JP4671440B2 - Reputation relationship extraction device, method and program thereof - Google Patents

Description

  The present invention relates to a technique for extracting relationships between elements of reputation information from Japanese text, and in particular, extracting a set of related evaluation objects and evaluation expressions.

  In recent years, research on techniques for extracting, organizing and presenting reputation information, which is information such as opinions and evaluations about a certain object, from input text data has been progressing. Here, as elements constituting reputation information, an evaluation object (information) that represents an object to be evaluated, an evaluation attribute that represents (specific evaluation items) the specifications (properties, characteristics, etc.) of the object to be evaluated, or a part thereof (Information), evaluation expressions (information) representing opinions and evaluations themselves, and evaluators (information) representing persons and organizations performing the evaluation.

  Then, each element of reputation information is extracted from the text data (for example, from the text “XX restaurant omelet is delicious but curry is bad”, each element of reputation information “evaluation target = ○○ restaurant, evaluation attribute = omelet rice” Curry, evaluation expression = delicious; bad ”is extracted), for example, as a technique for extracting the evaluation expression, an evaluation expression dictionary including a set of evaluation expressions (word information) and evaluation polarities of the expression is used. In particular, as a technique for extracting evaluation attributes, a method of creating an attribute dictionary composed of a set of evaluation attributes is generally used (Non-patent Document 1 (especially “3.4. 1 element extraction ”)).

  However, the relationship between each element of reputation information is extracted and linked (for example, the text “O restaurant's omelet rice is delicious but curry is bad”) and each element of reputation information “evaluation target = ○○ restaurant, evaluation Attribute = omelet rice; curry, evaluation expression = delicious; bad ", and associated reputation information" (evaluation object, evaluation attribute, evaluation expression) = (XX restaurant, omelet rice, delicious); (XX restaurant, curry, bad " ) ”Is not yet established (see Non-Patent Document 1 (particularly“ 3.4.2 Relationship Extraction ”)).

  On the other hand, there is a method described in Non-Patent Document 2 as a method for statistically determining the relationship between specific expressions (specific expression-specific expression) in text.

However, since the characteristic differs between the relationship between the unique expression and the specific expression and the relationship between the evaluation object and the evaluation expression in the reputation information, the method described in Non-Patent Document 2 is used as it is for the evaluation object in the reputation information and the evaluation expression. It is not effective to apply to the determination of relationship.
Takashi Inui, et al. “Research Trends on Analysis of Evaluation Information for Texts”, Natural Language Processing, Language Processing Society of Japan, July 2006, Vol. 13, no. 3, pp. 201-241 Toru Hirano, et al. “Extracting Semantic Relationships among Explicit Expressions in Texts” Proc. Of 13th Annual Conference of the Association for Natural Language Processing, March 2007

  Differences in characteristics between the relationship between the specific expression and the specific expression and the relationship between the evaluation object and the evaluation expression in the reputation information are mainly the following two points.

  The first difference is the number of relationships.

  In the relationship between specific expressions and specific expressions, a plurality of specific expressions are often related to a specific expression. For example, in the text “Mr. Yamada takes the Shinkansen from Shin-Osaka this morning, arrives in Tokyo at 10:00, and has lunch in Ginza.”, The specific expression “Yamada” is Yamada-Shin Osaka, Yamada-Tokyo, It is related to three unique expressions of Yamada-Ginza.

  On the other hand, in the relationship of evaluation object-evaluation expression in reputation information, the evaluation object related to a certain evaluation expression is one in principle, and the exception is when the evaluation objects are parallel. For example, in the text “Yamada-chan and Tanaka-chan are both cute”, the evaluation expression “cute” is related to both the evaluation object “Yamada” and the evaluation object “Tanaka”.

  The second difference is the order of appearance.

  There is almost no restriction on the order of appearance in the relationship between proper expression and proper expression. There could be both the text “I went to Ginza today. I ate French with Mr. Yamada” and the text “I ate with Mr. Yamada today. I went to French in Ginza.”

  However, in the evaluation object-evaluation expression relationship in reputation information, the evaluation expression rarely appears in the sentence ahead of the evaluation object, and the evaluation expression appears in the same sentence as the evaluation object or in the sentence behind it. There are many cases (ex. “Yamada-chan is cute”, “I met Yamada-chan today. After all, it is cute”). An exception is the case of inverted expressions. For example, in the text “It was cute. After all, Yamada-chan!”, The evaluation expression “cute” exists in the sentence ahead of the evaluation target “Yamada”.

  Further, the method described in Non-Patent Document 2 determines whether or not there is a relationship between each unique expression unit appearing in a document. Even when a specific expression representing the same entity appears multiple times, knowledge extraction is performed. From this point of view, it is desirable that the relations can be extracted in a collective manner rather than outputting the number of appearances for one entity.

  The present invention has been made in view of the above points, and there is basically one evaluation object related to a certain evaluation expression, and the position of the evaluation object related to a certain evaluation expression is the same sentence as the evaluation expression or its It is an object to provide a reputation relationship extraction apparatus, method and program capable of extracting a pair of a related evaluation object and an evaluation expression with high accuracy and low cost by utilizing the fact that it is often a forward sentence. And

The present invention relates to a reputation relationship extraction device for extracting a set of a related evaluation object and an evaluation expression from document-based language analysis information including at least word information, unique expression information, and evaluation expression information, An evaluation object and an evaluation expression are extracted from the language analysis information, and a combination generation unit that generates a combination of the evaluation object and the evaluation expression as an evaluation object-evaluation expression pair, and rearranges the evaluation object-evaluation expression pair in a predetermined order. A reordering unit, a discrimination result storage unit for storing an evaluation target-evaluation expression pair determined to have a relationship among all evaluation target-evaluation expression pairs, and an evaluation target included in the evaluation target-evaluation expression pair A model storage unit storing a plurality of models for determining whether or not there is a relationship with the evaluation expression, and the evaluation target-evaluation expression pair after the rearrangement is the first pair From the processing target pair setting unit that sequentially sets the processing target pair to the end pair and the discrimination result stored in the discrimination result storage unit, the evaluation target-evaluation expression pair set as the processing target pair is evaluated. For a determination target pair determination unit that determines whether a determination target pair should be determined whether or not there is a relationship between the target and the evaluation expression, and for the evaluation target-evaluation expression pair determined as the determination target pair, From the language analysis information, a general-purpose information extraction unit that extracts information about words between the pairs or dependency information between the pairs as features from the language analysis information, and an evaluation object-evaluation expression pair determined as the discrimination target pair, from the language analysis information statement distance between the pair, whether evaluation of the pair is included in the document title, extracts the evaluation expressions number and evaluation type between the pair as features And determine information extraction unit, evaluated the is determined that the determination target pairs - a model selection unit which classifies the evaluation expressions pairs, and feature of the extracted generic information extraction unit and Reputation information extraction section, the model selection extracting a model from the model storage unit by using the classification classification results in parts, the determination target pair determined as the evaluation target based on the extract out model - between evaluated and evaluation expression evaluation expression pairs A classifier that discriminates whether or not there is a relationship, a discrimination result storage unit that stores an evaluation object-evaluation expression pair that is discriminated to be related in a discrimination result storage unit, and an evaluation that is stored in the discrimination result storage unit And a pair output unit that outputs a target-evaluation expression pair.

  According to the present invention, there is a case in which there is in principle one evaluation object related to a certain evaluation expression, and the position of the evaluation object related to a certain evaluation expression may be the same sentence as the evaluation expression or a sentence in front of it. The relationship between the evaluation object-evaluation expression pair can be discriminated using the feature of many, and the related evaluation object-evaluation expression pair can be extracted according to each case.

  Hereinafter, the present invention will be described in detail with reference to the illustrated embodiments.

  FIG. 1 shows an example of an embodiment of a reputation relationship extraction device according to the present invention. The reputation relationship extraction device according to the present invention includes a pair generation processing unit 10, a discrimination result storage unit 20, a model storage unit 30, and a feature extraction / The determination processing unit 40 is configured. FIG. 2 is a flowchart of pair generation processing in the pair generation processing unit 10, and FIG. 3 is a flowchart of feature extraction / discrimination processing in the feature extraction / discrimination processing unit 40. The outline of the present invention will be described below.

  The pair generation processing unit 10 includes a combination generation unit 11 and a rearrangement unit 12. The language generation analysis information in document units stored in the storage unit or input from the input unit is used as an input, and the evaluation target and the evaluation expression Are output in a predetermined order as an evaluation object-evaluation expression pair.

  The discrimination result storage unit 20 stores information on evaluation target-evaluation expression pairs that are determined to be related among all evaluation target-evaluation expression pairs.

  The model storage unit 30 stores a plurality of models for determining whether or not there is a relationship between the evaluation object included in the evaluation object-evaluation expression pair and the evaluation expression.

  The feature extraction / discrimination processing unit 40 includes a processing target pair setting unit 41, a discrimination target pair determination unit 42, a general-purpose information extraction unit 43, a reputation information extraction unit 44, a model selection unit 45, and a classifier 46. Whether or not there is a relationship between the evaluation object and the evaluation expression among the evaluation object-evaluation expression pairs output from the pair generation processing unit 10, including the discrimination result storage unit 47 and the pair output unit 48. An evaluation object-evaluation expression pair to be discriminated is determined using the discrimination result storage unit 20, and feature extraction is performed on the pair to be discriminated from language analysis information, and a model is selected from the model storage unit 30. After performing the discrimination processing, storing the relevant evaluation object-evaluation expression pairs in the discrimination result storage unit 20 and processing all the evaluation object-evaluation expression pairs, the discriminant result storage unit 20 has a relationship. The evaluation target-evaluation expression pair Look out and outputs.

  The reputation relation extracting device of the present invention is composed of a computer device, a display means such as a monitor, an input means such as a keyboard, a storage means such as a hard disk or a memory, a communication device that can be connected to an external network, etc. Not shown).

  Hereinafter, the details of the processing in each unit described above will be described with examples.

<Language analysis information>
The input to this apparatus is language analysis information in document units stored in the storage means or input from the input means. Since the reputation information often describes the reputation of the evaluation target described in the document title, it is better to add the document title (which may not exist) to the language analysis information in addition to the text.

  The language analysis information includes at least word information, unique expression information, and evaluation expression information.

  The word information is information for each word divided into words obtained by performing a known morphological analysis process on text data, and includes at least a notation and a part of speech.

  The unique expression information is information related to each unique expression obtained by performing known unique expression extraction processing on the text data and the word information, and at least the position of the unique expression and its type (person name, place name, etc.) Including.

  The evaluation expression information is a known evaluation expression extraction process for text data, the word information, and the unique expression information (for example, “Nozomi Kobayashi, etc.” Extraction of attribute-evaluation value pair and opinion information using anaphora analysis method) "The processing using the evaluation value expression dictionary shown in the 11th Annual Conference of the Language Processing Society of Japan, March 2005, pp. 436-439"). Yes, including at least the position of the evaluation expression.

  Besides these, dependency analysis information obtained by performing known dependency analysis processing on text data and the word information may be included in the language analysis information.

  FIG. 4 shows an example of language analysis information including word information, unique expression information, evaluation expression information, and dependency analysis information.

  In the example of the present embodiment, the 0th sentence is always handled as a document title (when there is no document title, the 0th sentence is empty).

<Pair generation process>
The combination generation unit 11 acquires the language analysis information of the document unit stored in the storage unit or input from the input unit (FIG. 2: step S1), and extracts the evaluation object and the evaluation expression from this, Only combinations of specific expressions defined in advance are set as evaluation targets, and all combinations of evaluation targets and evaluation expressions are generated as evaluation target-evaluation expression pairs (FIG. 2: step S2).

  In the present embodiment, description will be made assuming that the types of specific expressions defined in advance are four types: person name, place name, organization name, and artifact name. The type of the unique expression is not fixed and may be determined in advance according to the application.

  In addition, all the combinations of the evaluation object and the evaluation expression are not generated as an evaluation object-evaluation expression pair, but the position of the evaluation object as viewed from the evaluation expression among all the combinations of the extracted evaluation object and the evaluation expression is Only combinations that fall within the preset sentence range, or all combinations of the extracted evaluation object and evaluation expression, only the combination that is included in at least one sentence that the evaluation object is set in advance, or both Only combinations to be performed may be generated as an evaluation target-evaluation expression pair. Note that the sentence setting described above specifies the sentence number (absolute value) indicating the position of each sentence, and the sentence range setting is the sentence number of the sentence including the evaluation object and the sentence sentence including the evaluation expression. This can be done by specifying the sentence distance (relative value) obtained from the difference from the number.

  Next, the rearrangement unit 12 sets the evaluation target-evaluation expression pair generated by the combination generation unit 11 in a predetermined order suitable for processing in the feature extraction / discrimination processing unit 20 described later, here, an evaluation expression unit that appears. Are rearranged in the order of appearance (step S3 in FIG. 3). At this time, the order of the pairs including the same evaluation expression is the order of the pair including the evaluation object in the same sentence as the evaluation expression (the order of the pair including the evaluation object in the same sentence as the evaluation expression is Sorted so that the pair containing the evaluation object in the sentence different from the evaluation expression is located after the pair containing the evaluation object that is closer to the evaluation expression), and further, the evaluation in the sentence different from the evaluation expression The order of the pairs including the target is as follows: the pair whose evaluation target is before the evaluation expression (before the pair including the evaluation target whose distance to the evaluation expression is closer), and the evaluation target after the evaluation expression A pair is rearranged so that a pair is positioned later (a pair including an evaluation object closer to the evaluation expression before). The rearranged evaluation object-evaluation expression pair is output, and the pair generation process ends.

  FIG. 5 shows an output example of the pair generation processing unit 10 generated from the language analysis information shown in FIG.

  FIG. 5A shows an example in which all combinations are generated. “X904i (artifact name)” (0th sentence (title)), “Tanaka Electric (organization name)”, which is one of the names of persons, places, organizations, and artifacts (2nd sentence), “X904i (artifact name)” (2nd sentence), “ABC cafe (organization name)” (4th sentence) are taken out (because “today” is a type of date, it is excluded) The evaluation expressions are sorted in the order of appearance “hot” (first sentence), “very popular” (second sentence), “very cool” (third sentence), and sorted according to the evaluation target position.

  FIG. 5B shows an example in which a combination is generated by adding the appearance position constraint. Here, only combinations in which the position of the evaluation target viewed from the evaluation expression is within a preset sentence range, specifically, a sentence. Only combinations in which the distance (= evaluation target sentence number−evaluation expression sentence number) is 0 or less (in other words, the position of the evaluation target is the same sentence as the evaluation expression or a sentence in front thereof) are output.

  The rearranged evaluation target-evaluation expression pairs generated by the pair generation processing unit 10 are temporarily stored in a storage device (not shown) together with the document unit language analysis information described above, and the feature extraction / discrimination processing unit 40 used for processing.

<Feature extraction / discrimination processing>
As described above, the discrimination result storage unit 20 stores information on the evaluation target-evaluation expression pair that is determined to have a relationship among all the evaluation target-evaluation expression pairs. The information includes at least evaluation object identification information, evaluation expression identification information, and the same sentence information.

  The evaluation object identification information may be a notation of the evaluation object itself, but may be in its standard form (for example, obtained from morphological analysis or obtained by referring to other external knowledge). For example, as an example of using the standard form as the evaluation object identification information, there is information called the standard form in the word information of the language analysis information, and the standard form of the word “Yokosuka” is “Yokosuka” (this Is obtained from the result of the morphological analysis process), and “Yokosuka” is stored as evaluation object identification information.

  The evaluation expression identification information may be any information as long as it can distinguish evaluation expressions at different positions. For example, when an ID is assigned to each evaluation expression, the ID may be used, or position information in a document may be used.

  The same sentence information is information that is Yes when the evaluation object and the evaluation expression of the evaluation object-evaluation expression pair exist in the same sentence, and No otherwise.

  In the example described later, the notation itself is used as the evaluation object identification information, and the sentence position (number) -word position (number) of the first word of the evaluation expression is used as the evaluation expression identification information. For example, in the pair of the arrangement order 1 in FIG. 5B, the evaluation expression “hot” is the third word of the first sentence, so the evaluation expression identification information is “1-3”, and the evaluation object identification information Becomes “X904i”.

  Hereinafter, processing by the feature extraction / discrimination processing unit 40 will be described in detail.

  First, in step S11, the processing target pair setting unit 41 sets the first pair as the processing target pair among the evaluation target-evaluation expression pairs output from the pair generation processing unit 10 (if the process moves from step S17, Among the evaluation target-evaluation expression pairs output from the pair generation processing unit 10, the next pair of the pair set as the processing target pair at that time is set as the processing target pair). Thereafter, the process proceeds to step S12.

  Next, in step S12, the evaluation target-evaluation expression pair set as the processing target pair based on the determination result stored in the determination result storage unit 20 by the determination target pair determination unit 42 is evaluated and evaluated. It is determined whether or not there is a determination target pair to be determined whether or not there is a relationship with the expression. Specifically, evaluation target identification information and evaluation expression identification matching the processing target pair are stored in the determination result storage unit 20 It is determined whether an evaluation object-evaluation expression pair including information is stored.

  If the determination is No and not stored, the processing target pair is determined as a determination target pair, and the process proceeds to step S13. If the determination is Yes and stored, the processing target pair is not a determination target pair. And the process proceeds to step S17.

  Here, as the determination condition in step S12, in addition to the condition whether or not the information on the pair that matches the processing target pair is stored in the determination result storage unit 20 described above, the determination result storage unit 20 stores the pair of processing target pairs. When it is determined whether or not an evaluation target-evaluation expression pair including evaluation expression identification information that matches the evaluation expression is stored, and if not stored, the processing target pair is determined as a determination target pair and stored Further determines whether the same sentence information of the evaluation target-evaluation expression pair stored in the determination result storage unit is Yes, and whether the sentence distance of the processing target pair is other than 0, and the same sentence information is Yes It is determined that the processing target pair is not a discrimination target pair only when the sentence distance is other than 0, and otherwise, the processing target pair is determined as a discrimination target pair. It may be a matter. In the case of this determination condition, when an evaluation object related to the same sentence as the evaluation expression is extracted, a plurality of parallel evaluation objects existing in the same sentence are continuously subjected to the discrimination process (step S15), and a plurality of evaluation objects can be extracted. However, the evaluation target in a sentence different from the evaluation expression skips the discrimination process and is treated as irrelevant. This is a process reflecting the feature that there is one evaluation object related to an evaluation expression in principle.

  In the following description, the latter condition is used as the determination condition in step S12.

  Next, in step S13, the general-purpose information extraction unit 43 and the reputation information extraction unit 44 perform general-purpose information extraction processing and reputation information extraction processing (respectively for the evaluation target-evaluation expression pair determined as the discrimination target pair). Together, this is called feature extraction processing), and the process proceeds to step S14.

  The details of the general-purpose information extraction process and reputation information extraction process in step S13 will be described below.

  The general-purpose information extraction processing simply applies information that can be used for relationship extraction in a general manner, for example, a sentence structure between NEs as shown in Non-Patent Document 2 and features representing context information to an evaluation target-evaluation expression pair. Words between pairs (character strings), words between pairs (parts of speech), number of words between pairs, number of unique expressions existing between pairs, destination clauses to be evaluated (strings), destination clauses to be evaluated ( Part of speech), dependency clause of evaluation expression (string), dependency clause of evaluation expression (part of speech), whether pair exists in the same phrase, distance of shortest path between clauses including pair, dependency structure Information, topic break information, omission information, etc. are extracted from language analysis information using existing techniques such as the method described in Non-Patent Document 2.

  FIG. 6 shows an example of features extracted by this general-purpose information extraction process, taking as an example the arrangement order 5 of FIG. 5B, “X904i—very cool”. In this example, the features of the general information include the words between the pairs (character strings), the number of words between the pairs, the number of unique expressions existing between the pairs, the evaluation target clauses (character strings), and the evaluation target relationships. The clause (part of speech), the dependency clause of the evaluation expression (character string), the dependency clause of the evaluation expression (part of speech), whether the pair exists in the same clause, and whether the evaluation target exists as abbreviation information or not? , Which case particles are to be evaluated are stored in the corresponding region? , Which case particles are to be evaluated are stored in the corresponding region? , What number in that area is remembered? , Is the evaluation target memorized in the highest case particle? , The route to the evaluation target.

  Whether the reputation information extraction process is included in at least the sentence distance between the pair and the document title as a feature useful for reputation from the language analysis information? The number of evaluation expressions between pairs and the type of evaluation object are extracted.

  Hereinafter, these features will be described using a specific example of the arrangement order 5, “X904i—very cool” in FIG. 5B.

  The sentence distance between the pairs is calculated from the evaluation object sentence number−the evaluation expression sentence number. In “X904i−very cool”, the evaluation expression sentence number = 3 and the evaluation object sentence number = 2. −3 = −1.

  Whether or not the document is included in the document title indicates whether or not the evaluation target exists in the document title. “X904i” of “X904i-very cool” exists in the second sentence, and the zeroth sentence ( Document title) does not exist, so No.

  The number of evaluation expressions between the pairs is the number of evaluation expressions existing between the evaluation object of the pair and the evaluation expression. In the example of “X904i—very cool”, there is no one, so it is zero.

  The evaluation target type is a type of specific expression of the evaluation target, and “X904i” is “artifact”.

  The features extracted in the reputation information extraction process are collectively shown in FIG.

  Next, in step S14, the model selection unit 45 classifies the evaluation target-evaluation expression pair determined as the determination target pair using information on the evaluation target-evaluation expression pair determined as the determination target pair. The process proceeds to step S15. In the present embodiment, classification is performed as to whether the evaluation object-evaluation expression pair is the same sentence or other than the same sentence (whether the evaluation object and the evaluation expression are included in the same sentence or other than the same sentence). And

  Next, in step S15, the classifier 46 outputs the features output from the general-purpose information extraction unit 43 and the reputation information extraction unit 44, and the evaluation target-evaluation expression pair classification information output from the model selection unit 45. Whether or not there is a relationship between the evaluation target and the evaluation expression of the evaluation target-evaluation expression pair determined as the determination target pair based on the extracted model. Is determined, and the process proceeds to step S16.

  Here, the model uses a result determined in advance as to whether or not there is a relationship between the evaluation object and the evaluation expression in the predetermined evaluation object-evaluation expression pair, and a text including the predetermined evaluation object-evaluation expression pair. Well-known machine learning using language analysis information created using the above-described existing technology and information extracted in advance from the pair generation processing unit 10, the general-purpose information extraction unit 43, and the reputation information extraction unit 44 Are created in advance and stored in the model storage unit 30. At this time, it is assumed that the discrimination result for the predetermined evaluation object-evaluation expression pair is based on human judgment. Each model may be configured according to the classification of the evaluation object-evaluation expression pair, such as “same sentence” and “other than the same sentence”, or may be configured without distinguishing the classification.

  The classifier 46 may be configured to output a numerical value representing the degree of relationship in addition to the determination result of whether or not there is a relationship. As machine learning, a well-known machine can be used. However, it is desirable to use a machine that can learn by directly inputting data of a tree structure or a graph structure.

  Next, in step S16, the determination result storage unit 47 stores the evaluation expression identification information, the evaluation object identification information, and the same sentence information related to the evaluation object-evaluation expression pair determined to be related by the classifier 46 as a determination result storage unit. 20. Thereafter, the process proceeds to step S17.

  If the classifier 46 outputs a numerical value indicating the degree of relationship, it may be determined that the relationship is relevant only when the degree of relationship is greater than a predetermined threshold value set in advance. .

  Next, in step S <b> 17, the processing target pair setting unit 41 determines whether or not the current processing target pair is the last pair of the evaluation target-evaluation expression pair output from the pair generation processing unit 10. If Yes, the process moves to step S18, and if No, the process moves to step S11.

  Next, in step S18, the pair output unit 48 acquires the evaluation object-evaluation expression pair (information) determined to have a relationship in the document to be processed from the determination result storage unit 20, and displays the display means or The data is output to the storage device, and the process ends.

  FIG. 8 shows a processing result in each unit of the feature extraction / discrimination processing unit 40 when the output of the pair generation processing unit 10 is FIG. 5B.

  First, in step S11, the pair in the arrangement order 1 is set as a processing target pair. In step S12, nothing is stored in the determination result storage unit 20, so that it is determined as a determination target pair, and the process moves to step S13. In step S14, since the evaluation target and the evaluation expression do not exist in the same sentence, they are classified as other than the same sentence. Next, in step S15, a model other than the same sentence is extracted and discriminated to obtain a result that there is no relation. In step S16, nothing is performed, and the process proceeds from step S17 to step S11. Set the processing target pair.

  Similarly to the previous step, the process proceeds to step S12 and step S13. In step S14, the sentence is classified as the same sentence because it exists in the same sentence. In step S15, the model of the same sentence is extracted and determined. In step S16, “X904i” which is the notation as evaluation object identification information, “2-9” because it is the ninth word of the second sentence as evaluation expression identification information, evaluation object and evaluation expression as the same sentence information Since “S” is the same sentence, “Yes” is set. In the same manner, the process proceeds in the same manner, and the determination in step S12, the determination result in step S15, and the information stored in the determination result storage unit 20 are set when there is a relationship as shown in FIG.

  Here, in the process for the arrangement order 4, in step S12, the evaluation expression “super popular” and the evaluation target notation “X904i” at the position 2-9 are stored in the determination result storage unit 20 in the arrangement order 2 process. Since the information matches the determined information, it is determined as Yes (not a discrimination target pair).

  FIG. 9 shows an output example of the final feature extraction / discrimination processing unit 40 corresponding to the processing result of FIG. Here, only the notation of the evaluation object and the evaluation expression is output, but information such as each position (the evaluation object may be a plurality of evaluation objects, for example, the position of the evaluation object located in the shortest position from the evaluation expression), etc. It may be output.

The block diagram which shows an example of embodiment of the reputation relationship extraction apparatus of this invention Flow chart of pair generation processing in the pair generation processing unit Flow chart of feature extraction / discrimination processing in the feature extraction / discrimination processing unit Explanatory diagram showing an example of language analysis information Explanatory drawing which shows an example of the evaluation object-evaluation expression pair produced | generated by the pair production | generation process part Explanatory drawing which shows an example of the feature extracted in the general-purpose information extraction part Explanatory drawing which shows an example of the feature extracted by the information extraction part for reputation Explanatory drawing which shows an example of the processing result in a feature extraction / discrimination processing part Explanatory drawing which shows an example of the output of the final feature extraction / discrimination processing unit

Explanation of symbols

  10: Pair generation processing unit, 11: Combination generation unit, 12: Rearrangement unit, 20: Discrimination result storage unit, 30: Model storage unit, 40: Feature extraction / discrimination processing unit, 41: Processing target pair setting unit, 42 : Discrimination target pair determination unit, 43: general-purpose information extraction unit, 44: reputation information extraction unit, 45: model selection unit, 46: classifier, 47: discrimination result storage unit, 48: pair output unit 48.

Claims (13)

A reputation relationship extraction device that extracts a set of a related evaluation object and an evaluation expression from document-based language analysis information including at least word information, unique expression information, and evaluation expression information,
A combination generation unit that extracts the evaluation object and the evaluation expression from the document unit language analysis information, and generates a combination of the evaluation object and the evaluation expression as an evaluation object-evaluation expression pair;
A rearrangement unit for rearranging the evaluation object-evaluation expression pairs in a predetermined order;
A discrimination result storage unit for storing an evaluation subject-evaluation expression pair determined to have a relationship among all evaluation object-evaluation expression pairs;
A model storage unit that stores a plurality of models for determining whether or not there is a relationship between the evaluation object included in the evaluation object-evaluation expression pair and the evaluation expression;
A processing target pair setting unit for sequentially setting the evaluation target-evaluation expression pair after the rearrangement from the first pair to the last pair as a processing target pair;
Based on the determination result stored in the determination result storage unit, it should be determined whether or not the evaluation object-evaluation expression pair set as the processing object pair has a relationship between the evaluation object and the evaluation expression. A discrimination target pair determination unit for determining whether the pair is a discrimination target;
A general-purpose information extraction unit that extracts information on words between the pairs or dependency information between the pairs as features from the language analysis information for the evaluation target-evaluation expression pair determined to be the determination target pair;
For the evaluation object-evaluation expression pair determined to be the determination object pair, the sentence distance between the pair, whether the evaluation object of the pair is included in the document title from the language analysis information , the number of evaluation expressions between the pair, and the evaluation Reputation information extraction unit that extracts target types as features ,
A model selection unit for classifying the evaluation object-evaluation expression pair determined as the determination object pair;
A model is extracted from a model storage unit using the features extracted by the general-purpose information extraction unit and the reputation information extraction unit and the classification result classified by the model selection unit, and the discrimination is performed based on the extracted model. A classifier for discriminating whether or not there is a relationship between the evaluation object of the evaluation object-evaluation expression pair determined as the object pair and the evaluation expression;
A determination result storage unit that stores the evaluation object-evaluation expression pair determined to have a relationship in the determination result storage unit;
A reputation relationship extraction apparatus comprising: a pair output unit that outputs an evaluation object-evaluation expression pair stored in a discrimination result storage unit. It is characterized by comprising a combination generation unit that extracts evaluation targets and evaluation expressions from document-based language analysis information, and generates all combinations of the extracted evaluation objects and evaluation expressions as evaluation target-evaluation expression pairs. The reputation relationship extraction device according to claim 1. The evaluation target and the evaluation expression are extracted from the language analysis information of the document unit, and the position of the evaluation target viewed from the evaluation expression is within a preset sentence range among all combinations of the extracted evaluation target and the evaluation expression. The reputation relation extraction device according to claim 1, further comprising a combination generation unit that generates only combinations as evaluation target-evaluation expression pairs. The evaluation target and the evaluation expression are extracted from the linguistic analysis information of the document unit, and only the combination included in at least one sentence preset by the evaluation target is evaluated among all the combinations of the extracted evaluation target and the evaluation expression. The reputation relation extraction device according to claim 1, further comprising a combination generation unit that generates as an evaluation expression pair. It is determined whether or not an evaluation target-evaluation expression pair including an evaluation target and an evaluation expression matching the processing target pair is stored in the determination result storage unit. If not stored, the processing target pair is determined as a determination target pair. The reputation relationship extraction device according to claim 1, further comprising: a discrimination target pair determination unit that determines and determines that the processing target pair is not a discrimination target pair if it is stored. It is determined whether or not an evaluation target-evaluation expression pair including an evaluation expression that matches the processing target pair is stored in the determination result storage unit, and if not stored, the processing target pair is determined as a determination target pair, If stored, the evaluation object and evaluation expression of the evaluation object-evaluation expression pair stored in the determination result storage unit exist in the same sentence, and the evaluation object and evaluation expression of the processing object pair are different in the sentence The reputation relationship extraction device according to claim 1, further comprising: a discrimination target pair determination unit that determines that the processing target pair is not a discrimination target pair only when the processing target pair exists. A reputation relationship extraction method for extracting a set of a related evaluation object and an evaluation expression from document-based language analysis information including at least word information, unique expression information, and evaluation expression information,
A step of generating a combination of the evaluation object and the evaluation expression as an evaluation object-evaluation expression pair, wherein the combination generation unit extracts the evaluation object and the evaluation expression from the document unit language analysis information;
A rearrangement unit rearranging the evaluation object-evaluation expression pairs in a predetermined order; and
A processing target pair setting unit sequentially sets the rearranged evaluation target-evaluation expression pair from the first pair to the last pair as a processing target pair;
Based on the discrimination result stored in the discrimination result storage unit that stores the evaluation subject-evaluation expression pair that is discriminated as having a relation among all the evaluation object-evaluation expression pairs, the discrimination target pair determination unit Determining whether the evaluation object-evaluation expression pair set as a pair is a determination object pair to determine whether or not there is a relationship between the evaluation object and the evaluation expression;
A step in which a general-purpose information extraction unit extracts information about words between the pairs or dependency information between the pairs as features from the language analysis information for the evaluation target-evaluation expression pair determined to be the determination target pair;
For the evaluation object-evaluation expression pair determined to be the determination target pair, the reputation information extraction unit determines whether the sentence title between the pairs is included in the document title from the language analysis information , whether the evaluation target of the pair is included in the document title, Extracting the number of evaluation expressions between and the types of evaluation target as features ,
A step in which the model selection unit classifies the evaluation target-evaluation expression pairs determined as the determination target pairs;
Classifier, a feature that the extracted generic information extracting unit and the reputation information extracting unit, by using the classified classification result by the model selection unit, evaluation - evaluated included in the evaluation expressions pair Rating A model is extracted from a model storage unit that stores a plurality of models for determining whether or not there is a relationship with an expression, and an evaluation target that is determined as the determination target pair based on the extracted model − Determining whether there is a relationship between the evaluation target of the evaluation expression pair and the evaluation expression;
A step in which the discrimination result storage unit stores the evaluation object-evaluation expression pair discriminated to be related in the discrimination result storage unit;
Pair output unit, the determination result storage unit stored in the evaluation - Reputation relation extraction method characterized by comprising the step of outputting the evaluation expressions pair. The combination generation unit includes a step of extracting the evaluation object and the evaluation expression from the document unit language analysis information, and generating all combinations of the extracted evaluation object and the evaluation expression as an evaluation object-evaluation expression pair. The reputation relationship extraction method according to claim 7 . The combination generation unit extracts the evaluation object and the evaluation expression from the document unit language analysis information, and among the combinations of the extracted evaluation object and the evaluation expression, the position of the evaluation object viewed from the evaluation expression is set in advance. The reputation relationship extracting method according to claim 7 , further comprising: generating only combinations within the range as evaluation target-evaluation expression pairs. The combination generation unit extracts the evaluation object and the evaluation expression from the document unit language analysis information, and the evaluation object is included in at least one sentence set in advance among all combinations of the extracted evaluation object and the evaluation expression. The reputation relationship extracting method according to claim 7 , further comprising: generating only a combination as an evaluation object-evaluation expression pair. The evaluation target and evaluation that match the processing target pair in the determination result storage unit that stores the evaluation object-evaluation expression pair that is determined to be related among all the evaluation object-evaluation expression pairs. It is determined whether or not an evaluation target-evaluation expression pair including an expression is stored. If not, the processing target pair is determined as a determination target pair, and if stored, the processing target pair is determined as a determination target pair. The reputation relationship extracting method according to claim 7 , further comprising: a step of determining that it is not. The determination target pair determination unit includes an evaluation expression that matches the processing target pair in the determination result storage unit that stores the evaluation target-evaluation expression pair that is determined to be related among all the evaluation target-evaluation expression pairs. It is determined whether or not an evaluation target-evaluation expression pair is stored, and if not stored, the processing target pair is determined as a determination target pair. If stored, it is further stored in the determination result storage unit. The evaluation target-evaluation expression pair of the evaluation target and the evaluation expression are present in the same sentence, and the processing target pair is determined not to be the discrimination target pair only when the evaluation target of the processing target pair and the evaluation expression exist in different sentences. The reputation relationship extracting method according to claim 7 , further comprising a step of: A reputation relationship extraction program for causing a computer to execute each processing step of the reputation relationship extraction method according to claim 7 .

Images