JPWO2017154192A1 - Text analysis apparatus and text analysis method - Google Patents

Abstract

  The text analysis device that calculates the degree of association that the first text supports that the first expression has an effect on the second expression holds the first dictionary consisting of expressions that mean the action between the expressions When the third expression included in the first dictionary is included in the first text, each of the first expression and the second expression and an expression having a dependency relationship with the third expression in the first text A first feature amount indicating a degree of coincidence is calculated, and a relevance level is calculated based on the size of the first feature amount.

Description

  The present invention relates to a text analysis apparatus and a text analysis method.

  In recent years, techniques for extracting opinions from text are being developed. As background art of this technical field, there is JP-A-2007-133903 (Patent Document 1). This gazette states that “the opinion element storage means 3 includes a first set of notations that can indicate an object, a second set of notations that can indicate an evaluation point for the object, and an object for the point of interest. A third set of notations that can indicate a property and a fourth set of notations that can indicate a positive, negative, or neutral evaluation of the object are registered. The opinion expression location detecting means 6 detects the partial text expressing the opinion from the input text based on the result of the opinion element matching means 4. " (See summary).

JP 2007-133903 A

  For example, in the case where the first expression, the second expression, and the text are given, the first expression is estimated by estimating whether or not the text supports the action of the first word on the second word. And the opinion related to the second expression can be extracted from the text.

  Consider an example in which the first expression is “casino”, the second expression is “employment”, and the action is an acceleration action or a suppression action. The text “There is a research result that employment in other industries will decrease due to the establishment of casinos” supports that the casino, which is the first expression, is promoting the employment, which is the second expression. In addition, the text “There is a research result that employment in other industries will decrease due to the establishment of the casino.” Supports that the casino, which is the first expression, has a restraining effect on the employment, which is the second expression. .

  The technology described in Patent Document 1 includes a notation that can indicate an object, a notation that can indicate the focus of evaluation on the object, a notation that can indicate the nature of the object regarding the focus, Based on the comparison result between the notation that can indicate neutral evaluation and the text, the evaluation of the object and the property of the object that is the evaluation of the opinion are extracted from the text.

  However, in the above-described supportability estimation for the action, it is not always sufficient even if the estimation is performed using the characteristics of the words or phrases indicating the evaluation in the text as in the technique described in Patent Document 1. The estimation accuracy cannot be ensured. In the supportability estimation, it is effective to consider a syntactic connection such as a dependency relationship with an expression indicating an action included in the text.

  In order to solve the above problems, one embodiment of the present invention employs the following configuration, for example. A text analysis device for calculating a degree of relevance to which the first text supports that the first representation is acting on the second representation, comprising a processor and a storage device, the storage device comprising: , Holding a first dictionary consisting of expressions meaning actions between expressions, the processor accepting inputs of the first text, the first expression, and the second expression, and a third dictionary included in the first dictionary When the expression is included in the first text, each of the first expression and the second expression and the degree of coincidence of the meaning between the expression in the first text and the expression having a dependency relationship with the third expression A text analysis device that calculates one feature value and calculates the relevance based on a size of the first feature value.

  According to one aspect of the present invention, it is possible to estimate with high accuracy the degree to which the text supports the action of the first expression on the second expression.

  Problems, configurations, and effects other than those described above will be clarified by the following description of embodiments.

It is a block diagram which shows the structural example of the text analyzer in Example 1. FIG. 2 is an example of a dictionary in the first embodiment. 6 is an example of a primary feature amount extraction rule in the first embodiment. 4 is an example of a secondary feature value generation rule in the first embodiment. 6 is a flowchart illustrating an example of a text analysis process according to the first exemplary embodiment. 3 is an example of a display screen in the first embodiment. 10 is a flowchart illustrating an example of secondary feature value generation rule generation processing according to the second exemplary embodiment.

  Embodiments of the present invention will be described below with reference to the accompanying drawings. It should be noted that this embodiment is merely an example for realizing the present invention, and does not limit the technical scope of the present invention. In each figure, the same reference numerals are given to common configurations.

  This embodiment describes a text analysis device. The text analysis apparatus according to the present embodiment accepts input of the first expression, the second expression, and the text. The text analysis device estimates the degree to which the input text supports that the first representation acts on the second representation. In addition, the said effect | action shows the effect | action which opposes like the effect | action which consists of a promotion effect and a suppression effect, for example. In addition, an action composed of an approval action and an opposite action is an example of the action.

  In the present embodiment, an example will be described in which the text analysis apparatus estimates the degree to which the input text supports that the first expression gives a promoting action or a suppressing action to the second expression. Further, the text analysis apparatus may determine whether the input text supports that the first expression gives a promoting action or a suppressing action to the second expression according to the degree. Furthermore, when it is determined that there is support by text, the text analysis device may determine which of the promoting action and the suppressing action is given to the second expression by the first expression.

  In the present embodiment, the first expression is called an agenda word, and the second expression is called a value word. Note that each agenda word and value word are not necessarily one word, and may be expressions such as one or more words, one or more phrases, one or more sentences, or a combination thereof. Further, the text input to the text analysis device does not necessarily have to be one sentence, and may be one or more words, one or more phrases, one or more sentences, or a combination thereof.

  Further, when the text indicates that the agenda word has an action on the value word, the text supports the action that the agenda word has on the value word. A combination of information indicating whether or not the text supports the action of the agenda word on the value word and information indicating the strength of the support are called supportability.

  For example, a case where the agenda word is “casino” and the value word is “employment” will be described. When the text analysis apparatus accepts the input of the text “about 35,000 jobs are created by 42 casinos in Mississippi”, the text is the agenda word “Casino”. It is determined to indicate that the value word “employment” is being acted, specifically, to indicate that a promotion effect is being imparted.

  On the other hand, when the text analyzer accepts the input of the text “There is a research result that employment in other industries will decrease due to the establishment of the casino”, the text “Casino” which is the agenda word is the value word. It is determined that it indicates that the “employment” has an effect, and specifically indicates that it has an inhibitory effect.

  FIG. 1 shows a configuration example of a text analysis apparatus according to this embodiment. The text analysis apparatus 100 according to the present embodiment determines whether or not the input text supports the promoting action or the inhibiting action that the agenda word gives to the value word, and the strongness of the promoting action or the inhibiting action when it is supported. Estimate. In the present embodiment, an example in which a text analysis apparatus analyzes English text will be described.

  The text analysis apparatus 100 of the present embodiment is configured by a computer having a processor (CPU) 101, a memory 102, an auxiliary storage device 103, and a communication I / F (interface) 104, for example.

  The processor 101 executes a program stored in the memory 102. The memory 102 includes a ROM that is a nonvolatile storage element and a RAM that is a volatile storage element. The ROM stores an immutable program (for example, BIOS). The RAM is a high-speed and volatile storage element such as a DRAM (Dynamic Random Access Memory), and temporarily stores a program executed by the processor 101 and data used when the program is executed.

  The auxiliary storage device 103 is a large-capacity non-volatile storage device such as a magnetic storage device (HDD) or a flash memory (SSD), and stores a program executed by the processor 101 and data used when the program is executed. To do. That is, the program is read from the auxiliary storage device 103, loaded into the memory 102, and executed by the processor 101.

  The text analysis apparatus 100 may have an input I / F 105 and an output I / F 108. The input I / F 105 is an interface that is connected to a keyboard 106, a mouse 107, and the like and receives input from an operator. The output I / F 108 is an interface to which a display 109, a printer, and the like are connected, and the execution result of the program is output in a form that can be visually recognized by the operator.

  The communication I / F 104 is a network interface device that controls communication with other devices in accordance with a predetermined protocol. The communication I / F 104 includes a serial interface such as USB.

  A program executed by the processor 101 is provided to the text analysis device 100 via a removable medium (CD-ROM, flash memory, etc.) or a network, and is stored in a nonvolatile auxiliary storage device 103 which is a non-temporary storage medium. . For this reason, the text analysis device 100 may have an interface for reading data from a removable medium.

  The text analysis apparatus 100 is a computer system configured on a single physical computer or a plurality of logically or physically configured computers, and operates in a separate thread on the same computer. It may be possible to operate on a virtual machine constructed on a plurality of physical computer resources.

  The processor 101 includes, for example, a syntax analysis unit 111, a primary feature quantity extraction unit 112, a secondary feature quantity generation unit 113, a machine learning unit 114, and an evaluation unit 115. For example, the processor 101 operates as a syntax analysis unit 111 by operating in accordance with a syntax analysis program loaded in the memory 102, and extracts a primary feature value by operating in accordance with a primary feature extraction program loaded in the memory 102. It functions as the unit 112. The same applies to the other units included in the processor 101.

  The syntax analysis unit 111 captures text via the communication I / F 104 or the input I / F 105. The syntax analysis unit 111 analyzes the captured text. The primary feature quantity extraction unit 112 refers to the dictionary 200 and the primary feature quantity extraction rule 300 and extracts one or more primary feature quantities of the analyzed text from the agenda word, the value word, and the analyzed text. . Details of the primary feature amount will be described later.

  The secondary feature value generation unit 113 refers to the secondary feature value generation rule 400 and generates a secondary feature value of the analyzed text from the extracted primary feature value. The machine learning unit 114 calculates a score for evaluating the supportability of the text from the calculated secondary feature amount. Note that the machine learning unit 114 may calculate the score using the calculated primary feature amount. The evaluation unit 115 evaluates supportability by the text from the calculated score.

  The auxiliary storage device 103 stores, for example, a dictionary 200, a primary feature quantity extraction rule 300, and a secondary feature quantity generation rule 400. The dictionary 200 includes one or more dictionaries, and each of the one or more dictionaries includes one or more expressions. Each expression included in the dictionary is not necessarily a single word, and may be, for example, one or more words, one or more phrases, one or more sentences, or a combination thereof. Hereinafter, in this embodiment, an example in which each expression included in the dictionary is one word will be described for convenience. The dictionary 200 is determined in advance, for example.

  The primary feature quantity extraction rule 300 indicates a rule for extracting the primary feature quantity of the text from the parsed text. The secondary feature value generation rule 400 indicates a rule for generating a secondary feature value of the text from the primary feature value extracted from the parsed text.

  In the present embodiment, the information used by the text analysis apparatus 100 may be expressed in any data structure without depending on the data structure. For example, a data structure appropriately selected from a table, list, database or queue can store the information. 3 to 5 described later show examples in which each data held in the auxiliary storage device 103 is expressed in a table structure.

  FIG. 2 is an example of the dictionary 200. The dictionary 200 includes a dictionary name field 201, a word field 202, and a code field 203. The dictionary name column 201 stores the name of the dictionary. The word column 202 stores words included in the corresponding dictionary.

  The code column 203 stores a code corresponding to each word. Hereinafter, in this embodiment, it is assumed that one word is stored in the word column 202. Note that the values included in the code field 203 are not limited to +1 and -1, and may be positive or negative values.

  The dictionary 200 includes an effect dictionary, a shifted dictionary, an absolutized dictionary, and a negated dictionary. The affect dictionary is a dictionary composed of words having a meaning that gives a promotion action or a suppression action from one expression to another expression. For example, the sign of the word indicating the promoting action is positive, and the sign of the word indicating the suppressing action is negative. In addition, when the text analysis apparatus 100 estimates supportability about actions other than the promoting action and the suppressing action, the dictionary 200 includes a dictionary including words indicating the action.

  The shifted dictionary consists of words that reverse the relationship. The sign of each word in the shifted dictionary is negative. For example, comparing the text "Casino suppresses health problem." With the text "Casino ban suppresses health problem.", The former suggests "Casino has a suppressive effect on health problems." On the other hand, the latter suggests that "the casino is promoting the health problem". That is, the word “ban” reverses the suppressive action and the promoting action from “casino” to “health problem”. Thus, “ban” is an example of a word that reverses the relationship.

  The absolutized dictionary is composed of words that fix the value polarity indicated by the phrase containing the word that is the destination. For example, if the word “health” indicates a positive value, “health problem” indicates a negative value. On the other hand, if the word “crime” shows negative value, “crime problem” also shows negative value. In other words, the word “problem” fixes the value polarity (positive / negative) of a phrase composed of “problem” and the related word regardless of the value polarity of the related word. Therefore, "problem" is an example of a word that fixes the value polarity to negative. In the absolutized dictionary, the sign of the word that fixes the value polarity to positive is positive, and the sign of the word that fixes the value polarity to negative is negative.

  The negated dictionary consists of words that indicate negation. The sign of each word in the negated dictionary is negative. For example, it is preferable that the sign of the partial negation expression such as “not always” or “not every” is larger than the sign of the complete negation such as “not”.

  The dictionary 200 may further include, for example, a believe dictionary, a state dictionary, a significant dictionary, an equivalent dictionary, a contrast dictionary, and the like. The believe dictionary consists of words representing permission or prohibition. The sign of the word representing permission is positive, and the sign of the word representing prohibition is negative. Specifically, the believe dictionary includes words such as “allow” having a positive sign and “prohibit” having a negative sign.

  The state dictionary is composed of words indicating that a speech or announcement is to be made, and includes words such as “show” and “reject”, for example. For example, the sign of the word that affirms or promotes the utterance or announcement content that is the predicate of the word such as the word “show” is positive, and the utterance or predicate that is the predicate of the word such as the word “reject” The sign of a word that denies or suppresses the content of the presentation is negative.

  When the dictionary 200 includes a state dictionary, the dictionary 200 may further include an authority dictionary. The authority dictionary stores a word indicating a subject who makes a statement or announcement, and stores a reliability indicating the reliability of the statement or announcement by the subject in the code field 203 instead of a code. For example, the authority dictionary stores, for example, a word “research” with a confidence level of 0.9, a word “professor” with a confidence level of 0.7, a word “person” with a confidence level of 0.1, and the like. The words stored in the authority dictionary are also called authority words.

  The significant dictionary is composed of words indicating good or bad, and includes words such as “effective” and “dangerous”, for example. For example, the sign of a good word such as “effective” is positive, and the sign of a bad word such as “dangerous” is negative.

  The equivalent dictionary is composed of words indicating equivalence relations, and includes words such as “viewed” and “regarded”, for example. For example, the sign in the equivalent dictionary for a word representing equivalence, such as the word "viewed", is positive, and the sign in the equivalent dictionary for a word representing non-equality, such as the predicate "isnt" (abbreviation for is not) is there. The contrast dictionary is composed of words indicating comparison relationships, and includes, for example, words such as “than” and “compared”. The signs of the words contained in the contrast dictionary are all positive.

  FIG. 3 is an example of the primary feature quantity extraction rule 300. First, details of the primary feature amount will be described. First, the primary feature amount will be described. The primary feature amount is typically expressed in a format of “d.ai-tgt.mod”.

  d represents a predicate dictionary name. The predicate dictionary is a dictionary that stores words that can be text predicates among the dictionaries included in the dictionary 200. For example, the effect dictionary, believe dictionary, state dictionary, significant dictionary, equivalent dictionary, and contrast dictionary are all predicate dictionaries, and the shifted dictionary, absolutized dictionary, authority dictionary, and negated dictionary are not predicate dictionaries.

  ai represents a dependency expression, and includes a subject and an object. Hereinafter, in this embodiment, the subject is represented as a1 and the object is represented as a2. The expression in the dependency relationship may be, for example, a noun or a noun phrase. For example, “affect.a1” indicates the subject of the word in the text when a word included in the effect dictionary, or a synonym or synonym of the word is included in a certain text.

  tgt represents an object and is one of agenda word o, value word v, authority word authority, inv (o), inv (v), or inv (authority). Note that inv indicates an adjective word, for example, inv (v) indicates a synonym of the value word v. An antonym of an agenda word is, for example, an alternative word. For example, if the agenda o is "Hillary Clinton", "Jeb Bush" and "Donald Trump" are other candidates in the 2016 US presidential election (Hillary Clinton is a candidate in the 2016 US presidential election) Is an example of inv (o). The auxiliary storage device 103 may further hold an antonym dictionary for specifying an antonym of each word.

  The primary feature amount may not include tgt. That is, the primary feature amount may be expressed in the format of “d.ai-mod”. For example, when tgt is an agenda word o, tgt may include a synonym, a synonym, a part of speech difference, a broader word, a lower word, and the like of the agenda word o. That is, at this time, tgt may be a word group having the agenda word o as a representative word. The same applies when tgt is the value word v or the authority word authority.

  mod includes, for example, dictionary names that are not predicate dictionaries such as shifted, absolutized, and negated, and unif. unif represents the degree of agreement between the words in ai and the meaning of tgt. The degree of coincidence is, for example, expressed by a value between 0 and 1, and ai is a character string that completely matches tgt (for example, even if there is a difference due to usage or only a difference due to single or double) 1.0 is included when a word is included, and 0.0 when ai consists only of a word that is completely different from tgt.

  When the synonym of tgt is included in ai, for example, the ratio of the IDF value with the original word calculated from a predetermined document or the like may be used as the degree of coincidence. However, when the ratio exceeds 1, the matching degree is 1. Since unif is an index indicating the connection between a predicate representing an action relationship and o or v, it is used to identify the presence or absence of the action relationship.

  shifted is a value indicating whether or not ai includes a word included in the shifted dictionary. For example, when ai includes a word included in the shifted dictionary, the value of shifted is a code in the shifted dictionary of the word, and when not included, the value of shifted is +1. Since shifted is an index indicating the presence of a word that reverses the relationship, it is used to identify the direction of the action when there is an action relationship, that is, whether the action is an accelerating action or an inhibitory action.

  absolutized is a value indicating whether or not ai includes a word included in the absolutized dictionary. For example, when ai includes a word included in the absolutized dictionary, the absolutized value is a code in the absolutized dictionary of the word, and when not included, the absolutized value is a null value and is not calculated. Since absolutized is an index indicating the presence of a word that fixes the polarity of value, it is used to identify the direction of action when there is an action relationship.

  Therefore, the primary feature value “d.ai-tgt.mod” is the word in the text of the predicate dictionary d, and the dependency relationship specified by ai includes the target tgt and / or It is a value indicating the feature corresponding to the target mod.

  For example, “affect.a1-o.unif” is a match between the subject (a1) of the word and the agenda (o) in the text when the word contained in the effect dictionary is included in the text. A value indicating the degree (unif).

  Note that “affect.a1-o.unif” is not calculated when the text does not include a word included in the effect dictionary. If the text includes a plurality of words included in the effect dictionary, “affect.a1-o.unif” is calculated for each of the plurality of words. The same applies to other primary feature quantities using the effect dictionary and primary feature quantities using other predicate dictionaries.

  Further, for example, “aff.a1-o.shifted” is a feature amount indicating whether or not a word included in the effect dictionary is included in the text, and there is a word that inverts the agenda in the subject of the word in the text It is.

  The primary feature amount may be expressed in the format “NP-tgt.mod”. NP indicates a noun phrase. That is, the primary feature “NP-tgt.mod” is a value indicating a feature corresponding to the noun phrase and / or the target mod when there is a noun phrase including the target tgt in the text. For example, “NP-v.shifted” is a value indicating whether or not the noun phrase includes a word in a shifted dictionary related to the value word when the text includes a noun phrase including the value word.

  Furthermore, the following primary feature amounts may be generated. For example, the primary feature quantity “sta.tgt.authority” indicates the reliability in the authority dictionary of the target tgt of the word when the text includes a word included in the state dictionary. The primary feature “sta.tgt.authority” is an index that indicates the presence of a predicate that represents a statement and the reliability of the target of the predicate, and thus identifies the strength of support, that is, the strength of promoting or suppressing Used to do.

  The primary feature “d.negated” indicates whether the text includes a word included in the dictionary d and whether the word is associated with a negative expression such as “not”. The value when negative expression is involved is -1, and the value when negative expression is not is 1. The primary feature “d.negated” is an index indicating the presence of a negative word, and is used to identify the direction of action when there is an action relationship.

  Further, in natural language processing, a generally used feature value may be generated as a primary feature value. For example, a value indicating the length of the text (for example, the number of words or characters in the text), a value indicating the number of noun phrases in the text, or the like may be generated as the primary feature amount.

  Here, some specific examples of primary features for the text t1: “The research showed that casino ban suppresses health problem.”, The agenda o: casino :, and the value word: health will be described.

  The word “suppresses” included in t1 is a word indicating a suppression action, and completely matches “suppress” included in the effect dictionary, and the code in the effect dictionary of “suppress” is −1. At t1, the subject a1 of “suppresses” includes the agenda word “casino”. Therefore, the value of the primary feature amount effect.a1-o.unif is 1. At t1, the object of “suppresses” includes the value word “health”. Therefore, the value of the primary feature amount effect.a2-v.unif is 1.

  Further, the word “ban” included in t1 is a word included in the shifted dictionary, and the code in the shifted dictionary of “ban” is −1. “ban” is associated with the agenda “casino” at t1. As described above, the agenda word “casino” is included in the subject of “suppresses”. Therefore, the value of the primary feature amount effect.a1-o.shifted is -1.

  Further, the word “problem” included in t1 is a word included in the absolutized dictionary, and the code of the “problem” in the absolutized dictionary is −1. "problem" is associated with the value word "health" at t1. Further, as described above, the value word “health” is included in the object of “suppresses”. Therefore, the value of the primary feature amount effect.a2-v.absolutized is -1.

  The word “research” included in the subject a1 of t1 is a word included in the state dictionary. Therefore, for example, when the reliability in the authority dictionary of “research” is 0.9, the value of the primary feature quantity sta.a1-authority is 0.9.

  Hereinafter, the primary feature quantity extraction rule 300 will be described. The primary feature quantity extraction rule 300 includes, for example, a primary feature quantity name field 301, a predicate dictionary field 302, a dependency field 303, a target field 304, and a calculation method field 305.

  The primary feature quantity name column 301 stores the name of the primary feature quantity. The predicate dictionary column 302 stores the name of the predicate dictionary d used for calculating the corresponding primary feature quantity. The dependency column 303 stores types of expressions having a dependency relationship, which are used for calculating the corresponding primary feature value. For example, each of the subject a1 and the object a2 is an example of an expression type having a dependency relationship.

  The target column 304 stores a target tgt used for calculation of the corresponding primary feature quantity. Specifically, the target column 304 stores, for example, the agenda word o, the value word v, or the authority word authority. The calculation method column 305 stores the calculation method of the corresponding primary feature quantity.

  The primary feature quantity extraction rule 300 has a format of {affect.ai-tgt.mod} in which i is 1 or 2, tgt is the agenda word o or the value word v, and mod is unif, shifted, or absolutized. The extraction rule of the primary feature-value represented by these is included. That is, the primary feature quantity extraction rule 300 is at least effect.a1-o.unif, effect.a1-o.shifted, effect.a1-o.absolutized, effect.a1-v.unif, effect.a1-v. shifted, effect.a1-v.absolutized, effect.a2-o.unif, effect.a2-o.shifted, effect.a2-o.absolutized, effect.a2-v.unif, effect.a2-v.shifted, And extraction rules for 12 types of primary feature values of effect.a2-v.absolutized.

  In addition, the primary feature quantity extraction rule 300 is, instead of the above-described 12 types of primary feature quantity extraction rules, tgt is the agenda word o or the value word v, and mod is unif, shifted, or absolutized, {NP -tgt.mod} may be included as a primary feature extraction rule. In this case, the primary feature quantity extraction rule 300 includes at least 6 of NP-o.unif, NP-o.shifted, NP-o.absolutized, NP-v.unif, NP-v.shifted, NP-v.absolutized. Includes extraction rules for types of primary features.

  When the primary feature quantity extraction rule 300 includes a primary feature quantity calculation rule using a predicate dictionary d, the primary feature quantity extraction rule 300 includes a primary feature quantity d.negated calculation rule for the predicate dictionary.

  Note that when the text analysis apparatus 100 determines only the presence / absence of the action as the supportability evaluation and does not identify the promotion action and the suppression action, the primary feature quantity extraction rule 300 determines that the primary feature quantity whose mod is shifted or absolutized. And the extraction rule of the primary feature value d.negated may not be included.

  The text analysis apparatus 100 can improve the accuracy of the supportability evaluation of the action by the text by using the above-described primary feature amount.

  FIG. 4 is an example of the secondary feature quantity generation rule 400. The secondary feature quantity generation rule 400 includes, for example, a secondary feature quantity name field 401, an exclusive secondary feature quantity name field 402, a score calculation method field 403, and a code calculation method field 404.

  The secondary feature amount is generally expressed in a format in which one or more target tgt is described in at least one of the left side and the right side of the predicate dictionary name such as “o affect v”. The secondary feature quantity of the format includes a word included in the effect dictionary, and the text includes the tgt described on the left side of the predicate dictionary name in the subject of the word, and the word Is a value indicating whether or not tgt described on the right side of the predicate dictionary name is included in the predicate.

  For example, the secondary feature amount “o affect v” is obtained when the word included in the effect dictionary includes the agenda word o in the subject of the word and the predicate of the word in the text. This is a value indicating whether or not the value word v is included.

  The secondary feature quantity “ov affect inv (v)” includes a word “o” and a value word “v” as the subject of the word when the word included in the effect dictionary is included in the text, and It is a value indicating whether or not a synonym of the value word v is included in the predicate of the word.

  Further, effect.sign described in FIG. 4 indicates a code of the word in the effect dictionary when the word included in the effect dictionary is included in the text. Also, val.sign shown in FIG. 4 indicates the sign of the value word v.

  The secondary feature quantity name column 401 stores the name of the secondary feature quantity. The exclusive secondary feature quantity name column 402 stores the names of secondary feature quantities that have an exclusive relationship with the corresponding secondary feature quantity. Specifically, a corresponding secondary feature value is generated only when the value of the secondary feature value stored in the exclusive secondary feature value name field 402 does not exist (in this embodiment, it is zero). Is done. For example, referring to the second line of the secondary feature value generation rule 400 in FIG. 4, the secondary feature value “o affect v” is zero when the value of the secondary feature value “ov affect inv (v)” is zero. Is calculated only for

  The score calculation method column 403 stores a method for calculating the score of the secondary feature amount. The code calculation method column 404 stores a code calculation method of the secondary feature quantity. The calculation methods stored in the score calculation method column 403 and the code calculation method column 404 use primary feature amounts. The score and sign of the secondary feature quantity include the primary feature quantity itself or a product of a plurality of primary feature quantities.

  The secondary feature quantity generation rule 400 may include a primary feature quantity as a secondary feature quantity. That is, a certain primary feature quantity itself may be a secondary feature quantity. When the primary feature amount itself is a secondary feature amount, for example, the score calculation method column 403 stores a calculation method of an absolute value of a value calculated by the calculation method indicated by the calculation method column 305 corresponding to the primary feature amount. The code calculation method column 404 stores a code calculation method for values calculated by the calculation method.

  FIG. 5 shows an example of the text analysis process. First, the syntax analysis unit 111 accepts text input via the communication I / F 104 or the input I / F 105, and the primary feature quantity extraction unit 112 receives one or more values via the communication I / F 104 or the input I / F 105. , The codes of the one or more value words, and one or more agenda words are received (S501).

  The syntax analysis unit 111 performs syntax analysis on the input text (S502). Specifically, the syntax analysis unit 111 specifies, for example, a word included in the input text and a part of speech of the word using a known syntax analysis method. The syntax analysis unit 111 further specifies, for example, the dependency between the specified words and the phrase included in the text and the part of speech of the phrase.

  Further, the syntax analysis unit 111 identifies a subject and an object of a verb (or verb phrase) included in the text using a predetermined rule, for example. The subject and the object are nouns or noun phrases. The predetermined rule is, for example, a rule for specifying the subject of the verb (or verb phrase) and the object from the arrangement of the phrase or the part of speech of the word in the sentence.

  Subsequently, the primary feature quantity extraction unit 112 extracts, from the analyzed text, for example, all primary feature quantities whose calculation methods are described in the primary feature quantity extraction rule 300 (S503). In addition, when a plurality of words included in the effect dictionary are included in the text, the primary feature amount extraction unit 112 uses, for example, a primary dictionary that uses an effect dictionary such as effect.a1-o.unif for each of the plurality of words. Extract the feature value. The same applies to other predicate dictionaries such as the state dictionary.

  The secondary feature quantity generation unit 113 uses the primary feature quantity extracted in step S503, for example, among the secondary feature quantities whose calculation methods are described in the secondary feature quantity generation rule 400, for example, an exclusive secondary feature quantity. The score and sign of the secondary feature quantity that has no or all exclusive secondary feature quantity values are zero (S504).

The machine learning unit 114 calculates the degree of association f _score (t) and the directionality f _sign (t) using the secondary feature amount calculated in step S504 (S505). The degree of relevance f _score (t) is an index indicating that the greater the value, the stronger the degree that the text supports the action of the agenda word on the value word. The directionality f _sign (t) indicates that the larger the value is, the more the agenda word supported by the text has an action on the value word. The smaller the value, the more the action. Indicates that the effect is in the suppression direction. Note that t represents the input text.

For example, the machine learning unit 114 multiplies each of the secondary feature value scores calculated in step S504 by a predetermined weight for each secondary feature value, and calculates the sum of the weighted secondary feature value scores. Relevance is calculated as f _score (t).

Further, for example, the machine learning unit 114 applies a predetermined weight for each secondary feature value to each code of the secondary feature value calculated in step S504, and sets the weighted secondary feature value code. Calculate the sum. The machine learning unit 114 further calculates, for example, a value obtained by multiplying the calculated _{sign of the} secondary feature value by g _sign (t) = state.sign × state.negated as the directionality f _sign (t). .

Note that state.sign is a code in the state dictionary of the predicate when the input text includes a word included in the state dictionary. If the input text does not include a word included in the state dictionary, g _sign (t) = + 1.

  State.negated is the code value of the word included in the negated dictionary related to the word included in the state dictionary when the input text includes a word included in the state dictionary. Note that sta.negated = + 1 when the word included in the negated dictionary related to the word included in the state dictionary does not exist in the text.

Expressions expressing negative utterances such as "reject" reverse the meaning of the whole text (state.sign = -1). Also, negative expressions that are related to expressions that express statements also reverse the meaning of the whole text (state.negated). Therefore, the machine learning unit 114 can more accurately calculate the direction of supportability of the entire text by multiplying the sum of codes separately from other secondary feature quantities by g _sign (t) at the end. it can. Note that the weight used when calculating the degree of association and the weight used when calculating the directionality in the same secondary feature quantity may be different from each other.

Subsequently, the evaluation unit 115 evaluates the supportability of the promotion / suppression action given to the value word by the agenda word based on the input text (S506). Specifically, for example, when the f _score (t) is equal to or greater than the first threshold, the evaluation unit 115 determines that the text indicates that the agenda word has a promoting action or a suppressing action on the value word. , F _score (t) is less than the first threshold, it is determined that the text indicates that the agenda word does not give the value word a promoting action or a suppressing action.

In addition, the evaluation unit 115 indicates that, for example, if f _score (t) is greater than or equal to the first threshold and f _sign (t) is greater than or equal to the second threshold, the agenda word has a promoting action on the value word Is determined to be indicated by the text. In addition, the evaluation unit 115 indicates that, for example, if f _score (t) is greater than or equal to the first threshold value and f _sign (t) is less than the second threshold value, the agenda word has an inhibitory action on the value word Is determined to be indicated by the text. The second threshold is 0, for example.

  For example, the evaluation unit 115 outputs the evaluation result to the display 109 or an external device such as another computer via the communication I / F 104 or the output I / F 108 (S507). The output contents of the evaluation result will be described later with reference to FIG.

In step S506, the evaluation unit 115 may determine only whether or not the text indicates that the agenda word has a promoting action or a suppressing action on the value word. In this case, the machine learning unit 114 does not have to calculate f _sign (t) in step S505, and the secondary feature value generation unit 113 does not have to calculate the sign of each secondary feature value in step S504. .

  In this case, the primary feature quantity extraction unit 112 may extract only the primary feature quantity included in the score calculation method column 403 of the secondary feature quantity generation rule 400 in step S503. In particular, when the scores of all secondary feature values are expressed only by primary feature values including unif, such as the scores of “o affect” and “ov affect inv (v)” in FIG. The unit 112 may extract only the primary feature amount including unif.

  FIG. 6 shows an example of the display screen of the present embodiment. The display screen 600 includes, for example, a value word agenda display area 601, a supportability estimation result display area 602, a text display area 603, a secondary feature quantity display area 604, a dictionary matching display area 605, and a primary feature quantity display area 606. Including.

  The value word agenda word display area 601 displays the input value word and the agenda word. Further, the value word agenda word display area 601 may function as an area for receiving input of a value word and an agenda word. The supportability estimation result display area 602 displays the supportability (not promotion, suppression, or neither promotion nor suppression) determined in step S506. A text display area 603 displays the input text. The text display area may function as an area for receiving text input.

  The secondary feature quantity display area 604 displays the score and code of the secondary feature quantity calculated in step S504. The secondary feature quantity display area 604 may function as an area for tuning the secondary feature quantity generation rule 400. In this case, the secondary feature quantity display area 604 includes an area for inputting, for example, a secondary feature quantity name, an exclusive secondary feature quantity name, a score calculation method, and a code calculation method.

  The dictionary matching display area 605 displays a list of words included in each dictionary among the words in the text, codes in the corresponding dictionary of each word, and the subject and object of each word.

  The primary feature quantity display area 606 displays the value of the primary feature quantity extracted in step S503. The primary feature quantity display area 606 may function as an area for tuning the primary feature quantity extraction rule 300. In this case, the primary feature quantity display area 606 includes, for example, an area for inputting a primary feature quantity name, a predicate dictionary, a dependency, a target, and a calculation method.

  As described above, the text analysis apparatus 100 according to the present embodiment can estimate the presence / absence of support with high accuracy by using the feature amount using the words included in the predicate dictionary and the dependency relationship between the words. Further, the text analysis apparatus 100 can estimate supportability with high accuracy by using a secondary feature value generated from a plurality of primary feature values.

  In addition, the text analysis apparatus 100 uses a feature amount such as shifted, absolutized, and negated, so that the direction of support when there is support (in this embodiment, whether the action is promotion or suppression). Can be estimated with high accuracy. Further, the text analysis apparatus 100 can estimate the strength of support with high accuracy by using the feature amount using the state dictionary and the authority dictionary.

  The text analysis apparatus 100 according to the present exemplary embodiment generates a secondary feature amount generation rule 400. FIG. 7 shows an example of the generation process of the secondary feature quantity generation rule 400. In the present embodiment, a rule for generating a secondary feature amount is generated, which is expressed in a format in which one or more target tgt is described in at least one of the left side and the right side of the predicate dictionary name d like “o affect v” An example of the processing will be described. Since the processes in steps S701 to S703 are the same as the processes in steps S501 to S503, description thereof will be omitted.

  The secondary feature quantity generation unit 113 determines one or more secondary feature quantities that can be generated from the one or more primary feature quantities generated in step S703, and names the determined secondary feature quantities as secondary feature quantity names. Stored in the column 401 (S704). The secondary feature amount in step S704 may be designated by the user.

  Further, the secondary feature quantity generation unit 113 selects a primary feature quantity combination that affects the supportability by using a known algorithm such as feature quantity selection, and secondary features that can be generated from the selected combination. The amount may be determined as a secondary feature amount in step S704.

  For example, the secondary feature quantity o affect v can be generated when the primary feature quantities effect.a1-o.unif and effect.a2-v.unif are extracted in step S703. As described above, effect.a1-o.unif is a primary feature amount indicating that the subject of the word in the text includes the agenda word o when the word included in the effect dictionary is included in the text. , Effect.a2-v.unif is a primary feature amount indicating whether or not the value word v is included in the predicate of the word in the text when the word included in the effect dictionary is included in the text.

  Therefore, when a word included in the effect dictionary is included in the text, it indicates whether the subject of the word includes the agenda word o and the word predicate includes the value word v in the text. The secondary feature amount o affect v is one of the secondary feature amounts calculated when the effect.a1-o.unif and the effect.a2-v.unif exist.

  Subsequently, the secondary feature quantity generation unit 113 determines a score calculation method and a code calculation method for each determined secondary feature quantity according to a predetermined rule, and the determined score calculation method is stored in the score calculation method column 403. And the determined code calculation method is stored in the code calculation method column 404 (S705). Note that the secondary feature quantity generation unit 113 may use the score and code calculation method designated by the user as the score and code calculation method in step S705.

  Hereinafter, an example of a process for determining a score and code calculation method using a predetermined rule will be described. First, an example of a score calculation method determination process will be described. When at least one of the secondary feature values tgt is the agenda word o, the secondary feature value generation unit 113 determines, for example, that the score is a value of d.ai-o.unif. When the agenda word o is included on the left side of the predicate dictionary name in the secondary feature quantity, ai in the score d.ai-o.unif is a1, and the agenda word o on the right side of the predicate dictionary name in the secondary feature quantity In the score d.ai-o.unif is a2.

  When the agenda word o is included on the left side and the right side of the predicate dictionary name in the secondary feature quantity, the secondary feature quantity generation unit 113 has scores of d.a1-o.unif and d.a2-o. It may be determined that it is a product of unif, the score may be determined to be the sum of d.a1-o.unif and d.a2-o.unif, and the score may be d.a1-o It may be determined that the value is either .unif or d.a2-o.unif. When all tgt included in the secondary feature amount are not the agenda word o, the secondary feature amount generation unit 113 determines that the score is 1.0, for example. The score determination method in the case where at least one of the secondary feature values tgt includes the acronym inv (o) is the same. That is, o in the determination process of the above-described score calculation method may be read as inv (o).

  When the secondary feature quantity tgt includes both o and inv (o), the secondary feature quantity generation unit 113 sets the score of the secondary feature quantity to, for example, d.ai-o corresponding to o. It may be determined to be the product of d.ai-inv (o) .unif corresponding to .unif and inv (o), or the d.ai-o.unif and d.ai-inv (o ) .unif and may be determined to be one of the values of d.ai-o.unif and d.ai-inv (o) .unif Also good.

  Hereinafter, an example of a code calculation method determination process will be described. The code is defined by, for example, a product of a plurality of primary feature values. The plurality of primary feature quantities include primary feature quantities for predicates in the predicate dictionary. For example, the secondary feature quantity generation unit 113 determines d.sign and d.negated as primary feature quantities for the predicate.

  When at least one of the secondary feature quantities tgt is the agenda word o, the plurality of primary feature quantities include primary feature quantities for o. For example, the secondary feature quantity generation unit 113 determines d.ai-o.shifted in each of the at least one o as a primary feature quantity for o. The ai of d.ai-o.shifted for o included on the left side of the predicate dictionary name is a1, and the ai of d.ai-o.shifted for o included on the right side of the predicate dictionary name is a2.

  Further, when at least one of the secondary feature values tgt is the acronym inv (o), the plurality of primary feature values include primary feature values for inv (o). For example, the secondary feature quantity generation unit 113 determines d.ai-inv (o) .shifted in each of the at least one inv (o) as a primary feature quantity for inv (o). The ai of d.ai-inv (o) .shifted for inv (o) included on the left side of the predicate dictionary name is a1, and d.ai-inv for inv (o) included on the right side of the predicate dictionary name. (o) .shifted ai is a2.

  When at least one of the secondary feature quantities tgt is the value word v, the plurality of primary feature quantities include primary feature quantities for v. For example, the secondary feature quantity generation unit 113 determines, for example, d.ai-v.shifted and c in each of the at least one v as primary feature quantities for v. C is a primary feature that takes 1 when effect.ai-v.absolutized is a null value and takes effect.ai-v.absolutized * val.sign when effect.ai-v.absolutized is not a null value. It is. Note that ai of d.ai-v.shifted for v included in the left side of the predicate dictionary name is a1, and ai of d.ai-v.shifted for v included in the right side of the predicate dictionary name is a2. The same applies to ai in d.ai-v.absolutized.

  When at least one of the secondary feature quantities tgt is the synonym inv (v) of the value word, the plurality of primary feature quantities include primary feature quantities for inv (v). For example, the secondary feature value generation unit 113 determines, for example, d.ai-inv (v) .shifted and c in each of the at least one inv (v) as primary feature values for inv (v). To do. C takes 1 when effect.ai-v.absolutized is a null value, and takes effect.ai-v.absolutized * val.sign * -1 when effect.ai-v.absolutized is not a null value. A primary feature. The ai of d.ai-inv (v) .shifted for inv (v) included on the left side of the predicate dictionary name is a1, and d.ai-inv for inv (v) included on the right side of the predicate dictionary name. (v) .shifted ai is a2. The same applies to ai in d.ai-inv (v) .absolutized.

  Hereinafter, the determination process of the calculation method of a score and a code will be described taking the secondary feature quantity o affect v as an example. Since o affect v includes one agenda word o on the left side of the predicate dictionary name, the secondary feature generation unit 113 determines that the score of o affect v is the value of effect.a1-o.unif. .

  The secondary feature quantity generator 113 selects a primary feature quantity effect.sign and effect.negated for the predicate dictionary. Moreover, since o affect v includes one agenda word o on the left side of the predicate dictionary name, the secondary feature quantity generation unit 113 selects effect.a1-o.shifted as the primary feature quantity for o. Moreover, since o affect v includes one value word v on the right side of the predicate dictionary name, the secondary feature quantity generation unit 113 has effect.a2-v.shifted, c, Select. The secondary feature value generation unit 113 determines the product of the primary feature values for which o affect v is selected, that is, effect.a1-o.shifted * affect.negated * affect.sign * affect.a2-v.shifted * c To do.

  Return to the description of the flowchart. After determining the secondary feature quantity score and code calculation method, the secondary feature quantity generation unit 113 determines an exclusive secondary feature quantity for each secondary feature quantity, and determines the determined exclusive secondary feature quantity. The name is stored in the exclusive secondary feature name column 402 (S706).

  For example, when the first secondary feature quantity and the second secondary feature quantity satisfy all the following three conditions, the secondary feature quantity generation unit 113 converts the second secondary feature quantity to the first secondary feature quantity. The exclusive secondary feature value is determined. The first condition is that the predicate dictionary name of the first secondary feature quantity matches the predicate dictionary name of the second secondary feature quantity.

  The second condition is that, for each tgt included on the left side of the predicate dictionary name of the first secondary feature quantity, the tgt or the synonym inv (tgt) of the tgt is on the left side of the predicate dictionary name of the second secondary feature quantity. It is included. Note that inv (inv (tgt)) = tgt. The third condition is that, for each tgt included on the right side of the predicate dictionary name of the first secondary feature quantity, the tgt or the synonym inv (tgt) of the tgt is on the right side of the predicate dictionary name of the second secondary feature quantity. It is included. Therefore, for example, the secondary feature quantity generation unit 113 compares all secondary feature quantities including the same predicate dictionary name and determines an exclusive secondary feature quantity of each secondary feature quantity.

  When the secondary feature quantity generation rule 400 includes the secondary feature quantity o affect v and ov affect v, the secondary feature quantity generation unit 113, for example, replaces ov affect v with o affect v according to the above-described conditions. To decide. Similarly, when the secondary feature quantity generation rule 400 includes the secondary feature quantity o affect v and ov affect inv (v), the secondary feature quantity generation unit 113, for example, according to the above-described condition, ov affect inv (v ) Is determined as the exclusive feature of o affect v.

  Note that the secondary feature quantity o v affect v is a feature quantity indicating the feature indicated by the secondary feature quantity o affect v and reflecting more information than o affect v. Therefore, the secondary feature quantity o v affect v can calculate a more accurate code than the secondary feature quantity o affect v.

  Therefore, when the secondary feature value generation unit 113 determines the exclusive secondary feature value as described above, if ov affect v exists (is not zero), the value of ov affect v is used for the supportability calculation. Therefore, the accuracy of the supportability evaluation is improved. Even if o v affect v does not exist (is zero), the text analysis apparatus 100 can evaluate support using the value of o affect v.

  The text analysis apparatus 100 according to the present embodiment can determine the secondary feature quantity capable of estimating the supportability with high accuracy by generating the secondary feature quantity generation rule 400 as described above.

  In addition, this invention is not limited to an above-described Example, Various modifications are included. For example, the above-described embodiments have been described in detail for easy understanding of the present invention, and are not necessarily limited to those having all the configurations described. Further, a part of the configuration of a certain embodiment can be replaced with the configuration of another embodiment, and the configuration of another embodiment can be added to the configuration of a certain embodiment. Further, it is possible to add, delete, and replace other configurations for a part of the configuration of each embodiment.

  Each of the above-described configurations, functions, processing units, processing means, and the like may be realized by hardware by designing a part or all of them with, for example, an integrated circuit. Each of the above-described configurations, functions, and the like may be realized by software by interpreting and executing a program that realizes each function by the processor. Information such as programs, tables, and files for realizing each function can be stored in a memory, a hard disk, a recording device such as an SSD (Solid State Drive), or a recording medium such as an IC card, an SD card, or a DVD.

  Further, the control lines and information lines indicate what is considered necessary for the explanation, and not all the control lines and information lines on the product are necessarily shown. Actually, it may be considered that almost all the components are connected to each other.

Claims (12)

A text analysis device that calculates a degree of association that is the degree to which the first text supports that the first expression acts on the second expression,
Including a processor and a storage device;
The storage device holds a first dictionary consisting of expressions meaning actions between expressions;
The processor is
Accepting input of the first text, the first representation and the second representation;
Each of the first expression and the second expression when the third expression included in the first dictionary is included in the first text, and an expression having a dependency relationship with the third expression in the first text; Calculating a first feature amount indicating a degree of coincidence of the meanings of
A text analysis device that calculates the degree of association based on a size of the first feature amount. The text analysis apparatus according to claim 1,
The first dictionary indicates whether the action meant by each expression included in the first dictionary is a first action or a second action,
The first action and the second action are opposite to each other,
The storage device
A second dictionary comprising an expression having a meaning to reverse the first action to the second action and to reverse the second action to the first action;
A third dictionary consisting of expressions that fix the value polarity indicated by the relation expression,
The processor is
Accepting an input of the value polarity of the second representation;
In the case where the third expression is included in the first text, the expression included in the second dictionary related to the first expression or the second expression has a dependency relationship with the third expression in the first text. A second feature amount indicating whether or not it is included in a certain expression;
In the case where the third expression is included in the first text, the expression included in the third dictionary related to the first expression or the second expression has a dependency relationship with the third expression in the first text. A third feature amount indicating whether or not it is included in a certain expression,
When it is determined that the relevance is greater than or equal to a predetermined threshold, the second feature value, the third feature value, the value polarity of the second expression, and the action that the third expression means are the first Based on whether it is an action or the second action, the direction indicating whether the action given to the second expression by the first expression is the first action or the second action A text analyzer that calculates. The text analysis apparatus according to claim 2,
The processor is
A plurality of features including one or more feature amounts of a feature amount group including the first feature amount, the second feature amount, and the third feature amount, when it is determined that the relevance is equal to or greater than the predetermined threshold; A text analysis device that calculates the directionality based on a product of quantities. The text analysis device according to claim 3,
The processor is
Whether the expression in the dependency relationship of the third expression in the first text includes at least one of the first expression and the second expression when the third expression is included in the first text. Specify a secondary feature value that indicates
When at least one of the first representation and the second representation in the secondary feature amount includes the first representation, the second feature amount in the first representation is included in the plurality of feature amounts,
When at least one of the first representation and the second representation in the secondary feature amount includes the second representation, the second feature amount and the third feature amount in the second representation, and the second representation A text analysis apparatus that includes the value polarity of the plurality of feature quantities in the plurality of feature amounts. The text analysis device according to claim 4,
The processor specifies a first secondary feature quantity and a second secondary feature quantity,
The first secondary feature amount is obtained when at least one of the first expression and the second expression is a subject of the third expression in the first text when the third expression is included in the first text. And whether or not the predicate of the third expression in the first text includes a second expression group consisting of at least one of the first expression and the second expression. At least one
The first secondary feature amount is obtained when at least one of the first expression and the second expression is a subject of the third expression in the first text when the third expression is included in the first text. And whether or not the predicate of the third expression in the first text includes a fourth expression group consisting of at least one of the first expression and the second expression. At least one
The third expression group includes the first expression group,
The fourth expression group includes the second expression group;
The processor calculates the directionality based on the product of the plurality of secondary feature amounts corresponding to the first secondary feature amount only when the value of the second secondary feature amount does not exist. Text analysis device. The text analysis device according to claim 3,
The storage device further holds a fourth dictionary consisting of negative expressions,
The plurality of feature amounts indicate whether or not the third text in the first text is related to an expression included in the fourth dictionary when the third text is included in the first text. A text analysis device including four feature quantities. The text analysis device according to claim 6,
The storage device further holds a fifth dictionary comprising a predicate expression indicating a statement or an announcement and a code indicating whether each of the predicate expressions has a positive meaning or a negative meaning,
The processor is
When the fourth expression included in the fifth dictionary is included in the one text, the sign of the fourth expression and the expression included in the fourth dictionary are related to the fourth expression in the first text. A text analysis device that calculates the directionality based on a fifth feature amount indicating whether or not there is. The text analysis apparatus according to claim 2,
The text analysis device, wherein the first action is a promoting action and the second action is a suppressing action. The text analysis apparatus according to claim 2,
Connected to the display device,
The processor is
Outputting the relevance to the display device;
A text analysis device that outputs the directionality to the display device when it is determined that the relevance is greater than or equal to a predetermined threshold. The text analysis apparatus according to claim 1,
The storage device further includes a fourth dictionary having an expression indicating that an utterance or an announcement has been made, and a noun that makes an utterance or an announcement, and a fifth dictionary for storing an announcement or an utterance reliability of each noun phrase Hold and
The processor is
When the fourth expression included in the fourth dictionary is included in the first text, the reliability of the noun or the noun phrase in the first text in the dependency relationship with the fourth expression in the fifth text is indicated. Calculate the sixth feature,
A text analysis device that calculates the relevance based on the sixth feature amount. The text analysis apparatus according to claim 1,
The text analysis device, wherein the expression having a dependency relationship with the third expression is a subject or predicate of the third expression. A text analysis method in which a text analysis device calculates a degree of relevance, which is the degree to which the first text supports that the first expression acts on the second expression,
The text analysis device maintains a first dictionary consisting of expressions meaning actions between expressions,
The text analysis method includes:
The text analyzer is
Accepting input of the first text, the first representation and the second representation;
Each of the first expression and the second expression when the third expression included in the first dictionary is included in the first text, and an expression having a dependency relationship with the third expression in the first text; Calculating a first feature amount indicating a degree of coincidence of the meanings of
A method of calculating the degree of association based on a size of the first feature amount.

Images