JPH0512331A - Valence structure analysis method - Google Patents

Abstract

(57) [Summary] [Purpose] It is possible to easily and accurately create a practical valence dictionary and classification code dictionary. [Structure] The first valence frame creation unit 2 creates a primary valence dictionary 3 by grouping dependency cases in which a common word is common based on surface similarity. The first classification code assigning unit 4 refers to the primary valence dictionary 3 and assigns the same classification code to words belonging to the same slot of the same valence frame to create the primary word classification code dictionary 5. The second valency frame creation unit 6 further divides the dependency cases having common accepting words into groups by considering not only the surface similarity but also the similarity between the primary word classification codes added to the related words. Then, the secondary valence dictionary 7 is created. The second classification code assigning unit 8 refers to the secondary valence dictionary 7 and
The secondary word classification code dictionary 9 is created by the same procedure as that of the classification code assigning unit 4.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a valence structure analysis method used for various natural language analyzes such as kana-kanji conversion in a Japanese word processor and language analysis of a text-to-speech synthesis system.

[0002]

2. Description of the Related Art Generally, in various natural language analysis, the analysis result is often not uniquely determined. For example,
In kana-kanji conversion in a Japanese word processor, since there are homophones (words that have the same reading but different notations),
Further, in the language analysis of the text-to-speech synthesis system, the analysis result may not be fixed because there are isomorphic words (words having the same notation but different readings).

In order to eliminate such ambiguity, as disclosed in, for example, the document "Natural Language Processing 62-6, pages 39 to 44" issued on July 24, 1987, the following is known. It is known that a valence pattern such as a verb and its governing case element is described as a valence pattern, and an analysis method based on this valence pattern is used.

In this kind of analysis method, the valence structure of each idiom is formalized into an appropriate framework, described as a valence pattern in a valence dictionary, and meanings are given to nouns that a case particle can take. It is designed to create classifications.

[0005]

By the way, in the conventional analysis method as described above, the valence valuation dictionary, that is, the valence valence pattern and the meaning classification are usually created manually. However, in natural language analysis, etc., many words are actually used, and even the same word has various meanings depending on the context, field of use, etc., so a practical valence valuation dictionary and semantic classification consistent with it are required. It is very difficult to create by hand. On the other hand, attempts have been made to automatically extract valence frames from dependency cases, but since the semantic classification of nouns used at that time is given top-down manually, it is not always possible to obtain an appropriate valence frame. Has not been done. As described above, conventionally, there is a drawback in that it is impossible to easily and accurately create a practical valence dictionary and a semantic classification consistent with the valence dictionary.

An object of the present invention is to provide a valence valence dictionary and a valence valence structure analysis method capable of easily and accurately creating a semantic valence dictionary consistent with the valence valence dictionary.

[0007]

In order to achieve the above object, the present invention provides the same valence as a valence frame creating means for grouping dependency cases having common words and creating a valence frame. And a classification code assigning means for creating a word classification code dictionary by adding the same classification code to words belonging to the same slot of the frame, and the valence frame creating means provides a dependency case in which the received words are common. In addition to the surface similarity, the similarity between classification codes added to related words is also taken into consideration for grouping to create a final valence frame.

The valence frame creating means is configured to sequentially update the valence frame by referring to the word classification code dictionary created by the classification code assigning means based on the valence frame created by itself. It is characterized by

Further, the classification code assigning means adds the same classification code to the words belonging to the same slot of the same valence frame, and at the same time adds the meaning classification code prepared by hand to each word to classify the words. It is characterized by creating a code dictionary.

[0010]

According to the present invention, the final valence frame is determined by taking into account the dependency case in which the dependency word is common, in addition to the surface similarity and the similarity between the classification codes added to the dependency word. create.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of an embodiment of the present invention.
In the present embodiment, regarding the dependency case database 1 that stores the dependency cases and the dependency cases stored in the dependency case database 1, the dependency cases that have the same accepting word are grouped from the surface similarity. A first valence frame creation unit 2 that divides this and creates it as a primary valence frame and describes it in the primary valence dictionary 3, and a primary bond created by the first valence frame creation unit 2. Valuation dictionary 3
To the first classification code assigning unit 4 for creating the primary word classification code dictionary 5 by assigning the same classification code to the words belonging to the same slot of the same valence frame, and the dependency case database 1. In addition, the primary word classification code dictionary 5 is also referred to, and the modification examples in which the common words are common are grouped in consideration of not only the surface similarity but also the similarity between the classification codes added to the related words. Shico 2
Refer to the second valence frame creation unit 6 created as the next valence frame and described in the secondary valence dictionary 7, and the secondary valence dictionary 7 created by the second valence frame creation unit 6. Then, a second classification code assigning unit 8 is provided which assigns the same classification code to the words belonging to the same slot of the same valence frame to create the secondary word classification code dictionary 9.

Next, the valence structure analysis processing in such a configuration will be described. Note that, in the following, for the sake of simplicity of explanation, only dependency cases centered on idioms will be treated as received words.

First, in the first valence frame creating unit 2, the degree of similarity d1 between dependency cases centered on a certain verb.
Ask for. The similarity d1 is defined so as to reflect the superficial similarity between the dependency cases (here, the similarity of the case particles taken by the denotation), and is defined as, for example, the following expression.

[0014]

[Equation 1]

Here, i is the type of case particle ("ni",
"Wa", "ga", "from", "to", "de", "yori", "to"), and n is the number of case particle types ("8"), and K _i is a dependency case. Further, δ (i) is a function that is “1” when the case slot, that is, the case particle i exists, and “0” when the case slot, that is, the case particle i does not exist, and w (i) is the case particle i. Represents the weight for. Here,
The case slot refers to a unit including not only a case particle but also a word (noun) existing before the case particle.

In this way, after the similarity d1 between the dependency cases centered on a certain declination is obtained, the first valence frame creating unit 2 collects several cases of d1 ≧ TH1. Create a group and create it as a primary valence frame. Note that TH1 is a predetermined threshold value. Such processing is performed for all the terms in the dependency case database 1 to create the primary valence dictionary 3.

Next, in the first classification code assigning section 4,
By referring to the primary valence dictionary 3, the same classification code is given to words (for example, nouns) belonging to the same case slot of the same valence frame. There are as many types of classification codes assigned to one word (noun) as there are frequencies in which the word (noun) appears in the dependency case database 1.

After the primary word classification code dictionary 5 is created by the first classification code assigning unit 4 as described above, the second valence frame creating unit 6 creates the same relationship in the dependency case database 1 by the second valence frame creating unit 6. For the receiving case, the similarity d2 between the dependency cases is obtained. In this case, the similarity d2 is
Not only the similarity of case particles taken by the term, but the similarity of the words (nouns) related to the adjectives through the case particles, that is, the similarity between the classification codes of related words, that is, the words (nouns) in case slots It is calculated as taking into consideration the similarity between classification codes. Specifically, it is determined whether or not the words (nouns) in the case slots are similar to each other, for example, based on the similarity dw of the following equation.

[0019]

[Equation 2]

Here, N (W ₁ ) and N (W ₂ ) are the number of meaning classification codes added to each word (noun) W ₁ and W ₂ , and N (W ₁₂ ) is the word (noun) W. It is the number of meaning classification codes common to ₁ and W ₂ . Second valency frame creation unit 6
Is the minimum value dw _min of the similarity dw between all the words (nouns) belonging to the case slot based on the equation 2
When the relationship of dw _min ≧ TH2 with respect to 2 is satisfied, it is determined that the words (nouns) in the case slot have similarity.

After determining the similarity of the words (nouns) in the case slot in this way, the second valence frame creating unit 6 defines the definition of δ (i) in Equation 1 in the case slot. Change the definition to "1" when the nouns have similarity, "0.5" when there is a case slot, and "0" when there is no case slot.
The similarity d2 is obtained by using the equation 1.

Thereafter, the cases of d2 ≧ TH1 are put together to form some groups, which are created as secondary valence frames.

As a result, in the dependency case in which a certain denotation is common, the similarity between the classification codes of words (nouns) in the same case slot is considered in addition to the superficial similarity based on the same case slot. Then, they can be divided into groups and created as a secondary valence frame, and a secondary valence dictionary 7 can be created.

After that, the second classification code assigning unit 8 adds the same classification code to the words (nouns) included in the same case slot of the same case frame in the secondary valence dictionary 7 to add the secondary word. The classification code dictionary 9 can be created.

Next, a concrete example of the above processing is shown in FIG.
This will be described with reference to (b) and (c). Now, for example, if the dependency case database 1 has dependency cases as shown in FIG. 2A, the first valence frame creating unit 2 will have a primary valence as shown in FIG. 2B. Create a frame. That is, the dependency case of FIG. 2A is a dependency case in which the verb "Kawasu" is a common verb, and the first valence frame creating unit 2 changes the verb from this dependency case. First, "wa" and "ga" are extracted as the case particles, and then "opinion", "body", and "criticism" are extracted as the nouns belonging to the same case particle "wa". In addition, "kai" is extracted as a noun belonging to the case slot of the same case particle "ga", thereby creating a primary valence frame as shown in FIG. 2 (b).

In FIG. 2B, the leftmost alphanumeric string represents the classification code assigned by the first classification code assigning unit 4 after the primary valence frame is created. The same classification code “0” is assigned to nouns belonging to the same case slot, for example, “opinion”, “body”, and “criticism”.
0610W ″ is given.

By the way, as can be seen from the dependency example of FIG. 2A, in the primary valence frame created as shown in FIG. 2B, the nouns "opinion" and "body" belonging to the same case slot Of "" and "criticism", "opinion" and "criticism" can be considered to have similar meanings, but "body" does not have similar meaning to the above two nouns.

The second valence frame creating unit 6 further determines whether or not the nouns in the same case slot have similarities, and by taking this similarity into consideration, the secondary valence frame is calculated. It is designed to be created.
The primary valence frame shown in FIG. 2 is updated to the secondary valence frame as shown in FIG. As can be seen from FIG. 2 (c), the noun “opinion” belonging to the same case slot to which the same classification code is added in FIG. 2 (b),
“Body” and “criticism” are the result of the similarity of nouns,
The nouns "opinion" and "criticism" are judged to constitute the same valence frame and the same classification code is added, but the noun "body" constitutes another valence frame. It is judged and another classification code is added.

As described above, according to the present embodiment, the valence frame and the classification code are automatically created in consideration of the surface similarity and the similarity between the classification codes added to the related words. Therefore, it becomes possible to easily create a valence dictionary and a semantic classification consistent with the valence dictionary without human intervention.

At this time, if the dependency case database 1 is composed of cases adapted to the field of use of the user,
Based on this, it is possible to create a highly accurate and practical valence dictionary and semantic classification adapted to the field of use of the user.

In the above embodiment, the second valence frame creating section 6 can also be configured to be able to refer to the secondary word classification code dictionary 9, and in this case, the second valence classification code dictionary 9 can be referred to. The valence frame creating unit 6 refers to the secondary word classification code dictionary 9 together with the dependency case database 1 when the secondary word classification code dictionary 9 is created by the second classification code assigning unit 8. , The second valence frame can be created again and the secondary valence dictionary 7 can be updated sequentially, and the secondary word classification code dictionary 9 can also be updated sequentially. Therefore, the accuracy of the valence dictionary 7 and the secondary word classification code dictionary 9 can be further improved by repeating such a series of processes several times.

Further, in order to actually eliminate the ambiguity such as the dependency case analysis, the finally created bond valence dictionaries 7 and 2 are used.
Use the next word classification code dictionary 9 to find the analysis solution you want to compare.
_{If 1} and K ₂ , the similarity between the valence frame in the valence dictionary and K ₁ and K ₂ is calculated in the same manner as in the valence frame creating unit 6, and the one with the higher similarity is determined to be the most plausible solution. Just do it.

Further, in FIG. 1, for convenience of description, the first valence frame creating section 2, the second valence frame creating section 6, the first classification code assigning section 4 and the second classification code assigning section. Although 8 and 8 are provided separately, they can be combined into one valence frame creating section and classification code assigning section. In this case, the primary word classification code dictionary 5 and the secondary word classification code dictionary 9 can also be combined into one as a word classification code dictionary, and when combined into one word classification code dictionary, the secondary word classification code dictionary 9 Can be regarded as an update of the primary word classification code dictionary 5.

Further, in the above embodiment, the classification code assigning section adds the same classification code to words belonging to the same slot of the same valence frame, but at the same time, each word is manually preliminarily added. By creating the word classification code dictionary by adding the prepared semantic classification code, it is possible to create a highly accurate valence valence dictionary and word classification code dictionary even when the number of dependency cases is relatively small.

[0035]

As described above, according to the present invention,
For the dependency case in which the common word is common, the final valence frame is created by taking into consideration the similarity between the classification codes added to the modification word as well as the surface similarity. It is possible to easily and accurately create a practical valence dictionary and a semantic classification that is consistent with it.

If the valence frame creating means is configured to sequentially update the valence frame by referring to the word classification code dictionary created by the classification code assigning means based on the valence frame created by itself. , The valence valence dictionary (valence valence frame) and the accuracy of the word classification code dictionary can be successively improved.

Further, the same classification code is added to the words belonging to the same slot of the same valence frame, and at the same time, the meaning classification code prepared by hand is added to each word to create the word classification code dictionary. If the classification code assigning means is configured, it is possible to create a highly accurate bond valence dictionary and word classification code dictionary even when the number of dependency cases is relatively small.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of the present invention.

2A, 2B, and 2C are diagrams showing a specific example of a valence structure analysis process.

[Explanation of symbols]

1 Dependency case database 2 First valence frame creation section 3 primary valence dictionary 4 First classification code assigning section 5 Primary word classification code dictionary 6 Second valency frame creation section 7 Secondary valence dictionary 8 Second classification code assigning section 9 Secondary word classification code dictionary

Claims (3)

[Claims] 1. A valence frame creating means for grouping dependency cases having common vocabulary words to create a valence frame, and adding the same classification code to words belonging to the same slot of the same valence frame. , A categorization code assigning means for creating a word categorization code dictionary, wherein the valence frame creating means classifies the dependency cases in which the vulnerable words are common, together with the surface similarity, to the categorization added to the dependency words. Grouped taking into account the similarity between chords,
A valence structure analysis method characterized by creating a final valence frame. 2. The valence frame creating means is configured to sequentially update the valence frame with reference to the word classification code dictionary created by the classification code assigning means based on the valence frame created by itself. The valence structure analysis method according to claim 1, wherein 3. The classification code assigning means adds the same classification code to words belonging to the same slot of the same valence frame, and at the same time adds a semantic classification code prepared in advance to each word to classify the words. The code dictionary is adapted to be created.
Or the valence structure analysis method described in 2.