KR102410715B1 - Apparatus and method for analyzing sentiment of text data based on machine learning - Google Patents

Abstract

The apparatus for analyzing sentiment analysis of machine learning-based text data according to an embodiment disclosed in this document includes a communication circuit, a memory, and a processor electrically connected to the communication circuit and the memory, and the processor is one or more external using the communication circuit. A topic label indicating the topic of the corpus collected among a plurality of topics together with a corpus collected from the document and a sentiment label indicating the sentiment of the corpus collected among a plurality of sentiments are acquired as a training data set, and , storing the acquired learning data set in the memory, and outputting sentiment information for the input corpus by inputting a plurality of training data sets of at least a specified number for each of all combinations of a plurality of subjects and a plurality of emotions into a specified language modeling algorithm By acquiring a single global sentiment classification model configured to

Description

Apparatus and method for sentiment analysis of text data based on machine learning

Embodiments disclosed in this document relate to an apparatus and method for analyzing consumer sentiment from text data based on machine learning.

With the development of communication technology, a large number of users can upload text through various types of media, and the amount of text is continuously increasing. Therefore, when analyzing multiple web pages containing text, such as social network service (SNS), blogs, and websites, it is possible to understand the interests, public opinion, opinions and emotions of various users by using the vast amount of information. have. The above-described text may be reprocessed into various types of information through analysis.

By analyzing a user's review of a specific product or service, the polarity (eg, positive or negative) of the user's sentiment toward the product or service can be identified. Analyzing the voice of customer (VOC) using natural language processing can quickly and efficiently quantify the emotions embedded in a large number of customer opinions.

The emotion inherent in the corpus may be determined based on a dictionary database or an artificial neural network. When emotion is analyzed based on an artificial neural network, machine learning may be performed using a corpus including customer opinions and a label indicating the emotion inherent in the corpus. When performing machine learning, a corpus associated with a specific topic may be used to derive accurate results. For example, machine learning may be performed using customer opinions related to a hotel, and a model derived by machine learning may provide excellent performance only for customer opinions related to a hotel. In other words, since accurate sentiment analysis results cannot be provided for new topics that have not been trained in advance, whenever a new topic appears, a large amount of training data corresponding to the new topic is collected again and machine learning is performed to develop a model corresponding to the new topic. There is a need to calculate

Embodiments of the present invention are to provide an analysis model capable of outputting an accurate sentiment analysis result for a customer opinion related to a topic that has not been learned in advance.

The apparatus for analyzing sentiment analysis of machine learning-based text data according to an embodiment disclosed in this document includes a communication circuit, a memory, and a processor electrically connected to the communication circuit and the memory, and the processor is one or more external using the communication circuit. A topic label indicating the topic of the corpus collected among a plurality of topics together with a corpus collected from the document and a sentiment label indicating the sentiment of the corpus collected among a plurality of sentiments are acquired as a training data set, and , storing the acquired learning data set in the memory, and outputting sentiment information for the input corpus by inputting a plurality of training data sets of at least a specified number for each of all combinations of a plurality of subjects and a plurality of emotions into a specified language modeling algorithm By acquiring a single global sentiment classification model configured to

According to an embodiment, the processor obtains a topic classification model configured to output topic information about an input corpus by inputting a plurality of training data sets into a specified language modeling algorithm, and divides the plurality of training data sets into a plurality of topics. It is configured to output sentiment information about the input corpus by classifying and inputting it into a designated language modeling algorithm, and obtains a plurality of local sentiment classification models corresponding to each of a plurality of subjects, and sets the analysis target corpus into a topic classification model and a global sentiment classification model. and by inputting the input to a local emotion classification model selected from a plurality of local emotion classification models, the emotion information on the analysis target corpus may be output.

According to an embodiment, the processor obtains similarity information for each of the analysis target corpus and the plurality of topics by inputting the analysis target corpus into the topic classification model, and classifies local emotions selected from a plurality of local sentiment classification models based on the similarity information model can be determined.

According to an embodiment, the processor may determine sentiment information on the analysis target corpus based on analysis results output by each of the global sentiment classification model and the selected local sentiment classification model.

In the method for analyzing emotion of text data based on machine learning according to an embodiment disclosed in this document, a subject label indicating the subject of the collected corpus among a plurality of subjects together with a corpus collected from one or more external documents and a plurality of emotions are collected Acquiring as a training data set a sentiment label representing the sentiment of the corpus, storing the acquired training data set, designating a plurality of training data sets greater than or equal to a specified number for each of a plurality of subjects and all combinations of a plurality of emotions Obtaining a single global sentiment classification model configured to output sentiment information for the input corpus by input to the language modeling algorithm, and inputting the analysis target corpus into the global sentiment classification model to provide sentiment information on the analysis target corpus can do.

According to the embodiments disclosed in this document, by using a global sentiment classification model calculated by evenly using various topics and emotional learning data sets, it is possible to accurately determine sentiment information on a corpus associated with various topics.

In addition, by utilizing the topic classification model for judging the subject of the corpus and the local sentiment classification model corresponding to a topic with high similarity to the corpus together with the global sentiment classification model, the accuracy of judgment on the emotional information of the corpus for a new topic is further improved. can be improved

In addition, various effects directly or indirectly identified through this document may be provided.

1 illustrates an operating environment of an apparatus for analyzing emotion of text data based on machine learning according to an embodiment.
2 is a block diagram illustrating a configuration of an apparatus for analyzing emotion of text data based on machine learning according to an embodiment.
3 is a diagram for explaining an exemplary operation of an apparatus for analyzing emotion of text data based on machine learning according to an embodiment.
4 is a diagram for explaining an exemplary operation of an apparatus for analyzing emotion of text data based on machine learning according to an embodiment.
5 is a diagram for explaining an exemplary operation of an apparatus for analyzing emotion of text data based on machine learning according to an embodiment.
6 is a diagram for explaining an exemplary operation of an apparatus for analyzing emotion of text data based on machine learning according to an embodiment.
7 is a diagram for explaining an exemplary operation of an apparatus for analyzing emotion of text data based on machine learning according to an embodiment.
FIG. 8 is a flowchart illustrating a method for analyzing sentiment of text data based on machine learning according to an embodiment of FIG. 3 .
In connection with the description of the drawings, the same or similar reference numerals may be used for the same or similar components.

Hereinafter, some embodiments of the present invention will be described in detail with reference to exemplary drawings. However, this is not intended to limit the present invention to specific embodiments, and it should be understood that various modifications, equivalents or substitutes of the embodiments of the present invention are included. In adding reference numerals to the components of each drawing, it should be noted that the same components are given the same reference numerals as much as possible even though they are indicated on different drawings. In addition, in describing the embodiment of the present invention, if it is determined that a detailed description of a related known configuration or function interferes with the understanding of the embodiment of the present invention, the detailed description thereof will be omitted.

1 illustrates an operating environment of an apparatus for analyzing emotion of text data based on machine learning according to an embodiment.

Referring to FIG. 1 , an apparatus for analyzing sentiment of text data according to an embodiment may be implemented in the form of a server 100 . The server 100 may crawl various texts from the external device 12 . The server 100 may collect text from various channels or platforms. For example, the server 100 may collect a corpus from external electronic documents such as social network service (SNS), blogs, various web sites, web pages, and web documents. The server 100 may collect a corpus including customer opinions on various topics.

The user terminal 11 may input a topic label indicating a topic of the corpus and a sentiment label indicating a sentiment contained in the corpus into the corpus collected by the server 100 . The subject label and the sentiment label may be assigned according to a user's input through the user terminal 11 . For example, when the corpus of “the room is clean and comfortable” is collected, the subject label “hotel” and the emotion label “positive” may be input by the user terminal 11 . The corpus, the subject label corresponding to the corpus, and the sentiment label may be stored in the server 100 as a training data set.

The server 100 may store a plurality of training data sets. The server 100 may perform machine learning by classifying a plurality of learning data sets by subject and emotion, and inputting more than a specified number of learning data sets for each classification into a language modeling algorithm. By evenly using a training data set including various topic labels and sentiment labels, a single global sentiment classification model that can respond to all topics can be generated. A topic classification model and a plurality of local sentiment classification models corresponding to each of various topics may be generated together with the global sentiment classification model.

When the above-described model is generated, the server 100 may receive the analysis target corpus, and may output emotional information on the analysis target corpus by inputting the received analysis target corpus into the model. Even if the analysis target corpus does not directly correspond to the pre-learned topic, accurate analysis results can be derived because learning on various topics is made.

Although the server 100 and the user terminal 11 are illustrated as being separated in FIG. 1 , the present invention is not limited thereto, and may be integrated into one computing device. In addition, in FIG. 1 , the server 100 and the user terminal 11 are each shown as one device, but the present invention is not limited thereto, and the server 100 and the user terminal 11 are implemented separately as two or more computing devices, respectively. it might be

In addition, at least some of the operations performed by the server 100 described with reference to FIG. 1 may be performed by the user terminal 11 . For example, an operation of collecting a corpus from an external document, an operation of storing a training data set, an operation of performing machine learning to generate a model, and an operation of outputting sentiment information on the corpus to be analyzed using the generated model etc. may be performed by the user terminal 11 .

2 is a block diagram illustrating a configuration of an apparatus 200 for analyzing machine learning-based text data according to an embodiment.

Referring to FIG. 2 , the apparatus 200 for analyzing text data according to an embodiment may include a communication circuit 210 , a memory 220 , and a processor 230 . The sentiment analysis apparatus 200 for text data may be implemented in the form of a server or may be implemented in the form of a user terminal 21 . When the sentiment analysis apparatus 200 of text data is implemented in the form of the user terminal 21, the function of the user terminal 21 shown in FIG. 2 may be performed by the sentiment analysis apparatus 200 of text data. .

The communication circuit 210 may communicate with the user terminal 21 and the external server 22 wirelessly or by wire. The communication circuit 210 may transmit/receive data to and from the user terminal 21 and the external server 22 .

Memory 220 may include volatile memory and/or non-volatile memory. The memory 220 may store various data handled by the sentiment analysis apparatus 200 of text data. For example, the memory 220 may store data processed inside the sentiment analysis apparatus 200 of text data, and may store data received from the user terminal 21 and the external server 22 .

The processor 230 may be electrically connected to the communication circuit 210 and the memory 220 . The processor 230 may control the communication circuit 210 and the memory 220 , and may perform various data processing and operations. Although the processor 230 is illustrated as a single configuration in FIG. 2 , it may be implemented as a plurality of separate components. The processor 230 may perform the following operations by executing software or instructions stored in the memory 220 .

According to an embodiment, the processor 230 may collect a corpus from one or more external documents using the communication circuit 210 . For example, the processor 230 may collect a corpus including reviews, reviews, or evaluations of various products or services from external electronic documents such as various web sites, web pages, and web documents. As another example, the corpus may be collected by the user terminal 21 .

According to an embodiment, the processor 230 may receive a subject label and a sentiment label for each of the collected corpus. The subject label may indicate the subject of the collected corpus among various subjects. Subjects are corpora-related services (e.g., accommodation, restaurants, cafes, movies, healthcare, education, finance, postal/delivery, tourism or follow-up services, etc.) or products (e.g., clothing, footwear, food, electronics). , furniture, games, software, mobile applications, automobiles, real estate, etc.). The value of the subject label may be a character that directly represents the subject of the corpus (e.g., accommodation, restaurant, clothing, or shoes, etc.), or it may be a code corresponding to the subject of the corpus (e.g. a, b, 0, or 1, etc.) . The emotion label may represent the emotion of the collected corpus among various emotions. Sentiment may include positive and negative (or positive/neutral/negative). A sentiment label can have a value of 0 (eg negative) or 1 (eg positive) (or -1 (negative), 0 (neutral), or 1 (positive)), and can be negative or positive (or negative, neutral, or positive). ) may be The user terminal 21 may receive an input of a subject label (eg, hotel) and a sentiment label (eg, 1) for a specific corpus (eg, the room is clean and comfortable) from the user. The processor 230 may receive the subject label and the sentiment label input from the user terminal 21 using the communication circuit 210 , and obtain the corresponding corpus, the subject label, and the sentiment label as a training data set.

According to an embodiment, the processor 230 may store the acquired training data set in the memory 220 . The processor 230 may utilize the training data set stored in the memory 220 for machine learning.

According to an embodiment, the processor 230 may perform machine learning by inputting a training data set to a specified language modeling algorithm. The processor 230 may secure a plurality of training data sets greater than or equal to a specified number for each of all combinations of a plurality of subjects and a plurality of emotions. For example, if there is a training data set including the subject hotel and clothing, and sentiment positive and negative, in order to correspond to all combinations for a given topic and emotion, n or more training data corresponding to the subject hotel and sentiment positive set, n or more training data sets corresponding to the subject hotel and emotional negative, n or more training data sets corresponding to the subject clothing and emotional positive, and n or more training data sets corresponding to the subject clothing and emotional negative are obtained. can The processor 230 may secure a uniformly mixed training data set for all combinations by using the subject label and the sentiment label. The processor 230 may input the secured training data set to the language modeling algorithm. The language modeling algorithm may be selected, for example, from one of open source algorithms such as BERT, KoBERT, GPT-3, Big Bird, ELMO or ULMFit. The configuration of the training data set will be described in detail with reference to FIG. 3 .

According to an embodiment, the processor 230 may acquire a single global sentiment classification model through the above-described machine learning. The processor 230 may obtain a global sentiment classification model configured to perform machine learning using the corpus and sentiment labels included in the training data set, and output sentiment information on the input corpus. The input corpus may be associated with any subject. Since a sufficient amount of mixed training data sets are input for each of the plurality of topics and all combinations of the plurality of emotions, it is possible to respond to all topics with one global sentiment classification model. In particular, the global emotional model can output highly reliable emotional information even on subjects that have not been learned in advance. The generation of the global sentiment classification model will be described in detail with reference to FIG. 4 .

According to an embodiment, the processor 230 may acquire a topic classification model by inputting a plurality of training data sets to a designated language modeling algorithm. The processor 230 may perform machine learning by using the corpus and topic labels included in the training data set, and obtain a topic classification model configured to output topic information about the input corpus. The generation of the topic classification model will be described in detail with reference to FIG. 4 .

According to an embodiment, the processor 230 may obtain a plurality of local sentiment classification models corresponding to each of the plurality of topics by classifying the plurality of learning data sets according to a plurality of topics and inputting them into a designated language modeling algorithm. The processor 230 performs machine learning using the corpus and sentiment labels included in the training data set corresponding to the specific topic, and the local sentiment classification model is configured to output sentiment information on the input corpus corresponding to the specific topic. can be obtained. For example, when a training data set for a hotel and clothing is provided, a local emotional classification model for the subject hotel and a local emotional classification model for the subject clothing may be obtained. The generation of the local sentiment classification model will be described in detail with reference to FIG. 4 .

According to an embodiment, the processor 230 may provide sentiment information on the analysis target corpus by inputting the analysis target corpus into the global sentiment classification model. The corpus to be analyzed may include textual customer comments. The processor 230 may calculate sentiment information on the analysis target corpus by inputting the analysis target corpus associated with a certain subject into the global sentiment classification model. Since sufficient learning is evenly mixed with respect to various subjects and emotions, sentiment information on a corpus associated with a pre-learned or non-pre-learned arbitrary topic can be accurately provided by a single global sentiment classification model. The output of emotion information using the global emotion classification model will be described in detail with reference to FIG. 5 .

According to an embodiment, in order to improve accuracy, the processor 230 inputs the analysis target corpus into a local sentiment classification model selected from a topic classification model, a global sentiment classification model, and a plurality of local sentiment classification models, thereby It is also possible to output emotional information.

For example, the processor 230 may obtain the analysis target corpus and similarity information for each of the plurality of subjects by inputting the analysis target corpus into the topic classification model. The similarity calculation using the topic classification model will be described in detail with reference to FIG. 6 . The processor 230 may determine one or more local emotion classification models respectively corresponding to one or more subjects having a high similarity to an analysis target corpus from among a plurality of local emotion classification models based on the similarity information. The processor 230 may determine sentiment information on the analysis target corpus based on the analysis results output by each of the global sentiment classification model and the selected local sentiment classification model. The processor 230 may calculate a final result by appropriately combining the analysis result by the global sentiment classification model and the analysis result by the selected local sentiment classification model.

For example, if two local emotion classification models corresponding to a subject similar to the analysis target corpus are selected, the processor 230 determines that the emotion analysis result by the global emotion classification model is “positive”, and returns to the first local emotion classification model. When the emotion analysis result is “positive” and the emotion analysis result by the second local emotion classification model is “negative”, Emotional information “positive” can be provided as the final result. As another example, when three local sentiment classification models corresponding to subjects similar to the analysis target corpus are selected, “positive” is derived as an analysis result by two models and “negative” is derived as an analysis result by the other two models In this case, the analysis result of a model having a subject similarity higher than a specified value among the used sentiment classification models can be provided as a final result. As another example, when the subject similarity (or the difference between the subject similarities) is smaller than a specified value, the analysis result of the global sentiment classification model may be provided as a final result.

3 is a diagram for explaining an exemplary operation of an apparatus for analyzing emotion of text data based on machine learning according to an embodiment.

Referring to FIG. 3 , the emotion analysis apparatus according to an embodiment may collect a plurality of learning data sets. The training data set may correspond to subjects such as hotels, restaurants, and clothing, and may correspond to emotions such as positive and negative. The training data set may include a corpus, subject labels, and sentiment labels.

The sentiment analysis apparatus may collect a plurality of corpora including corpus 1 to corpus 12. A subject label and a sentiment label may be generated by a user's input for each corpus. For example, for corpus 1 "The room is dirty and smelly", the subject label "Hotel" and the sentiment label "0" may be entered and stored.

The sentiment analysis apparatus may secure a specified number of training data sets or more for each of all combinations of subjects and emotions corresponding to the collected corpus. For example, n or more training data sets 310 including the topic label “Hotel” and the sentiment label “0”, and n or more training datasets 310 including the topic label “Hotel” and the sentiment label “1” (320). ), n or more training datasets 330 including the topic label “Restaurant” and the sentiment label “0”, n or more training datasets 340 including the topic label “Restaurant” and the sentiment label “1”; At least n training datasets 350 including the topic label “Clothing” and the sentiment label “0”, and n or more training datasets 360 including the topic label “Clothing” and the sentiment label “1” will be obtained. can Although the number of training data sets is illustrated as n or more in FIG. 3 , the present invention is not limited thereto, and the number of designated training data sets may be set differently for each combination.

By classifying a large number of corpora according to the input topic labels and sentiment labels, and obtaining a sufficiently mixed training data set for each of all combinations of classified topics and sentiments, various topics, in particular, not pre-trained through machine learning A single global sentiment classification model that can respond to a topic that is not yet available can be calculated.

4 is a diagram for explaining an exemplary operation of an apparatus for analyzing emotion of text data based on machine learning according to an embodiment.

Referring to FIG. 4 , the emotion analysis apparatus according to an embodiment may secure a learning data set 410 . The sentiment analysis apparatus may acquire the judgment model 430 by inputting the training data set 410 into the language modeling algorithm 420 .

The judgment model 430 includes a topic classification model 431 , a global sentiment classification model 432 , a first local sentiment classification model 433 , a second local sentiment classification model 434 , and a third local sentiment classification model 435 . may include Although it is illustrated that three local sentiment classification models are obtained in FIG. 4 , the present invention is not limited thereto, and local sentiment classification models corresponding to the number of topics included in the training data set 410 may be generated. 4, the judgment model 430 is shown to include a topic classification model 431 and local sentiment classification models 433, 434, and 435, but a topic classification model 431 and a local sentiment classification model 433, 434 and 435 correspond to optional components, and the decision model 430 may be implemented as a single global sentiment classification model 432 .

The topic classification model 431 may be derived by learning the first data set 411 including the corpus and the topic label among the training data sets 410 . The global sentiment classification model 432 can be derived by learning the second data set 412 including the corpus and sentiment labels when the number of training data sets 410 is sufficiently secured for each of all combinations of subject and sentiment. . The first local sentiment classification model 433 may be derived by learning the third data set 413 corresponding to the subject hotel. The second local sentiment classification model 434 may be derived by learning the fourth data set 414 corresponding to the subject restaurant. The third local sentiment classification model 435 may be derived by learning the fifth data set 415 corresponding to the subject clothing.

The sentiment analysis apparatus may acquire the analysis target corpus 440 . The emotion analysis apparatus may input the analysis target corpus 440 into the decision model 430 , and obtain the emotion information 450 output by the decision model 430 .

5 is a diagram for explaining an exemplary operation of an apparatus for analyzing emotion of text data based on machine learning according to an embodiment.

Referring to FIG. 5 , the sentiment analysis apparatus according to an embodiment may receive an analysis target corpus including corpus 1 to corpus 3 . The sentiment analysis apparatus may input a corpus to be analyzed as a global sentiment classification model.

The global sentiment classification model may be, for example, a model trained by a training data set associated with various topics such as “mobile app” and “electronic product”. However, since it is practically impossible to learn on all topics, the global sentiment classification model may be a model that has not been trained on the topic “game”. The global sentiment classification model can output analysis results for the input analysis target corpus regardless of the subject. For example, the global sentiment classification model can output emotional information “positive” for corpus 1 “The app design is clean and easy to use.” In corpus 2 “the finish quality is a bit poor and the phone grip is vague.” It is possible to output emotional information “negative” about the corpus 3 “a game with a lot of paid items… It is possible to output emotional information “Negative” for “Only the graphics are good.”

Even if the global sentiment classification model is not trained on the topic “game”, it may be a model trained on various topics such as “mobile app” or “software”, which are similar topics. Accordingly, it is possible to provide a highly reliable analysis result even for an unlearned subject “game”, thereby accurately providing emotional information on an analysis target corpus related to an arbitrary subject.

6 is a diagram for explaining an exemplary operation of an apparatus for analyzing emotion of text data based on machine learning according to an embodiment.

Referring to FIG. 6 , the sentiment analysis apparatus according to an embodiment may receive an analysis target corpus including corpus 1 to corpus 3 . The sentiment analysis apparatus may input a corpus to be analyzed as a subject classification model.

The subject classification model may be, for example, a model trained by a training data set associated with various subjects such as mobile apps, hotels, restaurants, food, clothing, and electronic products. The topic classification model may calculate a similarity between the input analysis target corpus and each of the plurality of learned topics. For example, the subject classification model can calculate a similarity of 71% between Corpus 1 “The app design is clean and easy to use” and the subject “Mobile app”, and Corpus 2 ”The finish quality is a bit poor and the grip on the phone is vague. The similarity between the subject “electronic products” can be calculated as 74%. Corpora 1 and Corpora 2 correspond to pre-learned topics, “mobile app” and “electronic product,” respectively, so that a particularly high degree of similarity can be calculated. As another example, the subject classification model is corpus 3 “games with a lot of paid items… It is possible to calculate a similarity of 51% between the graphic only” and the subject “mobile app”, and a similarity between Corpus 3 and the subject “Electronics” at 36%. The subject “game” of Corpus 3 may not correspond to the pre-learned subject, but it can be confirmed that the similarity is relatively high with the subjects “mobile app” and “electronic product” through similarity determination.

Even if the topic classification model is not trained on the topic “game,” it can identify “mobile apps” and “electronic products”, which are topics with relatively high similarity, through similarity calculation, so that the corpus to be analyzed can be matched with any topic. Even if it is related, it is possible to identify a subject similar to the analysis target corpus among pre-learned subjects.

7 is a diagram for explaining an exemplary operation of an apparatus for analyzing emotion of text data based on machine learning according to an embodiment.

Referring to FIG. 7 , the emotion analysis apparatus according to an embodiment may select a local emotion classification model corresponding to each subject having a high similarity to an analysis target corpus from among a plurality of subjects based on the similarity calculated by the subject classification model. . For example, the sentiment analysis device generates a local sentiment classification model (a mobile app sentiment classification model and an electronic product sentiment classification model) corresponding to each of two topics “mobile app” and “electronic product” that have high similarity to corpus 1 to 3 You can choose.

The sentiment analysis apparatus may input an analysis target corpus into each of the global sentiment classification model and the selected local sentiment classification model. The global sentiment classification model and the selected local sentiment classification model may output analysis results for each of the input analysis target corpus. For example, the global sentiment classification model may output emotional information “positive” to corpus 1, output emotional information “negative” to corpus 2, and output emotional information “positive” to corpus 3 . The global sentiment classification model can output analysis results with relatively high accuracy on any topic. The mobile app emotion classification model may output emotional information “positive” to corpus 1, output emotional information “negative” to corpus 2, and output emotional information “negative” to corpus 3 . The mobile app sentiment classification model can output relatively high-accuracy analysis results for the topic “mobile app”. The electronic product emotion classification model may output emotional information “negative” to corpus 1, output emotional information “negative” to corpus 2, and output emotional information “negative” to corpus 3 . The electronic product sentiment classification model can output a relatively high-accuracy analysis result for the subject “electronic product”.

The emotion analysis apparatus may output emotion information on each of the analysis target corpus as a final result based on the analysis results derived by the three emotion classification models. For example, for corpus 1, emotional information “positive” derived by two of the three models may be output, and for corpus 2, emotional information “negative” derived by all three models may be output. And, for corpus 3, emotional information “negation” derived by two of the three models may be output.

By using the local sentiment classification model corresponding to the subject with high similarity derived by the topic classification model together with the global sentiment classification model, it is possible to further improve the accuracy of sentiment analysis on the analysis target corpus.

8 is a flowchart illustrating a method for analyzing sentiment of text data based on machine learning according to an embodiment of FIG. 3 .

Hereinafter, it is assumed that the apparatus for analyzing the sentiment of the text data of FIG. 2 performs the process of FIG. 8 . In addition, in the description of FIG. 8 , an operation described as being performed by the loan service providing apparatus may be understood as being controlled by the processor 230 .

Referring to FIG. 8 , in step 810 , the apparatus for analyzing the sentiment of text data represents a subject label indicating a subject of the corpus collected among a plurality of subjects together with a corpus collected from an external document and a subject label indicating the subject of the corpus collected from among the plurality of emotions. Sentiment labels may be obtained as a training data set.

In step 820, the apparatus for analyzing the sentiment of text data may store the acquired training data set.

In step 830, the apparatus for analyzing the sentiment of text data outputs sentiment information for the input corpus by inputting a plurality of training data sets of a specified number or more for each of all combinations of a plurality of subjects and a plurality of emotions into a specified language modeling algorithm. It is possible to obtain a single global sentiment classification model configured.

In operation 840, the apparatus for analyzing the sentiment of text data may provide sentiment information on the analysis target corpus by inputting the analysis target corpus into the global sentiment classification model.

The embodiments of this document and the terms used therein are not intended to limit the technology described in this document to a specific embodiment, but it should be understood to include various modifications, equivalents, and/or substitutions of the embodiments. In connection with the description of the drawings, like reference numerals may be used for like components. The singular expression may include the plural expression unless the context clearly dictates otherwise. In this document, expressions such as “A or B”, “at least one of A and/or B”, “A, B or C” or “at least one of A, B and/or C” refer to all of the items listed together. Possible combinations may be included. Expressions such as "first," "second," "first," or "second," can modify the corresponding elements regardless of order or importance, and to distinguish one element from another element. It is used only and does not limit the corresponding components. When an element is referred to as being "connected (functionally or communicatively)" or "connected" to another element, the element is directly connected to the other element, or can be connected through

In this document, "adapted to or configured to", depending on the context, for example, hardware or software "suitable for," "having the ability to," "modified to, Can be used interchangeably with ""made to," "capable of," or "designed to." In some circumstances, the expression “a device configured to” may mean that the device is “capable of” with other devices or parts. For example, the phrase "a processor configured (or configured to perform) A, B, and C" refers to a dedicated processor (eg, an embedded processor) for performing the corresponding operations, or by executing one or more programs stored in a memory device; It may refer to a general-purpose processor (eg, CPU) capable of performing corresponding operations.

As used herein, the term “module” includes a unit composed of hardware, software, or firmware, and may be used interchangeably with terms such as, for example, logic, logic block, component, or circuit. can A “module” may be an integrally formed component or a minimum unit or a part that performs one or more functions. A “module” may be implemented mechanically or electronically, for example, known or to be developed, application-specific integrated circuit (ASIC) chips, field-programmable gate arrays (FPGAs), or It may include a programmable logic device.

At least a portion of an apparatus (eg, modules or functions thereof) or a method (eg, operations) according to an embodiment may be implemented as instructions stored in a computer-readable storage medium in the form of a program module. When the instruction is executed by the processor, the processor may perform a function corresponding to the instruction.

Each of the components (eg, a module or a program module) according to an embodiment may be composed of a singular or a plurality of entities, and some sub-components of the aforementioned sub-components may be omitted or other sub-components may be included. may include more. Alternatively or additionally, some components (eg, a module or a program module) may be integrated into one entity to perform the same or similar functions performed by each corresponding component before being integrated. Operations performed by a module, program module, or other component according to an embodiment are sequentially, parallelly, repetitively or heuristically executed, or at least some operations are executed in a different order, omitted, or other operations This can be added.

Claims (5)

In a machine learning-based sentiment analysis device for text data,
communication circuit;
Memory; and
a processor electrically coupled to the communication circuitry and the memory;
The processor is
A subject label indicating a subject of the collected corpus of a plurality of topics together with a corpus collected from one or more external documents using the communication circuitry and a sentiment of the collected corpus of a plurality of sentiments Acquiring a sentiment label representing
storing the acquired training data set in the memory;
Sentiment information on the input corpus by performing machine learning by inputting the corpus and sentiment labels included in the plurality of training data sets at least a specified number for each of the plurality of subjects and all combinations of the plurality of emotions into a designated language modeling algorithm Acquire a single global sentiment classification model configured to output
Obtaining a topic classification model configured to output topic information about the input corpus by performing machine learning by inputting the corpus and topic labels included in the plurality of training data sets into the designated language modeling algorithm,
By classifying the plurality of training data sets by the plurality of subjects, and performing machine learning by inputting a corpus and a sentiment label included in each of the classified plurality of training data sets into the designated language modeling algorithm, the input corpus Obtaining a plurality of local emotion classification models configured to output emotion information for and corresponding to each of the plurality of topics,
obtaining similarity information for the analysis target corpus and each of the plurality of subjects by inputting the analysis target corpus into the topic classification model;
selecting a local sentiment classification model corresponding to each of the one or more subjects having a relatively high degree of similarity to the analysis target corpus among the plurality of subjects from the plurality of local sentiment classification models based on the similarity information;
The apparatus, characterized in that by inputting the analysis target corpus into the global sentiment classification model and the selected local sentiment classification model, sentiment information on the analysis target corpus is provided. delete delete The method of claim 1,
The processor is
The apparatus, characterized in that the sentiment information for the analysis target corpus is determined based on the analysis results output by each of the global sentiment classification model and the selected local sentiment classification model. In the sentiment analysis method of machine learning-based text data performed by a processor included in a computing device,
obtaining, as a learning data set, a subject label indicating the subject of the collected corpus among a plurality of subjects together with a corpus collected from one or more external documents and a sentiment label indicating the emotion of the collected corpus among a plurality of emotions;
storing the acquired training data set;
Sentiment for the input corpus by performing machine learning by inputting the corpus and sentiment labels included in the plurality of training data sets at least a specified number for each of the plurality of subjects and all combinations of the plurality of emotions into a designated language modeling algorithm obtaining a single global sentiment classification model, configured to output information;
acquiring a topic classification model configured to output topic information about the input corpus by performing machine learning by inputting the corpus and topic labels included in the plurality of training data sets into the specified language modeling algorithm;
By classifying the plurality of training data sets by the plurality of subjects, and performing machine learning by inputting a corpus and a sentiment label included in each of the classified plurality of training data sets into the designated language modeling algorithm, the input corpus obtaining a plurality of local sentiment classification models configured to output sentiment information for , and corresponding to each of the plurality of subjects;
obtaining similarity information for the analysis target corpus and each of the plurality of subjects by inputting the analysis target corpus into the topic classification model;
selecting, from the plurality of local sentiment classification models, based on the similarity information, a local sentiment classification model corresponding to each of one or more subjects having a relatively high similarity to the analysis target corpus from among the plurality of subjects; and
and providing sentiment information on the analysis target corpus by inputting the analysis target corpus into the global sentiment classification model and the selected local sentiment classification model.

Images