JP3495912B2 - Search device with learning function - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a search device with a learning function that is installed in computer applications, office automation equipment, home electric appliances, and the like. It is a retrieval device with a learning function that makes it easy to retrieve the sentence requested by the user by storing it.

[0002]

2. Description of the Related Art In a conventional search, a user inputs at least one keyword, and the degree of similarity between the user's request and the document is determined by the number of keywords contained in the document, the type of the keyword, etc. Is evaluated and the document that meets the user's request is output as the correct answer.

However, the conventional search has two problems. The first point is that if the keyword entered by the user and the keyword contained in the document do not exactly match, it is not judged as "related", that is, not the same word, but similar words are ignored. The required document may not be retrieved depending on the content of the keyword input by the user. The second point is that if the accuracy of the search result is poor, there is no method for improving the result of the next search using the information such as the document selected by the user. However, the same result is returned.

As a method for making it possible to retrieve a document containing similar words, the context proposed in IEICE Technical Report AI 92-99 (1993) "Associative retrieval from a large-scale database" is published. There is a method of using a vector (which is called a feature vector in the present invention). With this method, it is possible to evaluate the semantic similarity between words and sentences even if the characters do not match. An example of using this feature vector for retrieval is, for example, Japanese Patent Laid-Open No. 19538/1994.
There is publication No. 8.

Further, as an invention aiming at improving the search efficiency, the result of the search is recorded and used in a later search, as disclosed in JP-A-7-
There is 182353 publication. The invention described in the publication discloses a technique for performing learning by recording a search result and performing a search by referring to a result recorded in a later search. That is, a search key composed of a plurality of keywords is input, and a document to be output is determined based on the keyword search and the positional relationship between the keywords. If the search key is not registered in the learning unit, the search key and the number of the output document are recorded in the learning unit. If the same search key is input in a later search, the number recorded in the learning unit is recorded. Output a document with.

[0006]

[Problems to be Solved by the Invention] Japanese Patent Application Laid-Open No. 7-18235
In the publication No. 3, the learning function is realized by recording the search key (keyword string) input in the past and the ID of the searched document, and the ID is recorded when the same search key is input later. Output the document as a correct answer document. However, since the learning content automatically records the ID of the retrieved document, it is not determined whether or not the retrieved document is the correct answer, and the learning does not improve the retrieval accuracy.

Since the search method is keyword search,
The learning effect does not appear unless you enter the same search key as the learned search key.

Further, in the publication, the learned content is first referred to at the time of search, and when the input keyword and the learned keyword completely match, the learned content is output. Therefore, even if there is more relevant content with respect to the input keyword such as addition to the database, it will not be searched.

In order to solve such a problem, the present invention provides:
We will improve the accuracy of subsequent searches by using the search method based on the feature vector and by learning the search results. By using the feature vector, it is possible to search for documents that include words that are semantically similar. Further, when the search result is not good, the search request sentence and the correct answer selected by the user are recorded, and the search accuracy is improved by referring to the recorded contents at the time of the subsequent search.

Further, in the present invention, the learned keyword is not prioritized, both the database and the learned contents are targeted for retrieval, and both contents are evaluated and the retrieval result is output. Therefore, more accurate evaluation can be performed.

[0011]

According to the present invention, a search request sentence when a search fails in order to solve the above problems,
A set of function names that the user considers to be correct is learned. In the subsequent search, the learning content is used for the search so that the function name that the user considers to be the correct answer is searched when the search request sentence similar to the search request sentence that failed in the past is input. To do.

According to the invention of claim 1, an input means for inputting a search request sentence, an output means for outputting a search result, a question history for recording the search request sentence and the ID information of the function name which the user considers to be the correct answer. Storage means, function name indicating function name, ID information indicating the function name, assumed request sentence as an example of search request sentence inquiring about a desired function, and description of function name to be displayed in the search result When the output rank of the function database that holds the sentence in correspondence with the function name that the user regarded as the correct answer in the search result is low, the search request sentence and the ID information of the function name that the user regarded as the correct answer A learning device for writing the question history into the question history storage means for learning.
Search means for searching a function name from a search request sentence input by referring to both the question history storage means and the function database.

In the invention of claim 1, the user inputs the search request sentence through the input means, the similarity between the search request sentence and the assumed request sentence of each item is evaluated by the search function, and the search result is output. . The user selects a function name that he considers fit. At this time, if the rank of the function name selected by the user from the search results is lower than the predetermined rank, the search request sentence and the information regarding the function selected by the user are recorded in the question history storage means, and learning is performed. To do. In the subsequent search, both the learning content and the function database recorded in the question history storage means are searched.

Here, it is preferable that the search request sentence input by the user at the time of learning is divided and recorded in a dictionary for learning because this is because the search request sentence input by the user is more important than the expected request sentence. This is because it is necessary to separately handle the search request sentence as a search target, and the learned search request sentence and the assumed request sentence are not semantically related to each other so as to prevent an adverse effect when the function name vector is created.

According to a second aspect of the present invention, the question history storage means erases learning contents that have become unnecessary due to a change in the learning state of the user.

According to the second aspect of the present invention, the learning content recorded in the question history storage means is put in an appropriate state. While using the search function, the user learns the function and the content of the explanation.
When the content of the explanatory note recorded in the function database is learned, the user will not issue a search request sentence regarding the learned explanatory note content. Therefore, the search request sentence recorded in the question history storage means need not be related to the contents learned by the user. Therefore, by discarding the learning content that has passed a time after being referred to, the learning content of the question history storage means can be matched with the learning state of the user.

According to a third aspect of the present invention, when the search request text recorded in the question history storage means is most similar to the input search request text as a result of the search by the search means, It is characterized in that the function database is referred to from the ID information of the learned function name, and the search is performed again based on the assumed request sentence corresponding to the ID information.

That is, this is a function for when it is determined that the search request text recorded in the question history recording means is most similar to the input search request text. In this case, the assumed request sentence is referenced from the function database using the learned ID information, the search function is searched again, and the function name having contents similar to the input search request sentence is searched. It is a feedback function.

According to the third aspect of the present invention, the function is executed when the retrieval request sentence recorded in the question history storage means is output to the first place in the retrieval result. In the question history storage means, a search request sentence in which the correct answer is output and whose rank is low is recorded. The search request sentence recorded in the question history storage means has a low degree of similarity with the assumed request sentence in the function database. Therefore, when the similarity between the search request text input by the user and the search request text learned in the question history storage means is high, the assumed request text recorded in the function database is not the search request text input by the user. The degree of similarity is low.

Therefore, the assumed request sentence is read by referring to the ID of the function name learned in combination with the search request sentence, the assumed request sentence vector based on the assumed request sentence is used as the search request sentence vector, and the association is reassociated. Perform a search. In the present invention, such a search method is called feedback search. As a result, it is possible to output an explanatory text closer to the content of the search request text.

According to a fourth aspect of the present invention, the search means performs the search based on the degree of coincidence between the keyword of the search request sentence and the keyword of the assumed request sentence.

In the fourth aspect of the present invention, first, an important keyword is extracted from the search request sentence input to the search function. Next, the expected request sentence in the function database is referenced, and the keyword is similarly extracted from the expected request sentence. With respect to the keywords extracted from the supposed request sentence, each function name is used as a unit of the keyword set, the degree of coincidence between both keyword sets is calculated, and the function names are output in descending order of the degree of coincidence as a search result. That is, the keyword set is extracted from the search request sentence, and the keyword set is extracted for each item also from the assumed request sentence of the item of each function name. Then, the degree of coincidence for each keyword set is calculated, and the function name of each item is output as a search result in the order of degree of coincidence.

According to a fifth aspect of the present invention, the search means performs the search based on an inner product result of the feature vector of the search request sentence and the feature vector of the assumed request sentence.

In the vector generation function, a vector representing the meaning of a sentence including a plurality of words is created by referring to a word vector dictionary in which words and feature vectors of words are recorded, and a search request sentence vector and a function name vector for each function are created. To create. By calculating the dot product of the vectors, the similarity between the request sentence vector and each function name vector is evaluated, and the function names corresponding to each function name vector are output in the order of similarity.

The feature vector used here is
IEICE Technical Report AI92-99 published by The Institute of Electronics, Information and Communication Engineers
(1993) "Associative search from large-scale database"
It is the context vector proposed in. That is, the "feature vector" in the present invention directly corresponds to the above "context vector". Japanese Patent Laid-Open No. 6-195388 discloses a search using this feature vector.

The feature vector indicates the degree of the relationship between the concept and the context of the words in the sentence, and is a vector expression of the degree of the semantic connection with many feature words. When the N concept classifications are characteristic words, the value of each element of the N-dimensional vector is associated with each characteristic word. Feature vector Xi = (xi
, Xi2, ..., xiN) is 0 ≦ xij
≦ Em. Em is a positive constant. When there is no relationship between the word i and the characteristic word j, xij = 0, and when there is a relationship, a large value is taken according to the degree of the relationship. For example, if the feature vector consists of five feature words (nature, city, noise, animal, green),
When the value of each element is a binary value of 0 or 1, the feature vector of the word “mountain” can be expressed as (1, 0, 0, 1, 1).

The invention of claim 5 is a method for evaluating the degree of similarity between a search request sentence and an assumed request sentence using a feature vector of a word. The request sentence vector and the assumed request sentence vector for each item are created by the vector generation function, and the similarity between the vectors is evaluated by the associative search function. The function name of each item is output as a search result in the order of similarity with the request sentence vector.

[0028]

BEST MODE FOR CARRYING OUT THE INVENTION The present embodiment will be described below by taking various functions of a word processor as an example. FIG. 1 is a schematic diagram of the apparatus of the present invention. This search device with a learning function has an input means 1 such as a keyboard for inputting a search request and a key input operation for selecting a function name from the search result, an output means 2 such as a display for displaying the search result, and a learning function. The control means 3 for controlling the attached search device, the function name 8 to be searched as illustrated in FIG. 3, the assumed request sentence, the function database 8 in which the explanation sentence is recorded, and the similarity between the search request sentence and the expected request sentence are evaluated. , A search function 7 for calculating the semantic similarity to the search request sentence for each function name.

The control means 3 records the search request sentence and the ID of the function name to learn the past search result when the rank of the function name judged by the user to be the correct answer is lower than the predetermined rank. A question history storage that allows you to refer to the ID recorded when a search request sentence similar to the search request sentence learned in the subsequent search is input and to display the explanation text of the correct answer function selected in the past The means 4, as a result of the search, when it is determined that the search request text recorded in the question history storage means 4 is similar to the input search request text, the assumed request text of the function name indicated by the ID is displayed. Data to be searched by the search function 7 by reading the learning content from the learning device 5 and the question history storage means 4 which perform a search again by referring to the function name vector as a search request sentence vector and integrating it with the function database 8. And a database creation function 6 to create an over nest.

FIG. 2 shows the data structure of the word vector dictionary 13. The word vector dictionary 13 records a word and a vector indicating the meaning and content of the word, and is referred to at the time of search as a part of the search function 7 (see FIG. 9). Each element of the vector indicates a relationship with a predetermined concept, which is 1 if the concept is related and 0 if not.

FIG. 3 shows the data structure of the function database 8. The function database 8 is a search target, and its data structure includes a function name (centering, shading, etc.) output as a search result, an ID of the function name, and an assumed request statement that is a request statement in which the function name is correct, It is composed of a description displayed when a function name is selected in the search result.

FIG. 4 shows the data structure of the question history storage means 4. In the present invention, the question history storage means 4 is used as a learning database in addition to the function database 8. Writing and erasing the learning content is performed in the question history storage means 4,
The function database 8 is not modified. When searching, separate databases are searched simultaneously. In the output result, the set of the search request sentence and the function name ID is recorded only when the rank of the function name regarded as the correct answer by the user is lower than the predetermined rank. At the time of search, it is handled in the same manner as the function database 8 and the recorded search request sentence is treated as an assumed request sentence.

FIG. 5 is a diagram showing an outline of a method of managing the learning contents in the question history storage means 4. The learning contents recorded in the question history storage unit 4 are managed in the learning order (time order). During the search, if the learning content is output in the search result as being similar to the search request sentence, the learning content is rewritten as newly learned. The number of learning contents that can be recorded in the question history storage unit 4 at the time of search is determined in advance, and when the number of learning contents exceeds the determined number, old contents are deleted.

When the search request sentence relating to the function learned by the user is no longer input, the learning content with the function as the correct answer is treated as old in the question history storage means 4. Therefore, it is not necessary to write old learning contents in the question history storage unit 4.

The operation performed by the question history storage means 4 is to write new learning contents, delete old learning contents, and rewrite when learning contents that have already been recorded appear in the higher ranks of the search results.

FIG. 6 is a flowchart showing a method of deleting learning contents. When the learning condition is determined by the learning device 5 in step S1 and the learning condition is satisfied,
In step S2, the search request sentence and the function name ID are passed to the question history storage unit 4.

At the stage when the storage capacity of the question history storage means 4 has a sufficient capacity, the learned contents are continuously recorded, and the process proceeds to step S4 where the learned contents are not deleted. When the storage capacity is exhausted, the process proceeds to step S3, and when learning is performed, old learning contents are deleted and the process proceeds to step S4.

FIG. 7 is a flow chart showing a rewriting method when the learning contents in the question history storage means 4 appear in the higher order of the retrieval results. It is considered that the learning content appears higher in the search results because the learning content is still important to the user, and is treated as newly learned.

In step S11, when the search request sentence of the question history storage means 4 appears higher in the search result, the same content is recorded in the question history storage means 4 as the newly learned content. If there is not enough capacity in the question history storage unit 4 in step S12, the process proceeds to step S13 as in the case of FIG. 6 and old learning contents are deleted.

Detect whether or not the learning condition is actually satisfied. The learning condition is that when the user selects a function name from the search result, if the rank of the selected function name is lower than a predetermined rank (if it is determined to be 10th, it is displayed below 11th rank). When the selected function name is selected) learning is performed. If this learning condition is satisfied, the ID of the function name selected from the search request sentence and the search result input by the user is written in the question history storage unit 4.

FIG. 8 is a flow chart of the feedback function. First, in the search result, it is checked whether or not the search request sentence recorded in the question history storage unit 4 comes first in the search result (step S21). If it is in the first place, the question history storage means 4 is referred to and the ID corresponding to the search request sentence output in the first place is taken out (step S).
22).

Then, the assumed request sentence having the function name indicated by the ID is passed to the search function 7 as a search request. The search function 7 performs a search as in the conventional case and outputs the search result. When the learning content recorded in the question history storage unit 4 is the first place in the search result before feedback, the semantic similarity between the input search request sentence and the learned search request sentence in the past. Is considered to be large.

In this case, since the learned past search request sentence has a low degree of semantic similarity to the assumed request sentence in the function database 8, the input search request sentence and the function name output after the second place The relationship between and is considered to be low. In such a case, by referring to the ID of the function name output to the first place by the above and performing a search again using the assumed request sentence, a function closer to the user's request is retrieved from the function database 8. be able to.

When a search request sentence is input by the user and a search is performed, the question history storage means 4 is referred to. A search request is extracted from the learning content recorded in the question history storage means 4 and treated as an assumed request sentence. The ID of the function name is used as it is. Then, the contents read from the function database 8 and the question history storage means 4 are integrated so that they can be handled as a search target of the associative search.

The method of using the keyword search in the search function 7 is described. The similarity between the word set extracted from the search request sentence and the word set extracted from each assumed request sentence is compared and output to the search result in order of similarity.

FIG. 9 is a diagram when the associative search is used for the search function 7. In the associative search, a word vector dictionary 13 in which a feature vector representing a word and a feature of the word is recorded, a vector generation function 12 that generates a vector representing the meaning of an input sentence by referring to the word vector dictionary 13, The associative search function 11 searches a sentence highly relevant to the search request sentence by taking the inner product and evaluating the semantic distance between the vectors.

When performing an associative search, a vector is created for each search request sentence and each function, and the associative search function 11 evaluates the semantic relevance between the search request sentence vector and the function name vector.

The search function 7 for searching the function name recorded in the function database 8 using the word vector dictionary 13 described in the present invention will be described below. As an example, a case where the search request sentence “I want to move the text to the center” is input will be described.

First, the word vector dictionary 13 is referred to, and the feature vector of the word included in the search request sentence is extracted. The criteria for extracting words is by longest match. Word vector dictionary 1
Most of the words recorded in 3 are nouns and verbs, so
In the case of an example sentence, the extracted word is "sentence, middle, move". Here, the concept indicated by the feature vector is (sentence, target, position, movement, insertion, enlargement), and the feature vectors V (sentence), V (middle), and V (movement) of the word extracted from the search request sentence are respectively V (Sentence) = (1, 1, 1, 1, 1, 0) V (middle) = (0, 0, 1, 1, 0, 0) V (movement) = (0, 1, 1, 1, 1 , 0). A feature vector of the search request sentence is created by adding all the vectors of the words included in the search request sentence.

In the case of this example, the sum V of the three feature vectors V
r becomes Vr = (1, 2, 3, 3, 2, 0). Finally, the search request sentence vector Vrn is obtained by normalizing the vector size to 10. Vrn in the case of the vector shown in the example is Vrn = (1.9, 3.8, 5.8, 5.8, 3.8, 0)
Becomes

Next, referring to the function database 8, a function name vector to be searched is created. First, as in the case of the search request sentence, an assumed request sentence vector is created for each assumed request sentence. As shown in FIG. 3, in the function database 8, a plurality of assumed request sentences correspond to one function name. Then, the vector of the assumed request sentence corresponding to each function name is added to create the function name vector of each function name defined in the function database 8.

Then, the semantic similarity between the search request sentence and the function name is evaluated by the search request sentence vector and the function name vector. As shown above, feature vectors are words and sentences,
Since there are many elements having values other than 0 between the two feature vectors, the word indicated by the two feature vectors indicates the semantic relationship with a predetermined concept.
It can be said that the degree of semantic similarity between sentences is high.

Therefore, the inner product of the two feature vectors is calculated, and the semantic similarity between the search request sentence and the function name is evaluated by the value of the inner product. When displaying the search result to the user, the function names are arranged in order of this value and displayed as the function related to the search request sentence.

[0054]

According to the present invention, a learning function is implemented to record a search result having a bad content, and if a search request sentence similar to the search request at that time is input later, the user can answer correctly during learning. The function name that was judged to be output.

Further, according to the present invention, the learning content in the question history storage means is put in an appropriate state. That is, although the result when the content is bad is recorded in the question history storage means, if the user learns about the content shown in the description, there is no need to search for the description, The content becomes less effective over time. Therefore, it is necessary to have a function of deleting unnecessary learning contents according to the learning state of the user. Unnecessary learning contents are those that do not appear in the top of the search results as time elapses after learning. Therefore, it is possible to keep only the important learning contents by managing the learning contents in chronological order and deleting old ones.

Further, according to the present invention, the present invention relates to a feedback function when the search request text in the question history storage means is the most similar to the input search request text. That is, when the search request sentence in the question history storage means is output to the first place in the search results, the ID of the corresponding function name
A more appropriate function name can be output by referring to the function database based on the information and creating a search request sentence from the content of the item indicated by the ID information.

[Brief description of drawings]

FIG. 1 is a schematic diagram of a search device with a learning function of the present invention.

FIG. 2 is a diagram showing an internal data structure of a word vector dictionary.

FIG. 3 is a diagram showing an internal data structure of a function database.

FIG. 4 is a diagram showing an internal data structure of a question history storage unit.

FIG. 5 is a schematic diagram showing a learning content management method in the question history storage means.

FIG. 6 is a flowchart of data management during learning in the question history storage means.

FIG. 7 is a flow chart of data management in the question history storage means when the learning content is ranked higher in the results.

FIG. 8 is a flowchart of a feedback function.

FIG. 9 is a diagram showing a configuration of a search function.

[Explanation of symbols]

1 Input means 2 output means 3 control means 4 Question history storage means 5 learning devices 6 Database creation function (database creation means) 7 Search function (search method) 8 function database

─────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-4-54564 (JP, A) Tomohiro Aikawa et al., Study on group information sharing support based on FISH usage history analysis, IPSJ research report, Japan , Information Processing Society of Japan, March 14, 1996, Volume 96, No. 26, pp. 31 to 36 (96-GW-16-6) (58) Fields investigated (Int.Cl. ⁷ , DB name) G06F 17/30 JISST file (JOIS)

Claims (5)

(57) [Claims] 1. An input unit for inputting a search request sentence, an output unit for outputting a search result, a question history storage unit for recording the search request sentence and ID information of a function name that the user regarded as a correct answer, and a function. The function name indicating the name of the function, the ID information indicating the function name, the assumed request sentence as an example of the search request sentence inquiring about the desired function, and the explanation sentence of the function name to be displayed in the search result are associated with each other. And a function database held by the user, and when the output rank of the function name that the user regarded as the correct answer in the search result is low, the search history and the ID information of the function name that the user regarded as the correct answer are stored in the question history storage means. A learning device for performing learning by writing to the question history storage device; and a search device for searching for a function name from a search request sentence input with reference to both the question history storage device and the function database. That learning function with a search device. 2. The retrieval device with a learning function according to claim 1, wherein the question history storage means erases learning contents that are no longer needed due to a change in the learning state of the user. 3. The function learned when the search request text recorded in the question history storage means and the input search request text are the most similar as a result of the search by the search means. 2. The search device with a learning function according to claim 1, wherein the function database is referenced from the ID information of the name, and the search is performed again based on the assumed request sentence corresponding to the ID information. 4. The search device with a learning function according to claim 1, wherein the search means performs the search based on a degree of coincidence between the keyword of the search request sentence and the keyword of the assumed request sentence. 5. The search device with a learning function according to claim 1, wherein the search means performs the search based on an inner product result of the feature vector of the search request sentence and the feature vector of the assumed request sentence.