JPH0769914B2 - Japanese phrase interdependency analysis device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a Japanese syntax analysis device for analyzing the structure of a Japanese sentence read from an input device.

(Prior art) Conventionally, as dependency analysis in Japanese sentences, the function of each bunsetsu is classified into several types to several tens of types according to a so-called phrase structure grammar, and according to this classification, the related bunsetsu comes before the receiving bunsetsu. The dependency relation is regular and semantic, using the general rule of dependency that the dependency phrase can have only one dependency phrase and the dependency relation does not intersect. According to the empirical rule that the clauses are located as close as possible to each other, the dependency pattern of each sentence is determined by trial and error. well known.

In this mode, shouldn't we treat non-regular dependencies (eg "in a circle with a radius of 5 cm" where "radius" acts as a so-called nominative for "5 cm") as an exception? , I was supposed to deal with each by exception handling.

(Problems to be solved by the invention) The empirical law used in the above-mentioned conventional technique is an irregular dependency (eg, in "writing a circle with a radius of 5 cm", "radius is" If you change the system to allow "5 centimeters" (acting as a so-called nominative) (eg, let a phrase with "ga" relate to a phrase with "no"), The side effect that accompanies becomes a problem (eg, in a sentence with the pattern "~ ga ~ no ~", the dependency of "~ ga" → "~ no" always occurs).

Therefore, in order to certify non-normal dependency, exception handling must be prepared for each non-regular expression to be certified, and even if it is prepared, what kind of exception handling is required? It is difficult to control whether it will start under conditions,
If the start timing is wrong, there is a problem that unnecessary analysis may be ambiguous or a necessary analysis result may not be obtained.

(Object of the Invention) An object of the present invention is to solve the above-mentioned drawbacks of the prior art and to simultaneously improve the accuracy of dependency analysis of Japanese sentences and simplify the control thereof.

(Means for Solving Problems) Inputting a Japanese sentence that is divided into bunsetsu consisting of one or more words and expressing a group of expressions, and the dependency relation between the bunsetsus forming the Japanese sentence. In a Japanese sentence analysis device that outputs, the combination of the type of the related source phrase and the type of the related destination phrase, the dependency attribute between the two related phrases, and the evaluation value related to the peculiarity of the combination of both related phrases are set in advance. Dependency peculiarity evaluation dictionary which holds the set peculiarity evaluation value as a pair, type of dependency bunsetsu, dependency attribute of the bunsetsu pair, and between the dependency source bunsetsu and the dependency destination bunsetsu The type of the "jumped bunsetsu" that is located and the closeness evaluation value set in advance for each combination of the type of the change source bunsetsu, the dependency attribute, and the type of the "jumped bunsetsu" are paired. And a dependency closeness evaluation dictionary held by From the sex evaluation dictionary, the presence / absence of the dependency between the phrases of the input Japanese sentence is determined, the peculiarity evaluation value of the phrase pair determined to be present is obtained, and the set of the phrase and the peculiarity evaluation value are obtained. With regard to the dependency peculiarity evaluation processing unit that outputs as a pair and the phrase pair output from the dependency peculiarity evaluation processing unit, refer to the dependency dependency closeness evaluation dictionary and refer to the target phrase in the dictionary. When there is no pair, the degree of closeness of the phrase pair is set to the highest value, and when there is a target phrase pair in the dictionary, the closeness evaluation value is set as the degree of closeness of the phrase pair, With respect to the dependency pattern candidates of the input sentence, which are obtained based on the dependency closeness evaluation processing unit that outputs and the dependency relations between the phrases extracted by the dependency speciality evaluation processing unit that are determined to have the dependency, And the peculiarity evaluation value of the dependency clause pair The sum of sed evaluation value is achieved by Japanese phrase inter dependency analysis apparatus characterized by having a dependency setting processing unit outputs as the evaluation value of the pattern candidate.

(Operation) In the present invention, in view of the fact that the conventional device directly determines the dependency according to the feature of the clause, the evaluation value for the peculiarity of the dependency and the evaluation for the closeness of the dependency By using it together with the value, it is possible to easily recognize the non-regular dependency and to simplify the control of the process.

Embodiment An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a basic block diagram of the entire Japanese parsing system to which the present invention is applied.

With respect to the Japanese sentence divided into words input from the input device 1, the dependency peculiarity evaluation processing unit 2 using the dependency peculiarity evaluation dictionary 6 determines whether or not there is a dependency between phrases, and the dependency is established. If it is possible, the evaluation value of the dependency attribute and the peculiarity of the dependency is obtained.

Next, the dependency closeness evaluation processing unit 3 using the dependency closeness evaluation dictionary 7 obtains an evaluation value of the closeness between clauses for which the dependency can be established.

Then, the dependency recognition processing unit 4 determines the dependency pattern for the input sentence, calculates the evaluation value of each dependency pattern based on the two types of evaluation values, and outputs the result from the output device 5 in the order of the evaluation value. Output.

FIG. 2 is an example of an input Japanese sentence divided into words, FIG. 3 is a configuration example of a dependency peculiarity evaluation dictionary, FIG. 4 is an operation example of a dependency peculiarity evaluation processing unit, and FIG. 6 is a configuration example of the dependency closeness evaluation dictionary, FIG. 6 is an operation example of the dependency closeness evaluation processing unit, FIG. 7 is an operation example of the dependency dependency qualification processing unit, FIG. FIG. 9 is a flowchart of the dependency speciality evaluation processing unit, FIG. 10 is a flowchart of the dependency closeness evaluation processing unit, and FIG. 11 is a flowchart of the dependency recognition processing unit.

The operation of each processing unit 2-4 in FIG. 1 for the example of FIG. 2 will be described below with reference to FIGS.

The sentence "When I was five, my family moved to Tokyo" in Fig. 2 is also an example of an irregular relation.

First, the operation of the dependency speciality evaluation processing unit 2 is as shown in FIG. 9, and an example of the operation will be described with reference to FIGS. 3 and 4.

First, a brief description of the peculiarities of dependency and the measure of dependency closeness is given.

(Speciality of dependency) The possibility of establishing dependency between clauses is defined in advance as a numerical value from grammatical information, and the one registered in the dictionary is used as the evaluation value. For example, "noun + ga (me)" generally refers only to "terms (run (verb), cold (adjective), quiet (adjective verb))" or "noun + (boy)" However, in the case of "I am / 5 / hour", "Noun +
Of course, in the case of "I / 5 / children / met," the dependency of "I → 5" does not hold. “Noun + ga”
There is a grammatical problem in allowing to be associated with "noun + no". Therefore, it is dealt with by giving a low evaluation to a dependency that is difficult to be grammatically established (a measure of the closeness of the dependency) that is located between the source bunsetsu and the destination bunsetsu. , The propriety of jumping is defined by a numerical value in advance for the phrase in which the source bunsetsu jumps, and the registered one is used as the evaluation value in the dictionary. For example, in "I am / walking / he is / running", there are grammatical references to "I ga" such as the phrase "walking" and "running", but in the case of "i → walking", jump over. There is no clause, but "I run →"
In the case of, since "walking" and "he" jumps over, "I" gives "-20" to the point that "I" jumps over "he".

Hereinafter, the criteria used to set the evaluation value will be described.

In this processing unit, by applying the condition of the dependency establishment as shown in FIG. 3 to the grammatical information of the words forming the bunsetsu, the bunsetsu pair in which the dependency can be established is examined, and at the same time, the dependency Find all attributes and peculiarity evaluation values. Here, the peculiarity evaluation value has a maximum score of 100, and the smaller the peculiarity evaluation value, the more the special dependency.

For example, as shown in FIG. 4, since the rule of rule No. 302 can be applied to the clause pair of clause No. 1 and clause No. 2, the dependency attribute is "case relation" and the peculiarity evaluation value is " 80 ".

Since there is no rule that can be applied to the pair of bunsetsu No.1 and bunsetsu No.3, dependency cannot be established.

Similarly, the dependency attribute and the peculiarity evaluation value in FIG. 4 are obtained.

Next, the operation of the dependency closeness evaluation processing unit 3 is as shown in FIG. 10, and an example of the operation will be described with reference to FIGS. 5 and 6.

In this processing unit, by applying the conditions as shown in FIG. 5 to the grammatical information of the words that form the bunsetsu, the dependency peculiarity evaluation processing unit 2 determines that the bunsetsu pair is likely to be affected. Calculate dependency closeness.

Here, the closeness evaluation value of the phrase pair for which the dependency is satisfied is
It is obtained by setting the initial value to 100 and decreasing according to the closeness evaluation value of FIG. 5 according to the jumping clause. That is, it indicates that the higher the closeness evaluation value of the bunsetsu pair, the higher the dependency of the bunsetsu pair.

For example, as shown in FIG. 6, in the bunsetsu pair of bunsetsu No. 1 and bunsetsu No. 2, since there is no bunsetsu located between them, the closeness evaluation value is 100.

In the clause pair of clause No.1 and clause No.6, clause No.2 is placed between them.
~ Four clauses of clause No.5 are located, and there is no applicable rule when skipping clause No.2. Therefore, the degree of closeness evaluation of clause pairs is not reduced, but clause No.3 is In case of jumping over, rule 50 is applicable because rule No. 503 applies.
When jumping over No.4, rule No.502 is applicable, so 10 is deducted.When skipping over the last clause No.5, there is no applicable rule, so no penalty is given. .1 and bunsetsu No. 6 bunsetsu pair evaluation value is 40
Becomes

Comparing these results, it can be seen that bunsetsu No. 2 which causes non-normal dependency is more preferable than bunsetsu No. 6 which causes regular dependency as a destination of bunsetsu No. 1. Similarly, the closeness evaluation value in FIG. 6 is obtained.

Finally, the operation of the dependency approval processing unit 4 is as shown in FIG. 11, and an example of the operation will be described with reference to FIGS. 7 and 8.

The present processing unit uses the peculiarity evaluation value obtained by the dependence peculiarity evaluation processing unit 2 and the closeness evaluation value obtained by the dependence closeness evaluation processing unit 3 to perform the modification with the evaluation value on the input sentence. Generate a pattern. Here, the simple sum of the peculiarity evaluation value and the closeness evaluation value is used as the evaluation value of each clause pair.

For example, as shown in FIG. 7, there are cases where the clause No. 1 is related to the clause No. 2 and the clause No. 6 is related to the clause No. 1. When these two are compared, the clause No. It can be seen that the evaluation of 2 is higher.

There can be three dependent clauses from the clause No.2,
Since the clause of No. 3 can relate to the clause of No. 6, only the clause of No. 2 is related to the clause of No. 3.
Will be limited to.

In the same manner, the dependency of the phrase pair is expanded and the result as shown in FIG. 8 is obtained.

The evaluation value in FIG. 8 is the sum of the peculiarity evaluation value and the closeness evaluation value of the clause pairs that make up each dependency pattern (in the example of this figure, the clause that is the dependency source is the clause number). Since there are five from 1 to clause No. 5, the highest evaluation value is
1000).

(Effects of the Invention) As described above, according to the present invention, it is possible to recognize even a non-formal dependency in Japanese while keeping the control of processing simple. Therefore, it is possible to input in a system for analyzing Japanese sentences. The tolerance for Japanese sentences can be increased.

[Brief description of drawings]

FIG. 1 is a basic block diagram of a Japanese parsing system to which the present invention is applied, FIG. 2 is an example of an input Japanese sentence divided into words, FIG. 3 is a configuration example of a dependency speciality evaluation dictionary, and FIG. Is a configuration example of the dependency speciality evaluation processing unit, FIG. 5 is a configuration example of the dependency closeness evaluation dictionary, FIG. 6 is an operation example of the dependency closeness evaluation processing unit, and FIG. 7 is a dependency recognition processing unit. Operation example, FIG. 8 is an output example of the dependency determination processing unit, FIG. 9 is a flowchart of the dependency specificity evaluation processing unit, FIG. 10 is a flowchart of the dependency closeness evaluation processing unit, and FIG. 11 is a dependency. It is a flowchart of a receiving authorization processing part. 1 ... Input device, 2 ... Dependency peculiarity evaluation processing unit, 3
...... Dependency closeness evaluation processing unit, 4 ... Dependency recognition processing unit, 5 ... Output device, 6 ... Dependency peculiarity evaluation dictionary, 7 ... Dependency closeness evaluation dictionary.

Claims (1)

[Claims] 1. A Japanese sentence in which a Japanese sentence, which is composed of one or more words and represents a group of expressions and which is divided into units, is input, and a dependency relation between the phrases constituting the Japanese sentence is output. In the analysis device, a peculiarity evaluation in which a combination of the type of the dependent bunsetsu and the type of the dependent bunsetsu, the dependency attribute between the both bunsetsus, and the evaluation value regarding the peculiarity of the combination of the both bunsetsus are set in advance It is located between the dependency peculiarity evaluation dictionary that holds a pair of values, the type of the dependency source phrase, the dependency attribute of the phrase pair, and the dependency source phrase and the dependency destination phrase. A relation that holds a type of "jumped bunsetsu", a type of the relation source bunsetsu, a dependency attribute, and a closeness evaluation value preset for each combination of "jumped bunsetsu" types Before the reception closeness evaluation dictionary and the peculiarity evaluation dictionary, It is determined whether or not there is a dependency between the phrases of the input Japanese sentence, the peculiarity evaluation value of the phrase pair judged to be present is obtained, and the pair of the phrase and the peculiarity evaluation value are output as a pair. With respect to the dependency peculiarity evaluation processing unit and the phrase pair output from the dependency peculiarity evaluation processing unit, the dependency closeness evaluation dictionary is referred to, and if there is no target phrase pair in the dictionary, Closeness of clause pairs is 1
If the highest value is set and the target phrase pair exists in the dictionary, the closeness evaluation value is set as the closeness of the phrase pair, and the dependent closeness evaluation processing unit outputs the closeness evaluation value. Regarding the dependency pattern candidates of the input sentence, which are found based on the dependency relations between the phrases extracted by the peculiarity evaluation processing unit, the peculiarity evaluation value and the degree of closeness of the dependency phrase pairs forming each pattern candidate And a dependency setting processing unit that outputs the sum of evaluation values as an evaluation value of the pattern candidate.