WO2014017023A1 - Cause expression extraction device, cause expression extraction method, and cause expression extraction program - Google Patents

Abstract

Provided is a cause expression extraction device, whereby, if inputted data is supplied which relates to an event or state, and this inputted data includes a cause expression which describes a cause leading to this event or state, it is possible to extract the cause expression from the inputted data. A polarity determination means (91) determines, when a result expression which is an expression which represents a result and a cause expression candidate which is a candidate for a cause expression which describes a cause of this result are supplied, a polarity which is a category for sorting an impression which a person receives from an expression, with respect to the result expression and the cause expression candidate, respectively. A cause expression determination means (92) determines, using the polarity of the result expression and the polarity of the cause expression candidate, whether the cause expression candidate is a cause expression which describes the cause of the result.

Description

Cause expression extraction apparatus, cause expression extraction method, and cause expression extraction program

The present invention relates to a cause expression extraction apparatus, a cause expression extraction method, and a cause expression extraction program for extracting an expression describing a cause that has led to a certain event or state from input expressions.

Hereinafter, in this specification, explanation will be given mainly using Japanese sentences and words as examples.

It is useful in various aspects to identify the cause that led to a certain event or condition. For example, when repairing a faulty product, correct repair can be performed for the first time by identifying the cause of the malfunction that has occurred. Further, for example, when conducting a questionnaire analysis, if the detailed reason why the respondent stated his / her opinion can be known, the analysis result can be read more accurately.

Thus, it is highly useful to identify the cause that led to the event or condition. Therefore, a technique for automatically extracting an expression corresponding to a cause (hereinafter referred to as a cause expression) from a text has been proposed (see, for example, Patent Document 1).

In the technique described in Patent Document 1, machine learning is performed based on correct answer data that manually specifies the location of the cause expression in the text. As an example of such correct answer data (an example in Japanese is given), text data such as “an error has occurred due to lack of memory (memori busoku de era- ga hassei shita)” can be considered. This example includes the cause expression “memory shortage (memori busoku)”. Further, as an example of text data not including the cause expression, for example, text data such as “an error has occurred in the printing module (insatsu moju-ru de era- ga hassei shita)” can be considered. Note that the sentence “An error occurred due to lack of memory (memori busoku de era- ga hassei shita)” in Japanese corresponds to the sentence “An error occurred because of memory shortage.” In English. And “memory shortage” in Japanese corresponds to “memory“ shortage ”in English. Also, the sentence “An error occurred in the printing module (insatsu moju-ru de era- ga hassei shita)” in Japanese corresponds to the sentence “An error occurred in the printing module.” In English.

Further, in the technique described in Patent Document 1, an expression that becomes a clue (hereinafter referred to as a clue expression) when extracting the cause expression is prepared in advance, and in machine learning, the presence or absence of the occurrence is identified. Use as one. Examples of clue expressions (for example in Japanese) include “cause of (ga （genninn de)”, “due (niyori)”, “de (de)”, and the like. “Because of (ga genninn de)”, “by (niyori)”, and “de (de)” in Japanese correspond to “because of” in English.

In the technique described in Patent Document 1, a pair of expressions that are likely to be causal relations (hereinafter referred to as causal relation pairs) is prepared in advance, and the presence or absence of appearance is one of the features in machine learning. Use as Examples of causal pairs include “killing people (hito wo korosu) – arrest (taiho)”, “election violation (senkyo ihan) – arrest (taiho)”, and the like. In addition, “kill people (hito wo korosu)”, “arrest (taiho)”, and “election violation (senkyo ihan)” in Japanese are “kill a person”, “arrest”, “election offenses” in English, respectively. ".

Note that the feature used when predicting the cause expression by machine learning is called “feature”.

・ Creating correct data is expensive. For this reason, the technique described in Patent Document 1 uses a clue expression that has a relatively high accuracy in determining the cause expression for the purpose of reducing the amount of correct data. However, as shown in the following example, just because a clue expression is included does not necessarily include a cause expression.

Cue expression: “de” in Japanese
"An error occurred due to insufficient memory. (Memori busoku de era-ga hassei shita.)"
"An error occurred in the print module. (Insatsu moju-ru de era-ga hassei shita.)" ... Cause expression is not included.
"An error occurred in text search. (Tekisuto kensaku de era-ga hassei shita.)" ... Cause expression is not included.
This Japanese sentence corresponds to the sentence “An error occurred when searching text” in English.
“A traffic accident occurred during drunk driving. (Inshu unntenn de koutsuu jiko ga okotta.)”… Includes cause expressions.
This Japanese sentence corresponds to the sentence “A traffic accident occurred because of drunken driving” in English.
“A traffic accident occurred in Tamagawa. (Tamagawa de koutsuu jiko ga okotta.)”… Cause expression is not included.
This Japanese sentence corresponds to the sentence “A traffic accident occurred at Tamagawa” in English.

Clue expression: “kara” in Japanese
“Poverty has progressed due to deregulation of employees.” (Including cause expression)
This Japanese sentence corresponds to the sentence “Poverty permeated because of the spread of irregular employment of staff” in English.
"Poverty has progressed from rural areas. (Nousonnbu kara hinnkonn ga susunda.)" ... Cause expression is not included.
This Japanese sentence corresponds to the sentence “Poverty permeated from rural districts” in English.

Clue expression: “tameni” in Japanese
“I'm sick because I have fever.” (Netsu ga aru tameni byouketsu suru.)
This Japanese sentence corresponds to the sentence “I am off sick because of getting a fever.” In English.
“I'm sick to get rid of my cold (kaze wo naosu tameni byouketsu suru.)”
・ ・ ・ Cause expression is not included. (Represents purpose.)
This Japanese sentence corresponds to the sentence “I am off sick in order to cure a cold.” In English.

Therefore, in the technique described in Patent Document 1, accuracy is increased by using a causal relationship pair. In the above example, in order to appropriately determine the cause expression, ideally, a causal relationship pair exemplified below may be given.

"Insufficient memory (memori busoku)-Error (era-)"
"Drinking driving (inshu unntenn)-traffic accident (koutsuu jiko)"
"Denormalization (hiseikika)-poverty (hinnkonn)"
"There is a fever (netsu ga aru)-sick leave (byouketsu)"

In addition, the “memory shortage”, “error”, “drinking driving”, “traffic accident”, “denormalization”, “poverty”, “having fever”, and “disease absence” shown here It corresponds to “memory shortage”, “error”, “drunken driving”, “traffic accident”, “the spread of irregular employment”, “poverty”, “getting a fever”, “be off sick” in English.

Also, emotional expression analysis is a technology that automatically classifies impressions that human beings generally receive for certain expressions. This classification category is called polarity. Examples of polarities of expression include “positive (good impression)”, “negative (bad impression)”, and “neutral (other than that)”.

Non-Patent Document 1 describes reversal of the polarity of expression.

Note that clue expressions are also described in Non-Patent Document 2.

Moreover, Non-Patent Document 3 is given as an example of emotion expression analysis.

JP 2009-157771 A (paragraphs 0032, 0069, 0072, 0073, etc.)

As described above, the technique described in Patent Document 1 uses a causal relationship pair. However, it is not easy to prepare a causal relationship pair in advance. In order to collect pairs of expressions, it is necessary to select expressions from a large number of candidates in the order of the square of the number of expressions. Therefore, it is difficult to perform this operation manually. Even when large-scale data such as a web corpus is used, a problem of data sparseness is likely to occur due to the property of handling a pair of expressions, and there is a high possibility that selection cannot be performed with high accuracy. In addition, even if the number of expressions is reduced using a synonym dictionary that represents the same expression, a thesaurus that defines the subordinate concept, and a thesaurus that defines the subordinate concept, it is still necessary to support the number of square orders. It is still not easy to prepare causal relationship pairs in advance.

Therefore, the present invention can extract a cause expression from input data when input data relating to a certain event or condition is given and the input data includes a cause expression describing the cause of the event or condition. It is an object to provide a cause expression extraction device, a cause expression extraction method, and a cause expression extraction program.

The cause expression extraction apparatus according to the present invention classifies impressions received by a human being from a result expression that is an expression representing a result and a cause expression candidate that is a candidate of a cause expression that describes the cause of the result. The cause expression candidate describes the cause of the result using the polarity determination means for determining the polarity of the case category for each of the result expression and the cause expression candidate, and the polarity of the result expression and the cause expression candidate. And a cause expression determination means for determining whether or not the cause expression has been made.

In addition, the cause expression extraction method according to the present invention gives an impression that a human being receives from an expression when a result expression that is a result expression and a cause expression candidate that is a cause expression candidate that describes the cause of the result are given. The polarity that is the category for classification is determined for each result expression and cause expression candidate, and the cause expression candidate describes the cause of the result by using the polarity of the result expression and the cause expression candidate. It is characterized by determining whether it is an expression.

In addition, the cause expression extraction program according to the present invention provides a computer with a result expression that is a result expression and a cause expression candidate that is a cause expression candidate that describes the cause of the result. Using the polarity determination process for determining the polarity, which is the category when the impression received is classified, for each of the result expression and the cause expression candidate, and the polarity of the result expression and the cause expression candidate, And cause expression determination processing for determining whether or not the cause expression describes the cause of the result.

According to the present invention, when input data relating to a certain event or state is given, and the input data includes a cause expression describing the cause of the event or state, the cause expression can be extracted from the input data. .

It is a block diagram which shows the example of the cause expression extraction apparatus of the 1st Embodiment of this invention. It is a flowchart which shows the example of the process progress of the 1st Embodiment of this invention. It is a block diagram which shows the example of the cause expression extraction apparatus of the 2nd Embodiment of this invention. It is a flowchart which shows the example of the process progress of the 2nd Embodiment of this invention. It is a schematic diagram which shows the example of the modification relationship in a text. It is a block diagram which shows the example of the minimum structure of the cause expression extraction apparatus of this invention.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following description, the present invention will be described mainly using Japanese sentences and words as examples. However, the present invention can also be applied to languages other than Japanese.

Embodiment 1. FIG.
FIG. 1 is a block diagram illustrating an example of a cause expression extraction apparatus according to the first embodiment of this invention. The cause expression extraction apparatus of this embodiment includes an input unit 1, a data processing device 2 that operates under program control, and an output unit 4. The data processing device 2 includes emotion analysis means 21 and cause expression determination means 22.

The input means 1 is an input interface through which input data is input. In the first embodiment, a combination of a result expression that is an expression representing a result of a causal relationship and a cause expression candidate that is a candidate for an expression (cause expression) describing the cause of the result is input via the input unit 1. Are input to the data processing device 2. The result expression represents, for example, an event or a state that is a result of the causal relationship. Result expressions include expressions that express the state of the mind (for example, emotional expressions such as “okoru” and “hotto suru” in Japanese), and expressions that express evaluations and opinions (for example, Japanese) "Beautiful (utsukusii)", "Good (yoi)", etc.) may be included. “Okoru”, “hotto suru”, “utsukusii”, “yoi” in Japanese are “get angry”, “breathe freely”, “ Corresponds to “beautiful” and “good”.

The result expression and the cause expression candidate may be a character string, or may be structured data such as a morphological analysis result or a syntax analysis result.

The emotion analysis means 21 determines the polarity for each of the result expression and the cause expression candidate input via the input means 1.

As already explained, polarity is a classification category of impressions that human beings generally receive for expression. The polarity includes “positive”, “negative”, and “neutral”. The following examples are given as examples of polarity.

“Lack of memory” – negative “drunken drive” – negative “error occurred” – negative “text search” – neutral “San Francisco” – neutral “happy birthday” – positive

For example, the emotion analysis unit 21 stores in advance a dictionary that lists expressions corresponding to “negative”, expressions corresponding to “neutral”, and expressions corresponding to “positive”, and input expressions (result expressions) And the cause expression candidate) and the expressions listed in the dictionary may be subjected to pattern matching to determine the polarity of the input expression. At this time, the emotion analysis means 21 may determine whether or not the character strings and the dependency structures match, and may perform pattern matching gently using a regular expression, a synonym dictionary, and a thesaurus.

Also, prepare a large amount of data with polarity given to the document, build a model that statistically determines the polarity of the expression from the data, and depending on the input expression and the model, The polarity of the input expression may be determined. As an example of data in which a polarity is given to a document, for example, a customer review published on the Internet can be cited.

Here, although two methods are listed as polarity determination methods, the polarity determination may be performed by combining these two methods. Moreover, the emotion analysis means 21 may acquire the polarity of expression by a method other than the above. For example, when there is a server that provides a polarity determination result for each expression, the emotion analysis unit 21 may query the server for the polarity corresponding to the input expression and acquire the polarity of the expression input from the server. Good.

The cause expression determination unit 22 determines whether or not the cause expression candidate is a cause expression representing the cause of the result represented by the result expression, using the polarity of the result expression and the polarity of the cause expression candidate. Preferably, the cause expression determination unit 22 determines whether the cause expression candidate is the cause expression based on the consistency between the polarity of the result expression and the polarity of the cause expression candidate. For example, the cause expression determination unit 22 may determine that the cause expression candidate is the cause expression when the polarity of the result expression matches the polarity of the cause expression candidate.

The cause expression determination means 22 may calculate not only a discrete determination result indicating whether or not the cause expression candidate is a cause expression, but also a value indicating how much the cause expression candidate is a cause expression.

The cause expression determination means 22 may additionally refer to the presence / absence of a clue expression, the learning result obtained by machine learning, and the like when determining whether the cause expression candidate is a cause expression.

The output unit 4 is an output device (for example, a display device) that outputs the determination result of the cause expression determination unit 22. The cause expression determination unit 22 causes the output unit 4 to output (for example, display) the result of determining whether the cause expression candidate is the cause expression.

The emotion analysis means 21 and the cause expression determination means 22 are realized by a CPU of a computer that operates according to a cause expression extraction program, for example. In this case, for example, the CPU may read the cause expression extraction program and operate as the emotion analysis means 21 and the cause expression determination means 22 according to the program. The cause expression extraction program may be recorded on a computer-readable recording medium.

Further, the emotion analysis means 21 and the cause expression determination means 22 may be realized as separate hardware.

Next, the process progress of this embodiment will be described. FIG. 2 is a flowchart showing an example of processing progress of the first embodiment of the present invention.

When the combination of the result expression and the cause expression candidate is input via the input means 1, the emotion analysis means 21 determines the polarity of the result expression and the cause expression candidate (step S1).

Next, the cause expression determination unit 22 uses the result expression polarity derived in step S1 and the cause expression candidate polarity to cause the cause expression candidate to indicate the cause of the result represented by the result expression. It is determined whether or not (step S2). For example, if the cause expression determination unit 22 matches the polarity of the result expression and the polarity of the cause expression candidate, the cause expression candidate is a cause expression representing the cause of the result represented by the result expression. You may judge. And when the polarity of a result expression and the polarity of a cause expression candidate do not correspond, you may determine with a cause expression candidate not being a cause expression. In step S2, the cause expression determination unit 22 causes the output unit 4 to output (for example, display) the determination result.

So far, the polarity of the expression has not been used to extract the cause expression. However, for example, the cause of a negative event or condition can be considered to be a negative event or condition. In the present invention, this is used to determine whether the cause expression candidate is a cause expression representing the cause of the result represented by the result expression using the polarity of the result expression and the polarity of the cause expression candidate. To do. Therefore, if the cause data describing the cause of the result is included in the input data (that is, if the cause expression candidate describes the cause of the result), the cause expression can be accurately extracted. .

For example, it is considered that defects (results) of products sent to call centers and causes of defects often have negative polarity. Therefore, it is possible to accurately extract the cause expression of the cause of the failure from the past cases accumulated in the call center. Here, the past case of the call center has been described as an example, but the input data may not be related to the past case of the call center.

In the present invention, it is not necessary to prepare causal relationship pairs in advance. The expression for which the polarity is to be determined may be a single expression such as “memory shortage (memori busoku)”, “error (era-)”, or the like. Then, the cost of the polarity determination processing for the expression is lower than the cost of constructing a causal relationship pair such as “memory shortage (memori busoku) —error (era-)”. That is, the polarity determination process for the expression can be realized more easily than the construction of the causal relationship pair. Therefore, in the present invention, it is possible to extract the cause expression without constructing a causal relationship pair.

Embodiment 2. FIG.
FIG. 3 is a block diagram illustrating an example of a cause expression extraction device according to the second exemplary embodiment of the present invention. Elements similar to those in the first embodiment are denoted by the same reference numerals as those in FIG. 1, and detailed description thereof is omitted.

The cause expression extraction device of the present embodiment includes an input unit 1, a data processing device 2 that operates by program control, a storage device 3, and an output unit 4. The data processing device 2 includes emotion analysis means 21, cause expression determination means 22, syntax analysis means 23, and clue matching means 24.

The storage device 3 stores a clue dictionary 31. The clue dictionary 31 is a set of expressions (cue expressions) that serve as clues when extracting cause expressions. A set of clue expressions may be determined in advance and stored in the storage device 3 as a clue dictionary 31. Note that clue expressions are often based on function words. Examples of clue expressions include “cause (ga genninn de)”, “due (niyori)”, “de (de)”, and the like. As already stated, “ga genninn de”, “by (niyori)”, and “de (de)” in Japanese correspond to “because of” in English.

In the second embodiment, text is input via the input means 1. Preferably, the text as input data is a single sentence.

The parsing means 23 parses the input text and obtains an analysis result. Specifically, the syntax analysis means 23 specifies the modification relationship (dependency relationship) of each expression in the text.

The clue collating means 24 collates the clue expression in the clue dictionary 31 with the text, and based on the location in the text corresponding to the clue expression and the modification relation of the expression in the text, the result expression and cause from the text Extract expression candidates.

In Japanese, in the case of clue expressions using function words, the expression immediately before the clue expression corresponds to the cause expression candidate, and the destination of the phrase including the cause expression candidate is often the result expression. Even in the case of the cue expressions “ga cause (ga genninn de)”, “dori (niyori)”, and “de (de)”, the cause expression candidates are immediately before the cue expression. The destination of the included phrase is the result expression. When these clue expressions are used, the clue collating unit 24 extracts a portion immediately before the clue expression in the text as a cause expression candidate, and extracts a destination of a phrase including the cause expression candidate as a result expression.

However, for all clue expressions, the cause expression candidate is immediately before the clue expression, and the destination of the cause expression candidate is not necessarily the result expression. For example, suppose the word “cause is (genninn ha)” in Japanese is used as a clue expression. In this case, the cause of the error is insufficient memory. (Era- ga hassei shita genninn ha memori busoku dearu.) Etc. Become. Therefore, in the clue expression dictionary 31, for each clue expression, information on which position the cause expression candidate appears on the basis of the clue expression (hereinafter referred to as cause expression candidate position information), and at which position. Information regarding whether the result expression appears (hereinafter referred to as result expression position information) may also be included. When the clue matching unit 24 detects the clue expression in the text, the clue collation means 24 refers to the cause expression position information and the result expression position information related to the clue expression, and extracts the result expression and the cause expression candidate from the text. Also good. Note that the Japanese sentence “The cause of the error is insufficient memory (era- ga hassei shita genninn ha memori busoku dearu.)” In this paragraph is “The cause of the error occurrence is” Equivalent to the sentence “memory shortage.”

The emotion analysis means 21 determines the polarity of the result expression and the cause expression candidate extracted from the text by the clue matching means 24, respectively.

The cause expression determination unit 22 determines whether or not the cause expression candidate is a cause expression representing the cause of the result represented by the result expression, using the polarity of the result expression and the polarity of the cause expression candidate.

The emotion analysis means 21 and the cause expression determination means 22 are the same as the emotion analysis means 21 and the cause expression determination means 22 in the first embodiment.

The syntax analysis unit 23, the clue collation unit 24, the emotion analysis unit 21, and the cause expression determination unit 22 are realized by a CPU of a computer that operates according to a cause expression extraction program, for example. Further, the syntax analysis unit 23, the clue collation unit 24, the emotion analysis unit 21, and the cause expression determination unit 22 may be realized as separate hardware.

Next, the process progress of this embodiment will be described. FIG. 4 is a flowchart showing an example of processing progress of the second embodiment of the present invention.

When a text is input via the input means 1, the syntax analysis means 23 parses the input text and obtains an analysis result (step S11). That is, the modification relationship of each expression in the text is specified.

Next, the clue collating means 24 collates the clue expression and the text, and based on the location in the text corresponding to the clue expression and the modification relationship of the expression in the text, the result expression and the cause expression candidate from the text. Is extracted (step S12).

Next, the emotion analysis means 21 determines the polarity of the result expression and the cause expression candidate (step S13).

Then, the cause expression determination means 22 is a cause expression that represents the cause of the result represented by the result expression, using the polarity of the result expression derived in step S1 and the polarity of the cause expression candidate. It is determined whether or not there is (step S14). The cause expression determination means 22 causes the output means 4 to output the determination result in step S14.

The polarity determination in step S13 is the same as step S1 in the first embodiment. The determination in step S14 is the same as that in step S2 in the first embodiment.

In the second embodiment, if the cause expression corresponding to the result expression is described in one text, the cause expression can be extracted with high accuracy. In other words, even if the cause expression candidate and the result expression are not specified and input, if the text is input, if the result expression and the cause expression are included in the text, the cause expression can be accurately extracted. Can do.

In the second embodiment, a result expression and a cause expression candidate are extracted from one text (for example, one sentence). In the present invention, both the result expression and the cause expression candidate are required to determine whether or not the cause expression candidate is the cause expression. Therefore, the second embodiment has a high affinity with the present invention.

In the second embodiment, the input data is preferably a single sentence. However, the input data in the second embodiment is not necessarily a single sentence. The input data in the second embodiment may be a text having a length that includes a clue expression and includes a result expression and a cause expression candidate.

Hereinafter, an example corresponding to the first embodiment will be described with reference to FIG. First, a set of a result expression that is an expression representing a result of a causal relationship and a cause expression candidate that is a candidate of an expression describing the cause of the result is input via the input unit 1. In this example, the result expression is “an error has occurred (era- gashassei shita)”, and the cause expression candidate is “the memory is insufficient (memori ga fusoku shite ita)” (hereinafter, the first 1)), and the result expression is “Error occurred (era- ga hassei shita)” and the cause expression candidate is “Text search performed (tekisuto kensaku wo okonatta)” , The second set) is taken as an example. “Error occurred (era- ga hassei shita)” in Japanese corresponds to “An error occurrd.” In English. “Memory shortage (memori ga fusoku shite ita)” corresponds to “The memory was insufficient.” In English. “Text search performed (tekisuto kensaku wo okonatta)” in Japanese corresponds to “Text search was executed.”.

The emotion analysis means 21 determines the polarity of the result expression and the cause expression candidate of each group. In this example, it is assumed that the polarity of the result expression “an error has occurred (era- ga hassei shita)” is determined to be “negative”. Further, it is assumed that the polarity of the cause expression candidate “memori ga fusoku shite ita” is “negative”. Further, it is assumed that the polarity of the cause expression candidate “text search is performed (tekisuto kensaku wo okonatta)” is “neutral”.

The cause expression determination means 22 uses the polarity to determine whether or not the cause expression candidate is truly a cause expression. In the present embodiment, a case is shown in which the cause expression determination means 22 also derives a score (hereinafter referred to as “cause score”) that represents the cause expression likelihood of the cause expression candidate as the probability that the cause expression candidate is the cause expression.

In the first set, the polarities of the result expression ("error occurred (era- ga hassei shita)") and the cause expression candidate ("memory was insufficient (memori ga fusoku shite ita)") were both negative Match. Therefore, the cause expression determination unit 22 is a cause expression corresponding to “an error has occurred (era- ga hassei shita)” as a cause expression candidate “memory is insufficient (memori ga fusoku shite ita)”. Is determined. In the second set, the result expression (“error occurred (era- ga hassei shita)”) is negative in polarity, and the cause expression candidate (“text search (tekisuto kensaku wo okonatta)”) The polarity is neutral and the polarities do not match. Therefore, the cause expression determination means 22 determines that the cause expression candidate “text search (tekisuto kensaku wo okonatta)” is not a cause expression corresponding to “an error has occurred (era- ga hassei shita)”. . The cause expression determination means 22 then causes the cause expression candidate “memori ga fusoku shite ita” to correspond to the result expression “error has occurred (era- ga hassei shita)”. The probability that the cause expression “text search (tekisuto kensaku wo okonatta)” is higher than the accuracy that is the cause expression corresponding to the result expression “error occurred (era- ga hassei shita)” .

There are various methods for increasing the accuracy. When the polarity of the result expression matches the polarity of the cause expression candidate, 1 is added to the cause score of the cause expression candidate, and the polarity of the result expression and the cause expression When the polarity of the candidate is different, the cause score of the cause expression candidate may not be changed. In this case, for the first set of cause expression candidates (“memory is insufficient (memori ga fusoku shite ita)”), 1 is added to the initial value (set to 0) of the cause score, The cause score is 1. On the other hand, the cause score of the above-mentioned second group of cause expression candidates (“text search was performed (tekisuto kensaku wo okonatta)”) remains the initial value 0.

The cause expression determination unit 22 outputs a determination result as to whether or not each cause expression candidate is a cause expression (for example, the output unit 4 displays the determination result). In this example, a cause expression candidate whose cause score is 1 (“memory is insufficient (memori ga fusoku shite ita)”) is output as a cause expression. For a cause expression candidate having a cause score of 0 (“text search performed (tekisuto kensaku wo okonatta)”), the fact that it is not a cause expression is output. The cause expression determination unit 22 may output not only a discrete determination result indicating whether the cause expression candidate is a cause expression but also a continuous value such as a cause score.

The cause expression determination unit 22 refers to the polarities of the result expression and the cause expression candidate when determining whether or not the cause expression candidate corresponds to the cause expression. It is preferable to use machine learning. Machine learning is a method of automatically constructing a model that makes a prediction close to the correct answer data based on the correct answer data and predicting a new input based on the model. In the extraction of the cause expression, for the combination of the result expression and the cause expression candidate, correct data that indicates whether or not the cause expression candidate is really the cause expression is prepared, and prediction that can correctly classify the correct answer data Learn the model. Then, when a new result expression and a cause expression candidate are given as input data, it is predicted based on the model whether or not the cause expression candidate is truly a cause expression. In general, in machine learning, a feature to be used (referred to as a feature) is selected in advance from among various features of input data, and a prediction model is learned based on the feature and subsequent prediction is performed. The features include, for example, words and word sequences that appear in the result expression and cause expression candidates, part of speech and meaning classification of each word, and whether or not a clue expression that easily represents the cause appears in the cause expression candidates.

In the present invention, this feature uses whether or not the polarity of the result expression matches the polarity of the cause expression candidate. It has been observed that the causes leading to negative events and conditions are often negative, and conversely, the causes leading to positive events and conditions are often positive. Therefore, by using the polarity consistency between the result expression and the cause expression candidate as the polarity, it can be determined with high accuracy whether or not the cause expression candidate is the cause expression. As described above, other features may be used together.

Note that even if the polarities appear to be reversed in the two expressions at first glance, the polarities may actually match. An example of the result expression in Japanese is “koto naki wo eta” and the cause expression “memori ga fusoku shite inakatta”. Note that “koto を 得 naki wo eta” in Japanese corresponds to “No harm is done.” In English. “The memory was not insufficient (memori ga fusoku shite inakatta)” in Japanese corresponds to “The memory was not insufficient.” In English. The polarity of the resulting expression (“koto naki wo eta”) is positive. The cause expression (“memory was not insufficient (memori ga fusoku shite inakatta)”) contains the word “memory is insufficient (memori ga fusoku)”. The word “insufficient memory (memori ga fusoku)” is used with a negative word, resulting in a positive polarity. Thus, when an expression is used with a negative word or a paradoxical expression, it is often handled that the polarity is reversed. Therefore, in the above example, the polarities of “I got nothing (koto naki wo eta)” and “Not enough memory (memori ga fusoku shite inakatta)” are positive, respectively, “Memori ga fusoku shite inakatta” ”causes“ koto naki wo eta ”. In addition, for example, when a negative expression appears in a description regarding a person / organization or the like having a conflict of interest, the same polarity reversal can occur in the expression. The expression “raibarusha seihin no memori ga fusoku shiteita)” contains the negative expression “memory shortage (memori ga」 fusoku) ”, but with this expression, The expression “raibarusha”, which represents those with conflicting interests, has also appeared. As a result, the overall polarity of the description, “the memory of the rival product was insufficient (raibarusha seihin no memori ga fusoku shiteita)” changes from negative, and is unlikely to be a cause expression of negative events and conditions. The inversion of the polarity as described above is described in Non-Patent Document 1. In addition, the Japanese sentence “raibarusha seihin no memori ga fusoku shiteita” illustrated here is equivalent to the sentence “The memories of competitor's products are insufficient.” In English To do. And Japanese “raibarusha” is equivalent to English “competitor”.

In the present embodiment, since the polarity of the result expression and the polarity of the cause expression candidate are used to determine whether or not the cause expression candidate is the cause expression, the cause expression candidate is the cause compared to the case where no polarity is used. Whether or not it is an expression can be determined with high accuracy.

Hereinafter, an example corresponding to the second embodiment will be described with reference to FIG. In this embodiment, a case where “cause (ga genninn de)”, “due (niyori)”, and “de (de)” are included in the clue dictionary 31 will be described as an example.

In this embodiment, as input data, a sentence “an error occurred due to insufficient memory (memori busoku de era- ga hassei shita)” and “an error occurred in text search (tekisuto kensaku de era- ga hassei) Take the case where the sentence "shita)" is given. Hereinafter, the former sentence is referred to as a first sentence, and the latter sentence is referred to as a second sentence. The first sentence corresponds to the sentence “An error occurred because of memory shortage.” In English. The second sentence is equivalent to a sentence “An error occurred when searching text.” In English.

The parsing means 23 parses the input text and obtains an analysis result. Specifically, the modification relation of each expression in the text is specified. FIG. 5 shows the modification relationships in the text in the above sentences. An arrow indicated by a solid line in FIG. 5 indicates the destination of the phrase. For example, the phrase “error is (era- ga)” is related to the clause “hassei shita”.

Next, the clue collating unit 24 collates the clue expression in the clue dictionary 31 with the text, and based on the location in the text corresponding to the clue expression and the modification relation of the expression in the text, the result is obtained from the text. Extract expressions and causal expression candidates. In this example, the clue collating means 24 detects the clue expression “de” in both of the two input sentences. Then, the clue matching unit 24 extracts the expression “memory shortage (memoriokubusoku)” immediately before the clue expression “de (de)” as a cause expression candidate for the first sentence, and uses the phrase including the cause expression candidate. “Error occurred (era- ga hassei shita)”, which is related to a certain “memory shortage (memori busoku de)”, is extracted as a result expression. Further, the clue matching unit 24 extracts the expression “text search (tekisutoskensaku)” immediately before the clue expression “de (de)” as a cause expression candidate for the second sentence, and uses the phrase including the cause expression candidate. “Error occurred (era- ga hassei shita)”, which is the destination of a certain “text search (tekisuto kensaku de)”, is extracted as a result expression.

Here, the cause expression candidate and the result expression are limited to one clause such as “occurs (hassei shita)” or includes multiple clauses such as “error occurred (era- ga hassei shita)”. There are several policies regarding how to do this. “Hassei shita” corresponds to “occurred” in English. It may not be preferable to have only one phrase as a cause expression candidate or result expression. For example, when the expression corresponding to the main word has no polarity, such as “hassei shita”, the polarity may not be determined as a whole unless the essential case information is considered. For example, in the example of “hassei shita”, the polarity differs depending on what has occurred. Therefore, it is preferable that the syntax analysis unit 23 determines the essential case of each expression. And it is preferable that the clue collation means 24 extracts a cause expression candidate and a result expression as an expression including an essential case.

When the clue matching means 24 extracts the result expression and the cause expression candidate from the first sentence and the second sentence, the emotion analysis means 22 determines the polarity of the result expression and the cause expression candidate of each sentence. The polarity determination process by the emotion analysis means 22 is the same as the polarity determination process in the first embodiment. Here, it is assumed that the polarity of the result expression “error has occurred (era- ga hassei shita)” is determined to be “negative”. Further, it is assumed that the polarity of the cause expression candidate “memory shortage (memori busoku)” is “negative”. Further, it is assumed that the polarity of the cause expression candidate “text search (tekisuto kensaku)” is “neutral”.

Subsequently, the cause expression determination means 22 determines whether or not the cause expression candidate in each sentence is a cause expression for the result expression by using the polarities related to the result expression of each sentence and the cause expression candidates. The determination process by the cause expression determination unit 22 is the same as the determination process by the cause expression determination unit 22 of the first embodiment. The cause expression determination means 22 is the cause expression candidate of the first sentence (“memory shortage (memori busoku)”) is the cause expression of the result expression of the first sentence (“error occurred (era- ga hassei shita)”). Judge that there is. Further, it is determined that the cause expression candidate of the second sentence (“text search (tekisuto kensaku)”) is not the cause expression of the result expression of the second sentence (“error occurred (era- ga hassei shita)”).

Then, the cause expression determination means 22 outputs a determination result as to whether or not the cause expression candidate of each sentence is a cause expression. The cause expression determination means 22 not only outputs a determination result relating to whether or not the cause expression candidate is a cause expression, but also provides a continuous value such as a cause score for each cause expression candidate as in the first embodiment. And the value may also be output.

In the second embodiment, the cause expression candidate and the result expression are extracted from the inputted sentence, the polarities are determined with respect to the cause expression candidate and the result expression, and the cause expression candidate is converted into the result expression based on the polarity. It is determined whether or not the corresponding cause expression is present. Therefore, even if a pair of cause expression candidates and result expressions is not obtained in advance, the cause expression can be accurately extracted from the input sentence.

Next, the minimum configuration of the present invention will be described. FIG. 6 is a block diagram showing an example of the minimum configuration of the cause expression extraction apparatus of the present invention. The cause expression extraction apparatus of the present invention includes polarity determination means 91 and cause expression determination means 92.

When polarity determination means 91 (for example, emotion analysis means 21) is given a result expression that is an expression representing a result and a cause expression candidate that is a candidate for a cause expression that describes the cause of the result, human beings can recognize from the expression. The polarity, which is a category in the case of classifying received impressions, is determined for each of the result expression and the cause expression candidate.

The cause expression determination unit 92 (for example, the cause expression determination unit 22) uses the result expression polarity and the cause expression candidate polarity to determine whether the cause expression candidate is a cause expression describing the cause of the result. judge.

With such a configuration, input data related to an event or state is given, and when the input data includes a cause expression describing the cause that led to the event or state, the cause expression can be extracted from the input data. .

In addition, a clue expression storage unit (for example, the storage device 3) that stores a clue expression that serves as a clue to extract the cause expression, and a syntax analysis unit (for example, a syntax analysis unit 23) that performs syntax analysis on the input text. And a result expression / cause expression candidate extracting means (for example, a clue matching means) for extracting the result expression and the cause expression candidate from the text based on the location in the text to be used as the clue expression and the parsing result. It may be.

This application claims priority based on Japanese Patent Application No. 2012-166075 filed on July 26, 2012, the entire disclosure of which is incorporated herein.

Although the present invention has been described with reference to the embodiments and examples, the present invention is not limited to the above embodiments and examples. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

Industrial applicability

The present invention is preferably applied to a cause expression extracting device that extracts an expression describing a cause that has led to a certain event or state from input expressions. For example, FAQ creation support for creating FAQs (Frequently Asked Questions) from a set of past cases that are a set of questions and answers to questions about products and services, text mining to extract important knowledge from text data, The present invention can be used for processing to extract an expression describing the cause of a certain event or state in a question and answer system that answers a language question in a natural language.

DESCRIPTION OF SYMBOLS 1 Input means 2 Data processor 3 Memory | storage device 4 Output means 21 Emotion analysis means 22 Cause expression determination means 23 Syntax analysis means 24 Clue matching means 31 Clue dictionary

Claims (6)

1. Given a result expression that is an expression that represents a result and a cause expression candidate that is a candidate for a cause expression that describes the cause of the result, the polarity that is a category in the case of classifying human impressions from the expression, Polarity determination means for determining each of the result expression and the cause expression candidate;
   A cause expression determination unit that determines whether the cause expression candidate is a cause expression describing the cause of the result by using the polarity of the result expression and the polarity of the cause expression candidate. Cause expression extraction device.
2. A clue expression storage means for storing a clue expression as a clue to extract a cause expression;
   A parsing means for parsing the input text;
   The cause according to claim 1, further comprising: a result expression / cause expression candidate extraction unit that extracts a result expression and a cause expression candidate from the text based on a location in the text corresponding to the clue expression and a parsing result. Expression extraction device.
3. Given a result expression that is an expression that represents a result and a cause expression candidate that is a candidate for a cause expression that describes the cause of the result, the polarity that is a category in the case of classifying human impressions from the expression, Determine for each of the result expression and the cause expression candidate,
   A cause expression extraction method characterized by determining whether or not the cause expression candidate is a cause expression describing the cause of the result by using the polarity of the result expression and the polarity of the cause expression candidate.
4. Memorize clue expressions that are clues to extract the cause expression,
   Parses the entered text,
   The cause expression extraction method according to claim 3, wherein a result expression and a cause expression candidate are extracted from the text based on a location in the text corresponding to the clue expression and a syntax analysis result.
5. On the computer,
   Given a result expression that is an expression that represents a result and a cause expression candidate that is a candidate for a cause expression that describes the cause of the result, the polarity that is a category in the case of classifying human impressions from the expression, Polarity determination processing for determining each of the result expression and the cause expression candidate, and
   Cause for causing cause expression determination processing to determine whether the cause expression candidate is a cause expression describing the cause of the result, using the polarity of the result expression and the polarity of the cause expression candidate Expression extraction program.
6. In a computer equipped with a clue expression storage means for storing a clue expression as a clue to extract a cause expression,
   A parsing process for parsing the input text, and
   The cause according to claim 5, wherein a result expression / cause expression candidate extraction process for extracting a result expression and a cause expression candidate from the text is executed based on a location in the text corresponding to the clue expression and a parsing result. Expression extraction program.

Images