classification, the object classification method provided by the present application may simultaneously give consideration to the text feature information, image feature information and other feature information, thereby being able to improve classification accuracy.

Description

   Identification method of emotional tendency, object classification method and data processing system   

The present invention relates to the field of data processing technology, and in particular, to a method for identifying emotional tendencies, a method for classifying objects, and a data processing system.

At present, the problem of classifying objects is involved in many technical fields. Generally speaking, objects are classified according to the text of the objects, and the objects are classified into two categories: the first category or the second category. In the text of an object, press punctuation to turn the text into multiple short texts.

Due to the rich meaning of Chinese characters, the same short text may correspond to different categories in different contexts. For example, taking the object as the user's evaluation text of the clothes as an example, the first user evaluates as "dim clothes color, just right", and the second user evaluates as "dim clothes color, not bright". The two objects mentioned above have the same short text "Dark clothes". If classified by text, two short texts will be classified into one category, but the two should correspond to different categories.

It can be seen that in different contexts, the "dim clothes color" in the first user evaluation corresponds to positive emotions and should be divided into the first category; the "dim clothes color" in the second user evaluation corresponds to negative emotions and should be divided For the second category. Therefore, the emotional tendency corresponding to short text is usually used to determine the category of objects.

In order to determine the emotional tendency of short texts, the traditional way is to manually view and determine the emotional tendency of short texts. Although the accuracy of manual annotation to determine the sentiment of short text is high, it is not efficient and cannot be applied to the processing of batch short texts.

The applicant of the present invention discovered during the research process that the processor can automatically recognize the emotional tendency of short text using a processor. The specific implementation process can be as follows: before the processor executes, the emotional lexicon is constructed. The emotional lexicon contains many positive words, such as "clothing", "large screen", "pretty", "quick", "suitable", "beautiful", etc. The emotional lexicon also contains many negative words, such as "clothing", "ugly "," Slow "," Small screen ", etc.

In order to process the object to be processed, first, the object to be processed is divided according to punctuation marks, and a short text is arranged between two adjacent punctuation marks, so that the object to be processed is divided into several short texts to be processed. For example, taking "the clothes are suitable and the mother likes it" as an example, after dividing according to the punctuation marks, you can get two short texts "the clothes are suitable" and "the mother likes it". Each short text of the object to be processed is a short text to be processed.

Referring to FIG. 1, a flowchart for determining a sentiment tendency of a short text to be processed by a processor, the execution process specifically includes the following steps:

Step 1: The processor performs segmentation on the short text to be processed, and obtains the segmentation result.

According to the preset word segmentation rules, the short text to be processed is divided into several words, and some words are the result of word segmentation.

For example, taking the short text to be processed as "suitable for clothing" as an example, the results obtained after segmentation are "suitable", "very", and "suitable". Taking the short text to be processed as "the mobile phone screen is very large", the segmentation results obtained after the segmentation are "mobile phone", "screen", "very" and "large".

Since segmentation of short texts to be processed is not the focus of the present invention, the specific implementation of the preset segmentation rules will not be described in detail here.

Step 2: Match the word segmentation results to the sentiment lexicon and match them according to the sentiment matching rules.

Step 3: Determine the emotional tendency corresponding to the short text to be processed.

Match the segmentation result with the sentiment lexicon and sentiment rules. If the segmentation results in the segmentation result correspond to positive emotions and do not contain negative words, then it is determined that the short text corresponds to positive emotions. If the emotional words in the segmentation result all correspond to negative emotions and do not contain negative words, then it is determined that the short text corresponds to negative emotions.

The processor can automatically execute the process shown in FIG. 1, so that the emotional tendency of the short text to be processed can be automatically determined. However, the applicant of the present invention found during research that although the above-mentioned automatic processing process can recognize the emotional tendency of the short text to be processed to a certain extent, the emotional tendency of the short text to be processed obtained by the above processing process may be inaccurate.

For example, taking the user evaluation on Taobao as an example, because Taobao has many categories (such as clothing category, electronic equipment category, mother and baby category, etc.), each category has a corresponding item. User reviews. The applicant found during research that short texts containing the same sentiment words in different categories may correspond to different sentiment tendencies.

For example, under the category of electronic devices, a short text is "large screen", and the short text's emotional tendency is positive emotion. Under the clothing category, a short text is "big clothes", and the short text's emotional tendency is negative emotion. From the above examples, it can be seen that under two different categories, the two short texts are "large", so the two short texts contain the same emotional words, but the two short texts have different emotional tendencies.

Because the processor in FIG. 1 automatically determines the emotional tendency of the short text, the processor adopts the same processing method for all objects, that is, the existing processing process does not separately process the emotional tendency of the short text from the perspective of the object category. Therefore, it is inaccurate to determine the emotional tendency of the short text in the prior art.

Therefore, the present invention provides a method for identifying an emotional tendency, so that the emotional tendency of a short text to be processed can be accurately determined.

In order to achieve the above object, the present invention provides the following technical features: a method for identifying emotional tendencies, including: determining a category identifier corresponding to short text to be processed; wherein a text between two punctuation marks adjacent to a text is called a short text; Determine the implementation of the sentiment estimation model corresponding to the category identifier; if the implementation of the sentiment estimation model is that all categories correspond to one sentiment estimation model, determine the characteristics corresponding to the short text to be processed Each feature in the feature set includes: a segmentation of the short text to be processed and a category identifier to which the short text to be processed belongs; based on a pre-trained sentiment estimation model, combined with the short text to be processed Feature set to estimate the sentiment degree of the short text to be processed; wherein, the sentiment estimation model includes: outputting a positive sentiment degree obtained after training according to at least two types of short text samples with emotional tendencies And negative emotion degree model; based on the positive emotion degree and negative emotion degree corresponding to the short text to be processed, determining the The sentiment tendency corresponding to the short text to be processed; if the implementation of the sentiment estimation model is a category corresponding to an sentiment estimation model, the feature set corresponding to the short text to be processed is determined; The features include: the word segmentation of the short text to be processed; the sentiment estimation of the short text to be processed according to the sentiment estimation model corresponding to the category identifier, and the feature set of the short text to be processed; The sentiment estimation model is: a model that outputs positive sentiment and negative sentiment after training according to the short text samples corresponding to the category identifier and sentimental short text samples; based on the short text to be processed The corresponding positive emotion degree and negative emotion degree determine the emotional tendency corresponding to the short text to be processed.

Preferably, after determining the emotional tendency corresponding to the short text to be processed, the method further includes: outputting the emotional tendency corresponding to the short text to be processed.

A method for identifying an emotional tendency includes: determining a feature set corresponding to short text to be processed; wherein a text between two punctuation marks adjacent to a text is called short text; each feature in the feature set includes: the The word segmentation of the short text to be processed and the category identifier to which the short text to be processed belong; according to a pre-trained sentiment estimation model and the feature set of the short text to be processed, the sentiment estimation of the short text to be processed; The sentiment estimation model includes: a model outputting positive sentiment and negative sentiment obtained after training according to at least two types of short text samples with emotional tendencies; and based on the corresponding short text to be processed The positive emotion degree and the negative emotion degree determine the emotion tendency corresponding to the short text to be processed.

Preferably, the determining a feature set corresponding to the short text to be processed includes: obtaining a category identifier corresponding to the short text to be processed, and a segmentation result obtained after performing a word segmentation operation on the short text to be processed; Each word segmentation result in the word segmentation result is combined with the category identifier to obtain each feature; the set of each feature is determined as the feature set of the short text to be processed.

Preferably, the determining a feature set corresponding to the short text to be processed includes: obtaining a category identifier corresponding to the short text to be processed, and a segmentation result obtained after performing a word segmentation operation on the short text to be processed; Each segmentation in the segmentation result is combined with the category identifier to obtain each feature; the n-ary language model is used to combine the features to obtain several combined features; each feature and the number of the combined features are The set is determined as a feature set of the short text to be processed.

Preferably, the use of the n-ary language model to combine the features to obtain several combined features includes: using a binary language model to combine the features to obtain several combined features.

Preferably, the emotion degree estimation model based on the pre-trained emotion degree estimation combined with the feature set of the short text to be processed, and the emotion degree estimation of the short text to be processed includes inputting the feature set to the emotion degree estimation. Model; after the estimation by the sentiment estimation model, the positive sentiment and the negative sentiment corresponding to the short text to be processed are output.

Preferably, determining the emotional tendency corresponding to the short text to be processed based on the positive emotion level and the negative emotion level corresponding to the short text to be processed includes: determining both the positive emotion level and the negative emotional level. The larger emotion degree of the person; judging whether the larger emotion degree is greater than the preset reliability; if the larger emotion degree is greater than the preset reliability, determining the emotional tendency corresponding to the short text to be processed and the The emotional tendencies of larger emotions are consistent.

Preferably, the sentiment estimation model includes a model that outputs maximum positive sentiment and negative sentiment by using a maximum entropy model and trained according to a feature set of several short texts corresponding to at least two category identifiers.

Preferably, after determining the emotional tendency corresponding to the short text to be processed, the method further includes: outputting the emotional tendency corresponding to the short text to be processed.

A method for identifying emotional tendency includes: determining a feature set and a category identifier corresponding to short text to be processed; wherein a text between two punctuation marks adjacent to a text is called short text; each feature in the feature set The method includes the word segmentation of the short text to be processed, an emotion level estimation model based on the sentiment degree estimation model corresponding to the category identifier, and a feature set of the short text to be processed, and the emotion. The degree estimation model is: a model that outputs positive sentiment and negative sentiment after training according to the short text samples corresponding to the category identifier and several short text samples with emotional tendencies; based on the corresponding short text to be processed The positive emotion degree and the negative emotion degree determine the emotion tendency corresponding to the short text to be processed.

Preferably, the determining a feature set corresponding to the short text to be processed includes: obtaining a segmentation result obtained after performing the word segmentation operation on the short text to be processed; using an n-gram language model to perform segmentation and combination on each segmentation to obtain several combinations Word segmentation; a set of each word segmentation and several combined word segmentation is determined as a feature set of the short text to be processed, and one word segmentation corresponds to one feature.

Preferably, the determining a feature set corresponding to the short text to be processed includes: obtaining a segmentation result obtained after performing the word segmentation operation on the short text to be processed; and determining the segmentation result as a feature of the short text to be processed Set, one participle corresponds to one feature.

Preferably, after determining the emotional tendency corresponding to the short text to be processed, the method further includes: outputting the emotional tendency corresponding to the short text to be processed.

An emotion tendency recognition system includes: a data providing device for sending a plurality of objects; a processor for receiving a plurality of objects sent by the data providing device, and constructing a sentiment estimation model based on short texts of the plurality of objects , And use the sentiment estimation model to determine the sentiment tendency of short text to be processed.

Preferably, the processor is further configured to construct a correspondence relationship between the sentiment estimation model and the category identifier to which the object belongs.

Preferably, the system further includes a receiving device; the processor is further configured to output the emotional tendency of the text to be processed; and the receiving device is configured to receive the emotional tendency of the text to be processed.

An emotional tendency recognition system includes: a data providing device for sending several objects; a model building device for receiving several objects sent by the data providing device; and constructing an emotion degree estimation based on short texts of the objects A model, and sends the sentiment estimation model; a processor is configured to receive the sentiment estimation model and use the sentiment estimation model to determine an emotional tendency of a short text to be processed.

Preferably, the model construction device is further configured to construct a correspondence between the sentiment estimation model and a category identifier to which the object belongs, and send the correspondence to the processor.

Preferably, the system further includes a receiving device; the processor is further configured to output the emotional tendency of the text to be processed; and the receiving device is configured to receive the emotional tendency of the text to be processed.

An object classification method includes: determining feature information of an object to be processed; wherein the feature information includes text feature information and image feature information, and the text feature information includes an emotional tendency of a short text; according to a pre-trained category The recognition model performs category recognition on the feature information of the object to be processed. The category recognition model is a classifier of the first category and the second category, which is obtained by training based on the feature information of a plurality of object samples.

Preferably, the feature information further includes: feature information of a first subject constructing the object; and / or, feature information of the second subject attached to the object.

Preferably, classifying the feature information based on a pre-trained class recognition model includes: inputting the feature information to the class recognition model; determining a first class match corresponding to the object to be processed The degree of matching with the second category; comparing the degree of matching with the first category and the degree of matching with the second category; if the degree of matching with the first category is greater than the degree with matching the second category, determining that the category of the object to be processed is the first Category; if the second category matching degree is greater than the first category matching degree, it is determined that the category of the object to be processed is the second category.

Preferably, the method further includes: after determining that the object to be processed is a first category, adding the object to be processed to an object set; and sending the objects in the object set.

Preferably, the method further includes: receiving a plurality of object samples, the object samples originating from the object set, and satisfying a preset rule; adding the plurality of object samples to an existing object sample of the training category recognition model Medium; based on the updated existing object samples, the class recognition model is retrained.

A method for classifying user evaluations includes: determining characteristic information of user evaluations to be processed; wherein the characteristic information includes text characteristic information of user evaluations, image characteristic information of user evaluations, characteristic information of sellers, and characteristic information of buyers Moreover, the text feature information includes the sentiment tendency of short text; class recognition is performed on the feature information evaluated by the user to be processed according to a pre-trained gradient promotion decision tree model; wherein the category recognition model is based on: Classifier of the first type of user evaluation and the second type of user evaluation obtained after training the feature information of the user evaluation sample.

Preferably, the method further includes: after determining that the to-be-processed user evaluation is a first-type user evaluation, adding the to-be-processed user evaluation to a first-type user evaluation set; and sending the first-type user evaluation set.

Preferably, the method further includes: receiving a plurality of first-type user evaluations, the first-type user evaluations originating from the first-type user evaluation set; and adding the plurality of first-type user evaluations to a category recognition model Among the existing user evaluation samples; based on the updated existing user evaluation samples, the class recognition model is retrained.

An object classification system includes: a data providing device for sending a plurality of objects; a processor for receiving a plurality of objects sent by the data providing device, and obtaining and outputting a first category and A category recognition model of the second category; used to determine feature information of an object to be processed; wherein the feature information includes text feature information and image feature information, and the text feature information includes short-text emotional tendencies; The category recognition model performs category recognition on the feature information of the object to be processed; it is also used to output objects of the first category; and a data receiving device for receiving and using the objects of the first category.

An object classification system includes: a data providing device for sending a plurality of objects; a model building device for receiving a plurality of objects sent by the data providing device, and obtaining and outputting a first number after training based on characteristic information of the plurality of objects A category recognition model of a category and a second category, and send the category recognition model; a processor, configured to receive the category recognition model and determine feature information of an object to be processed; wherein the feature information includes text feature information and Image feature information, and the text feature information includes an emotional tendency of a short text; class recognition is performed on the feature information of the object to be processed according to the category recognition model; and is also used to output the object of the first category; data A receiving device for receiving and using the objects of the first category.

Through the above technical means, the following beneficial effects can be achieved: The present invention provides a method for identifying emotional tendencies. This method uses several short texts with emotional tendencies corresponding to the category as training samples, and obtains the feature set of short texts for training. , And get a sentiment estimation model. Since each feature contains the word segmentation and category identification of the short text, the sentiment estimation model constructed by the application fully considers the category to which the short text belongs. Therefore, the sentiment tendency of the short text to be processed determined based on the sentiment estimation model is also more accurate.

100‧‧‧ Information Providing Equipment

200‧‧‧ processor

300‧‧‧model building equipment

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only the present invention. For some embodiments of the invention, for those of ordinary skill in the art, other drawings can be obtained based on these drawings without paying creative labor.

FIG. 1 is a flowchart of determining an emotional tendency of a short text to be processed in the prior art; FIGS. 2a-2b are schematic structural diagrams of an emotional tendency recognition system according to an embodiment of the present invention; A schematic diagram of the correspondence between the estimation model and the category; Figures 4a-4c are flowcharts of constructing an emotion degree estimation model according to an embodiment of the present invention; and Fig. 5 is another method for constructing an emotion degree estimation model provided by an embodiment of the present invention Flowchart; Figures 6a-6b are flowcharts of constructing an emotion degree estimation model according to an embodiment of the present invention; Figure 7 is a flowchart of a method for identifying sentiment tendency according to an embodiment of the present invention; A flowchart of a method for identifying an emotional tendency according to an embodiment of the invention; FIG. 9 is a flowchart of a method for identifying an emotional tendency provided by an embodiment of the invention; FIG. 10 is a flowchart of a method for identifying an emotional tendency provided by an embodiment of the invention; 11a-11b are flowcharts of a method for identifying an emotional tendency according to an embodiment of the present invention; FIG. 12 is a flowchart of an object classification method according to an embodiment of the present invention; and FIG. 13 is a flowchart provided by an embodiment of the present invention A flowchart of still another object classification method; FIG. 14 is a flowchart of still another object classification method according to an embodiment of the present invention; FIG. 15 is a flowchart of still another object classification method according to an embodiment of the present invention; FIG. 17 is a schematic structural diagram of another object classification system according to an embodiment of the present invention; and FIG. 18 is a flowchart of a scenario embodiment of an object classification method according to an embodiment of the present invention.

In the following, the technical solutions in the embodiments of the present invention will be clearly and completely described with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making progressive creations fall within the protection scope of the present invention.

In order to accurately determine the sentiment tendency of the short text to be processed, the present invention proposes a technical method for constructing an emotional degree estimation model to estimate the positive sentiment and negative sentiment corresponding to the short text to be processed using the sentiment estimation model. Among them, the positive sentiment degree is used to indicate the degree to which the short text to be processed belongs to the positive sentiment. Similarly, the negative sentiment degree is used to indicate the degree to which the short text to be processed belongs to the negative sentiment. After determining the positive sentiment and negative sentiment, the emotional tendency of the short text to be processed can be further determined.

In order to enable those skilled in the art to understand the application scenario of the present invention more clearly, referring to FIG. 2a or FIG. 2b, an emotional tendency recognition system is provided for the present invention.

The emotional tendency recognition system provided in FIG. 2 a specifically includes: a material providing device 100, and a processor 200 connected to the material providing device 100.

The data providing device 100 is configured to send several objects to the processor 200. The processor 200 is configured to construct an emotion degree estimation model according to short texts of several objects, and use the emotion degree estimation model to determine an emotional tendency of short texts to be processed.

The present invention also provides another recognition system for emotional tendencies (see Fig. 2b).

The emotional tendency recognition system provided in FIG. 2b specifically includes: a material providing device 100, a model building device 300 connected to the material providing device, and a processor 200 connected to the model building device. The model building device 300 may be a processing device having a processing capability.

Among them, the material providing device 100 is used to send several objects to the model building device 300. The model construction device 300 is configured to construct an emotion degree estimation model according to short texts of several objects, and send the emotion degree estimation model to the processor 200. The processor 200 is configured to determine an emotional tendency of a short text to be processed by using an emotion degree estimation model.

In the emotion tendency recognition system proposed in FIG. 2a and FIG. 2b, both the processor 200 and the model construction device 300 can execute the process of constructing an emotion degree estimation model, and the process of constructing the emotion degree estimation model by the two is the same . Therefore, the processor 200 or the model construction device 300 is collectively referred to as a processing device, so that the processing device is used to uniformly represent the processor 200 or the model construction device 300 in the process of constructing an emotion degree estimation model described below.

The system shown in FIG. 2a and FIG. 2b may further include a receiving device (not shown in the figure) connected to the processor. After the processor determines the emotional tendency of the short text to be processed, the processor is further configured to output the emotional tendency of the text to be processed; and the receiving device is configured to receive the emotional tendency of the text to be processed, so that the receiving device can Use the emotional tendency of the text to be processed to perform other processing.

The process of constructing a sentiment estimation model is described below. Since the category of short text is not considered in the process of determining the sentiment tendency of the short text to be processed in the prior art, the sentiment tendency determined in the prior art is not accurate. Therefore, the present invention considers the category of short texts in the process of constructing the sentiment estimation model by the processing device, so that the constructed sentiment estimation model can accurately determine the positive sentiment and negative sentiment of the short text to be processed.

The present invention proposes three implementation modes for processing equipment to construct an emotion degree estimation model. Refer to FIGS. 3a to 3c, which are schematic diagrams of a category and an emotion degree estimation model in the three implementation modes.

The first implementation manner: all categories correspond to a sentiment estimation model (see FIG. 3a). The second implementation method: each category corresponds to an emotion estimation model (see FIG. 3b). The third implementation: an implementation between the first and the second implementation (see Figure 3c); assuming there are N categories, the third implementation can build M emotions Estimation model, where M is a non-zero natural number, and 1 <M <N.

The detailed implementation process of these three implementation methods is described in detail below:

The first implementation method: all categories correspond to a sentiment estimation model.

In order to accurately determine the sentiment corresponding to the short text in each category, this implementation method builds a corresponding sentiment estimation model for all categories.

Referring to FIG. 4a, the process of estimating the sentiment degree corresponding to all categories includes the following steps:

Step S401: Determine a short text sample for constructing the sentiment estimation model.

a) Obtaining several objects in each category sent by the data providing device, and segmenting each object to obtain a short text set of each object.

The data providing device can send the objects in each category to the processing device, and the processing device can obtain multiple objects in each category. In order to facilitate subsequent processing, the processing device may segment each object according to punctuation marks, thereby segmenting each object into multiple short texts.

For example, taking the object of Taobao user evaluation as an example, a user under the clothing category evaluates "the clothes are very suitable, and the mother likes it", then after dividing according to the punctuation marks, you can get two short texts "the clothes are suitable" And "My mother loves it." Target short text. For example, a user under the category of electronic devices evaluates that "the mobile phone screen is large and the appearance is very beautiful", after dividing according to punctuation marks, two short texts "the mobile phone screen is large" and "the appearance is very beautiful" can be obtained.

b) Filter out short text samples from all short texts used to build the sentiment estimation model.

Through experiments, it is found that the execution process shown in FIG. 1 has a high accuracy rate for determining that a short text belongs to positive emotions and a low accuracy rate that determines that a short text belongs to negative emotions.

Therefore, in this step, the processing device may execute the process shown in FIG. 1 for each short text. If the process shown in FIG. 1 is performed, it is determined that one short text corresponds to a positive emotion. Then, it is determined that the short text can be used to construct an emotion degree estimation model, and the short text corresponds to positive emotion.

If the process shown in FIG. 1 is used, it is determined that a short text corresponds to a negative emotion. Then, further confirmation is performed manually. If a short text is negative emotion after manual confirmation, it is determined that the short text can be used to construct an emotion degree estimation model, and the short text corresponds to the negative emotion.

If a short text is positive emotion after manual confirmation, it indicates that the characteristics of the short text are not obvious, and it is not suitable as a short text for constructing a sentiment estimation model. The short text is therefore discarded.

Step S402: Determine a feature set corresponding to each short text.

In step S401, the segmentation result of each short text can be obtained by using the process shown in FIG. 1 (see step 1 in FIG. 1 for details, and details are not described herein again). Then, further determine the feature set corresponding to each short text.

This step can be implemented in two ways. The difference between the two methods is that the feature set determined in the first way includes combined features, while the feature set determined in the second way does not include combined features.

Since the feature set corresponding to each short text is determined to be consistent, a process of determining the feature set of the target short text is described in detail by taking a target short text as an example.

Referring to FIG. 4b, the specific process of the first execution manner of determining the feature set of the target short text:

Step 411: Obtain a category identifier corresponding to the short text to be processed and a segmentation result obtained after performing a word segmentation operation on the short text to be processed.

The processing device has obtained the segmentation result of the target short text in step S301. Since the target short text is consistent with the category of the object to be processed, the processing device may determine the category identifier of the object to be processed as the category identifier of the target short text.

Taking the target short text as the clothing category and "clothes are very large" as an example, the segmentation results corresponding to the target short text are "clothes", "very" and "big". Assuming the clothing category is identified as "16", the target The corresponding category identifier of the short text is "16".

Taking the target short text belongs to the category of electronic devices and the screen size is large as an example, the segmentation results corresponding to the target short text are "screen", "very" and "large", assuming that the electronic device category identifier is "10" The corresponding category identifier of the target short text is "10".

Step 412: Combine each participle and the category identifier to obtain each feature.

Since the word segmentation corresponding to the short text in different categories may be consistent, in order to fully consider the impact of the category on the short text, the present invention combines each word segmentation and category to obtain each feature.

Because features include category identifiers, and different category identifiers are different, using features can accurately distinguish word segmentation of different categories. In this way, the trained sentiment estimation model can accurately distinguish the same word segmentation in different categories.

Continuing the above example, taking the target short text “Clothing is very large” as an example, each feature corresponding to the target short text may be “Cloth 16”, “Very 16”, and “Large 16”. Taking the target short text “large screen” as an example, the respective features corresponding to the target short text may be “screen 10”, “very 10”, and “big 10”. From the perspective of features, the processing equipment can distinguish the segmentation "big 16" and "big 10" as two different features, and the two features belong to different categories.

In this example, the combination of the word segmentation and the category identifier is that the word segmentation comes first and the category target identifier follows, and the category tag comes first and the word segmentation comes after. Of course, there can be other combinations of word segmentation and category identification, which are not limited here.

Step 413: Perform n-element combination on each feature to obtain several combined features.

Because, through the research process, it was found that some features have a fixed match, such as "no color difference", "no color fade", "no pilling" and so on. For this fixed collocation, since both words are vocabularies of negative emotions, but the two words are superimposed to express positive emotions, if such words are separated, it will cause certain misjudgments. Therefore, this embodiment can perform feature combination.

Specifically, the features of each short text are combined using an n-gram language model. n is a non-zero natural number, and a meta in the n-gram language model corresponds to a participle in short text. The feature combination of the n-gram language model is specifically: merging adjacent n features together and n-1 features together until 2 features are merged together.

Taking n = 2 as an example, if the features of the target short text are "clothes 16", "very 16", and "big 16", after combining features using a binary language model, the combined feature is obtained as "clothes 16 and 16" "And" very 16 big 16 ".

Taking n = 3 as an example, if the features of the target short text are "clothes 16", "very 16", and "big 16", then the ternary language model is used to combine the features, and the combined feature is "clothes 16 to 16" Big 16 "," Clothing 16 is 16 "and" Very 16 Big 16 ".

Step 414: Determine each feature and a set of several combined features as the feature set of the target short text.

Continuing the above embodiment, taking the feature combination of the binary language model as an example, the feature set of the target short text finally obtained includes: "clothes 16", "very 16", "big 16", "clothes 16 very 16" And "very 16 big 16".

Referring to FIG. 4c, a specific process of a second execution manner for determining a feature set of a target short text:

Step 421: Obtain a category identifier corresponding to the short text to be processed and a segmentation result obtained after performing a word segmentation operation on the short text to be processed.

Step 422: Combine each participle and the category identifier to obtain each feature.

The execution process of steps S421 and S422 in FIG. 4c is consistent with steps S411 and S412 in FIG. 4b, and details are not described herein again.

Step 423: Determine the feature set as the feature set of the target short text.

The step of performing feature combination is missing in the execution process of FIG. 4c. Therefore, the feature set determined in step S422 can be directly determined as the feature set of the target short text.

Taking the target short text as "large clothes" as an example, the feature set of the target short text finally obtained after execution according to Fig. 4c includes: "clothes 16", "very 16", and "big 16".

Then return to FIG. 4a and proceed to step S403: determine the emotional tendency of each feature in the feature set corresponding to each short text, and the positive and negative emotion degrees of each feature, and compare each feature and the emotional tendency and positive feature corresponding to each feature The sentiment and negative sentiment are used as input parameters of the sentiment estimation model.

During the execution of the embodiment of FIG. 1 in step S401, the emotional tendency of the short text has been determined. The emotional tendency of each feature is consistent with that of short text. Therefore, when the short text corresponds to the positive emotion, it is determined that each feature in the feature set corresponds to the positive emotion; when the short text corresponds to the negative emotion, it is determined that each feature in the feature set corresponds to the negative emotion.

Taking a feature as an example, the process of determining the positive sentiment and negative sentiment of a feature is described in detail. The processing device can obtain a large number of the same feature, and the emotional tendencies corresponding to the feature may be the same or may be different.

Therefore, the processing device can count the total number of the features, and count the first number belonging to positive emotions and the second number belonging to negative emotions. The positive emotion degree of the feature is determined according to the proportional relationship between the first quantity and the total quantity; the negative emotion degree of the feature is determined according to the proportional relationship between the first quantity and the total quantity.

Step S404: Training is performed according to a preset classifier model, and an emotion degree estimation model obtained after training is obtained.

The preset classifier model may include a maximum entropy model, a support vector machine, a neural network algorithm, and so on. Relevant technical means have been provided for the training process, which will not be repeated here.

The following describes the second implementation of the processing device to construct the sentiment estimation model. In the second implementation, an emotional estimation model is constructed for each category. Therefore, since there is only one Category, so word segmentation is equivalent to features in the second implementation, so there is no need to combine word segmentation and category identification in the second implementation.

The construction process of the sentiment estimation model corresponding to each category is consistent. Therefore, taking a target category as an example, the process of constructing a target sentiment estimation model corresponding to the target category is described in detail.

Referring to FIG. 5, the process of constructing the target sentiment estimation model includes the following steps:

Step S501: Determine a short text sample for constructing a target sentiment estimation model.

a) Obtaining several objects under the target category sent by the data providing device, and segmenting each object to obtain a short text set of each object.

b) The short texts used to construct the sentiment estimation model are filtered out of all short texts.

The specific execution process of step S501 is similar to the execution process of step S401, and details are not described herein again.

Step S502: Determine a feature set corresponding to each short text.

In step S501, the segmentation result of each short text can be obtained by using the process shown in FIG. 1 (see step 1 in FIG. 1 for details, and details are not described herein again). Then, further determine the feature set corresponding to each short text. This step can be implemented in two ways. The difference between the two methods is that the feature set determined in the first way includes combined features, while the feature set determined in the second way does not include combined features.

Since the feature set corresponding to each short text is determined to be consistent, a process of determining the feature set of the target short text is described in detail by taking a target short text as an example.

Referring to FIG. 6a, the specific process of the first execution mode for determining the feature set of the target short text:

Step 601: Obtain a segmentation result corresponding to the target short text, and each segmentation corresponds to a feature.

Step 602: Perform n-element combination on each feature to obtain several combined features.

Step 603: Determine each feature and a set of several combined features as the feature set of the target short text.

Taking the short text to be processed as "large clothes" and taking the feature combination of the binary language model as an example, the feature set of the target short text finally obtained in this embodiment includes: "clothes", "very", "big", "Clothing is very" and "large."

Referring to FIG. 6b, a specific process of a second execution manner for determining a feature set of a target short text:

Step 611: Obtain a segmentation result corresponding to the target short text, and each segmentation corresponds to a feature.

Step 612: Determine the segmentation result as a feature set of the target short text.

The step of performing feature combination is missing in the execution process of FIG. 6b, so the set of each feature determined in step S611 can be directly determined as the feature set of the target short text.

Taking the target short text as "large clothes" as an example, the feature set of the target short text finally obtained after execution according to Fig. 6b includes: "clothes", "very", and "large".

Then return to FIG. 5 and proceed to step S503: determine the emotional tendency of each feature in the corresponding feature set of each short text under the target category, as well as the positive and negative emotional degrees of each feature, and compare each feature and each feature under the target category The corresponding sentiment tendencies, positive sentiments and negative sentiments are used as input parameters of the target sentiment estimation model.

During the execution of the embodiment of FIG. 1 in step S501, the emotional tendency of each short text has been determined. The emotional tendency of each feature is consistent with that of short text. Therefore, when the short text corresponds to a positive emotion, it is determined that each feature in the feature set corresponds to a positive emotion; when the short text corresponds to a negative emotion, it is determined that each feature in the feature set corresponds to a negative emotion.

Step S504: Perform training according to a preset classifier model, and obtain a target emotion degree estimation model obtained after training.

The preset classifier model may include a maximum entropy model, a support vector machine, a neural network algorithm, and so on. Relevant technical means have been provided for the training process, which will not be repeated here.

Fig. 5 is a process for constructing a sentiment estimation model for a category, and Fig. 3 is a process for constructing the sentiment estimation models for all categories. The processing steps of the two are very similar. The specific implementation process of 4 is not repeated here.

In the second implementation manner, each category corresponds to a sentiment estimation model. Therefore, in order to avoid confusion, after a sentiment estimation model is constructed, the processing device will also construct a mapping between the sentiment estimation model and the category identifier, so that subsequent processors can accurately determine the relationship between Emotional estimation model corresponding to category.

The third implementation manner of the processing device to construct the sentiment estimation model is described below.

In a third implementation manner, it may include: an emotion degree estimation model corresponding to two or more categories, and / or, an emotion degree estimation model corresponding to one category. For the construction process of the emotion estimation models corresponding to two or more categories, reference may be made to the embodiment shown in FIG. 4. For the sentiment estimation model corresponding to a category, reference may be made to the embodiment shown in FIG. 5, and details are not described herein again.

With reference to Figs. 2a and 2b, if the processing device for constructing the sentiment estimation model is the processor 200 itself, after the processor 200 completes the sentiment estimation model, it can be used directly to use the sentiment estimation model. Determine the emotional orientation of the short text to be processed.

When the processing device is the model construction device 300, the model construction device 300 sends the sentiment estimation model to the processor 200, so that the processor 200 uses the sentiment estimation model to determine the sentiment tendency of the short text to be processed.

The process of determining the emotional tendency of the short text to be processed by the processor 200 according to the sentiment estimation model is described below. Since there are three different implementations of the sentiment estimation model, the execution process of the processor 200 is not the same under different implementations. Therefore, the following describes the processor implementation in different implementations of the sentiment estimation model. Implementation process.

    The first:    

In the case where the emotion degree estimation model is implemented in the first implementation manner (all categories correspond to one emotion degree estimation model), the processor 200 adopts the following manner to determine the emotion tendency of the short text to be processed.

Referring to FIG. 7, a method for identifying an emotional tendency according to the present invention includes the following steps:

Step S701: Determine a feature set corresponding to the short text to be processed; wherein each feature in the feature set includes a word segmentation of the short text to be processed and a category identifier to which the text to be processed belongs.

Assuming that the first implementation method determines the feature set of short text in the process of determining the sentiment estimation model, the first execution method is also used in this step to determine the feature set of short text to be processed.

Referring to FIG. 8a, a first execution manner of determining a feature set corresponding to a short text to be processed specifically includes the following steps:

Step S801: Obtain a category identifier corresponding to the short text to be processed and a segmentation result obtained after performing a word segmentation operation on the short text to be processed.

Step S802: combining each participle in the segmentation result with the category identifier to obtain each feature.

Step S803: performing n-element combination on each feature to obtain several combined features.

Step S804: Determine each feature and a set of several combined features as the feature set of the short text to be processed.

For the execution process of FIG. 8a, refer to the execution process of FIG. 4a, and details are not described herein again.

Suppose that the first implementation method determines the feature set of short text in the second execution method in the process of determining the sentiment estimation model; then in this step, the second execution method is also used to determine the feature set of short text to be processed. .

Referring to FIG. 8b, a second execution manner of determining a feature set corresponding to the short text to be processed specifically includes the following steps:

Step S811: Obtain a category identifier corresponding to the short text to be processed and a segmentation result obtained after performing a word segmentation operation on the short text to be processed.

Step S812: Combine each participle in the segmentation result with the category identifier to obtain each feature.

Step S813: Determine the feature set as the feature set of the short text to be processed.

For the execution process of FIG. 8b, refer to the execution process of FIG. 4b, and details are not described herein again.

Then, return to FIG. 7 and step S702: according to the pre-trained sentiment estimation model and the feature set of the short text to be processed, perform the sentiment estimation on the short text to be processed; wherein the sentiment estimation model includes: A model that outputs positive and negative sentiment after training in two categories and several short text samples with emotional tendencies.

The processor inputs the feature set to the emotion degree estimation model, and outputs the positive emotion degree and the negative emotion degree corresponding to the feature set after being estimated by the emotion degree estimation model.

Step S703: Determine an emotional tendency corresponding to the short text to be processed based on the positive emotion level and the negative emotion level corresponding to the short text to be processed.

After determining the emotional tendency corresponding to the short text to be processed, the emotional tendency corresponding to the short text to be processed may also be output for use in other aspects.

After estimating the positive sentiment degree of the short text to be processed as positive emotion and the negative sentiment degree of the negative text to be processed in step S702, in order to further determine the emotional tendency of the short text to be processed, the positive sentiment degree and negative Emotional levels are compared. If the positive sentiment is greater than the negative sentiment, it is determined that the short text to be processed is a corresponding positive sentiment; if the negative sentiment is greater than the positive sentiment, it is determined that the short text to be processed corresponds to the negative sentiment.

In some cases, there is not much difference between positive affection and negative affection. Taking the probability expression as an example, the probability value of positive emotion is 0.51, and the probability value of negative emotion is 0.49. It can be understood that, because the degree of positive emotion and the degree of negative emotion are very close, it is theoretically impossible to accurately determine the emotional tendency of the short text to be processed. However, in this case, the emotional tendency of the short text to be processed is still determined according to the above paragraph, and an error will occur.

Therefore, referring to FIG. 9, the present invention provides the following ways to deal with the emotional tendency of short text.

Step S901: Determine a larger emotion degree between the positive emotion degree and the negative emotion degree.

Compare the positive sentiment and negative sentiment to determine the larger sentiment. If the degree of positive emotion is greater than the degree of negative emotion, the degree of positive emotion is determined to be a larger degree of emotion; if the degree of negative emotion is greater than the degree of positive emotion, the degree of negative emotion is determined to be a larger degree of emotion.

Step S902: It is determined whether the greater emotion degree is greater than a preset reliability.

In order to determine whether the larger emotional degree is credible, the present invention presets a preset reliability degree. The preset reliability is to determine the credibility of the larger emotion. Then, determine the magnitude of the larger emotion level and the preset reliability level.

Step S903: if the larger emotional degree is greater than the preset reliability, determine that the emotional tendency corresponding to the short text to be processed is consistent with the emotional tendency of the larger emotional degree.

If the greater sentiment is greater than the preset reliability, the credibility of determining the greater sentiment is higher. Therefore, the emotional tendency of the short text to be processed can be accurately determined. At this time, the emotional tendency of the short text to be processed is consistent with the emotional tendency of a larger emotional degree.

That is, if the larger emotion degree corresponds to the positive emotion degree, it is determined that the short text to be processed belongs to the positive emotion; if the larger emotion degree corresponds to the negative emotion degree, it is determined that the short text to be processed corresponds to the negative emotion.

Assuming the large sentiment is 0.8 and the preset reliability is 0.7, in this case, the sentiment tendency of the short text to be processed can be accurately determined.

Step S904: if the greater sentiment is not greater than the preset reliability, perform other processing procedures to determine the sentiment tendency of the text to be processed.

If the larger emotional degree is not greater than the preset reliability, the reliability of determining the larger emotional degree is low. Therefore, the emotional tendency of the short text to be processed cannot be accurately determined. Assuming that the large sentiment is 0.55 and the preset reliability is 0.7, in this case, the emotional tendency of the short text to be processed cannot be accurately determined.

In this case, some other processing may be performed to further determine the emotional tendency of the short text to be processed. This process is not the focus of the present invention and will not be repeated here.

The system shown in Figs. 2a and 2b may further include a receiving device (not shown in the figure) connected to the processor. After the processor determines the emotional tendency of the short text to be processed, the processor is further configured to output the emotional tendency of the text to be processed; and the receiving device is configured to receive the emotional tendency of the text to be processed, so that the receiving device can Take advantage of the emotional inclination of the text to be processed.

    The second:    

In the case where the emotion degree estimation model is implemented by using the second implementation manner, the processor 200 uses the following manner to determine the emotional tendency of the short text to be processed. Referring to FIG. 10, a method for identifying an emotional tendency according to the present invention specifically includes the following steps:

Step S1001: Determine a feature set and a category identifier corresponding to the short text to be processed.

It is assumed that in the process of determining the sentiment estimation model in the second implementation manner, the first execution manner is used to determine the feature set of short text; then in this step, the first execution manner is also used to determine the feature set of short text to be processed.

Referring to FIG. 11a, a specific process of determining a first implementation manner of a feature set of short text to be processed:

Step 1101: Obtain a segmentation result obtained after performing a segmentation operation on the short text to be processed.

Step 1102: Use the n-gram language model to perform segmentation and combination on each segmentation to obtain several combined segmentation.

Step 1103: Determine the set of each participle and several combined participles as the feature set of the short text to be processed, and one participle corresponds to one feature.

The execution process in FIG. 11a is similar to the execution process in FIG. 6a. For the specific execution process, refer to the execution process in FIG. 6a, and details are not described herein again.

Assuming that the second implementation method determines the feature set of short text in the process of determining the sentiment estimation model, the second execution method is also used in this step to determine the feature set of short text to be processed.

Referring to FIG. 11b, the specific process of the second execution manner for determining the feature set of the short text to be processed:

Step 1111: Obtain a segmentation result obtained after performing a segmentation operation on the short text to be processed.

Step 1112: Determine the segmentation result as the feature set of the short text to be processed, and one segmentation corresponds to one feature.

The execution process in FIG. 11b is similar to the execution process in FIG. 6b. For the specific execution process, refer to the execution process in FIG. 6a, and details are not described herein again.

Then return to FIG. 10 and proceed to step S1002: according to the sentiment degree estimation model corresponding to the category identifier, and combine the feature set of the short text to be processed, perform the sentiment degree estimation on the short text to be processed; The test model is a model that outputs the positive emotion degree and the negative emotion degree after training according to the feature set of several short text samples with emotional tendency corresponding to the category identifier.

In the second implementation, there are multiple sentiment estimation models. In order to obtain the sentiment estimation model suitable for the short text to be processed, multiple sentiment estimation models can be searched according to the category identifier to determine the sentiment estimation model corresponding to the category identification.

The processor inputs the feature set to the emotion degree estimation model, and outputs the positive emotion degree and the negative emotion degree corresponding to the feature set after being estimated by the emotion degree estimation model.

Step S1003: Determine an emotional tendency corresponding to the short text to be processed based on the positive emotion level and the negative emotion level corresponding to the short text to be processed. The execution process of this step is consistent with the execution process of step 703 in FIG. 7, and details are not described herein again.

In the system shown in FIG. 2a and FIG. 2b, a receiving device (not shown in the figure) connected to the processor may be further included. After the processor determines the emotional tendency corresponding to the short text to be processed, the processor is further configured to output the emotional tendency of the text to be processed; and the receiving device is configured to receive the emotional tendency of the text to be processed.

In the case where the emotion degree estimation model is implemented in the third implementation manner, the processor 200 stores the correspondence between the category identifier and the emotion degree estimation model in advance, and constructs each category identifier and the emotion degree estimation model in advance. The corresponding relationship of the construction method.

If the processor 200 receives a category identifier, it first judges the construction method of the emotion degree estimation model corresponding to the category identifier; if the emotion degree estimation model is constructed using the first implementation method, the adaptability is as shown in FIG. 4. The process shown determines the emotional tendency of the short text to be processed; that is, determines the feature set corresponding to the short text to be processed; wherein each feature in the feature set includes: the word segmentation of the short text to be processed and the to-be-processed The category identification to which the short text belongs; based on a pre-trained sentiment estimation model and the feature set of the short text to be processed, the sentiment estimation is performed on the short text to be processed; wherein the sentiment estimation model includes: A model that outputs positive and negative sentiment for two types of short text samples with sentiment tendency training; based on the positive and negative sentiment corresponding to the short text to be processed, determining the Emotional tendency corresponding to the short text to be processed.

If the sentiment estimation model is constructed using the second implementation method, the sentiment tendency of the short text to be processed is determined adaptively according to the process shown in FIG. 5. That is, a feature set corresponding to the short text to be processed is determined; wherein each feature in the feature set includes: a word segmentation of the short text to be processed; an emotion estimation model corresponding to the category identifier is combined with the Feature set for processing short texts, and sentiment estimation for short texts to be processed; wherein, the sentiment estimation model is obtained by training a number of short text samples corresponding to the category identifier with emotional tendencies. A model that outputs a positive sentiment and a negative sentiment; based on the positive sentiment and the negative sentiment corresponding to the short text to be processed, determines an emotional tendency corresponding to the short text to be processed. According to the embodiments shown in FIG. 7 and FIG. 10, it can be seen that the present invention has the following beneficial effects: The present invention provides a method for identifying emotional tendencies. This method uses several short texts with emotional tendencies for training and obtains the degree of emotion. Estimate model. Since each feature set contains the word segmentation and category identification of the short text, the sentiment estimation model constructed by the application fully considers the category to which the short text belongs. Therefore, the positive sentiment and negative sentiment of the short text to be processed determined based on the sentiment estimation model are more accurate than the prior art. Furthermore, the emotional tendency determined by using the positive emotion degree and the negative emotion degree is also more accurate.

The following uses the maximum entropy model as an example to describe in detail the training process for constructing the sentiment estimation model of the present invention. First, two matrices are constructed: matrix A and matrix B. Matrix A contains each feature and the positive sentiment and Negative emotion. Matrix B contains two classification results: positive emotions and negative emotions. For any feature a in matrix A, use b to represent its emotional tendency. f _i ( a , b ) indicates that (a, b) occurs together.

First calculate the expectation of f _i ( a , b ) in the training sample. Since there are no variables in the training model, the expected value is a constant after the calculation is completed. The specific calculation formula is as follows:

among them, ( f _i ) represents the expectation of f _i ( a , b ) in training sample i, ( a , b ) represents the empirical probability distribution of f _i ( a , b ) in the training sample.

The formula for the probability distribution of f _i ( a , b ) in the model is as follows:

among them, ( b ) represents the probability that the emotional tendency corresponding to the short text in the training sample is b, and p ( a | b ) represents the conditional probability of the feature a on the premise that the emotional tendency of the short text is b.

Then the calculation formula of f _i ( a , b ) in the maximum entropy model is:

In the maximum entropy model, the expectations of f _i ( a , b ) in the training samples should be consistent with the expectations of f _i ( a , b ) in the model. which is:

The Lagrange multiplier method is used to solve the optimal solution of the objective equation (2) under the constraint condition of formula (4). The optimal solution is as follows:

among them, Is the normalization factor such that , W _i is the weight of the feature f _i .

The formula (5) is substituted into the formula (1) to obtain the training result of the maximum entropy model, that is, the sentiment estimation model.

As shown in FIG. 12, the present invention provides an object classification method. Applied to the processor, in this embodiment, the emotion tendency of the short text of the object to be processed can be directly used to classify the object. It includes the following steps:

Step S1201: Determine short text information of the object to be processed, wherein the short text information includes an emotional tendency of the short text.

The processor can use punctuation to divide the object to be processed into several short texts, and each short text can determine its emotional tendency according to the process provided in FIG. 7 or FIG. 10 of the present invention, so that each short text in the object to be processed can be determined Emotional tendencies. In addition, the short text information may also include: the number of short texts that are positive emotions, the number of short texts that are negative emotions, the proportion of positive short texts, the proportion of negative short texts, and so on.

Step S1202: Perform category recognition on the short text information according to a pre-trained category recognition model; wherein the category recognition feature model is: the first category and the second category obtained by training based on short text information of several objects Classifier.

The class recognition model is a classifier that outputs the first class and the second class after training with short text information of several objects in advance. Specifically, a classification model such as a maximum entropy model, a neural network algorithm, or a support vector machine can be used to train short text information of several objects to obtain a category recognition model. Relevant technical means may adopt the training methods in the prior art, and details are not described herein again.

After the short text information of the object to be processed is obtained, the short text of the object to be processed is input to the category recognition model. After the category recognition model is processed, the category of the object to be processed can be determined.

In the actual process, it was found that for an object, in addition to the text, the object can also include an image. Taking the user evaluation whose object is the e-commerce system as an example, in addition to the text (character user evaluation), the user evaluation may also include an image of a product.

It can be understood that the object category determined by the short text information of the object alone is not accurate because the image feature information of the object is not considered; similarly, the object category determined by the image feature information of the object alone is not the same Accurate because the short text information of the subject is not taken into account. Therefore, in this embodiment, the short text information and the image feature information are combined, and the short text information and the image feature information are used to determine the object category together, thereby improving the accuracy of the object category.

The present invention further provides an object classification method. In this embodiment, multiple features of an object to be processed are used to classify the object. As shown in FIG. 13, it specifically includes the following steps:

Step S1301: Determine feature information corresponding to the object to be processed; wherein, the feature information includes short text information and image feature information, and the short text information includes emotional tendency of the short text.

The processor can use punctuation to divide the object to be processed into several short texts, and each short text can determine its emotional tendency according to the process provided in FIG. 7 or FIG. 10 of the present invention, so that each short text in the object to be processed can be determined Emotional tendencies. In addition, the short text information may also include: the number of short texts that are positive emotions, the number of short texts that are negative emotions, the proportion of positive short texts, the proportion of negative short texts, and so on.

The processor can process the image to obtain image feature information. The image feature information may include one or more of the following image features: image width, image height, number of faces in the image, number of sub-images included in the image, and whether the background of the image is Solid color, what is the percentage of the text area containing the image, the number of main colors in the significant area of the image, the number of main colors in the image, the score of the image psoriasis, the quality score of the image subject, the probability score of the image being a dummy model, Like the probability score of a real model, the image shows the probability score of product details, and so on.

Step S1302: classify the feature information according to a pre-trained class recognition model; wherein the class recognition feature model is: a classifier of the first class and the second class, which is trained based on the feature information of several objects .

The class recognition model is a classifier that outputs the first class and the second class after being trained in advance using short text information and image feature information of several objects. Specifically, a classification model such as a maximum entropy model, a neural network algorithm, or a support vector machine can be used to train short text information of several objects to obtain a category recognition model. Relevant technical means may adopt the training methods in the prior art, and details are not described herein again.

After obtaining the short text information of the object to be processed, the short text of the object to be processed is sent to the category recognition model, so as to determine the category of the object to be processed.

It can be understood that the more feature types in the feature information of the object to be processed, the more accurate the final result obtained. Therefore, in order to further improve the accuracy of the category of the object to be processed, the feature information may further include: the feature information attached to the to-be-processed object belonging to the first subject; and / or, the object to be processed belongs to the second subject. Feature information. Of course, other characteristic information can also be included, which will not be listed here one by one.

For example, taking the user evaluation as an example, the characteristic information attached to the first subject that belongs to the object to be processed is specifically: the characteristic information that belongs to the seller (the first subject) is attached to the product, for example, the credit rating of the seller and the sales volume of the seller Wait. The characteristic information belonging to the second subject attached to the object to be processed is specifically: the characteristic information belonging to the buyer (second subject) attached to the product, for example, the buyer's credit rating, the amount of non-default user evaluation data issued, and the release The number of user comments with pictures, and the percentage of user reviews with pictures.

After adding short text information, image feature information, and other feature information to the feature information, the feature information of the object will have multiple feature information. In order to comprehensively consider multiple feature information, this embodiment proposes to use a gradient boosting decision tree model to train several training samples to obtain a class recognition model.

Gradient lifting decision tree model is a lifting method based on decision tree. The gradient boosting decision tree model includes multiple decision trees. The reason for using multiple decision trees is to consider that for a single decision tree, it will cause overfitting due to excessive splitting and lose generalization ability; if there are too few splits, it will cause insufficient learning. full.

The following describes the training process of the gradient boosting decision tree model:

First, estimate the initial value F ₀ .

The initial value F ₀ may be a random value or may be equal to 0. The specific value may be determined according to actual conditions, and is not limited herein.

Second, iterate M times as follows to obtain M decision trees

A) Use the previous gradient boosting decision tree to update the estimated values of all the training samples corresponding to multiple feature information.

B) Randomly select some samples from all the training samples and use them as training samples for constructing the decision tree this time.

C) According to the features contained in the sample, calculate the information gain of each feature, and select the feature with the largest information gain for the first division. The left side represents the first category, and the right side represents the second category. Calculate the gradient of this time, and re-estimate the eigenvalue of the feature information of the sample in combination with the gradient.

The above steps are repeated J times to obtain the decision tree of J-level leaf nodes.

D) According to the obtained M decision trees, calculate the accuracy of the training samples on the decision tree, and use the accuracy rate as the weight of the decision tree.

Third, the M decision trees are linearly combined to obtain the final gradient boosted decision tree model.

The gradient boosting decision tree model includes multiple decision trees, which can be expressed as an addition model of multiple decision trees: F ( X ) = F ₀ + β ₁ T ₁ ( X ) + β ₂ T ₂ ( X ) + ... β _i T _i ( X ) ... + β _M T _M ( X ) ... Equation (6)

Among them, F ₀ is an initial value, T _i ( X ) represents the matching degree of the feature information of the object to be processed with a decision tree, β _i represents the weight of a decision tree, and M represents the total number of decision trees.

The use of multiple decision trees in the gradient boosting decision tree model is precisely to achieve good results in both training accuracy and generalization ability. As a boosting algorithm, the gradient boosting decision tree model naturally includes the idea of boosting: a series of weak classifiers are combined to form a strong classifier. It does not require each decision tree to learn too much, each tree learns a little knowledge, and then the knowledge learned by each decision tree is added up to form a powerful model.

The present invention further provides a method for classifying an object. As shown in FIG. 14, the method specifically includes the following steps:

Step S1401: Determine feature information corresponding to the object to be processed.

The feature information includes short text information, image feature information, feature information belonging to the first subject attached to the object to be processed, and feature information belonging to the second subject attached to the object to be processed. And, the short text information includes an emotional tendency of the short text.

Taking the object evaluation of the user as an example, this step may be: determining characteristic information of the user evaluation to be processed; wherein, the characteristic information includes text characteristic information of the user evaluation, image characteristic information of the user evaluation, seller characteristic information, Buyer characteristic information, and the text characteristic information includes the emotional tendency of the short text.

Step S1402: identify the feature information and a pre-trained gradient boosting decision tree model.

Taking the object as the user evaluation as an example, this step is to classify the feature information of the to-be-processed user evaluation based on a pre-trained gradient promotion decision tree model; wherein the class recognition model is based on several user evaluations. A classifier of the first type of user evaluation and the second type of user evaluation obtained after training on the feature information of the sample.

As shown in FIG. 15, specifically, this step includes the following steps:

Step S1501: The feature information is input to the category recognition model, that is, a gradient boosting decision tree model.

The gradient commission decision tree model has M trees, and the feature information is matched with the M trees, respectively, so as to obtain the category determined after matching each tree.

Step S1502: Determine a first category matching degree and a second category matching degree corresponding to the object to be processed.

The first category matching degree and the second category matching degree are determined according to the above formula 6.

First class matching degree F ₁ ( X ) = F ₀ + β ₁ T ₁ ( X ) + β ₂ T ₂ ( X ) + ... β _i T _i ( X ) ... + β _M T _M ( X ). Among them, T _i ( X ) represents the matching degree of feature information with a tree, and β _i represents the weight corresponding to the tree. If a tree determines that the feature information corresponds to the first category, the weight is β _i ; if a tree determines that the feature information corresponds to the second category, the weight is 0.

Second category matching degree F ₂ ( X ) = F ₀ + β ₁ T ₁ ( X ) + β ₂ T ₂ ( X ) + ... β _i T _i ( X ) ... + β _M T _M ( X ). Among them, T _i ( X ) represents the matching degree of the feature information with a tree, and β _i represents the weight corresponding to the tree. If a tree determines that the feature information corresponds to the second category, the weight is β _i ; if a tree determines that the feature information corresponds to the first category, the weight is 0.

Step S1503: Compare the first category matching degree and the second category matching degree. If the first category matching degree is greater than the second category matching degree, proceed to step S1504; if the second category matching degree is greater than the first category matching degree, proceed to step S1505.

Step S1504: Determine the category of the object to be processed as a first category.

Continue taking the object as the user evaluation as an example, then this step is to determine that the category of the user evaluation to be processed is the first category. The first category is high-quality user reviews, so this step is to determine the category of the user reviews to be processed as high-quality user reviews. Step S1505: Determine the category of the object to be processed as the second category.

Continue to take the object as the user evaluation as an example, this step is to determine the category of the user evaluation to be processed as the second category. The second category is inferior user evaluation, then this step is to determine the category of the user evaluation to be processed as inferior user evaluation.

After determining that the object to be processed is a first category, the object to be processed is added to an object set; and the objects in the object set are sent. The object set can be used by other devices. In the process of use, it can be screened again to determine multiple better object samples, and then send the object samples to the processor, so that the processor can use the better object samples to retrain the category. Recognize the model so that the category recognition model is more accurate. That is, the processor may receive multiple object samples, which are derived from the object set; add the multiple object samples to an existing object sample of the training category recognition model; based on the updated existing object Samples and retrain the category recognition model.

Continue to take the object as the user evaluation as an example, then the process is: after determining that the user evaluation to be processed is the first type of user evaluation, adding the user evaluation to be processed to the first type user evaluation set; and sending the The first type of user evaluation collection. The first user evaluation set can be used by users, and a better user evaluation can be determined in the first type of user evaluation set during use. A better user evaluation can then be sent to the processing device so that the processing device retrains the category recognition model. That is, the system can form a closed loop system.

That is, the processor receives a plurality of first-type user evaluations that are derived from the first-type user evaluation set; adding the plurality of first-type user evaluations to an existing category recognition model. In the user evaluation samples; based on the updated existing user evaluation samples, the class recognition model is retrained.

Referring to FIG. 16, the present invention provides an object classification system, including: a material providing device 100 for sending a plurality of objects.

The processor 200 is configured to receive a plurality of objects sent by the data providing device, and obtain and output category identification models of the first category and the second category after training according to the feature information of the objects; and used to determine the feature information of the object to be processed Wherein, the feature information includes text feature information and image feature information, and the text feature information includes short text's emotional tendency; class recognition of the feature information of the object to be processed according to the category recognition model ; Also used to output objects of the first category.

The material receiving device 400 is configured to receive and use the objects of the first category.

In the process of using the object collection, the data receiving device 400 may select a plurality of better object samples through screening again, and then send the object samples to the processor 200 so that the processor uses the better object samples to retrain the category. Recognize the model so that the category recognition model is more accurate.

Referring to FIG. 17, the present invention further provides an object classification system, including: a material providing device 100 for sending a plurality of objects.

A model building device 300 is configured to receive a plurality of objects sent by the data providing device, obtain and output category recognition models of the first category and the second category after training according to the feature information of the objects, and send the category recognition models. .

The processor 200 is configured to receive the category recognition model and determine feature information of an object to be processed; wherein the feature information includes text feature information and image feature information, and the text feature information includes emotion of short text Tendency; class recognition of feature information of the object to be processed according to the class recognition model; and also for outputting objects of a first class.

The material receiving device 400 is configured to receive and use the objects of the first category.

In the process of using the object collection, the data receiving device 400 may select a plurality of better object samples through screening again, and then send the object samples to the processor 200 so that the processor uses the better object samples to retrain the category. Recognize the model so that the category recognition model is more accurate.

The following uses a specific scenario embodiment to describe the object classification method in detail.

There are many user evaluations in the e-commerce system, and how to select high-quality user evaluations from the many user evaluations is a problem to be solved in this embodiment. Due to the large number and variety of user reviews in the e-commerce system, merchants need to spend a lot of time finding out good user reviews in the store, which invariably requires huge labor costs. At present, in the field of high-quality user evaluation and recognition, there are mainly two technologies commonly used in the industry: the first is a short text-based recognition technology; the second is an image feature-based recognition technology.

Recognition technology based on short text is relatively easy to implement, but it has some limitations: it does not pay attention to the image information posted by buyers in user evaluation. In actual scenarios, such as clothing, users are not only concerned about the text description part of the user evaluation, but also about the true appearance of the product, that is, the image feature information.

The recognition technology based on image features has significant effects, but it also has certain limitations. The high-quality user evaluation and recognition technology based on image features only uses the image information in user evaluation for identification, and does not care about the experience of the purchaser after the specific purchase, that is, short text information. Therefore, it can be seen that short text information and image feature information in user evaluation are equally important.

In addition, the applicant found that there are other features that can assist in determining quality user reviews. For example, seller characteristics and buyer characteristics. Therefore, in this embodiment, the above features are used as the basis for determining whether the user evaluation is a high-quality user evaluation or a low-quality user evaluation. For this reason, this embodiment proposes a machine learning method based on multiple feature fusion, that is, a gradient boosting decision tree model, to train several training samples, thereby obtaining a class recognition model.

As shown in FIG. 18, a flowchart for determining a quality user evaluation is provided for the present invention. From the figure, you can clearly determine the entire process of quality user evaluation. It consists of three parts:

    (1) Construction of user evaluation library    

To obtain a large number of user evaluations in the user evaluation server, firstly use a preprocessing rule to filter out some inferior user evaluations. The preprocessing rules can be: some requirements that images and texts need to meet in high-quality user evaluation, that is, using short text and a small number of dimensions in image features to filter a large number of user evaluations.

Specifically, the short texts in the high-quality user evaluation cannot be all negative emotions. Based on this, if the short texts in the user evaluations all correspond to negative emotions, it is determined as a non-high-quality user evaluation. There are also basic requirements for images in high-quality user evaluations. The resolution of the image reaches the preset resolution, the image is a non-conversation screen shot, the obvious advertising slogan in the image, and the watermark ratio is less than the preset value, etc.

The user evaluations in the user evaluation server that satisfy the above short text requirements and image feature requirements are put into a user evaluation database. For user evaluations that do not meet the requirements of short text and image features, these user evaluations are judged as high-quality user evaluations and are not placed in the user evaluation database.

The filtering of pre-processing rules can filter out some non-quality user evaluations. This can not only reduce the number of times the high-quality user evaluation recognition model is used, but also effectively filter out non-quality user evaluations, and improve the prediction accuracy of quality user evaluation recognition models Indeed.

    (2) Determine the quality user evaluation set    

The high-quality user evaluation recognition model is used to identify user evaluations in the user evaluation database. If the recognition result is a high-quality user evaluation, it is put into a high-quality user evaluation set.

    (3) Use a collection of high-quality user reviews.    

The data receiving device can obtain high-quality user evaluation from the high-quality user evaluation set, and use the high-quality evaluation in the actual application process. In the process of using the high-quality user evaluation in the high-quality user evaluation set by the data receiving device, the high-quality user evaluation in the high-quality evaluation set is re-screened according to the preset criteria, so as to select the high-quality user reviews that meet the preset criteria. Then, the high-quality user evaluation that meets the preset criteria is sent to the processor or model building device, so that the processor or model-building device iteratively updates the high-quality user evaluation recognition model.

    (4) Iterative update of high-quality user evaluation and recognition model.    

Utilize high-quality user evaluations that meet pre-set criteria and retrain the high-quality user evaluation identification model so that the high-quality user evaluation identification model can output high-quality user evaluations that meet user needs as much as possible.

Since the high-quality user evaluations selected in the high-quality user evaluation collection all meet the preset rules of the seller or the operating staff, these high-quality user evaluations are re-added into the user evaluation database, and the high-quality user evaluation identification model is updated and optimized so The high-quality user evaluation recognition model better identifies high-quality user evaluations that meet user expectations.

Based on the above process, it can be found that, in this embodiment, the user no longer needs to select one by one from the original user evaluation database, and only needs to select from the high-quality user evaluation set, which can quickly expect high-quality user evaluation, effectively reducing labor costs. . At the same time, the high-quality user evaluation model can effectively utilize the high-quality user evaluation provided by the merchant to iteratively update, thereby further identifying the high-quality user evaluation that meets the merchant's expectations.

If the functions described in the method of this embodiment are implemented in the form of a software functional unit and sold or used as an independent product, they can be stored in a computing device readable storage medium. Based on this understanding, the part of the embodiment of the present invention that contributes to the existing technology or the technical solution may be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for making one A computing device (which may be a personal computer, a server, a mobile computing device, or a network device, etc.) performs all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, etc. media.

Each embodiment in this specification is described in a progressive manner. Each embodiment focuses on differences from other embodiments, and the same or similar parts between the various embodiments may refer to each other.

The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims (30)

    A method for identifying sentiment tendency includes determining a category identifier corresponding to short text to be processed; wherein a text between two punctuation marks adjacent to a text is referred to as a short text; determining an emotion degree estimation corresponding to the category identifier Implementation of the model; if the implementation of the sentiment estimation model is that all categories correspond to one sentiment estimation model, then a feature set corresponding to the short text to be processed is determined; wherein each feature in the feature set includes : The word segmentation of the short text to be processed and the category identifier to which the short text to be processed belong; according to a pre-trained sentiment estimation model, combined with the feature set of the short text to be processed, the sentiment estimation of the short text to be processed ; Wherein the sentiment estimation model includes: a model that outputs positive sentiment and negative sentiment obtained after training according to at least two categories and several short text samples with sentiment tendency; based on the pending essay The corresponding positive sentiment and negative sentiment, determine the sentiment corresponding to the short text to be processed; if the sentiment is estimated The implementation of the type is that one category corresponds to one sentiment estimation model and determines the feature set corresponding to the short text to be processed; wherein each feature in the feature set includes: the word segmentation of the short text to be processed; The sentiment degree estimation model corresponding to the category identifier is combined with the feature set of the short text to be processed, and the sentiment degree estimation is performed on the short text to be processed. The emotion degree estimation model is: A model that outputs positive sentiment and negative sentiment obtained after training several short text samples with an emotional tendency; based on the positive sentiment and negative sentiment corresponding to the short text to be processed, determines the short processed text The corresponding emotional tendencies.         The method according to claim 1, wherein after determining an emotional tendency corresponding to the short text to be processed, the method further comprises: outputting an emotional tendency corresponding to the short text to be processed.         A method for identifying an emotional tendency includes: determining a feature set corresponding to short text to be processed; wherein a text between two punctuation marks adjacent to a text is called short text; each feature in the feature set includes: the The word segmentation of the short text to be processed and the category identifier to which the short text to be processed belong; according to a pre-trained sentiment estimation model and the feature set of the short text to be processed, the sentiment estimation of the short text to be processed; The sentiment estimation model includes: a model outputting positive sentiment and negative sentiment obtained after training according to at least two types of short text samples with emotional tendencies; and based on the corresponding short text to be processed The positive emotion degree and the negative emotion degree determine the emotion tendency corresponding to the short text to be processed.         The method according to claim 3, wherein the determining a feature set corresponding to the short text to be processed comprises: obtaining a category identifier corresponding to the short text to be processed, and obtaining the word segmentation operation after the short text to be processed is obtained The segmentation result of the combination; combining each segmentation in the segmentation result with the category identifier to obtain each feature; and determining the set of each feature as the feature set of the short text to be processed.         The method according to claim 3, wherein the determining a feature set corresponding to the short text to be processed comprises: obtaining a category identifier corresponding to the short text to be processed, and obtaining the word segmentation operation after the short text to be processed is obtained Word segmentation result; combining each segmentation word in the word segmentation result with the category identifier to obtain each feature; using an n-gram language model to combine the features to obtain several combined features; combining each feature with The set of several combined features is determined as a feature set of the short text to be processed.         The method according to claim 5, wherein the using the n-ary language model to combine the features to obtain several combined features includes: using a binary language model to combine the features to obtain Several combined features.         The method according to claim 3, wherein the estimating the sentiment of the short text to be processed according to the pre-trained sentiment estimation model and the feature set of the short text to be processed includes: inputting the feature set To the emotion degree estimation model; after the estimation by the emotion degree estimation model, the positive emotion degree and the negative emotion degree corresponding to the short text to be processed are output.         The method according to claim 3, wherein determining the emotional tendency corresponding to the short text to be processed based on the positive emotion level and the negative emotion level corresponding to the short text to be processed includes: determining the positive emotion level And the negative emotion degree; determine whether the larger emotion degree is greater than the preset reliability; if the larger emotion degree is greater than the preset reliability, determine the short text to be processed The corresponding emotional tendency is consistent with the larger emotional degree.         The method according to claim 3, wherein the sentiment estimation model includes: using a maximum entropy model and outputting a positive sentiment obtained after training according to a feature set of several short texts corresponding to at least two category identifiers And negative sentiment models.         The method according to claim 3, wherein after determining an emotional tendency corresponding to the short text to be processed, the method further comprises: outputting an emotional tendency corresponding to the short text to be processed.         A method for identifying emotional tendency includes: determining a feature set and a category identifier corresponding to short text to be processed; wherein a text between two punctuation marks adjacent to a text is called short text; each feature in the feature set The method includes the word segmentation of the short text to be processed, an emotion level estimation model based on the sentiment degree estimation model corresponding to the category identifier, and a feature set of the short text to be processed, and the emotion. The degree estimation model is: a model that outputs positive sentiment and negative sentiment after training according to the short text samples corresponding to the category identifier and several short text samples with emotional tendencies; based on the corresponding short text to be processed The positive emotion degree and the negative emotion degree determine the emotion tendency corresponding to the short text to be processed.         The method according to claim 11, wherein the determining a feature set corresponding to the short text to be processed includes: obtaining a segmentation result obtained after performing the word segmentation operation on the short text to be processed; and using an n-gram language model to perform each segmentation. Combine word segmentation to obtain several combined word segmentation; determine the set of each word segmentation and several combined word segmentation as the feature set of the short text to be processed, one word segment corresponds to one feature.         The method according to claim 11, wherein the determining a feature set corresponding to the short text to be processed comprises: obtaining a segmentation result obtained after performing a word segmentation operation on the short text to be processed; and determining the segmentation result as the The feature set of the short text to be processed is described, and one segmentation corresponds to one feature.         The method according to claim 11, wherein after determining the emotional tendency corresponding to the short text to be processed, the method further comprises: outputting the emotional tendency corresponding to the short text to be processed.         An emotion tendency recognition system includes: a data providing device for sending a plurality of objects; a processor for receiving a plurality of objects sent by the data providing device, and constructing a sentiment estimation model based on short texts of the plurality of objects , And use the sentiment estimation model to determine the sentiment tendency of short text to be processed.         The system according to claim 15, wherein the processor is further configured to construct a correspondence between an emotion degree estimation model and a category identifier to which the object belongs.         The system according to claim 15, wherein the system further comprises a receiving device; the processor is further configured to output the emotional tendency of the text to be processed; and the receiving device is configured to receive the text to be processed Emotional tendencies.         An emotional tendency recognition system includes: a data providing device for sending several objects; a model building device for receiving several objects sent by the data providing device; and constructing an emotion degree estimation based on short texts of the objects A model, and sends the sentiment estimation model; a processor is configured to receive the sentiment estimation model and use the sentiment estimation model to determine an emotional tendency of a short text to be processed.         The system according to claim 18, wherein the model construction device is further configured to construct a correspondence between the sentiment estimation model and a category identifier to which the object belongs, and send the correspondence to the processor.         The system according to claim 18, wherein the system further comprises a receiving device; the processor is further configured to output the emotional tendency of the text to be processed; and the receiving device is configured to receive the text to be processed Emotional tendencies.         An object classification method includes: determining feature information of an object to be processed; wherein the feature information includes text feature information and image feature information, and the text feature information includes an emotional tendency of a short text; according to a pre-trained category The recognition model performs category recognition on the feature information of the object to be processed. The category recognition model is a classifier of the first category and the second category, which is obtained by training based on the feature information of a plurality of object samples.         The method according to claim 21, wherein the feature information further comprises: constructing feature information of a first subject of the object; and / or, attaching feature information belonging to the second subject to the object.         The method according to claim 21, wherein the classifying the feature information according to a pre-trained class recognition model comprises: entering the feature information into the class recognition model; determining the pending processing The first category matching degree and the second category matching degree corresponding to the object; comparing the first category matching degree and the second category matching degree; if the first category matching degree is greater than the second category matching degree, determining the The category of the processing object is the first category; if the matching degree of the second category is greater than the matching degree of the first category, it is determined that the category of the object to be processed is the second category.         The method according to claim 23, further comprising: after determining that the object to be processed is a first category, adding the object to be processed to an object collection; and sending the objects in the object collection.         The method according to claim 24, further comprising: receiving a plurality of object samples, the object samples originating from the object set, and satisfying a preset rule; adding the plurality of object samples to a training category Among the existing object samples of the recognition model; based on the updated existing object samples, the class recognition model is retrained.         A method for classifying user evaluations, comprising: determining characteristic information of user evaluations to be processed; wherein the characteristic information includes text characteristic information of user evaluations, image characteristic information of user evaluations, characteristic information of sellers, and buyer's Feature information, and the text feature information includes the emotional tendency of short text; class recognition is performed on the feature information evaluated by the user to be processed according to a pre-trained gradient promotion decision tree model; wherein the category recognition model is: Classifiers of the first type of user evaluation and the second type of user evaluation obtained after training based on the feature information of several user evaluation samples.         The method according to claim 26, further comprising: after determining that the to-be-processed user evaluation is a first-type user evaluation, adding the to-be-processed user evaluation to a first-type user evaluation set; and sending the The first type of user evaluation collection.         The method according to claim 26, further comprising: receiving a plurality of first-type user evaluations, the first-type user evaluations originating from the first-type user evaluation set; and combining the plurality of first-type users The evaluation is added to the existing user evaluation samples of the category recognition model; based on the updated existing user evaluation samples, the category recognition model is retrained.         An object classification system includes: a data providing device for sending a plurality of objects; a processor for receiving a plurality of objects sent by the data providing device, and obtaining and outputting a first category and A category recognition model of the second category; used to determine feature information of an object to be processed; wherein the feature information includes text feature information and image feature information, and the text feature information includes short-text emotional tendencies; The category recognition model performs category recognition on the feature information of the object to be processed; it is also used to output objects of the first category; and a data receiving device for receiving and using the objects of the first category.         An object classification system includes: a data providing device for sending a plurality of objects; a model building device for receiving a plurality of objects sent by the data providing device, and obtaining and outputting a first number after training based on characteristic information of the plurality of objects A category recognition model of a category and a second category, and send the category recognition model; a processor, configured to receive the category recognition model and determine feature information of an object to be processed; wherein the feature information includes text feature information and Image feature information, and the text feature information includes an emotional tendency of a short text; class recognition is performed on the feature information of the object to be processed according to the category recognition model; and is also used to output the object of the first category; data A receiving device for receiving and using the objects of the first category.