Classification model training method, classification method, device, and medium
View Patent ↗Embodiments of this application disclose a classification model training method, a classification method, a device, and a medium. An initial classification model is first trained by using a first sample set including a large quantity of first samples, to obtain a pre-trained model, each first sample including a social text and an emoticon label corresponding to the social text; and the pre-trained model is then trained by using a second sample set including a small quantity of second samples, to obtain a social text sentiment classification model that uses a social text as an input and use a sentiment class probability distribution corresponding to the social text as an output. In this method, the model is trained by combining a large quantity of weakly supervised samples with a small quantity of supervised samples, to ensure that the model obtained through training has better model performance without increasing manually labeled samples.
1. A method performed at a computing device having one or more processors and memory storing a plurality of computer-readable instructions to be executed by the one or more processors, the method comprising:
training an initial classification model by using a first sample set, to obtain a pre-trained model, the first sample set comprising a quantity of weakly supervised first samples, each first sample comprising a first social text and an emoticon label in the first social text;
training the pre-trained model by using a second sample set, to obtain a social text sentiment classification model, the second sample set comprising a quantity of supervised second samples, each second sample comprising a second social text and a manually-added sentiment classification label corresponding to the second social text; and
applying a target social text to the social text sentiment classification model as an input to obtain a sentiment class probability distribution corresponding to the target social text as an output.
2. The method according to claim 1 , wherein the first sample set is generated by:
acquiring a plurality of social texts, and generating a social text set, the plurality of social texts comprising emoticons;
determining, for each emoticon in the social text set, a sentiment ability value of the emoticon according to a number of times the emoticon occurs in the social text set and a number of times the emoticon co-occurs with each sentiment word in a sentiment dictionary;
determining, for each social text in the social text set, an emoticon with a highest sentiment ability value comprised in the social text as an emoticon label corresponding to the social text; and
generating the first sample set according to the social texts and the emoticon labels corresponding to the social texts in the social text set.
3. The method according to claim 2 , wherein the generating the first sample set according to the social texts and the emoticon labels corresponding to the social texts in the social text set comprises:
extracting social texts corresponding to different emoticons from the social text set based on a principle that quantities of social texts corresponding to different emoticons are the same; and
generating the first sample set according to the extracted social texts and emoticon labels corresponding to the extracted social texts.
4. The method according to claim 2 , further comprising:
after determining the sentiment ability value of the emoticon, selecting, according to a descending order of the sentiment ability values of the emoticons, top M emoticons from the social text set as target emoticons, M being a positive integer; and
the determining an emoticon with a highest sentiment ability value comprised in the social text as an emoticon label corresponding to the social text comprises:
determining a target emoticon with the highest sentiment ability value in the social text as the emoticon label corresponding to the social text.
5. The method according to claim 1 , wherein the initial classification model comprises a word embedding layer, a convolutional layer, at least one feature extraction layer, and a classification layer, and the feature extraction layer comprises a plurality of convolutional layers and one pooling layer, and each convolutional layer is connected to all subsequent convolutional layers; and
the training the pre-trained model by using a second sample set, to obtain a social text sentiment classification model comprises:
performing functional modularization processing on the pre-trained model, to obtain a plurality of submodules, each submodule comprising at least one computing layer; and
sequentially training the submodule in a manner of freezing and unfreezing weights by using the second sample set, until all the submodules are in a convergence state, and using a model comprising a plurality of submodules in the convergent state as the social text sentiment classification model,
wherein the manner of freezing and unfreezing weights means that during training of a submodule, a weight of the submodule is unfrozen to train the submodule while weights of remaining submodules are frozen.
6. The method according to claim 1 , wherein the target social text comprises a microblog, and the sentiment classification label is used for representing a sentiment polarity of the microblog, classes of the sentiment polarity comprising a positive polarity, a negative polarity and a neutral.
7. The method according to claim 1 , further comprising:
determining a sentiment class of the target social text according to the sentiment class probability distribution corresponding to the target social text.
8. The method according to claim 7 , wherein the target social text is a chat text obtained through a chat interface and the method further comprises:
determining a reply content matching the sentiment class of the target social text; and
displaying the reply content on the chat interface.
9. The method according to claim 7 , wherein the target social text is an object comment text by a specified user and the method further comprises:
recommending an object to the specified user according to the sentiment class of the target social text.
10. The method according to claim 7 , wherein the target social text is related to a specified product and the method further comprises:
generating a public opinion report of the specified product according to respective sentiment classes of the target social text related to the specified product.
11. A computing device, comprising a processor and a memory,
the memory being configured to store a plurality of computer-readable instructions; and
the processor being configured to perform a plurality of operations according to the plurality of computer-readable instructions, the plurality of operations including:
training an initial classification model by using a first sample set, to obtain a pre-trained model, the first sample set comprising a quantity of weakly supervised first samples, each first sample comprising a first social text and an emoticon label in the first social text;
training the pre-trained model by using a second sample set, to obtain a social text sentiment classification model, the second sample set comprising a quantity of supervised second samples, each second sample comprising a second social text and a manually-added sentiment classification label corresponding to the second social text; and
applying a target social text to the social text sentiment classification model as an input to obtain a sentiment class probability distribution corresponding to the target social text as an output.
12. The computing device according to claim 11 , wherein the first sample set is generated by:
acquiring a plurality of social texts, and generating a social text set, the social texts comprising emoticons;
determining a sentiment ability value of each emoticon according to a number of times the emoticon occurs in the social text set and a number of times the emoticon co-occurs with each sentiment word in a sentiment dictionary;
determining, based on the social texts in the social text set and the sentiment ability values of the emoticons, an emoticon with the highest sentiment ability value comprised in the social text as an emoticon label corresponding to the social text; and
generating the first sample set according to the social texts and the emoticon labels corresponding to the social texts in the social text set.
13. The computing device according to claim 12 , wherein the generating the first sample set according to the social texts and the emoticon labels corresponding to the social texts in the social text set comprises:
extracting social texts corresponding to different emoticons from the social text set based on a principle that quantities of social texts corresponding to different emoticons are the same; and
generating the first sample set according to the extracted social texts and emoticon labels corresponding to the extracted social texts.
14. The computing device according to claim 12 , wherein the plurality of operations further comprise:
after determining the sentiment ability value of the emoticon, selecting, according to a descending order of the sentiment ability values of the emoticons, top M emoticons from the social text set as target emoticons, M being a positive integer; and
determining a target emoticon with the highest sentiment ability value in the social text as the emoticon label corresponding to the social text.
15. The computing device according to claim 11 , wherein the initial classification model comprises a word embedding layer, a convolutional layer, at least one feature extraction layer, and a classification layer, and the feature extraction layer comprises a plurality of convolutional layers and one pooling layer, and each convolutional layer is connected to all subsequent convolutional layers; and
the processor is configured to perform the following operations according to instructions in the program code:
performing functional modularization processing on the pre-trained model, to obtain a plurality of submodules, each submodule comprising at least one computing layer; and
sequentially training the submodule in a manner of freezing and unfreezing weights by using the second sample set, until all the submodules are in a convergence state, and using a model comprising a plurality of submodules in the convergent state as the social text sentiment classification model,
wherein the manner of freezing and unfreezing weights means that during training of a submodule, a weight of the submodule is unfrozen to train the submodule while weights of remaining submodules are frozen.
16. The computing device according to claim 11 , wherein the target social text comprises a microblog, and the sentiment classification label is used for representing a sentiment polarity of the microblog, classes of the sentiment polarity comprising a positive polarity, a negative polarity and a neutral.
17. The computing device according to claim 11 , wherein the plurality of operations further comprise:
determining a sentiment class of the target social text according to the sentiment class probability distribution corresponding to the target social text.
18. A non-transitory computer-readable storage medium storing a plurality of computer-readable instructions that, when executed by a processor of a computing device, cause the computing device to perform a plurality of operations including:
training an initial classification model by using a first sample set, to obtain a pre-trained model, the first sample set comprising a quantity of weakly supervised first samples, each first sample comprising a first social text and an emoticon label in the first social text;
training the pre-trained model by using a second sample set, to obtain a social text sentiment classification model, the second sample set comprising a quantity of supervised second samples, each second sample comprising a second social text and a manually-added sentiment classification label corresponding to the second social text; and
applying a target social text to the social text sentiment classification model as an input to obtain a sentiment class probability distribution corresponding to the target social text as an output.
19. The non-transitory computer-readable storage medium according to claim 18 , wherein the first sample set is generated by:
acquiring a plurality of social texts, and generating a social text set, the social texts comprising emoticons;
determining a sentiment ability value of each emoticon according to a number of times the emoticon occurs in the social text set and a number of times the emoticon co-occurs with each sentiment word in a sentiment dictionary;
determining, based on the social texts in the social text set and the sentiment ability values of the emoticons, an emoticon with the highest sentiment ability value comprised in the social text as an emoticon label corresponding to the social text; and
generating the first sample set according to the social texts and the emoticon labels corresponding to the social texts in the social text set.
20. The non-transitory computer-readable storage medium according to claim 19 , wherein the generating the first sample set according to the social texts and the emoticon labels corresponding to the social texts in the social text set comprises:
extracting social texts corresponding to different emoticons from the social text set based on a principle that quantities of social texts corresponding to different emoticons are the same; and
generating the first sample set according to the extracted social texts and emoticon labels corresponding to the extracted social texts.