JPH0682377B2 - Emotion information extraction device - Google Patents

Description

TECHNICAL FIELD The present invention relates to an apparatus and system for analyzing a Japanese sentence and extracting emotional information included in the Japanese sentence.

(Prior Art) Regarding the technology for analyzing Japanese sentences, morphological analysis means,
A technique of generating a semantic structure of a sentence by a syntactic analysis unit and a semantic analysis unit is known.

For example, the sentence "Beautiful flowers have bloomed." Is analyzed by morphological analysis.
Is a noun, "ga" is a case particle, "sakui" is a continuous form of the verb "blooming", and "ta" is an ending form of the auxiliary verb of completion.
It is parsed as follows: "Beautiful flower" is a noun phrase with a nominative, "ga" is attached to the main part, "Sakuta" is a predicate verb phrase, and "Beautiful flower has bloomed" is a sentence. . Then, the semantic analysis is performed such that the action of "blooming" has occurred in the "flower" having the attribute of "pretty".

(Problems to be solved by the invention) In the dialog processing between the computer and the user, in addition to the processing by logical information that has been conventionally performed, the user's state such as the ability and the psychological state of the user. The processing adapted to is required. However, even if the analysis up to the semantic analysis is performed,
There is no known technique for extracting user emotion information.

On the other hand, Japanese is characterized by many words that express emotions. In particular, many adverbs, adjectives, and adjective verbs directly express emotions. For example, the adjective "interesting" includes the feeling of "joy," and the word "exciting" includes the feeling of "sorry." The emotional elements of such words serve as clues when extracting emotional information from Japanese sentences.

An object of the present invention is to provide a method and apparatus for analyzing a Japanese sentence and extracting emotional information included in the Japanese sentence.

(Means for Solving Problems) The emotion information extraction device of the present invention is a morphological analysis unit that divides an input Japanese sentence into word strings, a syntactic analysis unit that parses the divided word strings, and a word. A word dictionary in which emotional elements of each unit are classified and registered, an emotional information conversion rule for converting the emotional elements based on a syntactic analysis result, and an emotional information conversion rule for the emotional elements of the word string extracted from the word dictionary. And emotion information extraction means for converting the information according to the following.

(Operation) By dividing the input Japanese sentence into word strings by the morpheme analysis means, it becomes possible to refer to the dictionary information in word units. By further parsing,
It is possible to obtain the dependency relation and sentence pattern between words and phrases. Emotional elements are registered in the words in the word dictionary.
For each word in the syntactic analysis tree obtained by the syntactic analysis, by referring to the word dictionary, the emotional element registered in the word can be extracted. The emotion information conversion rule is a rule for converting an emotion element extracted from a word according to sentence pattern information obtained by syntax analysis. The emotion element extracted from the word is converted into emotion information as a sentence by being converted by the emotion information conversion rule.

(Example) Hereinafter, the detail of this invention is demonstrated according to an Example.

FIG. 1 is a diagram showing an embodiment of the present invention. In the figure, 1 is an input unit for inputting Japanese sentences, 2 is a morphological analysis unit, 3 is a syntax analysis unit, 4 is a word dictionary in which emotional elements are registered, 5 is emotion information extraction unit, and 6 is a result of parsing emotional elements. Emotion information conversion rules for conversion according to the following, and 7 indicates an output display unit.
The morphological analysis unit 2 is. The Japanese sentence input from the input unit 1 is divided into word strings by referring to the word dictionary 8. The syntactic analysis unit 3 generates a syntactic analysis tree from the morphologically analyzed word string by a known syntactic analysis method, and further determines the type of sentence. The emotion information extraction unit 5 refers to the word dictionary 4 for each word in the syntactic analysis tree obtained by the syntactic analysis unit 3 and extracts the emotion element for each word, and further the syntactic analysis 3
Using the sentence type and emotion information conversion rule 6 obtained in step 1, the emotion element is converted according to the sentence type and output to the output unit 7.

FIG. 2 is a diagram showing an example of a word dictionary in which emotional elements are registered. In this dictionary, words that directly express emotions, such as adverbs, adjectives, and adjective verbs, are classified and registered according to the emotions they have. The emotion element section contains a name indicating the kind of emotion that the word has. For example, the adjective "interesting" is used when you are happy and interested,
"Fun" and "interest" are registered as emotional elements of "interesting". Multiple emotional elements can be registered. Emotion types include "joy, comfort, interest, self-confidence, expectation, anxiety, dissatisfaction, discomfort, confusion, pity, sadness, and anger." In addition, in the section of the grammatical information of the adverb, the rules regarding the response of the adverb are registered. Many verbs are a rule for confirming a particular emotion, as it represents a particular emotion when paired with a corresponding expression.

FIG. 3 is a diagram showing emotion information conversion rules. This rule gives a conversion rule when the emotion of the word situation changes to another emotion, such as when the word expressing the emotion is related to the negated word. In the cancellation column,
The conversion result when the emotional element is converted by the negated word is registered. The conversion result for the emotional element when the sentence is a question sentence is registered in the question sentence column. If it is blank, emotional information is not extracted.

Next, an embodiment of an apparatus that specifically realizes this method will be described with reference to the drawings. FIG. 4 is a diagram showing an apparatus according to one embodiment of the present invention. Input keyboard 41, display device 42, CP
The U43, the main storage device 44, and the auxiliary storage device 45 are connected via a bus 46. The auxiliary storage device 45 stores a word dictionary. A morpheme analysis program, a syntactic analysis program, and an emotion information extraction program are placed in the main storage device 44. In addition, a work area necessary for analyzing an input sentence and a dictionary buffer for fetching necessary data from a word dictionary. Area is provided. The processing program in main memory is CPU4
Performed by 3. The keyboard 41 is used for inputting instructions and Japanese sentences to the processing program. The display device 42 is used to display the input from the keyboard and the output of the processing result.

FIG. 5 is a flow chart showing the operation of the processing program. At input 501, one Japanese sentence is input. The morphological analysis 502 divides the input Japanese sentence into word strings. The syntactic analysis 502 generates a syntactic analysis tree from a word string divided into word units, and determines the type of sentence. Morphological analysis 50
Details of 2 and the syntax analysis 503 are known and will not be described. Below 504 is a part where emotion information is extracted. At 504, phrases that are not subject to emotion information extraction, such as quotation phrases, are removed from the syntax analysis tree. Phrases that are excluded from emotion information extraction include a quotation phrase enclosed in “”, a phrase related to the case particle “to” in the quotation, and a phrase related to the auxiliary particle “ka”. 505
Then, the word is extracted while tracing the parse tree obtained in 504. Finish when there are no more words to retrieve. 50
In step 6, it is checked whether or not the emotional element is registered in the extracted word, and if it is extracted, it is extracted and the processes in and after 507 are performed. If not, the process returns to 505. At 507, it is checked whether or not the focused word is related to the canceled word, and if so, the emotion element extracted from the word is converted according to the emotion information conversion rule. If it is not related to the negated word, it is not converted. In addition, the adverb corresponding to the word of cancellation is not converted. In 508, when the input sentence is a question sentence, the result of 507 is converted according to the emotion information conversion rule. The result obtained by the processing of 507 to 508 is temporarily stored in 509 and returns to 505. When all the words have been extracted, the process moves to 511. 511 checks rule 51 for conflicting stored results.
Check with 0 and remove any inconsistencies. The inspection rule 510 is a table in which, when emotional elements A and B are given, whether or not they conflict with each other is registered. For example, joy and interest do not conflict, joy and sadness do not. As for the method of removal, first, if there is emotion information obtained from the predicate, this is prioritized and those that contradict this are excluded. If emotional information cannot be obtained from the predicate, the emotional information obtained from the words directly related to the predicate is given priority, and contradictory information is excluded. If there is a contradiction, the one appearing in the latter 3 will take precedence. At 512, the final result obtained at 511 is output to the display device, and the processing ends.

Next, the operation will be described using a specific example. As input
Suppose the sentence "The speech yesterday was interesting." A parse tree is generated by the morphological analysis 502 and the syntactic analysis 503, and the result is that the main part is "Lecture of yesterday", the predicate is "interesting", and the type of sentence is a plain sentence. Next, in 504, quotation phrases and the like are removed. However, since there is no corresponding phrase, the process proceeds to the processing of 505 and below. In 505 and 506, a word in which an emotional element is registered is searched for while tracing a parse tree. Here, if the emotional element is registered in the word "interesting" and nothing is registered in other words, the emotional element is extracted from the word "interesting",
The result of "joy and interest" is obtained from Fig. 2. next
In 507, I check whether "interesting" is related to the canceled word, but since there is no canceled word, the previous result is passed to the next as it is. Further, in 508, processing is performed in the case of an interrogative sentence, but since the input sentence is a plain sentence and not an interrogative sentence from the result of the syntax analysis 503, the previous result of "joy and interest" is stored as an intermediate result. After checking the presence or absence of emotional elements in 505 and 506 for the remaining words in the input sentence, it is checked in 511 whether there are any conflicting results. Since "joy" and "interest" do not conflict, the result saved in 512 is entered,
The result is that the emotional information in the sentence "Yesterday's talk was interesting." Is "joy and interest."

If the input sentence is the sentence "Yesterday's lecture was not interesting.", The word "interesting" is related to the negated "no", so "joy and interest" extracted from "interesting" For the emotional element, conversion is performed when there is a word of cancellation. From FIG. 3, the result is that the cancellation of "joy" is "sorry, dissatisfied" and the cancellation of "interest" is "dissatisfied". The result of "sorry, dissatisfied" is obtained when one is duplicated.
If the input sentence is an interrogative sentence such as "Is the lecture yesterday interesting, wasn't it?""
The emotional information is obtained and is output as a result.

In the description of the present embodiment, it is described that a Japanese sentence is input from the input unit keyboard. However, for example, a Japanese sentence previously input to an external storage device or another input means such as a communication line or a voice recognition device is used. Similar processing can be performed on the input Japanese sentence. Moreover, although the output is described as being displayed on the display device, the output destination can be passed to another processing unit that performs processes such as voice synthesis and machine translation instead of the display device.

(Effect of the invention) By extracting emotion information of Japanese sentences, a response adapted to the emotion of the interlocutor can be generated in dialogue processing in natural language, and translation quality can be improved in machine translation. In a Japanese sentence generator such as a reader, it is possible to improve the quality of the generated voice such as how to put intonation.

[Brief description of drawings]

FIG. 1 is a block diagram showing the method of the present invention, FIG. 2 is a diagram showing an embodiment of a word dictionary in the present invention, FIG. 3 is a diagram showing emotion information conversion rules, and FIG. FIG. 5 is a diagram showing the configuration of the embodiment apparatus, and FIG. 5 is a flow chart showing the processing algorithm in the present invention. In the figure, 1 ... input section, 2 ... morphological analysis section, 3 ... syntax analysis section, 4 ... word dictionary, 5 ... emotion information extraction section, 6 ... emotion information exchange rule, 7 ... output section Are shown respectively.

Claims (1)

[Claims] 1. A morphological analysis unit that divides an input Japanese sentence into word strings, a syntactic analysis unit that parses the divided word strings, and a word dictionary that classifies and registers emotional elements in word units. Emotion information conversion rules for converting the emotion elements based on the result of syntactic analysis, and emotion information extraction means for converting the emotion elements of the word string extracted from the word dictionary according to the emotion information conversion rules. Emotion information extraction device.