JP2008242626A - Term registration apparatus - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To register unregistered terms in appropriate semantic categories by increasing the accuracy of determining semantic categories even if sentence patterns have ambiguity in terms of semantic categorization. <P>SOLUTION: A sentence pattern certainty calculation part 3 calculates a certainty of a sentence pattern in consideration of relationships between semantic categories of a plurality of terms constructing the sentence pattern. A term/semantic category similarity calculation part 6 calculates similarities between an unregistered term and a plurality of semantic categories in a thesaurus in accordance with the sentence pattern certainty. A term registration part 8 specifies the semantic category that has the highest similarity with the unregistered term by referring to the calculations of the term/semantic category similarity calculation part 8, and registers the unregistered term in the semantic category. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a term registration device that extracts terms from a corpus and registers the terms in a thesaurus.

In recent years, with the increase in the number of business documents accumulated electronically, there has been a demand for technology for searching and analyzing business-dependent documents and feeding back the search results and analysis results to the business. It is coming.
As an example, there is a technique of checking a description that leads to a defect by searching for a corresponding part of a new design document using terms related to a past defect recorded in the defect report.
As another example, when analyzing a defect report, the problem is analyzed by analyzing the correlation (text mining) between terms related to hardware such as devices and components and terms related to attributes such as current and voltage. There is a technology that feeds back information on the attributes of components that lead to the design.

To be able to search for domain-specific terms from business documents in a specific domain, a thesaurus with the semantic classification of the domain-specific terms (the semantic classification to which the term belongs, It is necessary to use a dictionary describing relationships and the like.
Generally, because the terms used in the target document are different for each applicable field and the semantic classification to which the term belongs is different, the thesaurus of other existing fields cannot be used. You need to create a possible thesaurus.
When creating a thesaurus applicable to the relevant field, it is inefficient to manually create a thesaurus individually. Therefore, it is desired to develop a technique for registering terms in the thesaurus with high accuracy.
Here, the “term” is a concept including a word and a compound word.

For example, the following Patent Documents 1 and 2 disclose a term registration device that registers terms in a thesaurus.
That is, the following Patent Document 1 describes the semantic classification of a thesaurus for registering unregistered words by calculating the similarity of unregistered words (terms) based on the appearance frequency of “verb + case” and “noun”. A technique for selecting (registered node) is disclosed.
However, in the term registration device disclosed in Patent Document 1, when there is a ambiguity of semantic classification in the sentence pattern of “verb + case”, the semantic classification of a thesaurus for registering unregistered words should be selected with high accuracy. I can't.
Here, as the ambiguity of the semantic classification, for example, when there is a sentence pattern “fixing X” consisting of the verb “fix” and case “to”, the term “X” belongs to the semantic classification [part]. And a situation that may belong to a semantic classification [signal].

Patent Document 2 below discloses a method of constructing a thesaurus of words by obtaining a distance between words based on co-occurrence relation data between words.
Specifically, the similarity of “noun” is calculated based on the appearance frequency of “verb + case” and “noun”, and if “verb” belongs to a plurality of semantic categories, “verb” Is resolved, and verbs having each semantic classification are treated as different words.
For another word, recalculate the appearance frequency of “verb + case” and “noun” and recalculate the semantic classification to which “noun” belongs, thereby increasing the calculation accuracy of the semantic classification to which “noun” belongs. I am doing so.
In Patent Document 2, the ambiguity of the semantic classification to which the “verb” belongs is eliminated by clustering only the distances between the elements of the set constituting the sentence pattern.
For example, in the sentence pattern “fix X”, the ambiguity is resolved by clustering by the distance between the elements of the set of terms X taking the case “”.

JP 2005-326952 A (paragraph numbers [0028] to [0029], FIG. 1) JP 2001-331515 A (paragraph numbers [0023] to [0049], FIG. 1)

  Since the conventional term registration device is configured as described above, in Patent Document 1, if there is ambiguity of semantic classification in a sentence pattern of “verb + case”, the semantic classification of a thesaurus for registering unregistered words However, in Patent Document 2, when the “verb” belongs to a plurality of semantic categories, the ambiguity of the semantic category to which the “verb” belongs is resolved, and the semantic category is selected. Accuracy can be increased. However, when recalculating the semantic classification to which the “noun” belongs, new information other than the “verb + case” sentence pattern is not used, so the amount of corpus data available is small and sufficient statistics can be obtained. If the difference in the distance between the elements of the term X in the sentence pattern is small due to the absence of the ambiguity, or if the ambiguity cannot be resolved without using a plurality of cases, the meaning ambiguity However, there is a problem that the accuracy of selecting the semantic classification cannot be improved even if the semantic classification to which the “noun” belongs is recalculated.

For example, by using a sentence pattern “fix X” from an electrical related technical document (for example, a technical document including sentences such as “fix unit to base” and “DWN fix to address”). When extracting terms, the term “unit” belonging to the semantic category [component] and “DWN” belonging to the semantic category [signal] are extracted.
At this time, regarding the term “unit” and the term “DWN”, which are terms that take the case “”, if the distance between the terms is calculated, the distance between the terms becomes small, and the ambiguity of the semantic classification cannot be eliminated. Sometimes.
In this case, the ambiguity of the semantic classification can be eliminated by calculating the distance between the terms, including the term “base” and the term “address” that take the case “ni” as well as the term “ni”. However, in Patent Document 2, such a distance calculation between terms is not performed.

  The present invention has been made to solve the above-described problems. Even when there is ambiguity of semantic classification in a sentence pattern, the accuracy of selecting semantic classification is improved, and an unregistered term is appropriately classified. An object of the present invention is to obtain a terminology registration device that can be registered in the Internet.

  The term registration device according to the present invention includes a term extracting means for extracting a term from a corpus using a sentence pattern expressing a restriction on a relationship between terms, and a plurality of terms constituting the sentence pattern belong to Extracting terms using sentence pattern certainty calculation means that calculates the certainty of the sentence pattern in consideration of the relationship between the semantic classifications, and sentence pattern certainty calculated by the sentence pattern certainty calculation means There is provided a term / similar classification similarity calculation means for calculating the similarity between an unregistered term extracted by the means and a plurality of semantic classifications in the thesaurus, and the term registration means calculates the similarity between the terms and the semantic classification. By referring to the calculation result of the means, the semantic classification having the highest similarity with the unregistered term extracted by the term extracting means is specified, and the unregistered term is registered in the semantic classification. That.

  According to the present invention, the term extracting means for extracting the term from the corpus using the sentence pattern expressing the restriction of the relationship between the terms, and the meaning to which the plurality of terms constituting the sentence pattern belong Extracted by term extraction means using sentence pattern certainty calculation means that calculates the certainty of the sentence pattern in consideration of the relationship between classifications, and sentence pattern certainty calculated by the sentence pattern certainty calculation means A term / semantic classification similarity calculation means for calculating similarity between a registered unregistered term and a plurality of semantic classifications in the thesaurus, and the term registration means calculates the similarity between the term / semantic classification calculation means By referring to the results, the semantic classification having the highest similarity with the unregistered terms extracted by the term extraction means is identified, and the unregistered terms are registered in the semantic classification. Even if the ambiguity of meaning classification exists, to improve the selection accuracy of the semantic classification, the effect that can be registered in the proper semantic classification unregistered terms.

Embodiment 1 FIG.
FIG. 1 is a block diagram showing a term registration device according to Embodiment 1 of the present invention. In FIG. 1, a corpus 1 is a memory storing document data of an electronic business document.
The term extraction unit 2 performs a process of extracting a term from the corpus 1 using a sentence pattern that expresses a constraint on the relationship between terms, that is, a sentence pattern that expresses a constraint on a dependency relationship between terms. . The term extracting unit 2 constitutes term extracting means.

The sentence pattern certainty calculation unit 3 considers the relationship between semantic classifications to which a plurality of terms constituting the sentence pattern (sentence pattern used by the term extraction unit 2) belongs, and the certainty of the sentence pattern The process of calculating is performed. The sentence pattern certainty calculator 3 constitutes a sentence pattern certainty calculator.
The term similarity calculation unit 4 uses the sentence pattern certainty calculated by the sentence pattern certainty calculation unit 3 to perform a process of calculating the similarity between terms. The inter-term similarity calculation unit 4 constitutes inter-term similarity calculation means.
The term similarity DB 5 is a database for storing the similarity between terms calculated by the term similarity calculation unit 4.

The term / semantic similarity calculation unit 6 uses the similarity between terms calculated by the inter-term similarity calculation unit 4, and uses the unregistered terms extracted by the term extraction unit 2 and a plurality of meanings in the thesaurus 9. A process of calculating the similarity between the classifications is performed. The term / semantic classification similarity calculation unit 6 constitutes a term / semantic classification similarity calculation means.
The term / semantic classification similarity DB 7 is a database that stores the similarity calculated by the term / meaning classification similarity calculation unit 6.

The term registration unit 8 refers to the calculation result of the term / semantic classification similarity calculation unit 6, and among the plurality of semantic classifications in the thesaurus 9, the term registration unit 8 has a similarity to an unregistered term extracted by the term extraction unit 2. A process of specifying the highest semantic classification and registering unregistered terms in the semantic classification is performed. The term registration unit 8 constitutes term registration means.
The thesaurus 9 is a memory for storing a plurality of semantic categories and storing terms belonging to a plurality of semantic categories.

FIG. 2 is a block diagram showing hardware resources of a computer that implements the term registration device according to Embodiment 1 of the present invention.
The computer that implements the term registration device of FIG. 1 is composed of the following hardware resources.
・ Input device 11 consisting of a keyboard, mouse, etc.
A communication device 12 used for communication with other control devices
A control device 13 including a CPU (Central Processing Unit) 14 and a main storage device 15
The control device 13 is implemented with the term extraction unit 2, sentence pattern certainty calculation unit 3, inter-term similarity calculation unit 4, term / semantic classification similarity calculation unit 6, and term registration unit 8 of FIG. 1. The

1 stores a program showing the processing contents of the term extraction unit 2, the sentence pattern certainty calculation unit 3, the interterm similarity calculation unit 4, the term / semantic classification similarity calculation unit 6 and the term registration unit 8 in FIG. , A secondary storage device 16 that stores the term similarity DB 5, term / semantic category similarity DB 7, and thesaurus 9 in FIG. 1.
The secondary storage device 16 may be used as auxiliary storage in the course of calculation processing in the control device 13.
An output device 17 including a display device 18 such as a display and a printing device 19 such as a printer.
The output device 17 is used for output of the calculation process in the control device 13, the similarity, the contents of the thesaurus 9, and the like.

  The program and data indicating the processing contents of the term extraction unit 2, the sentence pattern certainty calculation unit 3, the term similarity calculation unit 4, the term / semantic classification similarity calculation unit 6 and the term registration unit 8 in FIG. When the program is recorded on the recording medium 20 such as a ROM, the recording medium driving device 21 reads the program and data from the recording medium 20, and the control device 13 reads the program and data read by the recording medium driving device 21. Records in the secondary storage device 16.

FIG. 3 is a flowchart showing the processing contents of the term registration device according to Embodiment 1 of the present invention.
FIG. 4 is an explanatory diagram showing an example of a corpus.
In FIG. 4, only the sentence is shown for simplification of explanation, but information such as a document number indicating the source of the sentence or a sentence number in the document may be added.

FIG. 5 is an explanatory diagram showing an example of a sentence pattern.
FIG. 5 shows an example in which the sentence pattern is described in the format of “Φ (cs) → semantic classification”.
Φ is a verb and cs is an example of a particle representing a case, and Φ (cs) indicates that “an element having a dependency relationship between the verb Φ and the case cs is extracted”.
The semantic classification is a label of the semantic classification (a set of terms) in the thesaurus 9.
FIG. 5 shows an example of a sentence pattern using one case for the sake of simplicity of explanation, but a sentence pattern using a plurality of cases may be used.
In the following description, for the sake of brevity, only the portion of Φ (cs) may be referred to as a sentence pattern. In addition, the label of the semantic classification is surrounded by [].

FIG. 6 is an explanatory diagram showing an example of term data extracted by the term extraction unit 2 of the term registration device according to Embodiment 1 of the present invention.
It is assumed that each term data is composed of “heading” and “frequency”.
FIG. 6A shows an example in which the term is a verb, and V represents the heading of the verb. FIG. 6B shows an example in which the term is a noun, and N indicates the heading of the noun.

FIG. 7 is an explanatory diagram showing the relationship between terms and cases extracted by the term extraction unit 2 of the term registration device according to Embodiment 1 of the present invention.
In Fig. 7, terms are expressed by verbs and adjectives, and case relationships are expressed by superficial case particles, but other parts of speech and deep cases (cases obtained by deep syntactic analysis and semantic analysis). The relationship may be

FIG. 8 is an explanatory diagram showing an example of the thesaurus 9 of the term registration device according to Embodiment 1 of the present invention.
In FIG. 8, nodes of the thesaurus 9 are semantic classification labels or terms. Items enclosed in [] are semantic classification labels, and those not enclosed in [] are terms.
“−” And “+” from each node indicate that the one that is traced to the lower right is the lower term of the node, and the one that is traced to the upper left is the upper term of the node.
In general, the semantic classification has a hierarchy, but in the following, for the sake of simplicity, the term in the leaf will be described as belonging to the immediately higher semantic classification.
Here, for the sake of simplicity, only the upper-lower relationship is shown, but a thesaurus including other relationships such as a synonym relationship and a partial-whole relationship may be used.

Next, the operation will be described.
The term extraction unit 2 extracts terms from the corpus 1 shown in FIG. 4 by using the sentence pattern shown in FIG. 5 that expresses the dependency relationship of terms (step ST1).
Here, as shown in FIG. 6, the term extraction unit 2 extracts the terms V and N of the terms that are verbs and the terms that are nouns, and also extracts the frequency of the terms.
Moreover, the term extraction part 2 extracts the relationship between a term and a case as shown in FIG.

As described above, the term extraction unit 2 extracts terms and the relationship between terms and cases, and these extractions are performed by performing morphological analysis and syntax analysis on the corpus 1. Since the implementation method of morphological analysis and syntax analysis is widely known, detailed description is omitted here.
However, the syntax analysis includes dependency analysis for analyzing that a clause relates to another clause in the sentence.
Here, the phrase is composed of independent words (nouns, verbs, etc.) and attached words (particles, auxiliary verbs, etc.). Typical cases of dependency analysis include a case where a noun + case particle relates to a verb, and a case where a clause of a noun + case particle has a parallel relationship with a clause of another noun.

When the term extraction unit 2 extracts terms from the corpus 1, the sentence pattern certainty calculation unit 3 includes a plurality of terms constituting the sentence pattern of FIG. 5 used by the term extraction unit 2. The certainty of the sentence pattern (Φ, cs) is calculated (step ST2).
When calculating the certainty of the sentence pattern (Φ, cs), the sentence pattern certainty calculation unit 3 uses the similarity between terms stored in the inter-term similarity DB 5 to use the sentence pattern (Φ, cs). Calculate certainty.
In the inter-term similarity DB 5, a predetermined initial value (similarity between terms) is stored in the initial state before the inter-term similarity calculation unit 4 calculates the inter-term similarity. ing.
Hereinafter, the sentence pattern certainty factor calculation process in the sentence pattern certainty factor calculation unit 3 will be described in detail.

ただし、Φは動詞、ｃｓは格、ｅ１、ｅ２は用語を表している。
ｃｓ'は、格ｃｓ以外の格である。
Ｗ（ｅｉ，Φ（ｃｓ），ｃｓ'）は、コーパス１の中で、文パターンΦ（ｃｓ）とともに格ｃｓ'で共起する単語の集合であり、文パターン（Φ，ｃｓ）で格ｃｓ'をとる「単語バグ」と称する。 First, in order to explain the certainty of the sentence pattern, the following formula (1) for calculating “the certainty between terms using the sentence pattern” will be explained.

However, Φ represents a verb, cs represents a case, and e1 and e2 represent terms.
cs ′ is a case other than the case cs.
W (ei, Φ (cs), cs ′) is a set of words co-occurring in the corpus 1 together with the sentence pattern Φ (cs) in the case cs ′, and the case cs in the sentence pattern (Φ, cs). Called “word bug”.

Dw (W1, W2) is a distance between the word bug W1 and the word bug W2, and is defined as described later.
For example, when Φ1 = fixed, cs = “on”, e1 = “unit”, e2 = “metal fitting”, the case cs ′ becomes cs1 ′ = “to” and cs2 ′ = “on”. cs1 ′ and cs2 ′ have cases other than “O”.
At this time, the word bug is as follows.
W1 (unit, fixed (to), to) = {base, millet}
W2 (fitting, fixing (to), to)) = {control panel, seijyoban}
W3 (unit, fix (with),) = {hook, clasp}
W4 (unit, fix (in), with) = {screw, screw}

ただし、Ｎ（ｗ｜Ｗ）は単語バグＷにおける単語ｗの頻度
Ｗ１とＷ２のいずれかが空集合の場合は、Ｄｗ（Ｗ１，Ｗ１）＝０とする。 The distance Dw (W1, W2) between the word bug W1 and the word bug W2 is defined by the following equation (2).

However, N (w | W) is the frequency of the word w in the word bug W. If either W1 or W2 is an empty set, Dw (W1, W1) = 0.

Here, D _{g (F)} (w1, w2) is the distance between the term w1 and the term w2, and is obtained in advance by the conventional technique as an initial value.
Alternatively, it may be obtained from the similarity between terms stored in the term similarity DB 5 (similarity between terms calculated by the term similarity calculation unit 4). Further, a predetermined set value may be used.
However, in the repetition phase, it is determined from the similarity of terms stored in the inter-term similarity DB 5.

For example, the “confidence between terms using sentence patterns” for the term “unit” and the term “metal fitting” is calculated as in the following equation (3).

ただし、ｅは用語、ｓｅｍＣａｔは用語が属する意味分類を表している。
また、ｅ_j∈ｓｅｍＣａｔは、意味分類が“ｓｅｍＣａｔ”であることが既知である用語（コーパス１のすべての用語）を表している。 Next, since the certainty of the sentence pattern is expressed in the form of I (Φ, cs) (e, semCat), the certainty between terms using the sentence pattern is used to Is calculated as follows.

However, e represents a term and semCat represents a semantic classification to which the term belongs.
Further, e _j εsemCat represents a term (all terms in the corpus 1) whose semantic classification is known to be “semCat”.

Here, when the sentence pattern certainty calculation unit 3 calculates the certainty of the sentence pattern, the calculation of the certainty between the terms in the corpus 1 whose semantic classification is known is shown. The certainty factor of the sentence pattern may be calculated by multiplying the certainty factor by an appropriate coefficient, or by performing another operation on the certainty factor between terms.
For example, in the corpus 1, when the terms “metal fitting” and “switch” whose semantic classification is known to be [component] exist, the term e that takes the sentence pattern Φ (cs) = fixed () The certainty that is [part] is calculated by the following equation (5).

When the sentence pattern certainty calculation unit 3 calculates the certainty of the sentence pattern as described above, the interterm similarity calculating unit 4 calculates the similarity between the terms using the certainty of the sentence pattern. The similarity between terms is stored in the term similarity DB 5 (step ST3).
For example, the similarity D (e1, e2) between terms is calculated by the following formula (6).

The expression (6) calculates the certainty factor I _{(Φ, cs)} (e1, e2) between terms using all sentence patterns Φ (cs) for the terms e1, e2, and all the calculation results. This is equivalent to seeking the sum of
That is, when Φ has “fix”, “wear”, and “attach”, and the case cs has “to”, “ni”, and “de”, calculate the following formula (7): It corresponds to.

When the similarity between terms / semantic classification calculation unit 6 calculates the similarity between terms, the similarity between terms (or the sentence calculated by the sentence pattern certainty calculation unit 3) is calculated. The degree of similarity between the unregistered term extracted by the term extracting unit 2 and a plurality of semantic categories in the thesaurus 9 is calculated using the certainty of the pattern), and the similarity is calculated between the term and the semantic category Store in DB7 (step ST4).
The calculation of the similarity between an unregistered term and a plurality of semantic classifications in the thesaurus 9 is performed by calculating the co-occurrence frequency of the unregistered terms and sentence patterns and the registered terms (already registered in the semantic classification of the thesaurus 9). It can be obtained by using the co-occurrence frequency of the term) and the sentence pattern.

ただし、ｅは用語を表し、ｓｅｍＣａｔは用語が属する意味分類を表している。
また、ｅ_j∈ｓｅｍＣａｔは、意味分類が“ｓｅｍＣａｔ”であることが既知である用語（コーパス１のすべての用語）を表している。 That is, the similarity between the unregistered term e and the semantic classification semCat can be calculated by the following equation (8).

However, e represents a term and semCat represents a semantic classification to which the term belongs.
Further, e _j εsemCat represents a term (all terms in the corpus 1) whose semantic classification is known to be “semCat”.

When the term / semantic classification similarity calculation unit 6 calculates the similarity between an unregistered term and a plurality of semantic classifications in the thesaurus 9, the term registration / symbol classification similarity calculation unit 6 Referring to the calculation result, among the plurality of semantic categories in the thesaurus 9, the semantic category having the highest similarity with the unregistered term extracted by the term extracting unit 2 is specified, and the unregistered term is defined as the semantic category (Step ST5).
That is, an unregistered term is registered as a lower node of the semantic classification label of the thesaurus 9 that is closest to the unregistered term.

When the processing of steps ST1 to ST5 ends as described above, the term registration device calculates the similarity between terms calculated by the term similarity calculation unit 4 and the term similarity calculation unit 4. The difference between the similarities between terms previously stored in the interterm term similarity DB 5 is calculated, and when the difference is equal to or greater than a predetermined threshold, or the term / semantic classification similarity calculation unit 6 calculates Calculate the difference between the similarity between the unregistered term and the semantic category and the similarity stored in the term / semantic category similarity DB 7 before being calculated by the term / semantic category similarity calculating unit 6; If the difference is equal to or greater than the predetermined threshold, the process returns to step ST2, and the processes of steps ST2 to ST5 are executed again.
On the other hand, if any difference is less than the predetermined threshold value, the series of processing ends.

As apparent from the above, according to the first embodiment, the term extraction unit 2 that extracts terms from the corpus 1 using the sentence pattern expressing the constraint of the relationship between terms, and the sentence pattern The sentence pattern certainty factor calculation unit 3 that calculates the certainty factor of the sentence pattern in consideration of the relationship between semantic classifications to which a plurality of constituent terms belong, and the sentence pattern certainty factor calculation unit 3 A term / semantic classification similarity calculation unit 6 is provided that calculates the similarity between an unregistered term extracted by the term extraction unit 2 and a plurality of semantic classifications in the thesaurus using the certainty of the sentence pattern. The term registration unit 8 refers to the calculation result of the term / semantic classification similarity calculation unit 8 to identify the semantic category having the highest similarity with the unregistered term extracted by the term extraction unit 2, and is not registered. In the above semantic classification Even if there is ambiguity of semantic classification in the sentence pattern, it has the effect of increasing the accuracy of selecting semantic classification and registering unregistered terms in appropriate semantic classification. .
That is, according to the first embodiment, when calculating the certainty of a sentence pattern, the relationship between the semantic classifications to which a plurality of terms constituting the sentence pattern belong (not only the case “ In addition, even if there is ambiguity of semantic classification in the sentence pattern, the accuracy of semantic classification selection is improved and unregistered terms are registered in the appropriate semantic classification. The effect which can be done is produced.

  Further, according to the first embodiment, the sentence pattern certainty calculation unit 3 calculates the certainty of the sentence pattern using the similarity between terms previously calculated by the interterm similarity calculation unit 4. Since it comprised, there exists an effect which can improve the calculation precision of the certainty of a sentence pattern.

Embodiment 2. FIG.
In Embodiment 1 described above, the term extracting unit 2 extracts a term from the corpus 1 using a sentence pattern that expresses a dependency relationship between terms, but the term extracting unit 2 uses the term pattern. Terms may be extracted from the corpus 1 using sentence patterns expressing constraints on similarity relationships.
Specifically, it is as follows.

FIG. 9 is an explanatory diagram showing an example of a corpus.
In FIG. 9, only the sentence is shown for the sake of simplification, but information such as a document number indicating the source of the sentence or a sentence number in the document may be added.
FIG. 10 is an explanatory diagram showing an example of a sentence pattern.
FIG. 10 shows an example in which the sentence pattern is described in a format of “Φ (cs) → semantic classification”.
In the first embodiment, Φ represents a verb, but in the second embodiment, Φ represents the relationship between terms by representing the surface layer.

FIG. 11 is a flowchart showing the processing contents of the term extraction unit 2 of the term registration device according to Embodiment 1 of the present invention.
FIG. 12 is an explanatory diagram showing the relationship between terms extracted by the term extraction unit 2 of the term registration device according to Embodiment 1 of the present invention.

Next, the operation will be described.
The term extraction unit 2 extracts terms from the corpus 1 in FIG. 9 using the sentence pattern in FIG. 10 expressing the constraints on the similarity relationship between terms.
Hereinafter, the term extracting process in the term extracting unit 2 will be described in detail.

First, the term extraction unit 2 performs morphological analysis and syntax analysis on the sentence stored in the corpus 1 (step ST11).
Since the implementation method of morphological analysis and syntax analysis is widely known, detailed description is omitted here.

When the vocabulary analysis and syntax analysis are performed on the sentence, the term extraction unit 2 extracts inter-term relationship data as shown in FIG. 12 as data matching the sentence pattern of FIG. 10 from the analysis result (step ST12). .
In FIG. 12, the relationship between terms is stored in the Rel column, one term in the sentence pattern is stored in the E1 column, and the other term in the sentence pattern is stored in the E2 column.

In the sentence pattern in the second embodiment, as shown in FIG. 12, there are sentence patterns such as “A or B”, “B of A”, “A and B”, and “B like A”.
For example, when the relationship between terms is “no”, there is ambiguity in terms of the relationship between semantic classifications. For example, there is an ambiguity such as an object / attribute relationship “<attribute> of <part> (eg, capacitor voltage)” and a partial overall relationship “part of component (eg, remote control button)”.

In this second embodiment, the sentence pattern confidence is confirmed by replacing the part described in the first embodiment as Φ is a verb with the relationship between terms stored in the Rel column of FIG. The degree calculation unit 3 calculates the certainty factor of the sentence pattern having the relationship Rel.
Similar to the first embodiment, the term / semantic classification similarity calculating unit 6 calculates the similarity between terms when the term similarity calculating unit 4 calculates the similarity between terms (or sentence pattern). Using the certainty factor of the sentence pattern calculated by the certainty factor calculation unit 3), the similarity between the relationship Rel extracted by the term extraction unit 2 and a plurality of semantic categories in the thesaurus 9 is calculated, and the similarity Are stored in the term / semantic classification similarity DB 7.

When the term / semantic classification similarity calculation unit 6 calculates the similarity between the relationship Rel and the plurality of semantic classifications in the thesaurus 9, the term registration unit 8 calculates the calculation result of the term / semantic classification similarity calculation unit 6. Referring to the above, the semantic category having the highest similarity with the relation Rel extracted by the term extraction unit 2 is identified from among the plurality of semantic categories in the thesaurus 9, and the unregistered term is registered in the semantic category. To do.
That is, an unregistered term is registered as a lower node of the semantic classification label of the thesaurus 9 having the closest distance to the relationship Rel.

  As is apparent from the above, according to the second embodiment, the term extraction unit 2 extracts terms from the corpus 1 using the sentence pattern of FIG. 10 expressing constraints on the similarity of terms. Since it comprised, even when the term which resembles is contained in corpus 1, there exists an effect which can register an unregistered term in an appropriate semantic classification.

Embodiment 3 FIG.
In the first embodiment, the sentence pattern certainty calculation unit 3 uses the similarity between terms stored in the interterm similarity DB 5 to calculate the certainty of the sentence pattern. When the appearance frequency is low, smoothing of the appearance frequency may be performed.
That is, in the third embodiment, among the terms extracted by the term extracting unit 2, for example, the certainty of the sentence pattern is corrected according to the similarity between terms that are verbs.
In the third embodiment, the sentence pattern certainty is corrected according to the similarity between terms that are verbs, but is not limited to the similarity between terms that are verbs. The certainty of the sentence pattern may be corrected according to the similarity between terms that are adjectives or the similarity between terms that are adjective verbs.

FIG. 13 is an explanatory diagram showing an example of a corpus.
FIG. 14 is a flowchart showing the processing contents of the sentence pattern certainty calculation unit 3 of the term registration device according to Embodiment 1 of the present invention.

Next, the operation will be described.
The term extraction unit 2 extracts terms from the corpus 1 shown in FIG. 13 using the sentence pattern shown in FIG. 5 that expresses constraints on the dependency relationship of terms as in the first embodiment (see FIG. 3). Step ST1).
Here, as shown in FIG. 6, the term extraction unit 2 extracts the terms V and N of the terms that are verbs and the terms that are nouns, and also extracts the frequency of the terms.
Moreover, the term extraction part 2 extracts the relationship between a term and a case as shown in FIG.

When the term extraction unit 2 extracts a term from the corpus 1, the sentence pattern certainty calculation unit 3 forms the sentence pattern of FIG. 5 used by the term extraction unit 2 as in the first embodiment. The certainty factor of the sentence pattern (Φ, cs) is calculated in consideration of the relationship between the semantic classifications to which the term “B” belongs (step ST2 in FIG. 3).
Moreover, the sentence pattern certainty calculation part 3 calculates the similarity between the verbs which comprise a sentence pattern from the similarity between nouns which are case elements (step ST21 of FIG. 14).

  In step ST2 of FIG. 3, the sentence pattern certainty calculation unit 3 extracts the sentence pattern Φ (cs) by “extracting an element having a dependency relationship between the verb Φ and the case cs” (hereinafter referred to as “a sentence pattern related to a noun”). In step ST21 of FIG. 14, the sentence pattern certainty calculator 3 extracts the sentence pattern Φ (cs) as “the noun Φ and the case cs are in a dependency relationship”. "(Hereinafter referred to as" sentence pattern related to verb "), the certainty factor is calculated.

When the sentence pattern certainty calculation unit 3 calculates the certainty of the sentence pattern related to the noun, the interterm term similarity calculating unit 4 uses the certainty of the sentence pattern to determine between the terms. The similarity is calculated, and the similarity between the terms is stored in the inter-term similarity DB 5 (step ST3 in FIG. 3).
In addition, when the sentence pattern certainty calculation unit 3 calculates the certainty of the sentence pattern related to the verb, the inter-term similarity calculation unit 4 uses the certainty of the sentence pattern, as in the first embodiment, to determine the verb. The similarity between the verbs is calculated, and the similarity between the verbs is stored in the inter-term similarity DB 5.

ただし、Ｗ１とＷ２のいずれかが空集合の場合は、Ｄｗ（Ｗ１，Ｗ１）＝０とする。 When the inter-term similarity calculation unit 4 stores the similarity between verbs in the inter-term similarity DB 5, the sentence pattern certainty calculation unit 3 corrects the certainty of the sentence pattern related to the noun according to the similarity between the verbs. (Step ST22 in FIG. 14).
In the first embodiment, in the corpus 1, W (e _i , Φ (cs), cs ′) is a sentence pattern (as a set of terms that co-occur with the sentence pattern Φ (cs) in the case cs ′. This is referred to as a “word bug” that takes a case cs ′ with Φ, cs), and the distance between the word bugs is defined by the following equation (9).

However, if either W1 or W2 is an empty set, Dw (W1, W1) = 0.

ただし、Σは、Φ_i以外のすべてのΦ_jに関して加算するものとする。
また、ｓｉｍ（Φ₁，Φ₂）は、文パターンΦ₁（ｃｓ）と文パターンΦ₂（ｃｓ）における用語Φ₁と用語Φ₂の類似度である。 At this time, in the first embodiment, N (w | W) is defined as the frequency of the word w in the word bug W, but in this third embodiment, N (w | W) is expressed by the following formula ( 10).

However, Σ is added for all Φ _j other than Φ _i .
Sim (Φ ₁ , Φ ₂ ) is the similarity between the term Φ ₁ and the term Φ _{2 in} the sentence pattern Φ ₁ (cs) and the sentence pattern Φ ₂ (cs).

For example, for each verb, the noun “unit” co-occurs with the case “o”, and the noun “control panel” co-occurs with the case particle “ni”.
Further, the similarity between the verb “fix” and the verb “wear” calculated by the inter-term similarity calculation unit 4 is 0.9, and the similarity between the verb “fix” and the verb “assign” is 0.8. To do.
W1 (e _i , Φ (cs), cs ′) = W (unit, fixed (to))
W2 (e _i , Φ (cs), cs ′) = W (unit, mounted (to))
W3 (e _i , Φ (cs), cs ′) = W (unit, assigned ())
In this case, N (control panel | fixed) is calculated as follows.
N (control panel | fixed) = frequency of word “control panel” in word bug W1
+ Frequency of word “control board” in word bug W2 × 0.9
+ Frequency of word “control board” in word bug W2 × 0.8

In addition, in the above calculation of N (w | W), in order to remove noise, addition of frequency multiplied by similarity sim (Φ _i , Φ _j ) of two items is performed, and similarity sim (Φ _i , Φ You may restrict | limit when _j ) is more than a predetermined threshold value.
Also, like the verbs “load” and “pile”, one has a detailed semantic classification (if it is in the lower hierarchy on the thesaurus), and the other takes a broad meaning (in the higher hierarchy on the thesaurus) In some cases, in the calculation of the frequency of the word w in the word bug W, the frequency multiplied by the word similarity of the detailed semantic classification may be added only to the word having a wide meaning. .
Since the subsequent steps are the same as those in the first embodiment, description thereof is omitted.

  As is apparent from the above, according to the third embodiment, since the configuration is such that the certainty of the sentence pattern related to the noun is corrected according to the similarity between terms extracted by the term extracting unit 2, the above-described embodiment As compared with the first aspect, there is an effect that it is possible to further improve the accuracy of selecting the semantic classification.

It is a block diagram which shows the term registration apparatus by Embodiment 1 of this invention. It is a block diagram which shows the hardware resource of the computer which implement | achieves the term registration apparatus by Embodiment 1 of this invention. It is a flowchart which shows the processing content of the vocabulary registration apparatus by Embodiment 1 of this invention. It is explanatory drawing which shows an example of corpus. It is explanatory drawing which shows an example of a sentence pattern. It is explanatory drawing which shows an example of the term data extracted by the term extraction part 2 of the term registration apparatus by Embodiment 1 of this invention. It is explanatory drawing which shows the relationship between the term extracted by the term extraction part 2 of the term registration apparatus by Embodiment 1 of this invention, and a case. It is explanatory drawing which shows an example of the thesaurus 9 of the term registration apparatus by Embodiment 1 of this invention. It is explanatory drawing which shows an example of corpus. It is explanatory drawing which shows an example of a sentence pattern. It is a flowchart which shows the processing content of the term extraction part 2 of the term registration apparatus by Embodiment 1 of this invention. It is explanatory drawing which shows the relationship between the terms extracted by the term extraction part 2 of the term registration apparatus by Embodiment 1 of this invention. It is explanatory drawing which shows an example of corpus. It is a flowchart which shows the processing content of the sentence pattern reliability calculation part 3 of the term registration apparatus by Embodiment 1 of this invention.

Explanation of symbols

  1 corpus, 2 term extraction unit (term extraction unit), 3 sentence pattern certainty calculation unit (sentence pattern certainty calculation unit), 4 inter-term similarity calculation unit (inter-term similarity calculation unit), 5 inter-term similarity DB, 6 Term / semantic classification similarity calculation section (term / semantic classification similarity calculation means), 7 Term / semantic classification similarity DB, 8 Term registration section (term registration means), 9 Thesaurus, 11 Input device , 12 communication device, 13 control device, 14 CPU, 15 main storage device, 16 secondary storage device, 17 output device, 18 display device, 19 printing device.

Claims (6)

  Considering the relationship between term extraction means that extracts terms from the corpus using sentence patterns expressing the relationship restrictions between terms and the semantic classification to which multiple terms that make up the sentence pattern belong The sentence pattern certainty factor calculating means for calculating the certainty factor of the sentence pattern and the unregistered information extracted by the term extracting unit using the certainty factor of the sentence pattern calculated by the sentence pattern certainty factor calculating unit. The term extraction is performed by referring to the calculation result of the term / semantic classification similarity calculation means for calculating the similarity between the term and the plurality of semantic classifications in the thesaurus, and the calculation result of the term / semantic classification similarity calculation means. A term registration device comprising: a term registration unit that identifies a semantic category having the highest degree of similarity with an unregistered term extracted by the means, and registers the unregistered term in the semantic category.   Considering the relationship between term extraction means that extracts terms from the corpus using sentence patterns expressing the relationship restrictions between terms and the semantic classification to which multiple terms that make up the sentence pattern belong The sentence pattern certainty factor calculating means for calculating the certainty factor of the sentence pattern and the sentence pattern certainty factor calculated by the sentence pattern certainty factor calculating means are used to calculate the similarity between terms. Similarity between unregistered terms extracted by the term extracting means and a plurality of semantic classifications in the thesaurus using the similarity calculating means and the similarity between terms calculated by the inter-term similarity calculating means Refer to the calculation results of the term / semantic classification similarity calculation means for calculating the degree and the term / semantic classification similarity calculation means, and the similarity to the unregistered term extracted by the term extraction means is the highest. Identify the semantic classification had the term registration apparatus that includes a term registering means for registering the unregistered terms in the semantic classification.   3. The term registration device according to claim 2, wherein the sentence pattern certainty calculating means calculates the certainty of the sentence pattern using the similarity between terms previously calculated by the inter-term similarity calculating means.   4. The term extraction unit according to claim 1, wherein the term extraction unit extracts a term from the corpus using a sentence pattern expressing a dependency relationship of terms. Term registration device.   4. The term according to any one of claims 1 to 3, wherein the term extracting means extracts a term from the corpus using a sentence pattern expressing a restriction of similarity relation of terms. Registration device.   6. The sentence pattern certainty factor calculating unit corrects the certainty factor of a sentence pattern according to the similarity between terms extracted by the term extracting unit. Terminology registration device.

Images