JP3024045B2 - Data retrieval device based on natural language - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a data retrieval apparatus based on a natural language, and in particular, retrieves text data such as an electronic book and an electronic dictionary, and an image database or a composite multimedia database based on a natural language. Related to the device.

[0002]

2. Description of the Related Art Conventionally, as a data search device using a natural language, there are a device using a full text search, a device using a keyword search, and a device performing a search based on the degree of coincidence of vectors.

[0003] Japanese Patent Application Laid-Open No. 4-340677 discloses a similar pattern search device in which a learning function is incorporated in an image search device for performing image search based on an input image.
This apparatus searches for a similar pattern image from an input pattern image based on the physical characteristics of the pattern. By configuring a means for calculating the degree of similarity from the physical characteristics by a neural network, the neural network It is possible to perform learning so as to reduce the difference between the similarity output from the above and the expected value of the given similarity.

As a natural language image search method, there is a keyword search method for adding a keyword to an image.
This is to add a plurality of keywords representing the features to an image and search for the image based on the degree of coincidence between the keyword given as an input sentence and the keyword added to the image.

[0005] An information database apparatus based on a keyword search having a learning function is disclosed in Japanese Patent Application Laid-Open No. 5-233707. This device determines the class to which the input sentence or keyword belongs,
Based on the learning result from the class, association with another keyword is performed, and document classification and keyword association support are possible. As a method of determining the class to which a sentence belongs,
The keyword is extracted from the input text, and the class to which the text belongs is determined from the feature vector converted based on the keyword. Since a neural network is used as a means for determining a class, learning adapted to a class classification given by a user is also possible. It should be noted that the feature vector used in this device is simply based on the frequency of appearance of the keyword, and can be regarded as a type of keyword search as a search method.

A document search method and a document search system having a learning function and performing a full text search are disclosed in Japanese Patent Laid-Open No. 5-3.
No. 42255. This method and system are used to determine the degree of matching between an input sentence and a search target sentence, to analyze the syntax of the input sentence, calculate the degree of connection between words in the input sentence, and connect the words between words in the search target sentence. The calculation of the degree, the calculation of the degree of connection with the word in the synonym dictionary, and the calculation of the degree of connection between the input sentence and the word of the search target sentence are sequentially executed. Since the means for calculating the degree of connection of each process is constituted by a neural network, it is possible to learn so that the difference between the degree of fitness obtained as an output and the degree of fitness given by the user is small.

[0007] AI92 issued by IEICE.
-99 (1993-1), "Associative search from large-scale document database", a document search method using a context vector is proposed. "Context vector" used here
Is a representation of the meaning of a word in a distributed manner, and associates each word with 100 or more characteristic words. The context vector of a document or a question sentence is obtained by normalizing the context vector of each important word extracted from the text.
Since the search is performed by calculating the inner product of the context vector of the question sentence and the context vector of each document, it is possible to search the ranking of the search results and search for similar sentence examples.

[0008]

When searching for image data, the above-described similar pattern search apparatus searches for similarities in the pattern represented as an image, and therefore cannot search for the displayed contents.

[0009] In the case of performing data search using an added keyword, it is assumed that an information database apparatus based on the above-described keyword search can search only by a word added to data as a keyword, and that a user inputs. It is practically impossible to add in advance all words including impression words and association words to images as keywords. Therefore, only a limited input sentence added as a keyword can be searched. Similarly, in the above-described document search method and document search system for performing a full-text search, search can be performed only by words or sentences appearing in a search target document.

As described above, any search method using a keyword or a pattern can search only in the range of a natural language explicitly expressed in search target data.
It is impossible to search according to human knowledge and sensitivity.

According to the search method based on the context vector, a similar sentence can be searched, but high-speed hardware is required, the cost of the apparatus is increased, and the scale becomes large.

Further, even when a neural network is used to provide a learning function to the device, high-speed and large-scale hardware is required, which increases the cost of the device and increases the scale. .

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and is characterized in that a data vector representing a feature of data and a data vector representing a feature of a question sentence are characterized according to human knowledge and sensitivity. To provide a data search device in a natural language that can be created based on a word vector that has been input and that can search for image data other than text data from input words or sentences according to human knowledge and sensitivity. Aim.

[0014]

According to Means for Solving the Problems] The present invention, the above objective, an input means for inputting a search target data as quality Toibun, a word dictionary for storing word vector representing a feature of a word, the wherein the vector generation means for creating a data vector based on the word vectors representing the features of the question, a database for storing data vector representing the feature of the search target data and the search target data, the data vector of the question from the coincidence degree between data vector search target data search means for searching for the search target data in the database, and output means for outputting a search result, the search
Judgment input means for inputting user judgment on search results
And the database of the question sentence according to the judgment of the user.
To the search target data
A data search device comprising: a learning unit that updates by adding or subtracting from a vector, wherein the vector generation unit includes :
Theory added to the image data based on the word vectors
This is achieved by the data retrieval device of claim 1 for creating a data vector representing a feature of a clear text .

[0015]

[0016]

[Action] In the data retrieval apparatus according to claim 1, the search target data by the input means is input as the quality Toibun, the question on the basis of the word vector by vector generation JP
Create a data vector representing the butterfly, questions by the search means
The search target data in the database is searched based on the degree of coincidence between the data vector of the sentence and the data vector of the search target data , and a search result is output by the output unit. By using a data vector created from word vectors characterized according to human knowledge, sensitivity, etc., data that matches the words contained in the question sentence is searched by searching the search target data based on the degree of matching of the data vectors. Not only that, it is also possible to perform a search in accordance with human knowledge, sensitivity, and the like. Further, the vector generation means attaches the image data based on the word vector .
By creating a data vector representing the features of the added explanation , a multimedia database including image data and the like can be searched in a natural language.

[0017]

[0018]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below in detail with reference to the drawings.

FIG. 1 is a block diagram of the present embodiment. The input means 1 comprises a keyboard or the like for inputting a question sentence, a database 2 to be searched for storing each data and a data vector representing its characteristics, An output unit 3 such as a display for outputting a result; a word dictionary 4 for storing word vectors representing the characteristics of each word; a vector generating unit 5 for generating a data vector from a natural language text; and a search unit 6 for executing a search And judgment input means 7 for inputting a user's judgment on the search result, and learning means 8 for performing learning according to the input judgment.

The specific configurations of the word dictionary 4, the vector generation unit 5 and the search unit 6 in this embodiment are the same as those of the system proposed in the above-mentioned "associative search from a large-scale document database". , "Data vector" of the present embodiment
Corresponds to a "context vector".

The operation of this embodiment will be described below with reference to the drawings.

First, the operation of the vector generation means 5 will be described with reference to FIG. In FIG. 2, for example, if the input sentence is a natural language text “lava released by a volcanic explosion”, in step S21, “volcanic”, “explosion”, “release”, “lava”
Two words are extracted. In step S22, a word vector of each word stored in the word dictionary is extracted as follows.

Volcano = (0,2,0,0,1,0,2,0,1) Explosion = (0,1,0,0,0,1,1,0,0) Emission = (1, (0,0,0,0,1,0,0,1) Lava = (0,2,0,0,0,2,1,0,0) In step S23, each word vector was added to obtain CV0. Thereafter, CV0 is normalized to a length of 10, and CV1 is obtained as follows.

CV0 = (1,5,0,0,1,4,4,0,2) | CV0 | = (1 ² +5 ² +... +2 ² ) ^1/2 = 7.93
7 CV1 = 10 × CV0 / | CV0 | = (1,6,0,0,1,5,5,0,3) In step S24, CV1 is output as a data vector corresponding to the input sentence. In this way, a normalized weighted sum of each word vector extracted from the natural language text is generated as a data vector of the text. Each data vector corresponding to the text data of the natural language stored in the database is generated by the above-described method using the vector generating means 5. The data vector corresponding to the question sentence is generated by the vector generating means 5 in the same manner as the question sentence as an input.

FIG. 3 shows an example of a method of storing a data vector corresponding to a natural language text. Here, an index for associating a natural language text with a data vector is simultaneously held. Here, the index means the address of the corresponding text data.

Next, a method of generating a data vector of image data in this embodiment will be described with reference to FIG.

As shown in FIG. 4, the image data in the present embodiment is obtained by adding a description to image data which is bit image data that can be displayed by the output means 3. First, in step S51, i representing an image number is initialized to 1. In step S52, the index of the image data of the i-th image is stored in the buffer IDX.
In step S53, a description corresponding to the i-th image is extracted. In step S54, the description is used as an input to obtain a data vector CV by the vector generation unit 5. In step S55, the obtained data vector C
V is stored in the data vector buffer DCVi. In step S56, 1 is added to i so as to indicate the next image. If the image data is over, the flow is not finished. If not, the process returns to step 52 and the above-described operation is repeated. FIG. 6 shows the data vectors and indices obtained by the above-described method in comparison with FIG. 3, and corresponding data vectors are stored for image data. Here, the index means the address of the corresponding image data.

The search means 6 calculates the distance between the data vector of the question sentence and the data vector in the database, and outputs a natural language text and image data corresponding to the index of the data vector whose distance is short as the search result. .

Hereinafter, the learning function of this embodiment will be described with reference to the drawings.

FIG. 7 is a flowchart showing the operation of the entire learning function of this embodiment. In step S71, the input unit 1
The question sentence is input, and the data is retrieved from the database 2 by the retrieval means 6 in step S72. In step S73, the search result is output to the output unit 3. In step S74, the user inputs a judgment on the search result from the judgment input unit 7, and in step S75, the learning unit 8 performs learning according to the judgment. If the search is to be ended in step S76, the flow ends; otherwise, step S7.
Return to 1.

In this embodiment, when the user determines that one of the retrieved data is correct from the input means 1 as an input of the determination in step S74, the user determines ×, and when the user determines that the data is incorrect, the user selects ×. Is configured to be input. Thus, in this embodiment, the input unit 1 has the function of the judgment input unit 7.

The operation of the learning means in step S75 will be described in detail with reference to FIG.

Let Q = (q1, q2,..., Qn) be the data vector of the question sentence obtained from the input question sentence. Let V = (v1, v2,..., Vn). Here, the learning means 8 updates V to V + εQ when the user inputs ○, and updates V to −Q when the user inputs X. Here, ε is a positive constant determined from the learning speed and accuracy.

FIG. 8 shows a specific example of the output result by the learning means. FIG. 8A shows an example in which a question sentence is displayed as a search result in descending order of score. In this example, a natural language text is a search target, but the image search is similarly displayed. The score here is an inner product of the data vector of the question sentence and the data vector of the retrieved data, which may be considered as a numerical value of the closeness of the distance between the two vectors. As a result of the first search, the question text "Lava released by volcanic eruption" was searched for text data "Original topography" as 82 points and "volcaniclastic rock" as 66 points. Here, the data vector of the question text is Q = (1,6,0,0,1,5,5,0,3), and the data vector of “volcaniclastic rock” is V = (5,3,2, 0,0,4,1,1,6), and 66 points are obtained as the inner product thereof. When the user selects “volcaniclastic rock”, determines that the data is correct for the question text, and inputs “○”, the learning means 8 sets ε = 0.2 based on the above equation. Find the data vector to be updated as follows.

V + 0.2 × Q = (5,4,2,0,0,
(5, 2, 0, 7) After that, when the same question sentence is input again, as shown in FIG. 8 (b), the score of "volcaniclastic rock" becomes 85 points from the calculation of the inner product, and is searched higher. Thus, learning according to the user's judgment is correctly performed. Conversely, when the user inputs x, the learning is similarly performed correctly.

As described above, the determination of the user is made by ○, whereby the burden on the user is reduced as compared with the conventional learning means.

In this embodiment, the description of the image is used to create the data vector of the image. However, for example, a method of creating the data vector from several words representing the impression of the image given by the user may be used. Conceivable. Further, the learning method according to the present invention is not limited to the method using the above-described formula. Furthermore, the structure of the word dictionary, the method of constructing the data vector, the method of searching, and the method of judgment by the user are not limited to the method of the above-described embodiment.

[0038]

According to the data search apparatus of the first aspect,
By using a data vector created from word vectors characterized according to human knowledge, sensitivity, etc., data that matches the words contained in the question sentence is searched by searching the search target data based on the degree of matching of the data vectors. Not only that, it is also possible to perform a search in accordance with human knowledge, sensitivity, and the like. Furthermore, the vector generation means creates a data vector representing the characteristics of the description added to the image data based on the word vector, so that a multimedia database including image data and the like can be searched in a natural language. .

[0039]

[Brief description of the drawings]

FIG. 1 is a configuration diagram of an embodiment of a data search device of the present invention.

FIG. 2 is a flowchart showing a method of generating a data vector by a vector generating means of the apparatus of FIG.

FIG. 3 is an example of a method for storing natural language text data in a database of the apparatus of FIG. 1;

FIG. 4 is a configuration example of image data of the apparatus of FIG. 1;

FIG. 5 is a flowchart illustrating a data vector generation method of image data by a vector generation unit of the apparatus of FIG. 1;

6 is an example of a method of storing image data and data vectors created by the method of FIG. 5;

FIG. 7 is a flowchart showing a learning method of the apparatus of FIG. 1;

FIG. 8 is a diagram showing a display example of a search result and a learning result of the apparatus of FIG. 1;

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-4-98463 (JP, A) JP-A-4-352279 (JP, A) JP-A-2-125363 (JP, A) Akutoshi, Ikeuchi, Obuchi "Construction of Encyclopedia Text Database Using Semantic Vectors", IPSJ Symposium, Vol. 93, no. 9, p. 227-234, 1993 (Heisei 5-12-2) (58) Field surveyed (Int. Cl. ⁷ , DB name) G06F 17/30 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. A input means for inputting a search target data as quality Toibun, a word dictionary for storing word vector representing a feature of a word, a data vector representing a feature of the question on the basis of the word vectors a vector generation means for creating, from the coincidence degree between the search target data and a database for storing de <br/> Tabekutoru representing the feature of the search target data, data vector of the search target data and data vectors of the question Search means for searching for search target data in the database, output means for outputting a search result ,
Judgment input for inputting a user's judgment on the search result
Means and data of the question sentence according to the judgment of the user
The value obtained by multiplying the vector by a predetermined value is
And a learning means for updating by adding or subtracting from the data vector.
It is added to the image data based on the word vectors
Creating a data vector that represents the characteristics of the written description
Characteristic data search device.