KR101326313B1 - Method of classifying emotion from multi sentence using context information - Google Patents

Abstract

The present invention relates to a method for classifying emotions from multiple sentences using context information, and more specifically, (1) extracting context information for each of multiple sentences constituting text, and (2) using extracted context information. Extracting a plurality of main sentences from the multiple sentences, (3) extracting an emotional feature for each of the extracted main sentences, and (4) extracting each of the main sentences using the emotion classifier Categorizing the emotions is characterized by their construction.
According to the method of classifying emotions from multiple sentences using context information proposed by the present invention, the main sentences are extracted from the multiple sentences constituting the text using the context information, the emotions are classified, and the classification is performed on the extracted main sentences. By combining these emotions, you can accurately classify emotions from multiple sentences that can be collected online and use them in your marketing strategy.

Description

Method of classifying emotions from multiple sentences using context information {METHOD OF CLASSIFYING EMOTION FROM MULTI SENTENCE USING CONTEXT INFORMATION}

The present invention relates to a emotion classification method, and more particularly, to a emotion classification method from multiple sentences using context information.

Due to the popularization of smart phones, social network services (SNS) such as Twitter and Facebook have become commonplace, and a lot of opinions on various topics are being expressed in real time. SNS is a service that strengthens existing networks and forms new ones to form a broad network of people. Many people create huge amounts of text information in the form of comments to each other through such services. have.

Recently, viral marketing, or word-of-mouth marketing strategies, has been widely used for major issues such as product reviews, movie reviews, and food ratings. It is recognized that it is very important to the marketing strategy to accurately determine the.

Therefore, research to find meaningful information from huge texts written by SNS users is of interest, and in particular, the emotions contained in sentences are very wide range of information. Research is being done (see Publication No. 10-2002-0042248), but very weak.

The present invention has been proposed to solve the above problems of the conventionally proposed methods, extracts the main sentence from the multiple sentences constituting the text using the context information, classifies the emotion to the extracted main sentence, An object of the present invention is to provide a method of classifying emotions from multiple sentences using context information, by accurately classifying emotions from multiple sentences that can be collected online by combining the classified emotions, and using them in a marketing strategy.

According to an aspect of the present invention for achieving the above object, a method for classifying emotions from multiple sentences using context information,

(1) extracting context information for each of the multiple sentences constituting the text;

(2) extracting a plurality of main sentences from the multiple sentences using the extracted context information;

(3) extracting an emotional feature for each of the extracted main sentences; And

And (4) classifying the emotions of each of the main sentences using an emotion classifier using the extracted emotion features.

Preferably,

(5) may further comprise combining the emotions of each of the main sentences classified.

Preferably, the step (1)

(1-1) calculating keyword information included in a sentence by using the following equation;

(Where S is the input sentence, k _i is the i th keyword included in the input sentence S, ω _i ∈ R is the keyword weight, and K is the number of keywords included in the input sentence S).

(1-2) calculating weights for positions of sentences in the text using the following equation;

(Where index (S _i ) is the index of the main sentence S _i , and T is the number of sentences in the text.)

(1-3) calculating the degree of emotional change between sentences using the following equation; And

(Where n (S _ps ) is the number of preceding sentences with the same emotion.)

(1-4) calculating context information of a sentence by using the following equation by using the calculated keyword information, a weight of a position of a sentence, and a degree of change between sentences.

. ≪ / RTI >

Preferably, the step (3)

(3-1) classifying words included in sentences by parts of speech using a morpheme analyzer; And

(3-2) extracting an emotional feature of the word using an emotional dictionary.

More preferably, in the step (3-2),

The emotion dictionary may be constructed by adding a domain-based emotion dictionary to a formal emotion dictionary based on a lexical dictionary.

More preferably, the emotion dictionary,

It may include emotional features for nouns, verbs, adjectives, adverbs, and emoticons.

Even more preferably, the emoticon,

It may be a reference emoticon, which is the simplest emoticon form among irregularly frequently used emoticons.

Even more preferably,

Irregular emoticons included in sentences may be converted into the reference emoticons using a Bayesian framework to extract emotional features.

According to the method of classifying emotions from multiple sentences using context information proposed by the present invention, the main sentences are extracted from the multiple sentences constituting the text using the context information, the emotions are classified, and the classification is performed on the extracted main sentences. By combining these emotions, you can accurately classify emotions from multiple sentences that can be collected online and use them in your marketing strategy.

1 is a flowchart of a method for classifying emotions from multiple sentences using context information according to an embodiment of the present invention.
2 is a detailed flowchart of step S100 of a method for classifying emotions from multiple sentences using context information according to an embodiment of the present invention.
3 is a detailed flowchart of step S300 of a method for classifying emotions from multiple sentences using context information according to an embodiment of the present invention.
4 is a flowchart illustrating a detailed flow of processing an irregular emoticon using a Bayesian framework in a method for classifying emotions from multiple sentences using context information according to an embodiment of the present invention.
FIG. 5 is a diagram illustrating an emotion classification performance test result of a sentence in a method of classifying emotions from multiple sentences using context information according to an embodiment of the present invention. FIG.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings, in order that those skilled in the art can easily carry out the present invention. In the following detailed description of the preferred embodiments of the present invention, a detailed description of known functions and configurations incorporated herein will be omitted when it may make the subject matter of the present invention rather unclear. In the drawings, like reference numerals are used throughout the drawings.

In addition, in the entire specification, when a part is referred to as being 'connected' to another part, it may be referred to as 'indirectly connected' not only with 'directly connected' . In addition, the term 'comprising' of an element means that the element may further include other elements, not to exclude other elements unless specifically stated otherwise.

1 is a flowchart of a method for classifying emotions from multiple sentences using context information according to an embodiment of the present invention. As shown in FIG. 1, in the emotion classification method using multiple contexts according to an embodiment of the present invention, the context information is extracted for each of the multiple sentences constituting the text (S100). Extracting a plurality of main sentences from the multiple sentences using context information (S200), extracting an emotional feature for each extracted main sentence (S300), and using the emotion classifier using the extracted emotional features, respectively. It may be configured to include a step of classifying the emotions of the main sentence (S400), and may further include a step (S500) of combining the emotions of each classified main sentence.

In step S100, context information is extracted for each of the multiple sentences constituting the text, and the detailed flow of step S100 will be described in detail with reference to FIG. 2.

2 is a detailed flowchart of step S100 of a method for classifying emotions from multiple sentences using context information according to an embodiment of the present invention. As shown in FIG. 2, step S100 of a method for classifying emotions from multiple sentences using context information according to an embodiment of the present disclosure includes calculating keyword information included in a sentence (S110), and a sentence in text. Calculating the weight of the position (S120), calculating the emotional variation between sentences (S130) and the calculated context information of the sentence by using the keyword information, the weight of the position of the sentence and the degree of change between sentences. It may include calculating (S140).

In step S110, keyword information included in a sentence is calculated using Equation 1 below.

Here, S is an input sentence, k _i is an i-th keyword included in the input sentence S, ω _i ∈ R is a keyword weight, and K represents the number of keywords included in the input sentence S. The keyword measures the frequency of the corresponding vocabulary through sentences previously collected for each domain, and assigns the frequency as a weight (0 to 1).

In step S120, a weight for the position of the sentence in the text is calculated. More specifically, since the author of a text generally expresses his or her opinion through the first sentence or the ending sentence of the text, the position of the sentence in the text is very important information for estimating the overall feeling of the text. By using Equation 2 below, the weight of the position of the sentence in the text is calculated.

Here, index (S _i ) is the index of the main sentence S _i , and T represents the number of sentences in the text.

In step S130, the degree of emotional change between sentences is calculated. More specifically, since the emotions of the whole sentence are often changed when the emotion of the sentence is sustained and suddenly the emotion changes, the emotional change between sentences is also very important information for estimating the emotion of the whole sentence. Use 3 to calculate the degree of emotional change between sentences.

Here, n (S _ps ) represents the number of preceding sentences having the same emotion.

In step S140, the context information of the sentence is calculated by Equation 4 below using the keyword information calculated in steps S110 through S130, the weight of the position of the sentence, and the degree of change between sentences.

In step S200, the main sentence is extracted from the multiple sentences using the context information extracted in step S100. That is, the main sentence is extracted through step S200, and the main sentences important for estimating the emotion of the entire text may be extracted, and a plurality of extracted main sentences may be extracted.

In step S300, an emotional feature is extracted for each of the main sentences extracted by step S200, and the detailed flow of step S300 will be described in detail with reference to FIG.

3 is a detailed flowchart of step S300 of a method for classifying emotions from multiple sentences using context information according to an embodiment of the present invention. As shown in FIG. 3, step S300 of a method for classifying emotions from multiple sentences using context information according to an embodiment of the present disclosure includes: classifying words included in sentences by parts of speech using a morpheme analyzer (S310). And extracting 320 an emotional feature for a word using the emotional dictionary.

In step S310, a word included in a sentence is classified by parts of speech using a morpheme analyzer. When the words are morphologically analyzed, it is possible to determine whether or not the same words are used from the matching stems, and the words can be classified by parts of speech based on this process.

In step S320, the emotion feature for the words classified in step S310 is extracted using the emotion dictionary. More specifically, the emotion dictionary may include words corresponding to emotions for each part-of-speech and emotion intensity of the words, and may extract the types of emotions and the intensity of emotions for the words as emotion characteristics. In this case, the emotion dictionary may include emotional features for nouns, verbs, adjectives, adverbs, and emoticons.

On the other hand, even the same vocabulary may show different emotions according to specific domains. For example, the word "light" has a negative meaning in the "personal" domain, while a positive meaning in the "communication" domain. In other words, the same vocabulary can be different emotions according to a specific domain, it is preferable that the emotional dictionary is built by adding a domain-based emotional dictionary to the existing formal emotional dictionary based on the lexical dictionary, through which a more accurate Emotion classification can be done.

Furthermore, in step S320, although the emoticons included in the sentence are very important factors in classifying the feelings of the sentence, they are written very irregularly, even though they have the same meaning due to the user's taste, typos, and other factors. There is a difficult problem to use as an emotional feature. For example, “^ _ ^” and “^ _________ ^” have the same meaning, but the number of “_” can be used differently according to the individual. The use of such irregular emoticons as it is is one of the factors that makes it difficult to classify the correct emotion.

To this end, it is preferable to convert the irregular emoticons included in the sentence into a reference emoticon included in the emotion dictionary to extract emotional features from the emoticon. Here, the "reference emoticon" is an emoticon included in the emotion dictionary, and refers to the simplest emoticon form among emoticons that are frequently used irregularly. That is, by converting an emoticon used irregularly into such a reference emoticon, accurate emotion classification can be performed from the irregular emoticon included in the sentence.

More specifically, an irregular emoticon included in a sentence may be converted into a reference emoticon using a Bayesian framework, and FIG. 4 shows Bayesian in a method of classifying emotions from multiple sentences using context information according to an embodiment of the present invention. A detailed flow of processing an irregular emoticon using a framework is illustrated. As shown in FIG. 4, after disassembling an emoticon, a probability distribution model is constructed through a normalization process using a histogram, and an optimal emoticon is extracted by calculating a similarity between an irregular emoticon and a reference emoticon and extracting an optimal reference emoticon. Can be processed.

In step S400, the emotion of each main sentence is classified through the emotion classifier using the emotion feature extracted in step S300. That is, the emotion feature extracted by step S300 is configured as a feature vector to classify the emotions of the sentence through an emotion classifier. In this case, the emotion classifier for sentence emotion classification may use a support vector machine (SVM).

In step S500, the emotions of the respective main sentences classified by step S400 may be combined, and through this, the emotions of the multiple sentences may be classified to finally estimate the overall emotion of the text.

[Experimental Example]

Emotion classification performance experiment of sentence

We randomly collect user-written posts from social network services (SNS) such as Twitter, Facebook, and Me2day by general, product review, travel, food, and movie domains, and then use each of the four methods Classification was performed.

That is, the method using only the existing formal dictionary (case 1), the method using only the emotion dictionary constructed by adding the domain-based emotion dictionary (case 2), the method using the context information and the existing formal dictionary (case 3), and the context information Emotion classification was performed using the method using case dictionary (case 4) constructed by adding domain-based emotion dictionary. Emotion classification performance of the sentence according to each method performed using F ₁ -measure using precision (“p”) and recall (“r”) according to Equations 5 to 7 below. It evaluated, and the result is shown in Table 1 and FIG.

domain Case emotion p r F1

Normal Case 1 Positive 0.5798 0.5644 0.5719 denial 0.6377 0.4891 0.5536 neutrality 0.6841 0.5991 0.6387 Case 2 Positive 0.6213 0.5891 0.6047 denial 0.6124 0.6401 0.6259 neutrality 0.7135 0.6787 0.6956



Product Review Case 1 Positive 0.6012 0.8181 0.6930 denial 0.6663 0.2513 0.3649 neutrality 0.5387 0.6578 0.5923 Case 2 Positive 0.6648 0.7273 0.6946 denial 0.6259 0.6211 0.6234 neutrality 0.9121 0.6806 0.7795 Case 3 Positive 0.8122 0.7301 0.7689 denial 0.6381 0.6114 0.6244 neutrality 0.7533 0.8101 0.7806 Case 4 Positive 0.8129 0.7013 0.7529 denial 0.6587 0.7759 0.7125 neutrality 0.8264 0.8585 0.8421



Travel Case 1 Positive 0.7512 0.7598 0.7554 denial 0.6602 0.3289 0.4390 neutrality 0.4451 0.6654 0.5333 Case 2 Positive 0.7146 0.8336 0.7695 denial 0.6657 0.3328 0.4437 neutrality 0.4281 0.5045 0.4631 Case 3 Positive 0.7498 0.7592 0.7544 denial 0.5722 0.6687 0.6166 neutrality 0.8007 0.6618 0.7246 Case 4 Positive 0.6599 0.8304 0.7353 denial 0.5431 0.5007 0.5210 neutrality 0.8704 0.5011 0.6360



food Case 1 Positive 0.7141 0.8401 0.7719 denial 0.506 0.1916 0.2779 neutrality 0.3754 0.4894 0.4248 Case 2 Positive 0.7271 0.8891 0.7999 denial 0.2035 0.2789 0.2353 neutrality 0.902 0.2531 0.3952 Case 3 Positive 0.8334 0.8136 0.8233 denial 0.5014 0.9042 0.6450 neutrality 0.8576 0.7234 0.7848 Case 4 Positive 0.7891 0.8341 0.8109 denial 0.5301 0.9012 0.6675 neutrality 0.8249 0.6402 0.7209



movie Case 1 Positive 0.6304 0.6681 0.6487 denial 0.7813 0.4285 0.5534 neutrality 0.2111 0.5131 0.2991 Case 2 Positive 0.6936 0.4462 0.5430 denial 0.7288 0.5898 0.6519 neutrality 0.3312 0.7366 0.4569 Case 3 Positive 0.5813 0.7777 0.6653 denial 0.6054 0.4284 0.5017 neutrality 0.8652 0.4809 0.6181 Case 4 Positive 0.6148 0.8876 0.7264 denial 0.9384 0.6278 0.7523 neutrality 0.5812 0.6857 0.6291

As shown in Table 1 and FIG. 5, it was confirmed that the effect of emotion classification performance is excellent when the context information and the emotion dictionary are used as compared with the case using only the emotion dictionary. In addition, it is confirmed that the emotion classification performance effect is better in the case of the emotion dictionary constructed by adding the domain-based emotion dictionary than the general emotion dictionary. Therefore, it can be seen that the method according to the present invention has an excellent effect on the performance of emotion classification from multiple sentences constituting text.

The present invention may be embodied in many other specific forms without departing from the spirit or essential characteristics of the invention.

S100: extracting context information for each of the multiple sentences constituting the text
S200: extracting a plurality of main sentences from the multiple sentences using the extracted context information
S300: extracting an emotional feature for each extracted main sentence
S400: classifying the emotions of each main sentence through the emotion classifier using the extracted emotion features
S500: step of combining emotions of each classified main sentence

Claims (8)

(1) extracting context information for each of the multiple sentences constituting the text;
(2) extracting a plurality of main sentences from the multiple sentences using the extracted context information;
(3) extracting an emotional feature for each of the extracted main sentences; And
(4) classifying the emotions of each of the main sentences using an emotion classifier using the extracted emotion features,
The step (1)
(1-1) calculating keyword information included in a sentence by using the following equation;

(Where S is the input sentence, k _i is the i th keyword included in the input sentence S, ω _i ∈ R is the keyword weight, and K is the number of keywords included in the input sentence S).
(1-2) calculating weights for positions of sentences in the text using the following equation;

(Where index (S _i ) is the index of the main sentence S _i , and T is the number of sentences in the text.)
(1-3) calculating the degree of emotional change between sentences using the following equation; And

(Where n (S _ps ) is the number of preceding sentences with the same emotion.)
(1-4) calculating context information of a sentence by using the following equation by using the calculated keyword information, a weight of a position of a sentence, and a degree of change between sentences.

Emotion classification method from multiple sentences using the context information, comprising a.
The method of claim 1,
(5) further comprising combining the emotions of each of the main sentences classified; emotion classification method from multiple sentences using context information.
delete 2. The method of claim 1, wherein step (3)
(3-1) classifying words included in sentences by parts of speech using a morpheme analyzer; And
(3-2) A method of classifying emotions from multiple sentences using context information, comprising extracting an emotion feature for the word using an emotion dictionary.
The method according to claim 4, wherein in step (3-2),
The emotion dictionary is constructed by adding a domain-based emotion dictionary to a formal emotion dictionary based on a lexical dictionary.
The method of claim 4, wherein the emotion dictionary,
Emotion classification method from multiple sentences using context information, characterized by including emotional features for nouns, verbs, adjectives, adverbs, and emoticons.
The method of claim 6, wherein the emoticon,
A method of classifying emotions from multiple sentences using context information, characterized in that the reference emoticon is the simplest emoticon among irregularly used emoticons.
The method of claim 7, wherein
Irregular emoticons included in the sentence is converted to the reference emoticon using a Bayesian framework to extract the emotional features, emotion classification method from multiple sentences using context information.

Images