JP2003085170A - Fixed form sentence clustering device and method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a fixed form sentence clustering device and a method capable of collecting, clustering, and outputting similar fixed form sentences as one class. SOLUTION: This fixed form sentence clustering device comprises a means 10 for dividing respective sentences in a corpus with every word, a word n gram dictionary means 20 for forming word n grams by linking the words obtained by division, and extracting and preserving the word n grams for satisfying a prescribed condition from the word n grams, a fixed form sentence extracting means 30 for forming a set of word n grams by combining a part or the whole in the preserved word n grams, and extracting a sentence becoming a prescribed threshold value or more in a rate including the set as a fixed form sentence form the corpus, a means 40 for preserving the extracted fixed form sentence and the set of word n grams, and a means 50 for calculating similarity of mutual classes to classes being a cluster of prescribed sentences included in corpus, and merging two classes having the highest similarity into one.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a clustering apparatus and method for collecting similar sentences from a plurality of fixed phrases and classifying them into classes.

[0002]

2. Description of the Related Art Conventionally, a sentence including a word string in which words are arranged in a fixed pattern (hereinafter, this pattern is referred to as a fixed pattern and this sentence is referred to as a fixed sentence) is extracted from a set of a plurality of sentences. In this case, it was extracted in sentence units. A method of extracting a fixed form sentence is described in, for example, Naoto Kato, “Machine Translation Method for Sentences Containing Fixed Form Pattern”, IPSJ Journal, Vo1.3
6, No. 9, pp. As disclosed in 2081-2090, 1995, it is roughly composed of the following three steps.

In step 1, the word n-gram dictionary (n is
A natural number of 1 or more. ) Is created. In step 2, for each sentence of the set of multiple sentences, the ratio including the word n-gram (hereinafter,
It is called the content rate. ) Is calculated. In step 3, a sentence whose content rate is equal to or higher than a threshold value is extracted as a fixed sentence.

Next, the above-mentioned conventional method will be described by taking as an example the case where a fixed phrase is extracted from economic news.
First, in step 1, a word n-gram that is a word string in which n words are connected is formed, and a word n-gram dictionary having the word n-gram as a constituent element is created. Here, the word n-gram dictionary is created by extracting, from the formed word n-grams, the word n-grams that frequently appear in the sentences in the corpus.

The word n-gram dictionary 1 created using economic news is composed of, for example, word n-grams W1, W2, W3 as shown in FIG. In fact, in economic news, many expressions include word strings having a fixed pattern arrangement, and a large number of word n-grams are extracted from economic news. The symbol “/” in FIG. 5A represents a word break.

Next, in step 2, the content rate, which is the rate including the word n-gram obtained in step 1, is calculated for each sentence in the economic news. For example, FIG. 5C shows an example in which the content rate is calculated using the word n-gram dictionary 1 shown in FIG. 5A and the four example sentences shown in FIG. 5B.

Example sentence 1, example sentence 2, and example sentence 4 shown in FIG.
For each word n-gram W in the word n-gram dictionary 1.
Each one contains 1, W2, W3, example sentence 3
Contains only W1. In the formula for calculating the content shown in FIG. 5C, the denominator is the total number of words constituting the sentence for which the content is to be calculated, and the numerator is each word n-gram W1, W2, W3 contained in the target sentence. It is the sum total of the number of constituent words (here, n = 3).

In step 3, sentences having a content rate equal to or higher than a predetermined content rate threshold value are extracted. For example, assuming that the threshold value is 0.5, the example sentence 1, the example sentence 2, and the example sentence 4 shown in FIG. 5B are extracted as the standard sentences, but the example sentence 3 is not extracted. Generally, fixed phrases are used when analyzing language phenomena manually. In such a case, it may be easier to process by collecting similar sentences.

[0009]

However, according to the conventional method of extracting a fixed sentence, the fixed sentence can be extracted only in units of sentences, not in units of a set of a series of similar sentences (hereinafter referred to as a class). There is a problem in that it is inefficient to look at all sentences when analyzing. In fact, tens to hundreds of similar fixed phrases may be obtained, and it is difficult to manually investigate all of them. Therefore, similar fixed phrases are easier to analyze when they are grouped as one class.

The present invention has been made to solve such a problem, and an object thereof is a fixed sentence clustering apparatus and method capable of collectively clustering similar fixed sentences into one class and outputting them. To provide.

[0011]

In view of the above points, the invention according to claim 1 divides each sentence included in a corpus, which is a set of predetermined sentences, into words constituting each sentence. The morpheme analysis means and the words obtained by dividing each sentence included in the corpus are connected based on a predetermined rule to generate a word n-gram (n is a natural number of 1 or more), and the word a word n-gram dictionary means for extracting and storing word n-grams satisfying a predetermined condition from n-grams;
The n-grams of a predetermined type in the n-grams stored by the gram dictionary means are combined to generate a set of word n-grams, and a sentence in which the ratio including the set of word n-grams is a predetermined threshold value or more is the corpus. A fixed sentence extracting means for extracting as a fixed sentence from among the sentences included in the fixed sentence, a fixed sentence saving means for saving the fixed sentence extracted by the fixed sentence extracting means and the set of the word n-gram included in the fixed sentence, A fixed sentence clustering means for calculating a similarity between the classes with respect to a class which is a set of predetermined sentences included in the corpus, and performing merging in which the two classes having the highest similarity are merged into one class. It has a configuration including.

With this configuration, since the fixed phrase clustering means is provided to detect and merge similar fixed phrases,
It is possible to realize a fixed sentence clustering device that can combine similar fixed phrases among the fixed phrases extracted from the corpus into one class and output them as a fixed phrase class.

The invention according to claim 2 is the method according to claim 1, wherein the predetermined condition satisfied by the word n-gram extracted by the word n-gram dictionary means is the word n-gram in the extracted sentence. The number of occurrences, which is the number of occurrences of, is greater than or equal to a predetermined number. With this configuration, the fixed phrase clustering means is provided to detect and merge similar fixed phrases, so that similar fixed phrases among the fixed phrases extracted from the corpus are combined into one class,
It is possible to realize a fixed sentence clustering device that can output as a fixed sentence class.

In the invention according to claim 3, a step of dividing each sentence included in the corpus, which is a set of predetermined sentences, into words constituting each sentence, and each sentence included in the corpus are divided. The words obtained by dividing the words are connected based on a predetermined rule to form a word n-gram (n is a natural number of 1 or more.).
And a word n-gram dictionary step of extracting and storing word n-grams satisfying a predetermined condition from the word n-grams, and a predetermined type of word in the word n-grams saved by the word n-gram dictionary step. A standard sentence extraction for generating a set of word n-grams by combining n-grams, and extracting a sentence having a ratio including the set of word n-grams equal to or more than a predetermined threshold value as a standard sentence from the sentences included in the corpus. A step, a step of saving a set of the fixed phrase extracted by the fixed phrase extraction step and the word n-gram included in the fixed phrase, and the class that is a set of predetermined sentences included in the corpus And a step of calculating the degree of similarity between the classes and merging the two classes having the highest degree of similarity into one class.

With this configuration, a step of detecting and merging similar fixed phrases is provided. Therefore, similar fixed phrases among the fixed phrases extracted from the corpus are combined into one class and output as a fixed phrase class. It is possible to realize a fixed sentence clustering method that enables

The invention according to claim 4 is the method according to claim 3, wherein the predetermined condition satisfied by the word n-gram extracted in the word n-gram dictionary step is the word n-gram in the extracted sentence. The number of occurrences, which is the number of occurrences of, is greater than or equal to a predetermined number. With this configuration, since a step of detecting and merging similar fixed phrases is provided, it is possible to combine similar fixed phrases among the fixed phrases extracted from the corpus into one class and output them as a fixed phrase class. A fixed-form sentence clustering method can be realized.

The invention according to claim 5 is the method according to claim 3 or 4, further comprising: a predetermined rule based on the generation of the word n-gram in the word n-gram dictionary step, While maintaining the order of the array of n, n (n is a natural number of 1 or more) is connected to generate a word n-gram. With this configuration,
Since a step of detecting and merging similar fixed phrases is provided, a fixed phrase clustering method capable of collecting similar fixed phrases among the fixed phrases extracted from the corpus into one class and outputting as a fixed phrase class. Can be realized.

[0018]

BEST MODE FOR CARRYING OUT THE INVENTION A fixed phrase clustering apparatus according to a first embodiment of the present invention will be described below with reference to the accompanying drawings. FIG. 1 shows a block configuration diagram of a fixed text clustering apparatus according to a first exemplary embodiment of the present invention. In FIG. 1, the fixed-form sentence clustering apparatus 100.
Is composed of a morpheme analysis unit 10, a word n-gram dictionary unit (n is a natural number of 1 or more) 20, a fixed sentence extraction unit 30, a fixed sentence storage unit 40, and a fixed sentence clustering unit 50.

The morphological analysis means 10 divides each sentence included in a publicly known corpus externally input to the standard sentence clustering apparatus 100 into words constituting each sentence (hereinafter referred to as morphological analysis). Is. The word n-gram dictionary means 20 connects the words obtained by dividing each sentence included in the corpus according to a predetermined rule to generate a word n-gram (n is a natural number of 1 or more), This is a means for extracting and saving word n-grams satisfying a predetermined condition from the word n-grams.

Here, the predetermined condition may be that the number of times the word n-gram appears in the extracted sentence (hereinafter, referred to as the occurrence number) is a predetermined number or more. The predetermined rule may be a rule that n words (n is a natural number greater than or equal to 1) are continuously formed to generate a word n-gram while maintaining the order of arrangement of the words forming each sentence.

The fixed phrase extraction means 30 combines word n-grams of a predetermined type in the word n-gram dictionary means 20 to generate a set of word n-grams, and a ratio including the set of word n-grams. Is a means for extracting a sentence having a predetermined threshold value or more from the sentences included in the corpus as a fixed sentence. Fixed phrase storage means 4
Reference numeral 0 is a means for storing a fixed sentence extracted by the fixed sentence extracting means 30 and a set of word n-grams included in the fixed sentence.

The standard sentence clustering means 50 calculates the degree of similarity between classes for a class which is a set of predetermined sentences included in the corpus, and sets the two classes having the highest degree of similarity as one class. It is a means for merging.

In the first embodiment of the present invention, the case where Japanese is the object of processing has been described, but the same applies when English, French and other foreign languages are the objects of processing. It is possible to extract fixed phrases and perform clustering. The fixed-form sentence clustering device for them is obvious from the above description, and therefore its explanation is omitted.

As described above, the fixed-form sentence clustering apparatus according to the first embodiment of the present invention is provided with fixed-form sentence clustering means to detect and merge similar fixed-form sentences. It is possible to cluster similar similar fixed phrases into one class and output them as a fixed phrase class.

FIG. 2 is a flow chart showing the flow of processing in the fixed text clustering method according to the second embodiment of the present invention. The configuration of the device for executing the boilerplate clustering method according to the second embodiment of the present invention is the same as the configuration of the boilerplate clustering device according to the first embodiment of the present invention shown in FIG. Therefore, the description thereof will be omitted.

In step S201, the morphological analysis means 10
Performs morphological analysis on all sentences in a known corpus. Specifically, when a corpus is input from the outside, each sentence included in the corpus is divided into words that compose it. In step S202, the word n-gram dictionary means 20.
Is a word obtained by dividing each sentence included in the above corpus based on a predetermined rule.
Is a natural number of 1 or more. ) Is generated, a word n-gram satisfying a predetermined condition is extracted from the word n-grams and stored.

Here, the predetermined condition may be that the number of times the word n-gram appears in the extracted sentence (hereinafter referred to as the number of occurrences) is a predetermined number or more. The predetermined rule may be a rule that n words (n is a natural number greater than or equal to 1) are continuously formed to generate a word n-gram while maintaining the order of arrangement of the words forming each sentence.

In step S203, the fixed phrase extraction means 30
Generates a set of word n-grams by combining word n-grams of a predetermined type using the word n-gram dictionary obtained in step S202, and determines the content rate of the word n-gram set for each sentence in the corpus. calculate. Step S204
Then, the fixed phrase extraction means 30 extracts a sentence whose content ratio calculated in step S203 is equal to or more than a predetermined threshold value as a fixed phrase from the sentences included in the corpus, and gives the fixed phrase together with the fixed phrase. Collect a set of n-grams to create a set of fixed phrases. Here, the fixed sentence or a set of fixed sentences extracted above may be stored in the fixed sentence storage unit 40.

In step S205, the fixed phrase clustering means 50 extracts C + 1 fixed phrases from the set of fixed phrases created in step S204, using a predetermined natural number as the number of classes of fixed phrases, C, and extracts each fixed sentence. A class consisting of one sentence corresponding to is created and the similarity between the classes is calculated. In step S206, the fixed phrase clustering unit 50 merges the two classes having the highest similarity among the classes obtained in step S205.

In step S207, the fixed-form sentence clustering unit 50 has an empty set of fixed-form sentences (hereinafter, referred to as a set of remaining fixed-form sentences) after a predetermined number of sentences are extracted in step S205 or step S208 described later. Judge whether it is a set or not. If it is determined in step S207 that the remaining set of fixed phrases is not an empty set, step S208
Then, the fixed-form sentence clustering unit 50 extracts one more fixed-form sentence from the remaining set of fixed-form sentences, creates a new class, and calculates the similarity between the classes.

When the processing in step S208 is completed,
The process returns to step S206 to repeat the above process.
When it is determined in step S207 that the remaining set of fixed phrases is an empty set, the fixed phrase clustering unit 50
The result of classifying fixed phrases into classes (hereinafter referred to as clustering) is output.

The processing in each of the above steps will be described below for Japanese. Also, in the following explanation,
It is assumed that the corpus input from the outside includes six sentences shown as example sentences in FIG. Step S
At 201, the morphological analysis unit 10 divides each sentence into words for the example sentences 1 to 6. The result of morphological analysis of example sentence 1 is shown in FIG.

In step S202, the word n-gram dictionary means 20 obtains a word n-gram satisfying a predetermined condition by using the words obtained by dividing each sentence included in the corpus. An example of the word n-gram generated for the corpus shown in this example is shown in FIG. The word 3 gram when n is 3 shown in FIG. 3C has a configuration in which three consecutive words (“present”, “ha”, and “,”) form a word string while maintaining their order. Has become.

The same processing as above is performed on all the sentences in the corpus, and the obtained word n-grams are totaled to calculate the appearance frequency. For example, it is calculated as shown in FIG. Here, as one of the above-mentioned certain conditions, for example, if "word n gram whose appearance frequency is 800 times or more" is used, the word n
The word 3 gram “current / wa /,” in FIG. 3 (d) is registered in the gram dictionary, but the word 4 gram “current / wa /, / yes” is not registered.

Similarly, word n-grams satisfying this condition are collected to create a word n-gram dictionary. For example, a word n-gram dictionary having entries as shown in FIG. 4 (a) can be obtained. In step S203, the fixed phrase extraction means 3
In the case of 0, a predetermined set of word n-grams is formed using the word n-gram dictionary obtained in step S202, and the content rate, which is the ratio including this set of word n-grams in each sentence in the corpus, is calculated.

For each sentence of Example 2 shown in FIG.
Using the word n-gram dictionary 2 shown in FIG.
The content rate is calculated as shown in (b). Below, FIG.
A description will be given using Example sentence 1 in Example 2 shown in (a). As illustrated in FIG. 4C, the example sentence 1 includes one word n-gram W1, W2, and W3 illustrated in FIG. 4A, respectively.
In the calculation of the content rate in FIG. 4B, the denominator is the total number of words constituting the sentence whose content rate is to be calculated, and the numerator is the total number of words constituting each word n-gram included in the sentence.

In the case of the example sentence 1, the total number of words forming the example sentence 1 is 14, and each word n-gram W included in the example sentence 1 is W.
The number of words forming each of 1, W2 and W3 is
Three, three and three. In step S204, the fixed phrase extraction unit 30 extracts, from the sentences in the corpus, the sentence whose content rate calculated in step S203 is equal to or more than a predetermined threshold as a fixed phrase, and extracts it together with the extracted fixed phrase. Output a set of word n-grams that gives.

In the following, the threshold value of the content rate is, for example, 0.
5 will be described. In this case, since the content rates of the example sentence 1, the example sentence 2, the example sentence 4, the example sentence 5, and the example sentence 6 are 0.5 or more, they are extracted as the standard sentences. However, example sentence 3
For, the content rate is 0.25 and is 0.5 or less, so it is not extracted as a fixed phrase. A pair of the extracted fixed sentence and a set of word n-grams giving it is made into a pair, and the pair is collected to form a set of fixed sentence.

The set of fixed phrases is constructed, for example, as shown in FIG. In FIG. 4C, the example sentence 1 is
It is shown to include a set of word n-grams composed of word n-grams W1, W2, W3. In step S205,
The fixed-form sentence clustering means 50 takes a predetermined natural number as the number of classes of fixed-form sentences C, extracts C + 1 sentences from the set of fixed-form sentences created in step S204, and classifies one sentence corresponding to each sentence. Create and calculate the similarity between each class.

In the following description, the number of classes C is 2. In step S205, for example, three sentences, example sentence 1, example sentence 2, and example sentence 5 are extracted from the set of fixed-form sentences generated in step S204, and each of them is classified into class 1 (class consisting of example sentence 1) and class 2 (example sentence 2). Class consisting of),
It is a constituent element of class 3 (class consisting of example sentence 5).
As the similarity between the classes further calculated in step S205, for example, the function Si defined in the following equation (1) is used.
Let us use m (x _i , x _j ).

[0041]

[Equation 1] here,

[Equation 2] Is.

S_i, S_jIs the i-th class, j-th class
It is a symbol that identifies a lath. Also, x_iIs the class s_iTo
It is a set of included word n-grams, and p types of word n-grams.
Mux_i ¹~ X_i ^pComposed by. f_iIs the word n
Each word n-gram that makes up a set of frames x_i ¹~ X_i ^pIs included in the sentence
Number of occurrences (hereinafter referred to as the number of occurrences) f_i ¹~ F_i ^pBy
Is a set of data that is composed of

Similarly, x _j is a word included in the class s _j
It is a set of n-grams, and q kinds of words n-grams x _j ^{1 to} x _j ^q.
Composed by. f _j is a set of data composed of the occurrence times f _i ¹ ~f _i ^p for each word n-gram x _j ¹ ~x _j ^q constituting the set of word n-grams. Further, the symbols and expressions located below the summation symbol Σ appearing in the numerator are the same word n-gram among the word n-grams forming the pairs x _i , x _j of the word n-grams included in the classes s _i , s _j. (For example, x _i ^r
And x _j ^s ), their occurrences (in this example, f
_i ^r and f _j ^s ) are added. The two classes shown in FIG. 4D will be described as an example.

The degree of similarity between the two classes shown in FIG. 4D is obtained by substituting each occurrence number shown in FIG. 4D into the above equation (1).

[Equation 3] Becomes

In step S206, the fixed text clustering means 50 merges the two classes having the highest similarity among the similarities obtained in step S205. In the above example 2, class 1 (class consisting of example sentence 1), class 2 (class consisting of example sentence 2), class 1
And class 3 (class consisting of example sentence 5) and class 2 and class 3 are calculated in similarity.

In the above-described example 2, since each word n-gram type and the number of appearances (1 in each case) appearing in class 1 and class 2 are the same, the similarity between class 1 and class 2 is 1.0. However, the similarity between the class 1 and the class 3 and the similarity between the class 2 and the class 3 are 0 because the respective word n-gram types that appear are completely different. Therefore, the similarity between class 1 and class 2 is judged to be the maximum,
Class 1 and class 2 are merged into one class (class consisting of example sentence 1 and example sentence 2).

The word n-grams of the class after merging include, for example, word n-grams of all kinds included in the two classes before merging, and the word n-grams of the same kind occur in two classes. It shall be the total number of times.
For example, the class s ₃ obtained by merging the class s ₁ and the class s ₂ shown in FIG. 4D is x ₃ = (W27, W441, W78, W84), f ₃ = (3,
3,1,1,3).

In step S207, the fixed-form sentence clustering means 50 uses step S205 or step S20.
In step 8, it is determined whether the set of fixed phrases after the predetermined number of sentences have been extracted is an empty set. When it is determined in step S207 that the "remaining fixed phrase set" is not an empty set, the fixed phrase clustering unit 50 extracts one more fixed phrase from the remaining fixed phrase set and newly creates it in step S208. Create different classes and calculate the similarity between each class.

When the sentence shown in Example 2 above is used, after the processing in step S207, for example, the example sentence 4 is taken out and is made a constituent element of class 4 (class consisting of example sentence 4), and the process in step S206 is performed. Compute the similarity with each of the C classes that exist later. That is, the similarity between class 1 and class 4 and the similarity between class 3 and class 4 are calculated, and then the process returns to step S206.

After step S206, the same processing as described above is repeated until the set of fixed phrases becomes empty. If it is determined in step S207 that the "set of remaining fixed texts" is an empty set, the fixed text clustering unit 50 outputs the result of clustering the fixed texts. For example, two classes, that is, class 1 including example sentence 1, example sentence 2, and example sentence 4 and class 5 including example sentence 5 and example sentence 6, are output.

In the second embodiment of the present invention, the case where Japanese is the object of processing has been described, but the same applies when English, French, and other foreign languages are the objects of processing. It is possible to extract fixed phrases and perform clustering. The fixed-form sentence clustering method for them is obvious from the above description, and therefore its explanation is omitted.

As described above, the fixed phrase clustering method according to the second embodiment of the present invention has a step of detecting and merging similar fixed phrases, and therefore, among fixed phrases extracted from the corpus. It is possible to cluster similar similar phrases into one class and output as a fixed phrase class.

[0053]

As described above, according to the present invention, the fixed phrases that are similar to each other among the fixed phrases extracted from the corpus are clustered into one class and can be output as the fixed phrase class. A sentence clustering apparatus and method can be realized.

[Brief description of drawings]

FIG. 1 is a block configuration diagram of a fixed sentence clustering device according to a first exemplary embodiment of the present invention.

FIG. 2 is a flowchart showing a flow of processing in a boilerplate clustering method according to a second embodiment of the present invention.

FIG. 3 is an explanatory diagram of processing in a boilerplate clustering method according to the second embodiment of the present invention.

FIG. 4 is an explanatory diagram of processing in a boilerplate clustering method according to the second embodiment of the present invention.

FIG. 5 is an explanatory diagram of processing in a conventional fixed text extraction method.

[Explanation of symbols]

10 Morphological analysis means 20-word n-gram dictionary means 30 Fixed phrase extraction means 40 Fixed phrase storage means 50 Fixed sentence clustering means 100 Fixed sentence clustering device

Claims (5)

[Claims] 1. A morphological analysis means for dividing each sentence included in a corpus, which is a set of predetermined sentences, into words constituting each sentence, and obtained by dividing each sentence included in the corpus. A word for generating a word n-gram (n is a natural number of 1 or more) by connecting the words in accordance with a predetermined rule, and extracting a word n-gram satisfying a predetermined condition from the word n-gram and storing the word. The n-gram dictionary means and the word n-gram stored in the word n-gram dictionary means are combined to generate a set of word n-grams, and a ratio including the set of word n-grams is predetermined. A fixed sentence extraction unit that extracts a sentence having a threshold value or more as a fixed sentence from sentences included in the corpus, a fixed sentence extracted by the fixed sentence extraction unit and the word n-gram included in the fixed sentence. Pair Fixed phrase storage means to save,
A fixed sentence clustering unit that calculates a similarity between the classes with respect to a class that is a set of predetermined sentences included in the corpus and performs merging with the two classes having the highest similarity as one class. A fixed-form sentence clustering device characterized by being provided. 2. The predetermined condition satisfied by the word n-gram extracted by the word n-gram dictionary means is that the number of occurrences of the word n-gram in the extracted sentence is a predetermined number or more. The fixed-form sentence clustering device according to claim 1. 3. A step of dividing each sentence included in a corpus, which is a set of predetermined sentences, into words constituting each sentence, and the word obtained by dividing each sentence included in the corpus. To generate a word n-gram (n is a natural number of 1 or more.) Based on a predetermined rule, and extract a word n-gram satisfying a predetermined condition from the word n-gram and store it. A combination of the dictionary step and a predetermined type of word n-gram in the word n-gram stored by the word n-gram dictionary step is combined to generate a set of word n-grams, and a ratio including the set of word n-grams is a predetermined threshold value. The fixed sentence extraction step of extracting the above sentence as a fixed sentence from the sentences included in the corpus, the fixed sentence extracted by the fixed sentence extraction step and the word n group included in the fixed sentence. Storing a set of rams, and calculating a degree of similarity between the classes, which is a set of predetermined sentences included in the corpus, and regards two classes having the highest degree of similarity as one class. A fixed-form sentence clustering method comprising a step of merging. 4. The predetermined condition satisfied by the word n-gram extracted in the word n-gram dictionary step is that the number of occurrences of the word n-gram in the extracted sentence is a predetermined number or more. The fixed sentence clustering method according to claim 3, wherein: 5. A predetermined rule based on generating the word n-gram in the word n-gram dictionary step is n pieces (n is a natural number of 1 or more) while maintaining the order of arrangement of the words forming each sentence. 5. The fixed sentence clustering method according to claim 3 or 4, wherein the rule is a concatenation to generate a word n-gram.