JP2002269138A - Method for generating retrieval condition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To generate such a retrieval condition that does not take retrieval time so much by taking an index characteristic into consideration and by making a retrieval condition generating means adjust the length of a retrieval word included in the retrieval condition to be generated. SOLUTION: This retrieval condition generating method generates a retrieval condition that can be processed by a retrieval system from a retrieval request sentence in which a retrieval request is described in a natural language, and has a division step for dividing the retrieval request sentence into words, a candidate selection step for removing undesired words from word division results and selecting retrieval word candidates, an adjustment step for adjusting the length of a retrieval word candidate in accordance with the index format of an object to be retrieved by using the generated retrieval condition to generate a retrieval word, and a connection step for connecting a plurality of retrieved words with an operator when there are the plurality of selected retrieval words.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a search condition generating method for generating a search condition from a search request given in a natural language.

[0002]

2. Description of the Related Art Normally, a condition in which keywords are combined by a logical operator such as AND or OR is used as an input to a document search system. However, generating a search condition using such an operator is a process unfamiliar to general users, and it has been difficult to generate a search condition suitable for the user's intention. To solve this problem,
A method has been proposed in which a user describes a search request in a natural language, and a document search system or its support system generates search conditions suitable for the document search system.

As examples of a conventional search condition generation method, Japanese Patent Application Laid-Open Nos. Hei 10-334115 and Hei 11-4
No. 5249 is known. In each of these methods, the input search request sentence is divided into words, unnecessary words are removed from the divided words, and the remaining words are combined with appropriate operators as necessary. Is common.

When implementing a document search, an index is generally used to speed up the search. When words are not explicitly indicated by spaces or the like as in Japanese, n-gram (n-character set) is used as an index unit from the search target document.
-Gram indexes are often used. n- for n = 2
In the gram index, a two-character set “rabbit”, “raising”, and “raising” is extracted from the document “raising a rabbit” (referred to as document 1), and the appearance of those characters in document 1 is recorded in an index file. Record it. At the time of search, n-gr
Am is extracted, and a document including the specified search word is determined using the appearance information of the n-gram. For example, if the search word is "Kai", the document including "Kai" can be easily determined because the search word itself is a two-character set. However, when the search word is shorter than the index unit n-gram such as “rabbit”, a document including any of n-grams including “rabbit” must be determined, and the search takes time ( Reference: Ogawa et al., "Efficient Document Retrieval Method Using n-gram Index", IEICE Transactions, Vol.
DI, No. 1 pp. 121-129,199
9).

[0005]

However, in the above-mentioned prior art, since the length of a word is not taken into consideration, there is a problem that it takes a long time to perform a search when the generated search condition is used. .

According to the present invention, the search condition generation means adjusts the length of the search word included in the search condition to be generated in consideration of the characteristics of the index. it can.

[0007]

According to the first aspect of the present invention,
A search condition generating method for generating a search condition that can be processed by a search system from a search request sentence in which a search request is described in a natural language, comprising a dividing step of dividing the search request sentence into words;
A candidate selecting step of selecting a search word candidate by removing unnecessary words from the word segmentation result, generating a search word by adjusting the length of the search word candidate according to the index format of the search target using the generated search condition A search condition generation method characterized by comprising an adjusting step and a combining step of combining selected search words with an operator when a plurality of search words are selected.

According to a second aspect of the present invention, in the search condition generating method according to the first aspect, the index format of the search target is ng
In the case of an n-gram index using ram (a set of n consecutive characters; n is 2 or more) as an index unit,
In the adjusting step, when the search word candidate having less than n characters is adjacent to another search word candidate in the search request sentence, the search word candidates having less than n characters are excluded, and only the search word candidates having n or more characters are searched. It is characterized by the following.

According to a third aspect of the present invention, in the search condition generating method according to the first aspect, the index format of the search target is ng
If the index is a ram index (where n is 2 or more), the adjusting step generates a word connecting the two search word candidates if the search word candidate is shorter than another search word candidate in the search request sentence, It is characterized in that the generated word is used as a search word.

According to a fourth aspect of the present invention, in the search condition generating method according to the first aspect, the index format of the search target is ng
If the index is a ram index (where n is 2 or more), the adjusting step is performed when the search word candidate having less than n characters and another search word candidate (referred to as a connection target candidate) are adjacent to each other in the search request sentence. A connected word is generated, and the generated word and a candidate for connection are set as search words.

According to a fifth aspect of the present invention, in the search condition generating method according to the fourth aspect, the combining step adjusts the importance of the generated word and the candidate for the connection target and combines them. .

According to a sixth aspect of the present invention, in the search condition generating method according to the third or fourth aspect, when the part of speech of a search word candidate having less than n characters is a prefix, the concatenation target candidate is prefixed in the search request sentence. It is characterized in that concatenation is performed only when it is after the letter.

According to a seventh aspect of the present invention, in the search condition generating method according to the third or fourth aspect, when the part of speech of a search word candidate having less than n characters is a suffix, the candidate to be connected is suffixed in the search request sentence. It is characterized in that concatenation is performed only when it is before a letter.

[0014]

FIG. 1 shows an embodiment of the present invention. In FIG. 1, a dividing step is a step of dividing a search request sentence into words, a candidate selecting step is a step of removing unnecessary words from a word division result and selecting a search word candidate, and an adjusting step is according to an index format used in the search. The step of adjusting the length of the search word candidates to generate a search word, and the step of combining, when there are a plurality of selected search words, combining them by an operator.

For example, assume that the search request sentence is "drug having antimalarial activity". In this case, first, words are divided into “anti”, “malaria”, “active”, “no”, and “drug” in the dividing step. As a method of dividing words into words, a morphological analysis, which is a basic method of Japanese analysis, may be employed.

In the next candidate selection step, a word suitable for the search is selected from the division results as a search word candidate. Generally, nouns are not selected because they are not suitable for retrieval. In this case, the particle “no” is deleted, and the other “anti”, “malaria”, “activity”, and “drug” are left as search word candidates.

The next adjustment step is a step of adjusting the search word candidates according to the index format. As described above, if the index format is an n-gram index, it takes a long time to search for a word of less than n characters.
If the index is m, it is better not to use “anti” as it is. Therefore, in the present embodiment, search word candidates with less than n characters are deleted, and the other search word candidates “malaria”, “activity”, and “drug” are output as search words. In the adjustment step, if a search word candidate having less than n characters appears alone, it is not deleted. This is to avoid that “bone”, which is a search word candidate, falls out of the search condition when a search request sentence such as “formation of bone” is given.

In the final combining step, if there are a plurality of search words, they are combined with an appropriate operator. Search operators include an AND operator and an OR operator.
If the OR operator (represented by '+') is used, the finally generated search condition is “malaria + activity +
Drug ”.

In the above-described method of deleting a search word candidate having less than n characters, the search word candidate is deleted, so that the search intention may not be correctly indicated in the search condition. In fact, the meaning differs from the presence or absence of the deleted "anti" in the above example, which could lead to a search for a document describing, for example, a drug that increases malaria activity.

In order to solve this problem, search word candidates having less than n characters (there are adjacent search word candidates) need not be deleted, but may be linked to adjacent search word candidates. At this time, the other connected search word candidate is deleted. In the above example, the search word "antimalarial" is obtained by connecting "anti" and "malaria" (note that "malaria" and "activity" are not less than n characters and are not connected). The search condition finally generated is “antimalarial + activity + drug”.

In practice, the second method is not sufficient. For example, it is conceivable that the expression “antimalarial” is used instead of “antimalarial” in the document, but such a document cannot be searched by the second method.

In order to solve this problem, if there is an expression that is the same as the search request sentence, the document is specified by using the same expression, and if there is an expression that is almost similar but not completely the same, the document is searched. What should I do? Therefore,
In the third method, as for a search word candidate having another search word candidate with less than n adjacent characters, the search word candidate connected with the adjacent search word candidate is searched and the other connected search word candidate is deleted at the same time. And select it as a search word. In the above example, “antimalaria”, which is a combination of “anti” and “malaria”, and “malaria”, which is a candidate for connection, are set as search words. In this case, the final search condition is “antimalarial + malaria + activity + drug”.

The third method is effective in a ranking search in which the similarity (or the degree of coincidence / fitness) of a search document with a search condition is calculated, and the documents are ordered based on the calculated value. It is meaningless to add the latter to the search word, because in the case of an exact match search without ranking, a document containing "antimalarial" always includes "malaria". on the other hand,
In the case of ranking search, similarity is calculated based on the appearance of each search word, so "antimalarial"
This is because the similarity differs when only one of the "malarias" is used and when both are used, resulting in a difference in the ranking result. For example, the similarity is calculated according to the following policy.

(1) The importance of a search word in a document is the number of appearances of the search word in the document.

(2) The similarity of a document to a search condition is the sum of the importance of the search word.

In this case, the similarity of the document “New antimalarial drug was released” to “antimalarial + malaria + activity + drug” indicates that the importance of “antimalarial” appears once in this document. Therefore, 1 and “malaria” are also 1, and the sum of the two is 2. However, since the appearance of "anti-malaria" and the appearance of "malaria" overlap, if the two are summed as they are, the similarity will unduly increase, which is a problem.

To solve this problem, in the ranking search, the importance of the search word may be adjusted using an appropriate operator. For example, as an operator for adjusting the importance of a search word to a value multiplied by a coefficient, an adjust operator (AD
JUST (0.5, malaria) is expressed, and the original importance of malaria multiplied by 0.5 is used as the importance.) A search generated by the third method Apply to words. In this case, the final search condition is "ADJ
EST (0.5, antimalarial) + ADJUST (0.
5, malaria) + activity + drug.

The second or third method has problems. For example, a search request sentence "strong antimalarial drug" is divided into "strong", "anti", "malaria", and "drug".
In the adjusting step, since “anti” is linked to other adjacent search word candidates, “strong anti” and “antimalarial” are generated. At this time, "antimalaria" is a meaningful word, but "strong anti-malaria" is a meaningless word,
It is better not to include it in the search condition.

In order to solve this problem, the part of speech may be used. Since the prefix has the role of modifying the following word, the prefix is connected only to the following search word candidates. In the above example, "anti" is a prefix, so this rule applies. Therefore, "strong anti-malaria" is not generated and only "anti-malaria" is generated.

On the other hand, since the suffix has the role of modifying the preceding word, the suffix is connected only to the preceding search word candidate. For example, a search request sentence “Genetic Algorithm” is divided into “Genetic”, “Target” and “Algorithm”, and “Target” is a suffix. Therefore, "genetic" is generated, but "target algorithm" is not generated.

[0031]

According to the search condition generating method of the first aspect, the search condition generating means adjusts the length of the search word included in the generated search condition in consideration of the index characteristics.
It is possible to generate a search condition that does not require a search time.

[0032] In the search condition generating method of claim 2,
By using the n-gram index to exclude search word candidates with less than n characters, it is possible to generate a search condition that does not require a search time.

In the search condition generating method according to the third aspect,
By linking a search word candidate having less than n characters with another adjacent search word candidate, it is possible to generate a search condition that meets the user's intention.

[0034] In the search condition generating method of claim 4,
By linking a search word candidate having less than n characters with another adjacent search word candidate and also performing a search word on the search word candidate itself to be connected, a search condition more suited to the user's intention can be generated. .

In the search condition generating method according to claim 5,
By adjusting the importance of the search word determined by the method of claim 4, it is possible to generate a search condition more suited to the user's intention.

In the search condition generating method according to claims 6 and 7, it is possible to generate a search condition more suited to the user's intention by using a part of speech of a search word candidate having less than n characters.

[Brief description of the drawings]

FIG. 1 shows an embodiment of the present invention, and is a flowchart of search condition generation.

Claims (7)

[Claims] 1. A search condition generating method for generating a search condition that can be processed by a search system from a search request sentence in which a search request is described in a natural language, comprising: a dividing step of dividing the search request sentence into words; A candidate selecting step of selecting a search word candidate by removing unnecessary words from, an adjusting step of adjusting the length of the search word candidate according to the index format of the search to be performed using the generated search condition, and generating a search word; A search condition generation method, comprising: when there are a plurality of selected search words, combining them with an operator. 2. The search target index format is n-gram (n
In the case of an n-gram index in which a set of consecutive characters; n is 2 or more) as an index unit, the adjustment step is performed when the search word candidate having less than n characters is included in another search word candidate and a search request sentence. 2. The search condition generation method according to claim 1, wherein when adjacent, a search word candidate having less than n characters is excluded, and only search word candidates having n characters or more are set as search words. 3. When the index format of the search target is an n-gram index (where n is 2 or more), the adjusting step is such that a search word candidate having less than n characters is adjacent to another search word candidate in the search request sentence. 2. The search condition generating method according to claim 1, wherein when the search condition is present, a word is generated by connecting the two, and the generated word is used as a search word. 4. If the index format of the search target is an n-gram index (where n is 2 or more), the adjusting step is performed when a search word candidate having less than n characters and another search word candidate (referred to as a connection target candidate) are used. 2. The search condition generation method according to claim 1, wherein when the search request sentences are adjacent to each other, a word is generated by connecting the two, and the generated word and the connection target candidate are used as search words. 5. The search condition generating method according to claim 4, wherein the combining step adjusts the importance of the generated word and the candidate to be connected and combines them. 6. The method according to claim 3, wherein when the part of speech of the search word candidate having less than n characters is a prefix, the connection is performed only when the connection target candidate is located after the prefix in the search request sentence. Or the search condition generation method according to 4. 7. The method according to claim 3, wherein when the part of speech of the search word candidate having less than n characters is a suffix, the connection is performed only when the connection target candidate precedes the suffix in the search request sentence. Or the search condition generation method according to 4.