KR102609132B1 - Method and apparatus for automatically constructing sentiment dictionary - Google Patents

Abstract

Disclosed is a method and device for automatically constructing an emotional dictionary.
According to an embodiment of the present invention, a method of automatically constructing an emotional dictionary includes the step of obtaining an opinion corresponding to the basis of the evaluation opinion from evaluation data containing evaluation opinions about an evaluation object - the evaluation Data belong to one of a plurality of evaluation classes according to the evaluation opinion -; extracting one or more emotion words representing emotion from words in the grounded opinion; and calculating the emotion of the token according to the association between one or more tokens containing one or more of the emotion words and the evaluation classes.

Description

Method and apparatus for automatically constructing a sentiment dictionary {METHOD AND APPARATUS FOR AUTOMATICALLY CONSTRUCTING SENTIMENT DICTIONARY}

The present invention relates to a method and device for constructing an emotional dictionary. More specifically, the present invention relates to a method and device for constructing an emotional dictionary. More specifically, the emotional dictionary can be constructed automatically without human intervention, so that an emotional dictionary can be constructed quickly and at low cost, and it is possible to construct an emotional dictionary according to the context. It concerns a method and device for building an emotional dictionary that can accurately reflect even changes.

Financial markets have semi-strong form efficiency in which publicly available information is reflected in market prices, but private information is not reflected in market prices.

Analyst reports serve as a channel for generating and supplying valuable market information in terms of financial market efficiency. For example, analysts interpret a company's disclosed information and play a role in increasing the speed at which this information is reflected in the stock price, and by discovering undisclosed information and providing it to the market, they can minimize information asymmetry between companies and investors. Therefore, it can be said that analyst reports function as a medium of information between the financial market and investors.

Despite the above important functions of analyst reports, the actual usability of analyst reports is limited for the following reasons.

First, analysts optimistically present quantitative indicators such as recommended opinions, target stock prices, and profit forecasts, minimize the presentation of sell opinions due to conflict of interest issues, and provide forecasts close to the average (representative) for fear of damage to reputation. There is a tendency to present (herding behavior).

Second, herding behavior may result in information inefficiency that weakens the information environment in the financial market, and it is also possible that the reliability of the analyst report itself may decrease because it is difficult for investors to determine whether the report is herding.

Third, due to the high labor intensity of individual analysts, the readability and communication power of the analyst report itself is low, making it difficult for investors to interpret the contents of the analyst report.

In order to increase the usability of analyst reports, research has been conducted on the quantitative and qualitative indicators described in analyst reports.

Examples of research on quantitative indicators include research on market reactions to quantitative indicators presented by analysts, and research on profit forecasts and forecast accuracy presented by analysts. Additionally, research on market reactions according to the sentiment (positive or negative) of analyst reports can be cited as a representative example of research on qualitative indicators.

In terms of methodology, most studies on qualitative indicators focus on measuring the sentiment of analyst reports using sentiment dictionaries or measuring the sentiment of analyst reports using machine learning techniques.

However, since conventional emotion dictionaries focus on the syntax of words and do not consider semantic relations of words, they have limitations in identifying the exact emotion of words. For example, even if it is the same word, the emotion of the word may vary depending on the topic or context in which it is used, but the conventional emotion dictionary judges the emotion in units of one specific word (1-gram), so it depends on the topic or context. It cannot reflect the fact that the emotion of a word may vary depending on the word.

Therefore, an N-gram unit emotional dictionary is needed to reflect the emotion of words that may vary depending on topic or context.

Meanwhile, in order to construct a conventional emotional dictionary, a person's subjective judgment is involved in the evaluation of words to be included in the emotional dictionary, and the process of labeling a data set using machine learning techniques by directly judging whether it is positive or negative (dictionary) learning process) is carried out. For this reason, there is a problem that even in the same analyst report, the emotion measurement results may vary depending on the emotion dictionary or machine learning model.

Therefore, it can be said that a new method for learning emotions objectively and cost-effectively is needed.

In order to solve the above-described problem, an embodiment of the present invention provides a method for constructing an emotional dictionary that is constructed without human subjective judgment and can reflect the accurate emotion of words that vary depending on the topic or context; and The purpose is to provide a device.

According to an embodiment of the present invention, a method of automatically constructing an emotional dictionary includes the step of obtaining an opinion corresponding to the basis of the evaluation opinion from evaluation data containing evaluation opinions about an evaluation object - the evaluation Data belong to one of a plurality of evaluation classes according to the evaluation opinion -; extracting one or more emotion words representing emotion from words in the grounded opinion; and calculating the emotion of the token according to the association between one or more tokens containing one or more of the emotion words and the evaluation classes.

According to another embodiment of the present invention, an apparatus for automatically constructing an emotional dictionary, comprising: an acquisition unit that obtains an opinion corresponding to the basis of the evaluation opinion from evaluation data containing an evaluation opinion about an evaluation object - The evaluation data belong to one of a plurality of evaluation classes according to the evaluation opinion; an extraction unit that extracts one or more emotional words representing emotions from among the words in the grounded opinion; and a calculation unit that calculates the emotion of one or more tokens containing one or more emotion words and the evaluation classes according to the association between the tokens.

According to one embodiment of the present invention, since a sentiment dictionary is constructed in units of tokens containing a plurality of words, the sentiment of words that may vary depending on the topic or context can be accurately determined.

In addition, according to another embodiment of the present invention, since an emotional dictionary is automatically constructed without the intervention of a person's subjective judgment, higher accuracy in emotional judgment can be ensured, and the emotional dictionary can be further updated using new words. It can be implemented efficiently.

The effects that can be obtained from the present disclosure are not limited to the effects mentioned above, and other effects not mentioned can be clearly understood by those skilled in the art from the description below. will be.

Figure 1 is a block diagram for explaining the structure of a device that automatically builds an emotional dictionary.
Figure 2 is a flow chart to explain an example of a method for automatically building an emotional dictionary.
Figures 3 and 4 are flowcharts for explaining various examples of methods for automatically building an emotion dictionary.
Figure 5 is a flow chart to explain an example of selecting a token based on the number of times included in evaluation data.
Figure 6 is a block diagram illustrating the structure of a device that determines the emotion of target data using an emotion dictionary constructed through a method of automatically constructing an emotion dictionary.
Figure 7 is a flow chart to explain an example of a method for measuring emotion.

Below, with reference to the attached drawings, embodiments of the present invention will be described in detail so that those skilled in the art can easily implement the present invention. However, the present invention may be implemented in many different forms and is not limited to the embodiments described herein. In order to clearly explain the present invention in the drawings, parts that are not related to the description are omitted, and similar parts are given similar reference numerals throughout the specification.

The terms used in the present invention are general terms that are currently widely used as much as possible while considering the functions in the present invention, but they may vary depending on the intention or precedent of a person working in the art, the emergence of new technology, etc. In addition, in certain cases, there are terms arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the description of the relevant invention. Therefore, the terms used in the present invention should be defined based on the meaning of the term and the overall content of the present invention, rather than simply the name of the term.

When it is said that a part "includes" a certain element throughout the specification, this means that, unless specifically stated to the contrary, it does not exclude other elements but may further include other elements. In addition, terms such as "... unit" and "module" used in the specification refer to a unit that processes at least one function or operation, which may be implemented as hardware or software, or as a combination of hardware and software. . In addition, when a part is said to be "connected" to another part throughout the specification, this includes not only the case where it is "directly connected," but also the case where it is connected "with another element in between."

Before a detailed description of the present invention, the terms mentioned in this specification and the scope to which the present invention can be applied are defined.

Evaluation data: Evaluation data refers to the evaluation results written by a user about a service or product or a specific object after using a service or product provided by a specific object or analyzing a specific object.

Additionally, the evaluation data corresponds to raw data used to build an emotional dictionary. Since the present invention builds an emotional dictionary through machine learning techniques using information included in evaluation data, the evaluation data may correspond to learning data.

Evaluation object: The evaluation object corresponds to the service, goods, or specific object evaluated through the evaluation results.

User: Users can be tourists, moviegoers, viewers, customers, analysts, employers, etc.

Specific target: A specific target can be a company, an employee, etc., and a service or good can be a product, movie, food, work provided by the employee to the employer, etc.

Therefore, evaluation data can include reviews written by tourists about tourist destinations, reviews by moviegoers written about movies, viewers' reviews written about video programs, analyst reports written about companies, customer reviews written about products, reviews written about employers. This may apply to the employer's evaluation, etc.

Additionally, evaluation data may exist in various forms, such as postings on SNS, postings on Internet bulletin boards, reports uploaded to a report management system provided by a securities company, and information uploaded to an Internet web page that shares personnel information.

Evaluation data may include ‘evaluation opinion’ and ‘supporting opinion’.

Evaluation opinion: Evaluation opinion corresponds to the final or comprehensive evaluation result of the evaluation object. For example, the evaluation opinion may correspond to a rating value included in a review, a recommendation opinion included in a report, or a target stock price, etc. Additionally, evaluation opinions may include actual market returns.

Evaluation opinions may be in the form of multiple grades that can be distinguished from each other, such as numerical data. As another example, evaluation opinions are differentiated in terms of emotional intensity, such as 'like' and 'angry', 'positive' and 'negative', 'recommend to buy' and 'recommend not to buy', 'expected to rise' and 'expected to fall', etc. It may be in a possible form. Additionally, the evaluation opinion may be divided into three or more intensities, such as 'I like it', 'It's heartwarming', 'I'm sad', 'I'm angry', and 'I want a follow-up article'.

Evaluation class: An evaluation opinion may be composed of a plurality of grades that can be distinguished from each other according to numerical value or emotional intensity, and each of these plurality grades may constitute an evaluation class. In a different embodiment, the evaluation class may be configured in a new form with a distinction different from the plurality of grades included in the evaluation opinion.

Ground opinion: Ground opinion refers to the evaluation result corresponding to the basis or reason for the evaluation opinion. For example, if the evaluation data is written as “a certain evaluation is appropriate for a certain reason,” the part corresponding to “a certain reason” corresponds to the supporting opinion mentioned in this specification.

Scope to which the present invention can be applied: As described above, evaluation data corresponding to raw data may have various types or forms. Accordingly, the present invention can be applied to all of the various types or forms that evaluation data can have.

For example, when the evaluation data corresponds to a review, the present invention can be used as a method of building an emotional dictionary specialized for reviews. As another example, when the evaluation data corresponds to the employer's evaluation of the employee, the present invention may be used as a method of constructing an emotional dictionary specialized for the evaluation of the employee. As another example, when the evaluation data corresponds to an analyst report, the present invention may be used as a method of constructing an emotional dictionary specialized for the analyst report.

Depending on the embodiment, the present invention can also be used as a method of constructing an emotional dictionary specialized for each company or industry. For example, if a report (evaluation data) contains the sentiment word (token) “rising oil prices,” that token will have a positive sentiment toward a specific industry (oil companies), but may have a positive sentiment toward another industry (oil companies). You may have negative feelings toward industries that produce goods using raw materials. In order to reflect these characteristics, the present invention can be configured to perform a separate learning process by dividing each industry group in which the emotional dictionary is used.

Through this specification, 'automatic construction of an emotion dictionary' and 'emotion measurement using a constructed emotion dictionary' are proposed. Below, the present invention will be described in detail focusing on an example where the evaluation data is an analyst report.

1. Automatic construction of emotional dictionary

Figure 1 is a block diagram for explaining the structure of a device 100 for automatically building an emotional dictionary (hereinafter referred to as 'construction device'), and Figure 2 is a method for automatically building an emotional dictionary (hereinafter referred to as 'construction method'). This is a flowchart to explain an example of). Figure 5 is a flow chart to explain an example of selecting tokens based on the number of appearances.

As shown in FIG. 1, the construction device 100 may be configured to include an acquisition unit 110, an extraction unit 120, and a calculation unit 130. Depending on the embodiment, the construction device 100 may further include a selection unit 140, and the calculation unit 130 may include a calculation unit 132 and an emotion calculation unit 134.

The acquisition unit 110 may obtain a grounded opinion and/or an evaluation opinion from analyst reports corresponding to the evaluation data (S210).

As previously explained, an analyst report includes recommendations for an evaluation target (e.g., a company) and a target stock price (rating opinion), and an opinion on the basis or reasons for which these recommendations and target stock price were derived (e.g., a company). supporting opinions) are also included.

Each of the analyst reports may belong to one of a plurality of evaluation classes. The recommended opinion and target stock price obtained from an analyst report may correspond to one of a plurality of grades, and the evaluation class to which the corresponding analyst report belongs may be determined depending on the grade to which this recommended opinion and target stock price belong. For example, evaluation classes may be divided into 'sell', 'hold' and 'buy', or 'downgrade', 'reiteration' and 'upgrade'.

The extraction unit 120 may extract one or more emotional words from words included in the grounded opinion (S220).

Emotion words refer to words that express emotion. Emotion words can be divided into positive words expressing positive emotions and negative words expressing negative emotions. Positive and negative words may be classified into more detailed groups, such as strong positive words, positive words, negative words, and strong negative words, depending on the intensity of emotion.

Depending on the embodiment, the extraction unit 120 may be configured to extract only words corresponding to a specific part of speech as emotional words or to not extract words corresponding to other parts of speech.

For example, the extraction unit 120 may be configured not to extract numbers and special characters from all words included in the supporting opinion. This is a result that reflects the fact that numbers and special characters are not used to express emotions, and is intended to implement emotional word extraction processing more efficiently by reducing the number of candidates extracted as emotional words in advance.

Correspondingly, the extraction unit 120 may be configured to extract only some or all of the parts of speech (nouns, verbs, adjectives, adverbs, negatives, etc.) that are mainly used to express emotions.

As another example, the extraction unit 120 may be configured not to extract words corresponding to other languages, including English, in order to build an emotional dictionary specialized for the Korean language.

The calculation unit 130 may calculate the emotion of the token according to the correlation between the token and the evaluation classes (S230). Here, calculating the emotion of a token may mean learning the emotion of the token.

A token may consist of one or more emotional words. For example, a token may be composed of one emotional word (1-gram token), or may be composed of multiple emotional words (n-gram token, where n is a natural number of 2 or more and represents the number of included emotional words). You can.

The reason for converting multiple emotional words into one token is that as the number of emotional words included in one token increases, the emotionality of the word, which may vary depending on the topic or context, can be more accurately reflected.

Meanwhile, the selection unit 140 of the present invention selects only tokens that appear more than a preset number of times in the analyst report among the tokens (S510), and the calculation unit 130 selects the tokens according to the correlation between the selected tokens and evaluation classes. , the emotion of the selected token can be calculated (S520). Here, the preset number refers to the minimum number of appearances that the token can be considered to reflect the actual sentiment of the analyst report.

Therefore, the greater the number of emotional words included in one token, the more accurately it can reflect the emotion of the word, which may vary depending on the topic or context. However, the number of times that token appears in analyst reports is relatively small. Therefore, the token may not be used to build the sentiment dictionary (i.e., it may be filtered).

To solve this problem, the selection unit 140 can adjust the preset number of times that is the standard for token filtering according to the number of emotional words included in a token. For example, the selection unit 140 may decrease the preset number as the number of emotional words included in a token increases, and may increase the preset number as the number of emotional words included in a token decreases.

The relationship between a token and evaluation classes may be the degree to which the token is included in each of the evaluation classes. Two examples of the association relationship can be used, and specific details about the two examples of the association relationship are shown in FIGS. 3 and 4, respectively.

1.1. Relationship 1

In the first example (Figure 3), the evaluation classes can be divided into two evaluation classes. The two evaluation classes are a positive class corresponding to an evaluation opinion with positive sentiment (e.g., upward recommendation) and a negative class corresponding to an evaluation opinion with negative sentiment (e.g., downward recommendation).

Accordingly, each of the analyst reports belongs to either the positive class or the negative class. An analyst report belonging to the positive class corresponds to a report with an upward recommendation, and an analyst report belonging to the negative class corresponds to a report with a downward recommendation.

The calculation unit 132 may compare and calculate the ratio (appearance ratio) between the number (number of appearances) of the token included in the recommendations upward reports and the number of the tokens included in the recommendation downgrade reports (S310 ).

That is, in the first example, the correlation is the ratio between the number of times the token appears in reports with upward recommendation opinions and the number of times the token appears in reports with downward recommendation opinions. The S310 process may be performed based on Bayesian probability.

The emotion calculation unit 134 may calculate the emotion of the corresponding token from positive emotion and negative emotion according to the ratio calculated by the calculation unit 132 (S320).

If the token appears more frequently in the report with an upward recommendation than in the report with a downward recommendation, the sentiment of the token may be calculated as a positive sentiment. On the other hand, if the token appears more often in the report with a lower recommendation opinion than in the report with an upward recommendation, the sentiment of the token may be calculated as a negative sentiment.

The calculation unit 130 can classify a list of tokens with positive sentiment and a list of tokens with negative sentiment by performing processes S310 and S320 on all tokens.

A test to verify the first example was conducted on a total of 510 analyst reports published on KOSPI 200 stocks between 2016 and 2018. Among the analyst reports, 242 are reports with upward recommendations, and the remaining 268 are reports with downward recommendations.

After removing English letters, numbers, and special characters from the main text (supporting opinions) in analyst reports, only nouns, adjectives, and adverbs were extracted as emotional words using KoNLPy morpheme analyzer, and then a total of 81,494 1-gram extracted emotional words were extracted. token and converted into 93,818 2-gram tokens.

Among the tokens, only tokens with an absolute frequency of appearance of 10 or more were selected, and the ratio (correlation) between the number of appearances was calculated for the selected torques based on Bayesian probability.

Afterwards, the emotions of the tokens were calculated according to the correlation, and as a result, the list of tokens calculated with positive emotions and the list of tokens calculated with negative emotions are shown in Tables 1 and 2, respectively.

1.2. Relationship 2

In the second example (FIG. 4), evaluation classes may be divided into a plurality of evaluation classes. A plurality of evaluation classes may be distinguished from each other according to the strength of the evaluation opinion.

The calculation unit 130 may calculate the probability that a token belongs to each of the evaluation classes based on the emotion of the token (S410). That is, in the second example, the correlation is the probability that the token is included in each of the evaluation classes. The S410 process may be performed based on Bayes probability or modified Bayes probability.

Specifically, among the evaluation classes, the calculation unit 132 calculates 'the ratio (first ratio) in which the corresponding token appears in analyst reports belonging to one (specific) evaluation class' and 'the remaining (other) evaluation classes. The ratio (second ratio) in which the corresponding token appears in analyst reports belonging to 'can be calculated (S412). The calculation unit 132 may perform the calculation process of S412 for each evaluation class and calculate ratios (first ratio and second ratio) for each evaluation class.

The emotion calculation unit 134 may calculate the probability corresponding to the calculated ratios as the emotion of the corresponding token (S414). Through this process, the probability (sentiment) that the token is included in each of the evaluation classes can be calculated.

The calculation unit 130 may classify a list of tokens with the highest probability of belonging to a specific evaluation class among a plurality of evaluation classes by performing process S410 on all tokens.

A test to verify the second example was conducted on analyst reports published on KOSPI 200 stocks between 2016 and 2018.

A total of four evaluation classes were formed through the relationship between the direction of evaluation opinions (target stock price) included in analyst reports (dTPRC) and the direction of cumulative abnormal returns (CAR). Here, the cumulative excess return is the value obtained by subtracting the market return from the stock return, and is the accumulated excess return between the date of publication of the relevant analyst report and the next day. The composition of the evaluation classes and the number of analyst reports belonging to each evaluation class are shown in Table 3 below.

In the direction of the target stock price in Table 3, (-) means that the target stock price presented by a specific analyst in the report published at time t-1 is higher than the target stock price presented in the report published at time t (i.e., the target stock price is lowered). means, and (+) means the opposite (i.e., raising the target stock price). Here, time t is the current time, and time t-1 is the time immediately before time t, so the report issued at time t-1 is the report issued immediately before the report issued at time t. In the direction of excess return in Table 3, (-) means that the rate of return is lower (the rate of return is lower for 2 days from the time of report publication), (+) means the opposite, and None means that there is no change in rate of return. do.

After removing English letters, numbers, and special characters from the main text (supporting opinions) in analyst reports, only nouns, adjectives, adverbs, verbs, and negative words are extracted as emotional words using the eKoNLPy morpheme analyzer specialized in the economic/financial field. The emotional words were converted into 1-gram tokens or 5-gram tokens.

Among the tokens, only tokens with an absolute frequency of occurrence of 10 or more were selected, and ratios (associations) were calculated based on modified Bayes probability for the selected torques.

Afterwards, the sentiment of the tokens was calculated according to the probabilities corresponding to the correlation, and as a result, the list of tokens for each evaluation class is shown in Tables 4 to 7, respectively.

2. Sentiment measurement using the constructed emotion dictionary

Figure 6 is a block diagram for explaining the structure of a device (hereinafter referred to as 'measurement device') 600 that measures the emotion of an emotion measurement target using an emotion dictionary constructed through the construction method described above, and Figure 7 is a flow chart to explain an example of a method for measuring emotion (hereinafter referred to as 'measurement method').

As shown in FIG. 6 , the measurement device 600 may be configured to include an emotion acquisition unit 610, an emotion extraction unit 620, and an emotion calculation unit 630.

The emotion acquisition unit 610 may acquire one or more emotion words from the target data (S710).

Target data refers to data corresponding to the target of emotion measurement. Target data may have various types and forms, such as the evaluation data described above. Hereinafter, the measurement device 600 will be described focusing on an example in which the target data corresponds to an analyst report.

The emotion acquisition unit 610 may obtain text data within the analyst report from the analyst report and obtain an emotion word representing emotion from one or more words included in the text data. The text data acquired by the emotion acquisition unit 610 may be an opinion corresponding to the basis for the evaluation opinion stated in the corresponding analyst report.

Depending on the embodiment, the emotion acquisition unit 610 may be configured to acquire only words corresponding to a specific part of speech as emotion words, or may not acquire words corresponding to other parts of speech.

For example, the emotion acquisition unit 610 may exclude numbers and special characters or exclude English from all words included in text data. Additionally, the emotion acquisition unit 610 may obtain all of the nouns, verbs, adjectives, adverbs, and negative words as emotion words, or may acquire only some of them as emotion words.

The specific part of speech acquired as an emotion word by the emotion acquisition unit 610 may match the part of speech used to construct the emotion dictionary. This is to measure the emotion more accurately in the step of measuring the emotion of the emotion word by comparing the emotion word acquired by the emotion acquisition unit 610 with the emotion dictionary.

The emotion extraction unit 620 may extract the emotion of each token containing one or more acquired emotion words (S720).

A sentiment dictionary constructed through a construction method can be used to extract sentiment from a token. That is, the emotion extraction unit 620 can compare each token with the emotion dictionary and extract the emotion of the token included in the emotion dictionary from among the tokens.

The extracted emotion can take two forms depending on how the emotion dictionary is constructed.

The first (sentiment type 1) is a case of comparing a token with the sentiment dictionary built by association 1. In this case, whether the token is a positive sentiment (included in the positive word list) or a negative sentiment (included in the negative word list) Emotions can be extracted in a form that indicates whether or not something is recognized.

The second (sentiment type 2) is a case of comparing a token with an emotion dictionary built by correlation 2. In this case, the emotion can be extracted in the form of a probability that the token belongs to each of a plurality of evaluation classes.

The emotion calculation unit 630 may calculate the emotion of the analyst report based on the extracted emotion (S730).

When the sentiment of tokens is extracted as sentiment type 1, the sentiment of the analyst report is calculated according to the ratio between the number of tokens with positive sentiment and the number of tokens with negative sentiment (frequency of appearance of positive and negative words in the analyst report). It can be.

For example, the sentiment of an analyst report can be calculated using Equation 1 below.

In Equation 1, POS represents the number of tokens with positive sentiment, NEG represents the number of tokens with negative sentiment, POS% represents the ratio between the number of tokens with positive sentiment among the total number of tokens, NEG% represents the ratio between the number of tokens with negative sentiment among the total number of tokens. OPN represents the sentiment of analyst reports.

According to Equation 1, the closer the OPN value is to -1, the more negative the analyst report is, and the closer the OPN value is to +1, the more positive the analyst report is. do.

When the sentiment of the tokens is extracted as sentiment type 2, a vector obtained by adding up the sentiments of the tokens (probability that the token belongs to each of a plurality of evaluation classes) for each evaluation class can be calculated as the sentiment of the analyst report.

For example, the sentiment of an analyst report can be calculated using Equation 2 below.

In Equation 2, Likelihood _doc represents the likelihood vector (sentiment of the analyst report) of the corresponding analyst report, and tb _i represents whether the i-th token appears in the analyst report. Accordingly, tb _i can have either a value of 1 (appears) or 0 (does not appear). l _n,i represents the probability that the i-th token belongs to the n-th evaluation class.

According to Equation 2, the sentiment of the analyst report (Likelihood _doc ) is It is calculated in vector form as follows.

3. Performance of emotional dictionary

For a total of 15,927 analyst reports, we compared and analyzed the results of measuring the sentiment of analyst reports using the sentiment dictionary constructed through the present invention and the results of measuring the sentiment of analyst reports using conventional sentiment dictionaries.

The degree of reflection of recommendations proposed in analyst reports was compared and analyzed, and the results are shown in Table 8.

In Table 8, KOSAC is a sentiment dictionary built by Seoul National University, KNU is a sentiment dictionary built by Kunsan National University, HV is a sentiment dictionary built by Harvard University, and LM is a sentiment dictionary built by Loughran & McDonald. Sentiment dictionary 1 is an emotion dictionary constructed based on relation 1 described above, and emotion dictionary 2 is an emotion dictionary constructed based on relation 2.

At recommendation opinion levels '1' and '2' indicating negative emotions, it can be seen that the emotions measured using Sentiment Dictionary 1 and Sentiment Dictionary 2 are more negative than the emotions measured using conventional emotion dictionaries.

At recommendation opinion level '3', which represents neutral emotion, it can be seen that the emotion measured using emotion dictionary 1 and emotion dictionary 2 is more neutral than the emotion measured using conventional emotion dictionaries.

At recommendation opinion levels '4' and '5', which indicate positive emotions, it can be seen that the emotions measured using Sentiment Dictionary 1 and Sentiment Dictionary 2 are more positive than the emotions measured using conventional emotion dictionaries.

We compared and analyzed the degree of reflection of changes in recommended opinions proposed in analyst reports, and the results are shown in Table 9.

In '-2' and '-2', which changed the recommendation opinion to negative, it can be seen that the emotions measured using Sentiment Dictionary 1 and Sentiment Dictionary 2 are more negative than the emotions measured using conventional emotion dictionaries.

At '0', which maintains the recommendation opinion, it can be seen that the emotion measured using emotion dictionary 1 and emotion dictionary 2 is more neutral than the emotion measured using conventional emotion dictionaries.

In '+1' and '+2', which changed the recommendation opinion to positive, it can be seen that the emotion measured using Sentiment Dictionary 1 and Sentiment Dictionary 2 is more positive than the emotion measured using conventional emotion dictionaries.

We compared and analyzed the degree of reflection of target stock price changes proposed in analyst reports, and the results are shown in Table 10.

In 'downgrade', where the target stock price is changed negatively, it can be seen that the emotions measured using Sentiment Dictionary 1 and Sentiment Dictionary 2 are more negative than the emotions measured using conventional emotion dictionaries.

At '0', which maintains the target stock price, it can be seen that the emotions measured using Sentiment Dictionary 1 and Sentiment Dictionary 2 are more neutral than the emotions measured using conventional emotion dictionaries.

In 'upgrade', which changes the target stock price positively, it can be seen that the emotions measured using Sentiment Dictionary 1 and Sentiment Dictionary 2 are more positive than the emotions measured using conventional emotion dictionaries.

As shown in the comparative analysis results of Tables 8 to 10, using the sentiment dictionary constructed through the present invention, sentiment for quantitative indicators such as the recommendation opinion level of the analyst report, change in recommendation opinion level, and change in target stock price, etc. can be identified more accurately.

The description of the present invention described above is for illustrative purposes, and those skilled in the art will understand that the present invention can be easily modified into other specific forms without changing the technical idea or essential features of the present invention. will be. Therefore, the embodiments described above should be understood in all respects as illustrative and not restrictive. For example, each component described as unitary may be implemented in a distributed manner, and similarly, components described as distributed may also be implemented in a combined form.

The scope of the present invention is indicated by the claims described below rather than the detailed description above, and all changes or modified forms derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present invention. do.

Meanwhile, the processes shown in FIGS. 2 to 5 and 7 can be implemented as computer-readable codes on a computer-readable recording medium. Computer-readable recording media include all types of recording devices that store data that can be read by a computer system. In other words, computer-readable recording media may be non-transitory media such as ROM, RAM, CD-ROM, magnetic tape, floppy disk, and optical data storage devices, and may also be non-transitory media such as carrier waves (e.g., It may further include transitory media such as transmission via the Internet) and data transmission medium. Additionally, computer-readable recording media may be distributed across computer systems connected to a network, and computer-readable codes may be stored and executed in a distributed manner.

100: A device that automatically builds an emotional dictionary
600: Device for measuring emotion

Claims (10)

As a method for automatically building an emotional dictionary,
A step in which the acquisition department acquires an opinion corresponding to the basis of the evaluation opinion from evaluation data containing an evaluation opinion about the evaluation object - the evaluation data belongs to one of a plurality of evaluation classes according to the evaluation opinion -;
extracting one or more emotion words representing emotion from words in the supporting opinion; and
A calculation unit calculating the emotion of the token according to an association between one or more tokens containing one or more of the emotion words and the evaluation classes,
The token is,
A method comprising two or more of the extracted emotional words. According to paragraph 1,
The evaluation classes are:
It includes a positive class corresponding to an evaluation opinion of positive emotion and a negative class corresponding to an evaluation opinion of negative emotion,
The calculating step is,
Comparing and calculating the number of tokens included in evaluation data belonging to the positive class and the number of tokens included in evaluation data belonging to the negative class among the evaluation data; and
A method characterized in that it includes a step where an emotion calculation unit calculates one of positive emotion and negative emotion as the emotion of the token, according to the result of the comparison operation. According to paragraph 1,
The evaluation classes are,
They are classified according to the strength of the above evaluation opinion,
The calculating step is,
A method characterized in that the calculation unit calculates the probability that the token belongs to each of the evaluation classes based on the sentiment of the token. According to paragraph 3,
The calculating step is,
Among the evaluation classes, the operation unit calculates the ratio of the token included in the evaluation data belonging to a specific evaluation class and the ratio of the token included in the evaluation data belonging to other evaluation classes excluding the specific evaluation class, and determines the evaluation class. Step of calculating for each field; and
A method characterized in that it includes the step of calculating the probability corresponding to the ratios calculated for each evaluation class as the emotion of the token. According to paragraph 1,
The emotional words are,
A method characterized by including one or more of nouns, verbs, adjectives, adverbs, and negatives, but not including numbers and special characters. delete According to paragraph 1,
The calculating step is,
A selection unit selecting tokens included in the evaluation data more than a preset number of times from among the tokens; and
A method characterized in that it includes a step of calculating the emotion of the selected token according to the correlation between the selected token and the evaluation classes. In clause 7,
The preset number of times is,
A method characterized in that the number of emotional words included in the token decreases as it increases. A device that automatically builds an emotional dictionary,
An acquisition unit that obtains an opinion corresponding to the basis of the evaluation opinion from evaluation data containing an evaluation opinion about the evaluation object - the evaluation data belongs to one of a plurality of evaluation classes according to the evaluation opinion - ;
an extraction unit that extracts one or more emotional words representing emotions from among the words in the grounded opinion; and
A calculation unit that calculates the emotion of the token according to a relationship between one or more tokens containing the one or more emotion words and the evaluation classes,
The token is,
A device comprising two or more of the extracted emotional words. A computer-readable recording medium recording a program for executing the method of any one of claims 1 to 5, 7, and 8 on a computer.

Images