JP7724506B2 - Question and answer search device and question and answer search program - Google Patents

Description

Article 30, Paragraph 2 of the Patent Act will be applied. Published at the "Software Symposium 2023 in Sendai" on June 12, 2023.

The present invention relates to a question and answer search device and a question and answer search program that search for question and answer sentences corresponding to a character string entered as a question.

In situations where responses to inquiries are provided, such as in the work of call center operators, systems are used that search for and output (for example, display on a monitor) responses corresponding to the content of the inquiry; a specific example of this is described in Patent Document 1. The system described in Patent Document 1 has a database that stores FAQs consisting of combinations of anticipated questions and their answers.

The system generates a search query by extracting keywords from the conversation between the asker and the person answering the question, and based on the search query, searches for a question and answer statement (FAQ) stored in a database that corresponds to the question from the asker and displays it on a monitor. By using this system, the person answering the question can refer to the question and answer statement displayed on the monitor and decide what answer to return to the asker.
Patent Document 1 discloses that an optimal question-and-answer sentence is searched for based on a search query by keyword matching or concept search between the search query and the question-and-answer sentence.

Here, since both the search query and the question and answer are character strings, one possible method for comparing the similarity in meaning between different character strings is to use cosine similarity, as described in Patent Document 2. Cosine similarity represents character strings as vectors in feature space, and quantitatively expresses the similarity in meaning between different character strings numerically based on the angle between the vectors. Considering that vectors represent the meaning of character strings, it is important that the meaning of a character string can be uniquely defined.

JP 2022-66489 A JP 2023-084644 A

For example, in the case of a call center operator, when a call center operator inputs keywords to search for answers through conversation with a questioner and searches for an appropriate question and answer from among the questions and answers stored in a database, that is, when searching for a question and answer based on a string of characters that does not form a sentence and whose meaning cannot be uniquely determined (for example, a string of three nouns entered with a space between them), simply using cosine similarity will not be enough to find the question and answer that corresponds to the question.

The present invention was made in light of these circumstances, and aims to provide a question-and-answer sentence search device and question-and-answer sentence search program that can reliably search for question-and-answer sentences appropriate to a question, even if the character string entered based on the question is not a complete sentence.

A question and answer search device according to a first aspect of the present invention that achieves the above-mentioned objective comprises a text storage unit that stores a plurality of question and answer sentences, each consisting of a pair of question text sentences and answer text sentences, and search means that searches for the question and answer sentence corresponding to an input string based on an input string consisting of one or more words or phrases given as a string representing a question. The device is equipped with a vector storage unit that stores a plurality of question vectors obtained by converting each of the question text sentences into a vector in a feature space, and vector conversion means that converts the input string into an input vector expressed as a vector in the feature space. The search means uses modified cosine similarity, which is a correction of cosine similarity, as the similarity of the question vector to the input vector, and preferentially selects a question and answer sentence corresponding to the question vector with a larger modified cosine similarity as the question and answer sentence corresponding to the input string. The modified cosine similarity increases as the degree of match between the words or phrases included in the input string that formed the basis of the input vector and the words or phrases included in the question text sentence that formed the basis of the question vector increases.

A question and answer search device according to a second invention that achieves the above object includes a text storage unit that stores a plurality of question and answer sentences, each consisting of a pair of question text sentences and answer text sentences, and search means that searches for the question and answer sentence corresponding to an input string based on an input string consisting of one or more words given as a string representing a question. The device also includes a vector storage unit that stores a plurality of answer vectors obtained by converting each of the answer text sentences into a vector in a feature space, and vector conversion means that converts the input string into an input vector expressed as a vector in the feature space. The search means uses modified cosine similarity, which is a correction of cosine similarity, as the similarity of the answer vector to the input vector, and preferentially selects a question and answer sentence corresponding to an answer vector with a larger modified cosine similarity as the question and answer sentence corresponding to the input string. The modified cosine similarity increases the higher the match rate between the words contained in the input string that formed the basis of the input vector and the words contained in the answer text sentence that formed the basis of the answer vector.

A question and answer text search device according to a third invention that meets the above objective is a question and answer text search device having a text storage unit that stores a plurality of question and answer texts, each consisting of a pair of question text texts and answer text texts, and a search means that searches for the question and answer texts corresponding to an input string based on an input string consisting of one or more words or phrases given as a string representing a question, and is equipped with a vector storage unit that stores a plurality of question vectors obtained by converting a plurality of the question text texts into vectors in a feature space and a plurality of answer vectors obtained by converting a plurality of the answer text texts into vectors in the feature space, and a vector conversion means that converts the input string into an input vector expressed as a vector in the feature space, and the search means calculates the similarity of the question vector to the input vector and the answer vector to the input vector. First and second modified cosine similarities, which are obtained by correcting the cosine similarity, are used as the similarity between the input string and the answer vector, and the question and answer sentences corresponding to the question vector with larger first modified cosine similarity and the answer sentences corresponding to the answer vector with larger second modified cosine similarity are preferentially selected as the question and answer sentences corresponding to the input string, and the first modified cosine similarity increases as the degree of match between the words included in the input string on which the input vector is based and the words included in the question text on which the question vector is based increases, and the second modified cosine similarity increases as the degree of match between the words included in the input string on which the input vector is based and the words included in the answer text on which the answer vector is based increases.

A fourth aspect of the present invention that achieves the above-mentioned objective is a question and answer search program that causes a computer to perform a process of searching, based on an input string consisting of one or more words given as a string representing a question, for a question and answer sentence that corresponds to the input string from a plurality of question and answer sentences, each consisting of a pair of question text sentences and answer text sentences, which are stored in a text storage unit. The program causes the computer to perform a process of converting the input string into an input vector represented by a vector in a feature space, and a process of using modified cosine similarity, which is a correction of cosine similarity, as the similarity of each of a plurality of question vectors, obtained by converting a plurality of question text sentences into vectors in the feature space, to the input vector, and preferentially selecting a question and answer sentence that corresponds to the question vector with a larger modified cosine similarity as the question and answer sentence that corresponds to the input string. The modified cosine similarity becomes larger the higher the degree of match between the words contained in the input string that formed the basis of the input vector and the words contained in the question text sentence that formed the basis of the question vector.

A fifth aspect of the present invention, which achieves the above-mentioned objective, provides a question and answer search program that causes a computer to perform a process of searching, based on an input string consisting of one or more words or phrases given as a string representing a question, for a question and answer sentence that corresponds to the input string from a plurality of question and answer sentences, each consisting of a pair of question text sentences and answer text sentences, which are stored in a text storage unit. The program causes the computer to perform the following processes: converting the input string into an input vector represented by a vector in a feature space; and using modified cosine similarity, which is a correction of cosine similarity, as the similarity of a plurality of answer vectors, each obtained by converting a plurality of answer text sentences into vectors in the feature space, to the input vector; and preferentially selecting a question and answer sentence that corresponds to an answer vector with a larger modified cosine similarity as the question and answer sentence that corresponds to the input string; the modified cosine similarity becomes larger the higher the degree of match between the words or phrases included in the input string that formed the basis of the input vector and the words or phrases included in the answer text sentences that formed the basis of the answer vector.

A sixth invention that achieves the above-mentioned objective is a question and answer search program that causes a computer to perform a process of searching, based on an input string consisting of one or more words given as a string representing a question, for a question and answer sentence that corresponds to the input string from a plurality of question and answer sentences that are each a pair of question text sentences and answer text sentences and are stored in a text storage unit, the process including: a process of converting the input string into an input vector expressed as a vector in a feature space; a process of using a first modified cosine similarity, which is obtained by correcting the cosine similarity, as the similarity to the input vector of a plurality of question vectors obtained by converting a plurality of the question text sentences into vectors in the feature space, and a process of preferentially selecting a question and answer sentence that corresponds to the question vector with a larger first modified cosine similarity as the question and answer sentence that corresponds to the input string; The computer is caused to perform a process of employing a second modified cosine similarity, obtained by correcting the cosine similarity, as the similarity to the input vector of a plurality of answer vectors obtained by converting the plurality of answer text sentences into vectors in the feature space, and preferentially selecting, as the question and answer sentence corresponding to the input string, a question and answer sentence corresponding to the answer vector with a larger second modified cosine similarity, wherein the first modified cosine similarity increases as the degree of match between the words contained in the input string on which the input vector is based and the words contained in the question text sentence on which the question vector is based, and the second modified cosine similarity increases as the degree of match between the words contained in the input string on which the input vector is based and the words contained in the answer text sentence on which the answer vector is based.

A question and answer search device according to a first aspect of the present invention includes a vector storage unit that stores multiple question vectors obtained by converting multiple question text sentences into vectors in a feature space, and vector conversion means that converts an input character string into an input vector expressed as a vector in the feature space. The search means uses modified cosine similarity, which is a corrected cosine similarity, as the similarity of the question vector to the input vector, and preferentially selects question and answer sentences corresponding to question vectors with larger modified cosine similarity as question and answer sentences corresponding to the input character string. The higher the degree of match between the words contained in the input character string that formed the basis of the input vector and the words contained in the question text sentence that formed the basis of the question vector, the larger the modified cosine similarity becomes. This makes it possible to search for question and answer sentences that take into account the similarity of their meanings as sentences and the consistency of their words, and enables stable search for question and answer sentences that are appropriate for a question even if the string entered based on the question is not a complete sentence.

Furthermore, the question and answer sentence search device according to the second invention and the question and answer sentence search device according to the third invention also have the same configuration as the question and answer sentence search device according to the first invention, and are therefore able to reliably search for a question and answer sentence that is appropriate for the question even if the character string entered based on the question is not a complete sentence.

Furthermore, the question and answer text search program according to the fourth invention, the question and answer text search program according to the fifth invention, and the question and answer text search program according to the sixth invention correspond to the question and answer text search device according to the first invention, the question and answer text search device according to the second invention, and the question and answer text search device according to the third invention, respectively, and therefore can reliably search for a question and answer text appropriate to the question even if the string of characters entered based on the question is not a complete sentence.

1 is an explanatory diagram of a question and answer text search device according to a first embodiment of the present invention; 10 is an explanatory diagram of a process of searching for a question and answer sentence by the question and answer sentence search device. FIG. FIG. 10 is an explanatory diagram of a question and answer text search device according to a second embodiment of the present invention. 10 is an explanatory diagram of a process of searching for a question and answer sentence by the question and answer sentence search device. FIG. FIG. 10 is an explanatory diagram of a question and answer text search device according to a third embodiment of the present invention. 10 is an explanatory diagram of a process of searching for a question and answer sentence by the question and answer sentence search device. FIG.

Next, embodiments of the present invention will be described with reference to the accompanying drawings to facilitate understanding of the present invention.
1 and 2, a question and answer sentence search device 10 according to a first embodiment of the present invention is an apparatus having a text storage unit 14 that stores a plurality of question and answer sentences 13, each of which is made up of a pair of question text sentence 11 and answer text sentence 12, and a search means 16 that searches, based on an input character string 15 made up of one or more words given as a character string representing a question, for a question and answer sentence 13 that corresponds to the input character string 15. A detailed description will be given below.

The question and answer text search device 10 is composed of software such as an operating system and a software program for searching questions and answers (hereinafter also referred to as the "question and answer text search program"), and hardware such as a computer or storage medium that has the software installed. Therefore, the text storage unit 14, search means 16, and other means and units provided in the question and answer text search device 10 are each composed of either software or hardware, or both.

The term "computer" as used here refers to, for example, tablet, notebook, and desktop personal computers, mobile terminals, and servers, and there is no limit to the number of computers that the question and answer search device 10 may have. Therefore, the question and answer search device 10 may have one computer or multiple computers.
In addition, in this embodiment, the question and answer search device 10 is used in a call center that returns answers to questions received from users of products or services, but the use is not limited to this. For example, the question and answer search device 10 may be used as a system in which users of products or services themselves input inquiries (questions) that they wish to inquire about and receive answers to those inquiries.

As shown in Figures 1 and 2, the question and answer search device 10 includes an input unit 17 to which an input string 15 is input by a user of the question and answer search device 10 (hereinafter simply referred to as "user") using an input device (not shown), and vector conversion means 19 capable of converting text sentences into vectors in a feature space (high-dimensional space, vector space). Note that a publicly known trained model is used as the method for encoding text sentences into vectors.

The input device may be a keyboard, a microphone, or the like. If the input device is a keyboard, the user can input the input string 15 by operating the keyboard. If the input device is a microphone, the user can input the input string 15 by speaking into the microphone. One input string 15 is a text sentence consisting of one or more phrases. A phrase means a word (e.g., "renewal," "certificate") or a phrase consisting of multiple words (e.g., "exceptional action," "certificate not submitted"). In this embodiment, if one input string 15 is made up of multiple phrases, the input string 15 is provided to the input unit 17 with a space of one character between each phrase.

Vector conversion means 19 connected to input unit 17 can acquire input character strings 15 from input unit 17, and based on the acquired input character string 15, generates an input vector 20 that represents the input character string 15 as a vector in feature space. In other words, vector conversion means 19 converts input character strings 15 into input vectors 20. By vectorizing input character strings 15, it is possible, for example, to quantitatively handle the similarity in meaning of each of multiple input character strings 15 as a sentence.

The question and answer sentences 13 stored in the text storage unit 14 are question text sentences 11 and answer text sentences 12, which are question and answer responses converted into character data. The question and answer sentences 13 are either prepared in advance as expected question and answer sentences before using the question and answer sentence search device 10, or obtained while performing work using the question and answer sentence search device 10. The text storage unit 14 stores each question and answer sentence 13 by linking an identification number to the question text sentence 11 and answer text sentence 12.

The search means 16 is connected to the input unit 17, the vector conversion means 19, and the text storage unit 14. The search means 16 can obtain the input character string 15 from the input unit 17, the input vector 20 from the vector conversion means 19, and the question text sentence 11 and the answer text sentence 12 (question and answer sentence 13) from the text storage unit 14.
The search means 16 is also connected to a vector storage unit 22 that stores a plurality of question vectors 21 that correspond to all of the question text sentences 11 stored in the text storage unit 14 .

Question vectors 21 are obtained by converting question text sentences 11 into vectors in feature space using vector conversion means (not shown), and multiple question vectors 21 are pre-stored in the vector storage unit 22. The vector storage unit 22 stores each question vector 21 by linking it to an identification number that is the same as the identification number stored in the text storage unit 14 and linked to each question text sentence 11. When a new question text sentence 11 is added to the text storage unit 14 along with its identification number, the question vector 21 corresponding to that question text sentence 11 is added to the vector storage unit 22 along with the same identification number.

In this embodiment, the vector storage unit 22 stores each question vector 21 in groups such that mutually similar question vectors 21 are included in the same group. This makes the question vectors 21 stored in the vector storage unit 22 targets for approximate nearest neighbor search, which will be described later. Assuming that the vector storage unit 22 stores hundreds of thousands of question vectors 21, for example, the number of question vectors 21 included in each group is around 1,000. Furthermore, the number of dimensions of the feature space of the question vector 21 is the same as the number of dimensions of the feature space of the input vector 20.

The search means 16 includes a nearest neighbor search unit 23 that acquires an input vector 20 from the vector conversion means 19 , and a similarity calculation unit 24 that calculates the similarity of a query vector 21 to the input vector 20 .
As shown in FIG. 2, when a new input character string 15 is provided from the input unit 17 to the vector conversion means 19, the vector conversion means 19 converts the input character string 15 into an input vector 20 and provides it to the nearest neighbor search unit 23 (search means 16).

Each time an input vector 20 is acquired from the vector conversion means 19, the nearest neighbor search unit 23 performs an approximate nearest neighbor search to select, from all the question vectors 21 grouped and stored in the vector storage unit 22, one group that contains many question vectors 21 with high cosine similarity (not modified cosine similarity) to the acquired input vector 20, and acquires all the question vectors 21 contained in that group.

In this embodiment, the nearest neighbor search unit 23 acquires multiple query vectors 21 from the vector storage unit 22 (i.e., each group contains two or more query vectors 21). Furthermore, in this embodiment, a random projection tree is used for the approximate nearest neighbor search, but this is not limited to this; for example, LSH (Locality Sensitive Hashing) or KD tree may also be used. The vector storage unit 22 stores the query vectors 21 in a format suitable for the approximate nearest neighbor search algorithm used.

As shown in FIG. 1, the similarity calculation unit 24 is connected to the input unit 17, vector conversion means 19, nearest neighbor search unit 23, and text storage unit 14. Each time an input string 15 is provided to the input unit 17, the similarity calculation unit 24 obtains the input string 15 directly from the input unit 17 or via memory (storage medium). Furthermore, each time the vector conversion means 19 converts the input string 15 into an input vector 20, the similarity calculation unit 24 obtains the input vector 20 directly from the vector conversion means 19 or via memory.

Furthermore, each time the nearest neighbor search unit 23 acquires multiple (here, T) question vectors 21 from the vector storage unit 22, the similarity calculation unit 24 acquires the multiple question vectors 21 from the nearest neighbor search unit 23, and further acquires, based on the identification numbers linked to the multiple acquired question vectors 21, the question text sentence 11 (i.e., the question text sentence 11 on which each acquired question vector 21 was based) and the answer text sentence 12 (i.e., the question and answer sentence 13 corresponding to each acquired question vector 21) from the text storage unit 14 corresponding to each acquired question vector 21.

The similarity calculation unit 24 calculates the similarity to the acquired input vector 20 for multiple question vectors 21 acquired from the nearest neighbor search unit 23. Therefore, when the number of question vectors 21 stored in the vector storage unit 22 is N, the search means 16 will select fewer than N question vectors 21 through the approximate nearest neighbor search as targets for calculating the modified cosine similarity.

The similarity calculation unit 24 uses modified cosine similarity, which is a corrected cosine similarity, as the similarity of the question vector 21 to the input vector 20, and preferentially selects a question-and-answer sentence 13 (a question-and-answer sentence 13 having the relevant question vector 21) corresponding to the question vector 21 with a larger modified cosine similarity as the question-and-answer sentence 13 that corresponds to the input character string 15 (treating it as a question-and-answer sentence 13 that is more likely to correspond to the input character string 15). An output unit 25 that can display (output) the question-and-answer sentence 13 on a monitor (an example of an output device), not shown, is connected to the similarity calculation unit 24.

In this embodiment, the similarity calculation unit 24 acquires T question-and-answer sentences 13 from the text storage unit 14, and selects, from the acquired T question-and-answer sentences 13, only P question-and-answer sentences 13 that correspond to the top P (T>P) question vectors 21 with the largest modified cosine similarities as targets for display on the monitor. The similarity calculation unit 24 further sorts the selected P question-and-answer sentences 13 so that the greater the modified cosine similarity corresponding to the question-and-answer sentence 13, the more likely it is to be displayed on the monitor as a question-and-answer sentence 13 that corresponds to the input character string 15 (for example, displayed at the top of the list of question-and-answer sentences 13), and provides the P question-and-answer sentences 13 to the output unit 25.

Here, if θ is the angle between the input vector 20 and the question vector 21 in the feature space, the cosine similarity is cosθ, and is affected only by the angle between the input vector 20 and the question vector 21. In contrast, the modified cosine similarity is affected not only by the angle between the input vector 20 and the question vector 21, but also by the degree of match between the words contained in the input string 15 acquired from the input unit 17 (the input string 15 that formed the input vector 20 acquired from the nearest neighbor search unit 23) and the words contained in the question text sentence 11 that formed the question vector 21 acquired from the nearest neighbor search unit 23.

Specifically, the higher the match rate between the words in the input character string 15 and the words in the question text sentence 11, the higher the modified cosine similarity; and the lower the match rate, the lower the modified cosine similarity. Note that "large" and "small" modified cosine similarity here refer to relative magnitudes. For example, even if the modified cosine similarity is set to the maximum value when the match rate between the words in the input character string 15 and the words in the question text sentence 11 is 100%, and the modified cosine similarity is set to the maximum value minus a predetermined value when the match rate is 80%, this also means that the higher the match rate between the words in the input character string 15 and the words in the question text sentence 11, the higher the modified cosine similarity, and the lower the match rate, the lower the modified cosine similarity (this also applies to other embodiments). In this embodiment, S, calculated using the following formula 1, is the modified cosine similarity.

S=cos(α×θ) Formula 1

Here, let Q be the total number of words contained in input string 15, let x be the number of words contained in input string 15 that are also contained in question text sentence 11, and let k be a coefficient in the range 0<k<1. Then, α in equation 1 can be expressed as equation 2 below.

α=1−k×(x/Q) Equation 2

Modified cosine similarities other than those expressed by Equation 1 and Equation 2 may also be used. For example, α may be α = 1.2 - k × (x/Q), α = 1 - (1/2) × tanh{π(x/Q)}, α = (1/2) × [1 - tanh{π × ((x/Q) - 1)}], or α = (3/4) - (1/4) × tanh[2π × {(x/Q) - (1/2)}].

Experimental verification has confirmed that by using modified cosine similarity as the similarity between the input vector 20 and the question vector 21, more questions and answers 13 that are useful to the user are displayed on the monitor compared to when cosine similarity is used. Here, the words contained in both the input string 15 and the question text sentence 11 may be limited to words whose strings of characters exactly match those contained in the input string 15 and the question text sentence 11, or in addition to words whose strings of characters exactly match, words whose strings of characters differ may also be included as long as the words in both sentences represent the same subject.

Furthermore, the search means 16 does not calculate the modified cosine similarity for the input vector 20 for all question vectors 21 stored in the vector storage unit 22, but limits it to a predetermined number of question vectors 21 in order to shorten the processing time from when the search means 16 acquires the input vector 20 from the vector conversion means 19 until when it provides P question and answer sentences 13 to the output unit 25. In this embodiment, there are a large number of question vectors 21 stored in the vector storage unit 22 (e.g., hundreds of thousands), and furthermore, when calculating the modified cosine similarity, the matching between the words in the input string 15 that is the basis of the input vector 20 and the words in the question text sentence 11 that is the basis of the question vector 21 is detected, so the effect of shortening the processing time is significant.

From the explanation so far, it can be said that the question and answer search program used in the question and answer search device 10 is a software program that causes a computer to perform the process of searching for a question and answer sentence 13 that corresponds to an input string 15 from multiple question and answer sentences 13 stored in the text storage unit 14, based on an input string 15 given as a string representing a question. It causes the computer to perform the process of converting the input string 15 into an input vector 20, and the process of using modified cosine similarity as the similarity of each of the multiple question vectors 21 to the input vector 20, and preferentially selecting a question and answer sentence 13 that corresponds to a question vector 21 with a larger modified cosine similarity as the question and answer sentence 13 that corresponds to the input string 15.

The question and answer text search device 10 calculates the modified cosine similarity between the input vector 20 and the question vector 21, but this is not limited to this. Next, with reference to Figures 3 and 4, we will explain the question and answer text search device 30 that calculates the modified cosine similarity between the input vector 20 and the answer vector 31, which is obtained by converting the answer text sentence 12 into a vector in feature space. Note that in the question and answer text search device 30, components that are similar to those in the question and answer text search device 10 are assigned the same reference numerals and detailed descriptions will be omitted.

As shown in Figures 3 and 4, the question and answer sentence search device 30 according to the second embodiment of the present invention includes an input unit 17, vector conversion means 19, text storage unit 14, and output unit 25, which are similar to those included in the question and answer sentence search device 10.
On the other hand, the search means 32 and the vector storage unit 33 of the question and answer sentence search device 30 are different from the search means 16 and the vector storage unit 22 of the question and answer sentence search device 10, respectively.

The vector storage unit 33 stores the answer vector 31, not the question vector 21. The search means 32 includes a nearest neighbor search unit 34 that retrieves the input vector 20 from the vector conversion means 19 and the answer vector 31 from the vector storage unit 33, and a similarity calculation unit 35 that calculates the modified cosine similarity of the answer vector 31 with respect to the input vector 20.

Each time an input vector 20 is acquired from the vector conversion means 19, the nearest neighbor search unit 34 performs an approximate nearest neighbor search to select, from all answer vectors 31 grouped and stored in the vector storage unit 33, one group that contains many answer vectors 31 with high cosine similarity to the acquired input vector 20, and acquires all answer vectors 31 contained in that group.

The similarity calculation unit 35 obtains the input character string 15 directly from the input unit 17 or via memory, and obtains the input vector 20 directly from the vector conversion means 19 or via memory. Furthermore, the similarity calculation unit 35 obtains multiple (here, T) answer vectors 31 from the nearest neighbor search unit 34, which has obtained the T answer vectors 31 from the vector storage unit 22, and obtains the question text sentence 11 and answer text sentence 12 corresponding to each obtained answer vector 31 (i.e., the question and answer sentence 13 corresponding to each obtained answer vector 31) from the text storage unit 14.

Then, the similarity calculation unit 35 calculates the modified cosine similarity for the acquired input vector 20 for the T answer vectors 31 acquired from the nearest neighbor search unit 34. Therefore, when the number of answer vectors 31 stored in the vector storage unit 33 is M, the search means 32 will select fewer than M answer vectors 31 through the approximate nearest neighbor search as targets for calculating the modified cosine similarity.

The modified cosine similarity calculated by the similarity calculation unit 35 is equivalent to the modified cosine similarity expressed by the above-mentioned formulas 1 and 2, where x in formula 2 is the number of words contained in the input string 15 that are contained in the answer text sentence 12 that is the basis of the answer vector 31. In other words, the modified cosine similarity increases the higher the degree of match between the words contained in the input string 15 that is the basis of the input vector 20 and the words contained in the answer text sentence 12 that is the basis of the answer vector 31.

Then, the similarity calculation unit 35 selects only the P question-and-answer sentences 13 corresponding to the top P (T>P) answer vectors 31 with the largest modified cosine similarities as the targets for display on the monitor, and further sorts the P selected question-and-answer sentences 13 so that the larger the modified cosine similarity, the more likely the question-and-answer sentence 13 is displayed on the monitor as a question-and-answer sentence 13 that corresponds to the input character string 15, and provides the sorted P question-and-answer sentences 13 to the output unit 25.

The question and answer sentence search program used by the question and answer sentence search device 30 is a software program that causes a computer to perform the process of searching for a question and answer sentence 13 that corresponds to an input string 15 from multiple question and answer sentences 13 stored in the text storage unit 14, based on an input string 15 given as a string representing a question. The program causes the computer to perform the process of converting the input string 15 into an input vector 20, and the process of using modified cosine similarity as the similarity of each of the multiple answer vectors 31 to the input vector 20, and preferentially selecting a question and answer sentence 13 that corresponds to an answer vector 31 with a larger modified cosine similarity as the question and answer sentence 13 that corresponds to the input string 15.

Furthermore, while the question and answer text search device 10 calculates only the modified cosine similarity between the input vector 20 and the question vector 21, and the question and answer text search device 30 calculates only the modified cosine similarity between the input vector 20 and the answer vector 31, this is not limited to this. With reference to Figures 5 and 6, a question and answer text search device 50 that calculates both the modified cosine similarity between the input vector 20 and the question vector 21 and the modified cosine similarity between the input vector 20 and the answer vector 31 will be described. Note that in the question and answer text search device 50, components similar to those in the question and answer text search device 10 or the question and answer text search device 30 are designated by the same reference numerals and detailed description thereof will be omitted.

As shown in Figures 5 and 6, the question and answer sentence search device 50 according to the third embodiment of the present invention includes an input unit 17, vector conversion means 19, text storage unit 14, and output unit 25, which are similar to those included in the question and answer sentence search device 10 and the question and answer sentence search device 30.
On the other hand, the search means 51 and vector storage unit 52 of the question and answer search device 50 are different from the search means 16 and vector storage unit 22 of the question and answer search device 10 and the search means 32 and vector storage unit 33 of the question and answer search device 30, respectively.

The vector storage unit 52 stores both the question vector 21 and the answer vector 31, linked to an identification number. The search means 51 includes a nearest neighbor search unit 53 that acquires the input vector 20 from the vector conversion means 19 and the question vector 21 and answer vector 31 from the vector storage unit 52, and a similarity calculation unit 54 that calculates both the similarity of the question vector 21 to the input vector 20 (hereinafter, this similarity will be referred to as "similarity P") and the similarity of the answer vector 31 to the input vector 20 (hereinafter, this similarity will be referred to as "similarity Q").

In this embodiment, the first modified cosine similarity, which is a corrected cosine similarity, is used as the similarity P, and the second modified cosine similarity, which is a corrected cosine similarity, is used as the similarity Q. In this embodiment, the first modified cosine similarity is the same as the modified cosine similarity used by the question-and-answer text review device 10, and the second modified cosine similarity is the same as the modified cosine similarity used by the question-and-answer text review device 30.

Here, the first modified cosine similarity should increase as the degree of match between the words contained in the input string 15 on which the input vector 20 is based and the words contained in the question text sentence 11 on which the question vector 21 is based, increases, and the second modified cosine similarity should increase as the degree of match between the words contained in the input string 15 and the words contained in the answer text sentence 12 on which the answer vector 31 is based, increases. In this embodiment, the first modified cosine similarity and the second modified cosine similarity are both expressed by the above Equation 1 and Equation 2 (i.e., they are calculated using the same formula except for the object represented by x). Note that the formula for calculating the first modified cosine similarity and the formula for calculating the second modified cosine similarity do not need to be the same and may be slightly different; for example, the value of the coefficient k in Equation 2 may be different.

The question vectors 21 are grouped and stored in the vector storage unit 52, and the answer vectors 31 are also grouped and stored in the vector storage unit 52.
Each time an input vector 20 is acquired from the vector conversion means 19, the nearest neighbor search unit 53 performs an approximate nearest neighbor search to select, from all question vectors 21 stored in the vector storage unit 52, a group that contains many question vectors 21 that have a high similarity to the acquired input vector 20, and acquires all question vectors 21 included in that group. At the same time, the nearest neighbor search unit 53 performs an approximate nearest neighbor search to select, from all answer vectors 31 stored in the vector storage unit 52, a group that contains many answer vectors 31 that have a high similarity to the acquired input vector 20, and acquires all answer vectors 31 included in that group.

The similarity calculation unit 54 obtains the input character string 15 directly from the input unit 17 or via memory, and obtains the input vector 20 directly from the vector conversion means 19 or via memory. Furthermore, the similarity calculation unit 54 obtains multiple (here, T1) question vectors 21 and multiple (here, T2) answer vectors 31 from the nearest neighbor search unit 53, which has obtained the T1 question vectors 21 and the T2 answer vectors 31 from the vector storage unit 22, and obtains the question text sentences 11 and answer text sentences 12 corresponding to each of the obtained question vectors 21, as well as the question text sentences 11 and answer text sentences 12 corresponding to each of the obtained answer vectors 31, from the text storage unit 14.

Then, the similarity calculation unit 54 calculates a first modified cosine similarity for the acquired input vector 20 for the T1 question vectors 21 acquired from the nearest neighbor search unit 53, and calculates a second modified cosine similarity for the acquired input vector 20 for the T2 answer vectors 31 acquired from the nearest neighbor search unit 53.

Therefore, when the number of question vectors 21 and the number of answer vectors 31 stored in the vector storage unit 52 are N and M, respectively, the search means 51 will use the approximate nearest neighbor search to select fewer than N question vectors 21 as targets for calculating the first modified cosine similarity, and fewer than M answer vectors 31 as targets for calculating the second modified cosine similarity.

Then, the similarity calculation unit 54 selects, from the calculated T1 first modified cosine similarities and T2 second modified cosine similarities (T1 + T2 modified cosine similarities), only P question-and-answer sentences 13 corresponding to the question vectors 21 and answer vectors 31 with the highest P values (T1 + T2 > P) as candidates for display on the monitor, and further sorts the P selected question-and-answer sentences 13 so that the larger the first modified cosine similarity or second modified cosine similarity, the more likely the question-and-answer sentence 13 is displayed on the monitor as a sentence that corresponds to the input string 15, and provides the sorted P question-and-answer sentences 13 to the output unit 25.

Therefore, the similarity calculation unit 54 will preferentially select, as the question and answer sentence 13 corresponding to the input character string 15, the question and answer sentence 13 corresponding to the question vector 21 with the larger first modified cosine similarity and the question and answer sentence 13 corresponding to the answer vector 31 with the larger second modified cosine similarity.

The question and answer search program used in the question and answer search device 50 is a software program that causes a computer to perform the following processes: search for a question and answer sentence 13 that corresponds to an input string 15 from multiple question and answer sentences 13 stored in the text storage unit 14, based on an input string 15 given as a string representing a question. The program causes the computer to perform the following processes: converting the input string 15 into an input vector 20; using a first modified cosine similarity as the similarity of each of the multiple question vectors 21 to the input vector 20, and preferentially selecting, as the question and answer sentence 13 that corresponds to the input string 15, question and answer sentences 13 that correspond to the question vector 21 with a larger first modified cosine similarity; and using a second modified cosine similarity as the similarity of each of the multiple answer vectors 31 to the input vector 20, and preferentially selecting, as the question and answer sentence 13 that corresponds to the input string 15, question and answer sentences 13 that correspond to the answer vector 31 with a larger second modified cosine similarity.

Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and all changes in conditions that do not depart from the gist of the present invention are within the scope of application of the present invention.
For example, in the first embodiment, the search means may calculate modified cosine similarities for all question vectors stored in the vector storage unit, and in the second embodiment, the search means may calculate modified cosine similarities for all answer vectors stored in the vector storage unit.

Furthermore, in the third embodiment, the search means can calculate the first modified cosine similarity for all question vectors stored in the vector storage unit, and calculate the second modified cosine similarity for all answer vectors stored in the vector storage unit.

Furthermore, both question vectors and answer vectors may be stored in the vector storage unit, and when selecting a priority question and answer sentence corresponding to an input character string, it may be possible to select by settings whether the target for calculating the modified cosine similarity (including the first modified cosine similarity and the second modified cosine similarity) for the input vector is only the question vector, only the answer vector, or both the question vector and the answer vector.

10: Question and Answer Search Device, 11: Question Text, 12: Answer Text, 13: Question and Answer, 14: Text Storage Unit, 15: Input String, 16: Search Means, 17: Input Unit, 19: Vector Conversion Means, 20: Input Vector, 21: Question Vector, 22: Vector Storage Unit, 23: Nearest Neighbor Search Unit, 24: Similarity Calculation Unit, 25: Output Unit, 30: Question and Answer Search Device, 31: Answer Vector, 32: Search Means, 33: Vector Storage Unit, 34: Nearest Neighbor Search Unit, 35: Similarity Calculation Unit, 50: Question and Answer Search Device, 51: Search Means, 52: Vector Storage Unit, 53: Nearest Neighbor Search Unit, 54: Similarity Calculation Unit

Claims (9)

A question and answer sentence search device having a text storage unit that stores a plurality of question and answer sentences, each of which is a pair of a question text sentence and an answer text sentence, and a search means that searches for the question and answer sentence corresponding to an input character string based on an input character string that is made up of one or more words or phrases given as a character string representing a question,
a vector storage unit that stores a plurality of question vectors obtained by converting the plurality of question text sentences into vectors in a feature space;
vector conversion means for converting the input character string into an input vector expressed as a vector in the feature space;
the search means employs a modified cosine similarity obtained by correcting a cosine similarity as the similarity of the question vector to the input vector, and preferentially selects the question and answer sentence corresponding to the input character string as the question and answer sentence corresponding to the input character string when the modified cosine similarity is greater;
The question and answer search device is characterized in that the modified cosine similarity increases as the degree of match between the words contained in the input string that formed the input vector and the words contained in the question text that formed the question vector increases. The question and answer text search device according to claim 1, wherein when the number of question vectors stored in the vector storage unit is N, the search means selects, by an approximate nearest neighbor search, a number of question vectors less than N as targets for calculating the modified cosine similarity. A question and answer sentence search device having a text storage unit that stores a plurality of question and answer sentences, each of which is a pair of a question text sentence and an answer text sentence, and a search means that searches for the question and answer sentence corresponding to an input character string based on an input character string that is made up of one or more words or phrases given as a character string representing a question,
a vector storage unit that stores a plurality of answer vectors obtained by converting the plurality of answer text sentences into vectors in a feature space;
vector conversion means for converting the input character string into an input vector expressed as a vector in the feature space;
the search means employs a modified cosine similarity obtained by correcting a cosine similarity as the similarity of the answer vector to the input vector, and preferentially selects the question and answer sentence corresponding to the answer vector with a larger modified cosine similarity as the question and answer sentence corresponding to the input character string;
A question and answer text search device characterized in that the modified cosine similarity becomes larger the higher the degree of match between the words contained in the input string that formed the input vector and the words contained in the answer text that formed the answer vector. In the question and answer text search device described in claim 3, when the number of answer vectors stored in the vector storage unit is M, the search means selects, by approximate nearest neighbor search, answer vectors fewer than M as targets for calculating the modified cosine similarity. A question and answer sentence search device having a text storage unit that stores a plurality of question and answer sentences, each of which is a pair of a question text sentence and an answer text sentence, and a search means that searches for the question and answer sentence corresponding to an input character string based on an input character string that is made up of one or more words or phrases given as a character string representing a question,
a vector storage unit that stores a plurality of question vectors obtained by converting each of the plurality of question text sentences into vectors in a feature space and a plurality of answer vectors obtained by converting each of the plurality of answer text sentences into vectors in the feature space;
vector conversion means for converting the input character string into an input vector expressed as a vector in the feature space;
the search means employs first and second modified cosine similarities, which are corrected cosine similarities, as the similarity of the question vector to the input vector and the similarity of the answer vector to the input vector, respectively, and preferentially selects, as the question and answer sentence corresponding to the input character string, a question and answer sentence corresponding to the question vector with a larger first modified cosine similarity and a question and answer sentence corresponding to the answer vector with a larger second modified cosine similarity;
the first modified cosine similarity increases as the degree of match between the words contained in the input string that formed the input vector and the words contained in the question text sentence that formed the question vector increases, and the second modified cosine similarity increases as the degree of match between the words contained in the input string that formed the input vector and the words contained in the answer text sentence that formed the answer vector increases. In the question and answer text search device described in claim 5, when the number of question vectors and the number of answer vectors stored in the vector storage unit are N and M, respectively, the search means selects, by an approximate nearest neighbor search, question vectors fewer than N as targets for calculating the first modified cosine similarity, and selects answer vectors fewer than M as targets for calculating the second modified cosine similarity. A question and answer search program that causes a computer to perform a process of searching, based on an input character string consisting of one or more words given as a character string representing a question, a question and answer sentence corresponding to the input character string from a plurality of question and answer sentences each consisting of a pair of question text sentences and answer text sentences, the question and answer sentences being stored in a text storage unit, the process comprising:
A process of converting the input character string into an input vector expressed as a vector in a feature space;
employing a modified cosine similarity obtained by correcting cosine similarity as a similarity to the input vector of a plurality of question vectors obtained by converting the plurality of question text sentences into vectors in the feature space, and preferentially selecting, as the question and answer sentence corresponding to the input character string, a question and answer sentence corresponding to the question vector with a larger modified cosine similarity;
a modified cosine similarity that increases as the degree of match between the words contained in the input string that formed the input vector and the words contained in the question text that formed the question vector increases. A question and answer search program that causes a computer to perform a process of searching, based on an input character string consisting of one or more words given as a character string representing a question, a question and answer sentence corresponding to the input character string from a plurality of question and answer sentences each consisting of a pair of question text sentences and answer text sentences, the question and answer sentences being stored in a text storage unit, the process comprising:
A process of converting the input character string into an input vector expressed as a vector in a feature space;
employing a modified cosine similarity obtained by correcting cosine similarity as each similarity to the input vector of a plurality of answer vectors obtained by converting the plurality of answer text sentences into vectors in the feature space, and preferentially selecting the question and answer sentence corresponding to the answer vector with a larger modified cosine similarity as the question and answer sentence corresponding to the input character string;
A question and answer search program characterized in that the modified cosine similarity increases as the degree of match between the words contained in the input string that formed the input vector and the words contained in the answer text that formed the answer vector increases. A question and answer search program that causes a computer to perform a process of searching, based on an input character string consisting of one or more words given as a character string representing a question, a question and answer sentence corresponding to the input character string from a plurality of question and answer sentences each consisting of a pair of question text sentences and answer text sentences, the question and answer sentences being stored in a text storage unit, the process comprising:
A process of converting the input character string into an input vector expressed as a vector in a feature space;
a process of adopting a first modified cosine similarity obtained by correcting cosine similarity as each similarity to the input vector of a plurality of question vectors obtained by converting the plurality of question text sentences into vectors in the feature space, and preferentially selecting a question and answer sentence corresponding to the question vector with a larger first modified cosine similarity as the question and answer sentence corresponding to the input character string;
adopting a second modified cosine similarity obtained by correcting the cosine similarity as each similarity to the input vector of a plurality of answer vectors obtained by converting the plurality of answer text sentences into vectors in the feature space, and preferentially selecting, as the question and answer sentence corresponding to the input character string, a question and answer sentence corresponding to the answer vector having a larger second modified cosine similarity;
the first modified cosine similarity increases as the degree of match between the words contained in the input string that formed the input vector and the words contained in the question text sentence that formed the question vector increases, and the second modified cosine similarity increases as the degree of match between the words contained in the input string that formed the input vector and the words contained in the answer text sentence that formed the answer vector increases.