JPH0769710B2 - Natural language analysis method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a natural language analysis method, and more particularly to a natural language analysis method for performing processing by utilizing surface words.

[0002]

2. Description of the Related Art Social needs for natural language processing systems are increasing and research is being actively conducted. For example, in machine translation technology, not only written word translation but also technical development of spoken language translation such as automatic translation telephone has been recently performed. Unlike the written word translation, to realize the spoken word translation, a robust process that constantly outputs an answer even with rough content is required.

[0003]

However, the conventional method for performing a precise analysis based on a complicated grammar is not suitable for the analysis of a spoken language that often deviates from the grammar,
Writing complicated grammar rules requires skill, and when the number of rules becomes huge, it becomes difficult to perform management work such as inconsistency detection between rules, generates a huge number of ambiguous structures, and collates grammar rules There are many problems such as complicated algorithms and long processing time.

Therefore, a main object of the present invention is to provide a natural language analysis method capable of robust processing capable of handling non-grammatical sentences.

[0005]

The invention according to claim 1 is
A natural language analysis method for grasping the structure of an input natural language sentence, in which a surface pattern described by surface words and variables is searched by an index of each word of the input natural language sentence and searched. The surface patterns are combined to match the input sentence from a large linguistic unit such as a compound sentence to a small linguistic unit such as a word, and the combination of the input sentence and the surface pattern that is successfully matched is used as a structure candidate. It is something that is done.

More preferably, in the surface layer pattern of claim 1, the input sentence is corrected when necessary so that it can be collated with the input sentence.

Still more preferably, the surface layer pattern of claim 1 has a word that embodies a variable as an example.

More preferably, the maximum likelihood structure is determined and output from the structure candidates based on the semantic distance between the example and the words of the input sentence.

[0009]

[0010]

[0011]

[0012]

The natural language analysis method according to the present invention searches the surface layer pattern described by the words and variables of the surface layer by the index of each word of the input natural language sentence,
The retrieved surface patterns are combined to match the input sentence from a large linguistic unit such as a compound sentence to a small linguistic unit such as a word, and a combination of the input sentence and the surface pattern that is successfully matched is structured. To be a candidate.

[0013]

1 is a schematic block diagram of an embodiment of the present invention. In FIG. 1, the natural language analysis processing method according to the present invention is realized by a morpheme analysis module 1, a surface layer pattern matching module 2, an analysis assistance module 3, and a maximum likelihood structure determination module 4. The morphological analysis module 1 refers to a built-in dictionary and adds information such as word division and part-of-speech / inflection. The surface pattern index is prepared from the word. For example, "tai"
Has an index to the surface pattern "X I want to". The surface pattern is a description of linguistic units such as simple sentences and noun phrases by surface words and variables, such as “X taino ga ga” and “Y of X”. In addition, X and Y show a variable. The surface patterns that can be matched with the input sentence are narrowed down by the index of each word of the input sentence.

The surface layer pattern matching module 2 combines the narrowed down surface layer patterns to create the structure of the input sentence. Surface patterns are classified by several linguistic units, from those with a large unit at the compound sentence level to those with a small unit at the word level, and the surface pattern is matched efficiently with the input sentence.

FIG. 2 is a view showing typical surface layer patterns to which the present invention is applied, arranged from the largest linguistic unit to the smallest.

Here, an algorithm for matching the input sentence with the surface pattern will be described. (0) Perform (1) for the largest linguistic unit among the surface layer patterns obtained by the morphological analysis module 1.

(1) The surface pattern is collated with the input sentence. If the collation is successful, (2) is performed. If it fails, execute (3).

(2) If the collation succeeded in (1) is the surface layer pattern having no variable, the collation of that portion succeeds and ends. Otherwise, execute (1) for each variable part of the surface layer pattern.

(3) The linguistic unit of the surface layer pattern is lowered and (1) is executed. If the linguistic unit cannot be lowered any more, the collation ends in failure.

This algorithm collates the surface pattern with the input sentence from the largest linguistic unit to the smallest linguistic unit. If the matching is successful for all the parts of the input sentence, the combination result of the matched surface patterns is used as the structure. For example, for the sentence "I want to apply for a meeting", first of all, the sentence pattern "X Tai naga ga", which has the largest linguistic unit among the surface patterns that can be matched, is the single sentence unit. Collate.

Next, the surface layer pattern "X to Y" of the case relationship unit is collated with "apply to the meeting" which embodies the variable part of "X want to". Furthermore, the variable X of “X to Y”,
It means that Y is matched with the surface pattern of each word "meeting" and "application", and the input sentence is completely matched with the combination of the surface patterns. As a result, ((I want to apply for (meeting))) is obtained as a structure.

The analysis assisting module 3 is activated when the surface sentence pattern cannot be collated with the input sentence in the state of the input sentence as it is, and corrects the input sentence according to the auxiliary rule so that the surface sentence pattern can be collated with the input sentence. The auxiliary rule performs operations such as normalization on the input sentence so that the surface pattern can be matched. For example, the surface pattern cannot be directly matched in the input sentence “it attends”. Therefore,
Correct the input sentence according to the following supplementary rules.

Pronoun verb → pronoun verb ((that, participate), ...) Pronoun verb ((this, send), ...) Auxiliary rules are described using grammatical categories such as part-of-speech,
A word example that embodies each grammar category is described as a matching condition. The auxiliary rule above complements a particle when the pronoun and the verb are consecutive. For example, the pronoun and verb word sequences "it participate" are complemented by the particle "ni" and are modified to "participate in it.""Sendthis"
For, the particle "" is complemented and corrected to "send this." By this supplementary rule, the input sentence "it attends" is complemented with the particle "ni" because of the close semantic distance between (it attends) and (it attends). Will be fixed. In addition, “X to Y” is newly searched as a collatable surface pattern from the complemented particle “ni”. After that, in the surface layer pattern matching module 2, the surface layer pattern "X to Y" is matched to the input sentence, and a structure of ((it) (attends)) is obtained.

The maximum likelihood structure determination module 4 finds the semantic distance between the input word and the example word of the surface pattern from the structures obtained by the surface pattern matching module 2, and selects the one with the smallest total distance. Determine the structure of.

FIG. 3 is a diagram showing the basic processing flow of the natural language analysis method by pattern matching according to the present invention. Next, a specific example of analyzing the following sentence (1) will be described with reference to FIG.

(1) "I would like a room of 10,000 yen at the Kyoto Hotel." First, the morphological analysis module 1 performs word division processing of the input sentence. Then, the surface pattern that can be matched is searched by the index from each word of the sentence (1). For example, the surface pattern "X of Y" is searched for from the word "no" of the input sentence.

The surface pattern matching module 2 tries to match the surface pattern retrieved by the morphological analysis module 1 with the input sentence, but the matching fails because the particle is missing.

Common noun verb → common noun verb ((room, please), ...) Then, the analysis assistance module 3
Complement the particle "o" to the input sentence and correct the input sentence as (1 ') "Please give me a room of 10,000 yen at Kyoto Hotel."

The surface pattern matching module 2 receives the result of the analysis assisting module 3 and tries to match the surface pattern with (1 '). First of all, among the narrowed down surface patterns, the surface pattern "Please say X" of the single sentence unit, which has the largest linguistic unit, is collated. Next, "the room for 10,000 yen at the Kyoto hotel" that satisfies the variable X of "Please give me X."
Try to match the surface pattern to. This expression can be matched with the surface layer pattern "X of Y" of the case relation unit, but since it has two "no", two structures are created for the input sentence as follows. (2) ((((10,000 yen) (room)) of ((Kyoto Hotel))) (3) ((((Kyoto Hotel) ((10,000 yen)) (room))) Please,) "Kyoto Hotel", "10,000 yen" for the variable "Y of X",
"Room" can be matched, and the entire input sentence is composed entirely of surface patterns.

The maximum likelihood structure determination module 4 calculates the semantic distance between the input word and the example word of the surface pattern,
The sum of the distances of the structure of (2) is 0.0, and the sum of the distances of the structure of (3) is 0.5. (2) having a small sum of distances is output as a structure for (1).

[0031]

As described above, according to the present invention, the surface pattern is searched by the index of each word of the input sentence, and the searched surface pattern is searched for from a large linguistic unit such as a compound sentence. The input sentence is matched to such a small linguistic unit to make a candidate for the structure of the input sentence, and the maximum likelihood structure is determined and output from the candidate by calculating the semantic distance between the input and the example. By doing so, it is possible to realize a robust natural language analysis mechanism that requires a structure even for sentences whose structure is difficult to explain grammatically, such as spoken language.

[Brief description of drawings]

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

FIG. 2 is a diagram showing a typical surface layer pattern arranged from the largest linguistic unit to the smallest.

[Fig. 3] Fig. 3 is a block diagram showing a flow of natural language analysis processing for an input sentence "Please give a room of 10,000 yen at Kyoto Hotel".

[Explanation of symbols]

 1 Morphological analysis module 2 Surface pattern matching module 3 Analysis auxiliary module 4 Maximum likelihood structure determination module

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Hitoshi Iida, 5 Seiraya-cho, Seika-cho, Soraku-gun, Kyoto Pref. A-T-R Co., Ltd. Automatic Translation Telephone Research Laboratory (56) Reference JP-A-63- 213068 (JP, A) JP-A-60-200369 (JP, A)

Claims (4)

[Claims] 1. A natural language analysis method for grasping the structure of an input natural language sentence, wherein a surface layer pattern described by surface words and variables by an index of each word of the input natural language sentence is used. Search, combine the searched surface patterns, and match the input sentence from a large linguistic unit such as a compound sentence to a small linguistic unit such as a word, and a surface layer that is successfully matched with the input sentence A natural language analysis method in which a pattern combination is used as a structure candidate. 2. The natural language analysis method according to claim 1, wherein the input sentence is corrected when necessary so that the surface layer pattern can be collated with the input sentence. 3. The natural language analysis method according to claim 1, wherein the surface layer pattern has a word that embodies a variable as an example. 4. The natural language analysis according to claim 3, wherein the maximum likelihood structure is determined and output from the structure candidates based on the semantic distance between the example and the words of the input sentence. Method.