Sentence retrieval apparatus, method and program

Abstract

To provide a sentence retrieval apparatus, a method and a program capable of selecting recommended words useful for narrowing retrieval results in interactive sentence retrieval.SOLUTION: The sentence retrieval apparatus comprises a receiving unit, a retrieval unit, an analysis unit, and a determination unit. The receiving unit receives a string entered by a user. The retrieval unit retrieves a sentence collection in which a plurality of sentences to be retrieved are accumulated based on the received character string, and acquires one or more retrieval result sentences corresponding to the received character string. The analysis unit performs sentence analysis on the retrieval result sentence to obtain a relationship between words constituting the retrieval result sentence. The determination unit determines one or more recommended words for narrowing down the retrieval result sentence to a sentence desired by the user based on the received character string, the retrieval result sentence, and the relationship.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a text search device, method, and program. [Background technology]

[0002] There is an interactive text search system that searches a collection of questions and answers in response to a user's query to select question-answer pairs that match the query. In this system, when two or more question-answer pairs are found as search results, suggested words are provided to narrow down the search results. The suggested words are determined based on the frequency of occurrence of each word in the search results. Therefore, the suggested words may not always be effective in narrowing down the search results. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2020-123131 Summary of the Invention [Problem to be solved by the invention]

[0004] The problem to be solved by the present invention is to provide a text search device, method, and program that are capable of selecting recommended words that are useful for narrowing down search results in an interactive text search. [Means for solving the problem]

[0005] A text search device according to an embodiment includes a receiving unit, a search unit, an analysis unit, and a determination unit. The receiving unit receives a character string input by a user. The search unit searches a collection of texts that includes multiple search target texts based on the received character string, and obtains one or more search result texts that match the received character string. The analysis unit performs text analysis on the search result texts to obtain dependency relationships between words that make up the search result texts. The determination unit determines one or more recommended words to be used to narrow down the search result texts to texts desired by the user, based on the received character string, the search result texts, and the dependency relationships. [Brief explanation of the drawings]

[0006] [Figure 1] A diagram showing an example of the configuration of a text search system [Figure 2] FIG. 1 shows an example of the configuration of a text search device. [Figure 3] FIG. 1 is a diagram showing an example of a data structure of a question and answer collection according to the first embodiment; [Figure 4] FIG. 10 is a diagram showing an example of a processing procedure for text search processing. [Figure 5] FIG. 5 is a diagram showing an example of an interactive screen in step S406 of FIG. 4. [Figure 6] FIG. 5 is a diagram showing an example of an interactive screen in step S410 of FIG. 4. [Figure 7] FIG. 1 is a diagram illustrating a sentence analysis process and a recommended word determination process according to the first embodiment. [Figure 8] A diagram showing the analysis results for the search result question "An error was displayed when selecting the approver for the work approval request." [Figure 9] FIG. 10 is a diagram illustrating a sentence analysis process and a recommended word determination process according to a second embodiment. [Figure 10] A diagram showing the analysis results of the search result question "What is the deadline for patent proposal applications?" [Figure 11] FIG. 10 is a diagram illustrating another example of a sentence analysis process and a recommended word determination process according to the third embodiment. [Figure 12] A diagram showing the analysis results of the search result question "I would like to cancel my work approval application in order to change the details of the application." [Figure 13] FIG. 10 is a diagram showing an example of operation according to the third embodiment; [Figure 14] FIG. 10 is a diagram showing another example of operation according to the third embodiment; [Figure 15] FIG. 10 is a diagram showing another example of operation according to the third embodiment; [Figure 16] FIG. 13 is a diagram showing an example of a data structure of a question and answer collection according to the fourth embodiment. [Figure 17] FIG. 10 is a diagram showing an example of operation according to the fourth embodiment; [Figure 18] FIG. 13 is a diagram showing an example of operation according to the fifth embodiment. [Figure 19] FIG. 10 is a diagram illustrating a sentence analysis process and a recommended word determination process according to a comparative example. DETAILED DESCRIPTION OF THE INVENTION

[0007] Hereinafter, a text search device, method, and program according to this embodiment will be described with reference to the drawings.

[0008] (First embodiment) Fig. 1 is a diagram showing an example of the configuration of a text search system 1 according to the first embodiment. The text search system 1 is a computer network system that performs an interactive text search, searching a collection of texts based on a query from a user and selecting texts that match the query. As shown in Fig. 1, the text search system 1 includes a text search device 100, a text database 200, and a client terminal 300.

[0009] The text database 200 is a computer that maintains a database of a text collection that stores multiple texts to be searched. The text format according to this embodiment is not particularly limited, and may be a combination of two types of text, a question text and an answer text, a single text, or any other text format.

[0010] The client terminal 300 is a computer used by a user of the text retrieval system 1. The client terminal 300 has hardware such as a processor, an input device, a display device, and a communication device, and functions as a user interface for the text retrieval system 1. For example, the client terminal 300 accepts a character string related to a query input by a user via the input device. The character string related to the query may be input as natural sentences or as words. The word may be input as a single word or as a word string including multiple words. The character string related to the query is transmitted to the text retrieval device 100. The client terminal 300 also receives search results, etc. from the text retrieval device 100 and displays them on a display device.

[0011] FIG. 2 is a diagram showing an example of the configuration of a text search device 100 included in the text search system 1. As shown in FIG. 2, the text search device 100 is a computer having a processing circuit 11, a storage device 12, an input device 13, a communication device 14, and a display device 15. Data communication between the processing circuit 11, the storage device 12, the input device 13, the communication device 14, and the display device 15 is performed via a bus. The text search device 100 functions as a server device of the text search system 1. Specifically, the text search device 100 receives a character string related to a user's query from a client terminal 300, searches a collection of texts stored in a text database 200 based on the received character string related to the query, and displays the search results on the client terminal 300.

[0012] The processing circuit 11 includes a processor such as a CPU (Central Processing Unit) and a memory such as a RAM (Random Access Memory). The processing circuit 11 includes a reception unit 111, a search unit 112, an analysis unit 113, a determination unit 114, a display control unit 115, and an overall control unit 116. The processing circuit 11 executes a text search program to realize the functions of the above-mentioned units 111 to 116. The text search program is stored in a non-transitory computer-readable recording medium such as the storage device 12. The text search program may be implemented as a single program that describes all the functions of the above-mentioned units 111 to 116, or may be implemented as multiple modules divided into several functional units. Furthermore, the above-mentioned units 111 to 116 may be implemented by an integrated circuit such as an application specific integrated circuit (ASIC). In this case, the units may be implemented on a single integrated circuit or individually on multiple integrated circuits.

[0013] The reception unit 111 receives a character string related to an inquiry from a user. Specifically, the reception unit 111 acquires a character string related to an inquiry input via the client terminal 300. Hereinafter, the character string related to an inquiry will be referred to as a query statement.

[0014] The search unit 112 searches a collection of sentences that has accumulated multiple sentences to be searched, based on the character string accepted by the acceptance unit 111, and acquires one or more sentences that match the accepted character string. Hereinafter, the sentences acquired as search results will be referred to as search result sentences.

[0015] The analysis unit 113 performs a sentence analysis on the search result sentences acquired by the search unit 112 to acquire dependency relationships between words that make up the search result sentences. The analysis unit 113 uses dependency analysis, syntactic analysis, and / or a graph-based ranking method for the sentence analysis.

[0016] The determination unit 114 determines one or more recommended words to be recommended for narrowing down the search result sentences to sentences desired by the user, based on the character strings received by the reception unit 111, the search result sentences acquired by the search unit 112, and the dependency relationships acquired by the analysis unit 113. The determination unit 114 determines, as the recommended word, any one of a word that modifies a word included in the query sentence among the words constituting the search result sentences, a subject or object of a base word among the words constituting the search result sentences, and a word in an adverbial modifier clause of the base word among the words constituting the search result sentences. The base word is a word located at the root node at the highest level that does not have a parent node (node, vertex) when the structure of a sentence is represented as a tree structure, and is the word that serves as the starting point of links (arrows) that represent all dependency relationships included in the sentence.

[0017] The display control unit 115 displays various information on the display device 15 or the display device of the client terminal 300. As an example, the display control unit 115 displays the search result sentences acquired by the search unit 112 and the recommended words determined by the determination unit 114.

[0018] The overall control unit 116 overall controls the text search process performed by the processing circuit 11 in accordance with the text search program. For example, the overall control unit 116 operates the reception unit 111, the search unit 112, the analysis unit 113, the determination unit 114, and the display control unit 115 in accordance with the processing procedure in accordance with the text search program.

[0019] The storage device 12 is configured by a ROM (Read Only Memory), an HDD (Hard Disk Drive), an SSD (Solid State Drive), an integrated circuit storage device, etc. The storage device 12 stores a text search program and the like.

[0020] Various commands are input from an operator via the input device 13. A keyboard, a mouse, various switches, a touch pad, a touch panel display, etc. can be used as the input device 13. An output signal from the input device 13 is supplied to the processing circuit 11.

[0021] The communication device 14 is an interface for performing data communication with an external device connected to the text search device 100 via a network. As an example, the communication device 14 performs data communication with the text database 200 or the client terminal 300.

[0022] The display device 15 displays various information. As the display device 15, a CRT (Cathode-Ray Tube) display, a liquid crystal display, an organic EL (Electro Luminescence) display, an LED (Light-Emitting Diode) display, a plasma display, or any other display known in the art can be appropriately used. The display device 15 may also be a projector.

[0023] The text search process performed by the text search device 100 will be described below. In the following description, the text to be searched is assumed to be a combination of a question text and an answer text corresponding to the question text. A search for a combination of a question text and an answer text is also called a Q&A search. Hereinafter, such a combination will be called a question-and-answer pair. In a Q&A search, the text database 200 stores a question-and-answer collection that accumulates various question-and-answer pairs as a collection of texts. In a Q&A search, the receiving unit 111 receives a query text in question format as a query text. The search unit 112 acquires a question text, an answer text, or a question-and-answer pair as a search result text. In the following embodiment, for the sake of concrete explanation, the search result text is assumed to be a question-and-answer pair.

[0024] FIG. 3 is a diagram showing an example of the data structure of a question and answer collection. As shown in FIG. 3, a question and answer collection is a collection of question and answer pairs, which are combinations of a question statement and an answer statement to that question statement. Each question and answer pair is assigned an ID number to uniquely identify that question and answer pair. For example, the question and answer pair with ID "1" has the question statement "What is the deadline for filing a patent proposal?" and the answer statement "The deadline for filing a patent proposal is...". The question statement and answer statement are usually entered manually, but may also be entered automatically by a computer.

[0025] FIG. 4 is a diagram showing an example of a processing procedure of a text search process by the text search device 100. As shown in FIG. 4, the receiving unit 111 receives a character string (query sentence) related to a query from a user (step S401). An example of the processing procedure of step S401 is as follows. First, the user inputs the query sentence via an input device of the client terminal 300. The client terminal 300 transmits the query sentence to the text search device 100. The communication device 14 of the text search device 100 receives the query sentence from the client terminal 300, and the receiving unit 111 accepts the received query sentence.

[0026] The search unit 112 searches the question and answer collection stored in the text database 200 based on the query sentence received in step S401 (step S402). Then, the search unit 112 obtains, as search result sentences, question and answer pairs that match the query sentence from among the multiple question and answer pairs stored in the question and answer collection.

[0027] An example of the processing procedure of step S402 is as follows. First, the search unit 112 performs morphological analysis on the query sentence to extract words contained in the query sentence, and sets a search word from among the extracted words. For example, a word belonging to a specific part of speech, such as a noun or a verb, among the extracted words is set as the search word. As another example, search word candidates that can be used may be registered for each question-and-answer pair in the question-and-answer collection, and a search word candidate registered in the question-and-answer collection among the extracted words may be set as the search word. As another example, the search unit 112 may set the entire query sentence as the search word. Next, the search unit 112 searches the question-and-answer collection using the search word. Specifically, the search unit 112 identifies a question-and-answer pair that includes the search word from among multiple question-and-answer pairs included in the question-and-answer collection, and acquires the identified question-and-answer pair from the question-and-answer collection as a search result sentence. A question-and-answer pair that includes the search word may be a question-and-answer pair that includes a question sentence that includes the search word, a question-and-answer pair that includes an answer sentence that includes the search word, or a question-and-answer pair in which both the question sentence and the answer sentence include the search word.

[0028] After step S402 is performed, the central control unit 116 determines whether the number of search result sentences acquired in step S402 is equal to or greater than a threshold (step S403). Specifically, in step S403, it is determined whether the number of question-and-answer pairs acquired as search result sentences is equal to or greater than a threshold. If the number is equal to or greater than the threshold, the search results, i.e., the search result sentences, need to be narrowed down, and steps S404 to S407 are performed. If the number is less than the threshold, steps S404 to S407 are omitted. Typically, the threshold is set to two, but it may be set to three or more.

[0029] If it is determined in step S403 that the number of search result sentences is equal to or greater than the threshold (step S403: YES), the analysis unit 113 performs a sentence analysis on the search result sentences (step S404). The target of the sentence analysis may be the question sentence in the question-and-answer pair that is the search result sentence, or the answer sentence, or both the question sentence and the answer sentence. The target of the sentence analysis can be arbitrarily set by the user, etc. In the following embodiment, the target of the sentence analysis is assumed to be a question sentence. The question sentence in the question-and-answer pair that is the search result sentence will be referred to as the search result question sentence. The sentence analysis acquires the dependency relationships between words included in each search result question sentence. Details of the sentence analysis will be described later.

[0030] When step S404 is performed, the determination unit 114 determines one or more recommended words to be recommended for narrowing down the search result sentences to sentences desired by the user, based on the query sentence received in step S401, the search result sentences acquired in step S402, and the dependency relationships acquired in step S404 (step S405). The determination unit 114 determines, as the recommended words, any one of a word having a dependency relationship with the search word among the words constituting the search result sentences, a subject or object of a base word among the words constituting the search result sentences, and an adverbial modifier clause of a base word among the words constituting the search result sentences. The details of determining the recommended words will be described later.

[0031] When step S405 is performed, the display control unit 115 displays the search result sentence acquired in step S404 and the recommended words determined in steps S404 and S405 (step S406). As an example, the display control unit 115 displays an interactive screen on the display device of the client terminal 300, on which the search result sentence and the recommended words are displayed.

[0032] FIG. 5 is a diagram showing an example of the interactive screen I1 in step S406. As shown in FIG. 5, the interactive screen I1 includes a user-side message display field I11 and a robot-side message display field I12. The robot refers to a chatbot realized by the text search device 100. The interactive screen I1 displays user-side messages and robot-side messages in a time-series interactive format. The message display field I11 displays the query received in step S401. The message display field I12 displays the recommended words I121 determined in step S405 and the search result text I122 acquired in step S402. It is preferable that the search result question text be displayed as the search result text I122.

[0033] In the example of FIG. 5, the query sentence is "I would like to apply." The sentence search device 100 determines the search word "application" based on the query sentence, and acquires seven search result sentences I122 based on the search word. The number of search result sentences I122 is equal to or greater than a threshold, and three recommended words I121, "work approval," "patent proposal," and "visit," are determined. The search result sentences I122 and recommended words I121 thus obtained are displayed in the message display field I12 so that they can be individually selected. A question sentence is typically displayed as the search result sentence I122. However, this embodiment is not limited to this, and an answer sentence may be displayed, or both the question sentence and the answer sentence may be displayed.

[0034] When the number of search result sentences I122 displayed in the message display field I12 is large and the user wants to narrow down the search result sentences I122, the user selects recommended words I121 via an input device or the like of the client terminal 300. On the other hand, when the user finds a search result sentence that applies to the user's query from the search result sentences I122 displayed in the message display field I12, the user selects the desired search result sentence I122 via an input device or the like of the client terminal 300.

[0035] After step S406 is performed, the overall control unit 116 determines whether a recommended word or a search result sentence has been selected (step S407). If a recommended word has been selected (step S407: recommended word), the overall control unit 116 causes the search unit 112 to execute step S408. The search unit 112 narrows down the search result sentences based on the recommended word selected in step S407 (step S408). Specifically, in step S408, the search unit 112 extracts question-and-answer pairs that include the selected recommended word from among the multiple question-and-answer pairs. As an example, the search unit 112 identifies search result question sentences that include the selected recommended word from among the multiple question-and-answer pairs, and extracts the question-and-answer pair to which the identified search result question sentence belongs.

[0036] After step S408 is performed, the central control unit 116 determines whether the number of search result sentences after narrowing down is equal to or greater than a threshold (step S409). Specifically, in step S409, it is determined whether the number of search result sentences after narrowing down is equal to or greater than a threshold. If it is equal to or greater than the threshold, the search result sentences need to be further narrowed down, so steps S404 to S407 are performed again on the search result sentences after narrowing down. If it is less than the threshold, steps S404 to S407 are omitted. Typically, the threshold is set to two, but it may be set to three or more.

[0037] In this manner, steps S404 to S409 are repeated until a search result sentence is selected in step S407, or until it is determined in step S409 that the number of search result sentences after narrowing down is not equal to or greater than the threshold value.

[0038] Then, if it is determined in step S403 that the number of search result sentences is not equal to or greater than the threshold (step S403: NO), if a search result sentence is selected in step S407 (step S407: search result sentence), or if it is determined in step S409 that the number of search result sentences after narrowing down is not equal to or greater than the threshold (step S409: NO), the overall control unit 116 causes the display control unit 115 to execute step S410. The display control unit 115 displays a response sentence corresponding to the search result sentence (hereinafter, a corresponding response sentence) (step S410). As an example, the display control unit 115 displays an interactive screen on which the corresponding response sentence is displayed on the display device of the client terminal 300.

[0039] FIG. 6 is a diagram showing an example of the dialogue screen I2 in step S410. The dialogue screen I2 shown in FIG. 6 is a continuation of the dialogue screen I1 shown in FIG. 5. In the example of FIG. 6, it is assumed that "patent proposal" is selected from the three recommended words I121, "work approval," "patent proposal," and "visit." In this case, the selected recommended word "patent proposal" is displayed in the message display field I13. In step S408, the search unit 112 narrows down the seven question-and-answer pairs based on the selected recommended word "patent proposal." In the example of FIG. 6, it is assumed that one question-and-answer pair is extracted as a result of the narrowing down. In this case, since the number of question-and-answer pairs after narrowing down is less than the threshold value (two) (step S409: NO), the corresponding answer sentence from the narrowed down question-and-answer pair is displayed in the message display field I14 on the robot side.

[0040] As described above, in the first embodiment, the modification relationships between words contained in the search result sentences are used to determine recommended words, so it is possible to present accurate recommended words that take into consideration the structure of the search result sentences and the grammatical relationships between the words. Because accurate recommended words are displayed, the user can easily and efficiently find the search result sentences they want.

[0041] When step S410 is performed, the text search process ends.

[0042] Next, the sentence analysis process in step S404 and the recommended word determination process in step S405 will be described in detail. First, the sentence analysis process and the recommended word determination process in a comparative example will be described. In the comparative example, recommended words are determined based on the frequency of occurrence of words included in the search result sentences.

[0043] FIG. 19 is a diagram illustrating a sentence analysis process and a recommended word determination process according to a comparative example. As shown in FIG. 18, the query is "I want to apply," and the search word is "application." Based on the search word "application," seven search result sentences (question-answer pairs) containing the character string "application" are acquired. The search result questions in each question-answer pair are: "An error occurred when selecting the approver for my work approval application," "I would like to apply for cancellation in order to change the contents of my work approval application," "What if the approver changes the file attached to the patent proposal application?", "What is the deadline for the patent proposal application?", "An error occurred when selecting the approver for my visit request," "I would like to change the participants in my participation request," and "An error occurred when logging in to the portal site for the application and cannot log in." The symbols attached to each question-and-answer pair shown in FIG. 19 represent ID numbers.

[0044] The determination unit 114 according to the comparative example determines recommended words based on the frequency of occurrence of each word included in each search result question. The specific processing procedure is as follows: First, the determination unit 114 breaks down each search result question into words and counts the number of occurrences of each word included in all the obtained search result question sentences. The determination unit 114 then calculates an entropy score for each word based on the number of occurrences. The entropy score y is an example of an index value for evaluating the value of the word. As an example, the entropy score y is calculated according to the following formula (1) based on the number of occurrences and the number of search results. The number of search results means the number of search result questions. In this example, the number is seven. The entropy score of each word included in the search result question sentence is as shown in the table in FIG. 19.

[0045] y=-p·log(p) (1) p = number of occurrences / number of search results

[0046] In this example, the determination unit 114 determines recommended words based on the entropy score of each word included in the seven search result questions. Typically, any number of recommended words whose entropy scores are in a specified order are selected from the words included in the seven search result questions. For example, in the example of FIG. 19, three words "approver," "error," and "change," which have the same number one entropy score, are selected as recommended words. Note that, among the characters included in each search result question in FIG. 19, the words for which the entropy score is to be calculated are shown in bold, and the recommended words are shown in gray.

[0047] However, when sentences with similar intentions are disjointed or additional expressions that are not closely related to the intention expressed by the sentence are inserted in the sentence, as in the search result question sentence exemplified in Figure 19, it is not possible to select appropriate recommended words based solely on the frequency of occurrence. Also, there are cases where the user is interested in a search result question sentence with a low frequency of occurrence, and in this case, it is also not possible to select appropriate recommended words based solely on the frequency of occurrence.

[0048] Therefore, the determination unit 114 according to this embodiment selects recommended words based on the modification relationships between the words that make up the search result question. Below, several examples of the sentence analysis process and the recommended word determination process according to this embodiment will be described.

[0049] [Example 1] The determination unit 114 according to the first embodiment determines recommended words from among words contained in the search result sentences that have a modification relationship with the search word.

[0050] 7 is a diagram illustrating a text analysis process and a recommended word determination process according to Example 1. The query sentence, search words, and search result sentences shown in FIG. 7 are assumed to be the same as those in the comparative example shown in FIG.

[0051] When the search unit 112 acquires search result question sentences, the analysis unit 113 performs a sentence analysis on each search result question sentence to acquire dependency relationships between words included in each search result sentence. The dependency relationships acquired include at least the presence or absence of a connection between words and their grammatical relationship. A connection means that two words are in a relationship of a source and a destination of a connection. A grammatical relationship is a grammatical relationship between a source and a destination of a connection, and means, for example, the relationship between a modified word and a modifier (a modification relationship) or the relationship between a subject and a predicate.

[0052] FIG. 8 is a diagram illustrating an analysis result of the search result question sentence "An error was displayed when selecting an approver for the work approval application" by the text analysis according to the first embodiment. As illustrated in FIG. 8, the analysis unit 113 performs morphological analysis on the search result question sentence to break it down into words and determine the part of speech of each word, i.e., noun, particle, verb, auxiliary verb, etc. The analysis unit 113 identifies a search word from among multiple words included in the search result question sentence. In the case of FIG. 8, the search word "application" is identified. Next, the analysis unit 113 performs text analysis such as dependency analysis on the search result question sentence to identify words that have a dependency relationship with the search word "application." The text analysis may be performed using dependency analysis, syntactic analysis, and / or a graph-based ranking method. In the case of the search result question sentence illustrated in FIG. 8, "work approval" is identified as a noun modifying the search word "application," and "approver" is identified as a noun modified by the search word "application." In addition, if there is a word that is in a subject or predicate relationship with the search word "application", that word is also identified as a word that is in a modification relationship with the search word "application".

[0053] 8 for each search result question sentence, and identifies words that have a dependency relationship with the search word "application." The determination unit 114 acquires, from each search result question sentence, words that have a dependency relationship with the search word "application" as recommended word candidates.

[0054] The number of connections of words identified as recommended word candidates can be set arbitrarily. The number of connections corresponds to the number of dependency relationships between the final dependency destination and the search word when words in dependency relationships are sequentially identified starting from the search word. In the example of Figure 8, the number of connections is set to 1. As a result, "work approval application" and "application approver" are selected as recommended word candidates as words connected to the search word "application." "Selection" has a connection number of "2," so it is not selected as a recommended word candidate. Note that in Japanese, the presence or absence of a particle makes a minor difference and has little effect on the meaning of a word. Therefore, particles are omitted when selecting recommended word candidates.

[0055] The determination unit 114 calculates a score for each recommended word candidate based on the number of occurrences of the recommended word candidate. The score can be calculated based on the number of occurrences of the recommended word candidate that appears in the search result question sentence. The score based on the number of occurrences can be calculated as an entropy score according to the above formula (1). Here, the score may be calculated using the number of occurrences of the word directly, or may be calculated using the occurrence probability obtained by dividing the number of occurrences by the number of search results. As another example, a score may be calculated based on the number of question-answer pairs to which the recommended word candidate belongs. Figure 7 shows an example of scores for nouns that are in a modifying relationship with the search word "application".

[0056] The determination unit 114 then selects, from among the recommended word candidates, those whose scores satisfy a selection criterion as recommended words. The selection criterion may be, for example, that the score is in a specified ranking, or that the score is equal to or greater than a threshold. For example, as shown in FIG. 7, the three nouns with the highest scores, "application for work approval," "application for patent proposal," and "application for visit," are selected as recommended words. The selected recommended words, "application for work approval," "application for patent proposal," and "application for visit," are displayed on the client terminal 300 or the like by the display control unit 115 to narrow down the search result question sentences.

[0057] When displaying the recommended words, the display control unit 115 may display the search result sentence (question sentence) shown in FIG. 7. At this time, the display control unit 115 may display the relationship between the recommended words and words other than the recommended words among the words constituting the search result sentence. For example, the display control unit 115 may display the recommended words so that they are visually distinguished from other words. In FIG. 7, the recommended words are displayed in gray, but they may also be displayed using visual effects such as other colors or fonts. This allows the user to easily check the recommended words in the search result sentence. As another example, the display control unit 115 may display words identified as recommended word candidates so that they are visually distinguished from other words. In FIG. 7, the words identified as recommended word candidates are displayed in bold, but they may also be displayed using visual effects such as other fonts or colors.

[0058] The display control unit 115 may display the relationship between the recommended words and other words. For example, as shown in FIG. 8, the display control unit 115 may display the text analysis results of the search result text. Specifically, arrows indicating the presence or absence of connections between words included in the search result text, and character strings or symbols indicating the grammatical relationships between words included in the search result text may be displayed. Furthermore, the search words and recommended words included in the search result text may be displayed so as to be visually distinguished from other words. By displaying the text analysis results, the basis for selecting the recommended words can be presented to the user. This improves the convenience of the user in selecting recommended words and therefore the convenience of narrowing down searches.

[0059] An upper limit may be set for the number of recommended words. The upper limit is not particularly limited and may be set to any value from about 1 to 5. Of course, the upper limit may be set to 6 or more. The determination unit 114 selects recommended words whose number is less than the upper limit. Specifically, if the number of recommended word candidates that satisfy the selection criteria does not exceed the upper limit, the determination unit 114 selects the recommended word candidates as recommended words. On the other hand, if the number of recommended word candidates that satisfy the selection criteria exceeds the upper limit, the determination unit 114 changes the selection criteria so that the number of recommended word candidates does not exceed the upper limit. For example, the determination unit 114 may increase the score ranking or the threshold. Then, the determination unit 114 selects recommended word candidates that satisfy the changed selection criteria as recommended words. By setting an upper limit, the number of recommended words displayed can be reduced, thereby improving the convenience of the user in selecting recommended words.

[0060] According to the first embodiment, by utilizing not only the number of occurrences of each word but also the grammatical relationships between words, it is possible to select more appropriate words as recommended words for a narrowed search. This is because words that have a dependency relationship with the search word "application" are modifiers of that word. In general, adding modifiers to a word refines its meaning. By analyzing the dependency relationships, it is possible to find words that grammatically modify the search word, and therefore it is possible to present appropriate recommended words for the query entered by the user.

[0061] [Example 2] The determination unit 114 according to the second embodiment determines, as a recommended word, a subject or an object of a base word (hereinafter referred to as a base word) among words included in the search result sentence. The second embodiment will be described below. Note that the description of the parts common to the first embodiment, such as the display process of the recommended words, will be omitted.

[0062] Fig. 9 is a diagram illustrating a sentence analysis process and a recommended word determination process according to Example 2. The query sentence, search words, and search result sentences shown in Fig. 9 are assumed to be the same as those in Fig. 7. When the search unit 112 acquires search result sentences, the analysis unit 113 performs sentence analysis on each search result question sentence to identify the subject or object of the base word.

[0063] FIG. 10 is a diagram illustrating an analysis result of the search result question "What is the deadline for filing a patent proposal?" according to the second embodiment. As shown in FIG. 10, the analysis unit 113 performs morphological analysis on the search result question to break it down into words and determine the part of speech of each word. The analysis unit 113 identifies a search word from among multiple words included in the search result question. In the case of FIG. 10, the search word "application" is identified. Next, the analysis unit 113 performs sentence analysis such as dependency analysis on the search result question to identify words that have a dependency relationship with the search word "application." The sentence analysis may be performed using dependency analysis, syntactic analysis, and / or a graph-based ranking method. Next, the analysis unit 113 identifies a base word based on the dependency relationship. The base word, also called a "root," is the word located at the top of a tree structure representing the dependency relationship. The base word corresponds to the word that the sentence essentially asks. For example, in the case of the search result question "What is the deadline for filing a patent proposal?", the base word is "when" because it asks about the deadline, such as "until when." The analysis unit 113 then identifies the subject or object of the base word. The subject or object of the base word represents the subject that the sentence is asking about. Note that the base word may match the search word, and this is also allowed.

[0064] The analysis unit 113 performs the sentence analysis shown in FIG. 10 on each search result question sentence, and identifies the subject or object of the base word for each search result question sentence. The determination unit 114 then selects the identified subject or object of the base word as a recommended word candidate. In this case, the determination unit 114 simultaneously acquires and concatenates not only words that are dependent on the base word, but also words that modify the word. This is because such words are easier for humans to understand. For example, in addition to "deadline," which is the subject of the base word "when," "application," which modifies "deadline," and "patent proposal," which modifies "application," are concatenated, and the word "patent proposal application deadline" is selected as a recommended word candidate. As in the first embodiment, the number of concatenations can be set arbitrarily.

[0065] The method according to the second embodiment makes it possible to obtain, as recommended word candidates, the target words that each search result question essentially asks. In other words, it is possible to present what the user wanted to inquire about using recommended words, and if the user's query is displayed among the recommended words, the user can select that word to accurately narrow down the search result question sentences. The determination unit 114 then selects, from the recommended word candidates, those whose scores satisfy the selection criteria as recommended words. The selection criteria may be the same as those in the first embodiment. The recommended words are displayed on the client terminal 300 by the display control unit 115.

[0066] [Example 3] The determination unit 114 according to the third embodiment determines, as a recommended word, an adverbial modifier clause of a base word among words included in a search result sentence. The third embodiment will be described below. Note that a description of the parts common to the first embodiment, such as the display process of recommended words, will be omitted.

[0067] Fig. 11 is a diagram illustrating a sentence analysis process and a recommended word determination process according to Example 3. The query, search words, and search result sentences shown in Fig. 11 are assumed to be the same as those in Fig. 7. When the search result sentences are acquired by the search unit 112, the analysis unit 113 performs sentence analysis on each search result sentence to identify adverbial modifying clauses of the base word.

[0068] FIG. 12 is a diagram illustrating an analysis result of the search result question sentence "I would like to apply for cancellation in order to change the content of my work approval application" by the sentence analysis according to the third embodiment. As shown in FIG. 12, the analysis unit 113 performs morphological analysis on the search result question sentence to break it down into words and determine the part of speech of each word. The analysis unit 113 identifies a search word from among multiple words included in the search result question sentence. In the case of FIG. 12, the search word "application" is identified. Next, the analysis unit 113 performs sentence analysis such as dependency analysis on the search result question sentence to identify words that have a dependency relationship with the search word "application." The sentence analysis may be performed using dependency analysis, syntactic analysis, and / or a graph-based ranking method. Next, the analysis unit 113 identifies a base word based on the dependency relationship. The analysis unit 113 then identifies an adverbial modifier clause of the identified base word. The adverbial modifier clause is an expression such as "because of," "when," "in the case of," or "because," and often expresses "reason, background, or explanation of the situation" in a sentence. For example, in the case of the search result question "I would like to apply for cancellation in order to change the content of my work approval application," "cancellation application" is identified as the base word, and "change the content of my work approval application" is the adverbial modifying clause. As in the first embodiment, the number of links can be set arbitrarily.

[0069] The analysis unit 113 performs the sentence analysis shown in FIG. 12 on each search result question sentence, and identifies an adverbial modifier clause of the base word for each search result question sentence. The determination unit 114 then selects the identified adverbial modifier clause of the base word as a recommended word candidate. As the recommended word candidate, a word string of multiple words constituting an adverbial modifier clause may be selected, or a word belonging to a specific part of speech, such as a noun or a verb, may be selected from the multiple words. The determination unit 114 then selects, as a recommended word, one of the recommended word candidates whose score satisfies a selection criterion. The selection criterion may be the same as that in the first embodiment. The recommended words are displayed on the client terminal 300 by the display control unit 115.

[0070] According to the third embodiment, it is possible to obtain the "reason, background, and explanation of the situation" as to why a question such as the search result question was asked. In other words, it is possible to respond with recommended words as to "why" the user wanted to inquire, and if the recommended words indicate the situation the user is having trouble with, the search result question can be narrowed down appropriately by selecting that word.

[0071] [others] Examples 1, 2, and 3 can be combined as appropriate. That is, the determination unit 114 may select one or more candidates that meet a selection criterion as recommended words from the recommended word candidates (words in a modification relationship with the search word) selected in Example 1, the recommended word candidates (subjects or objects of the base word) selected in Example 2, and the recommended word candidates (adverbial modifier clauses of the base word) selected in Example 3. The selection criterion may be set to a word with a score that is in a high specified order, a score that is equal to or greater than a threshold, or the like.

[0072] Note that various modifications are possible to the first embodiment. For example, the use of recommended words is not limited to only presenting them to the user. As an example, the search unit 112 may automatically narrow down search result question sentences based on the recommended words without displaying the recommended words using the display control unit 115. As another modification, although the text database 200 is a computer separate from the text search device 100, it may be included in the text search device 100. Similarly, although the client terminal 300 is a computer separate from the text search device 100, it may be included in the text search device 100.

[0073] [Summary] The text search device 100 according to the first embodiment includes a receiving unit 111, a search unit 112, an analysis unit 113, and a determination unit 114. The receiving unit 111 receives a character string (query sentence) input by a user. The search unit 112 searches a text collection that stores multiple texts to be searched, based on the query sentence, and acquires one or more search result sentences that match the query sentence. The analysis unit 113 performs text analysis on the search result sentences and acquires dependency relationships between words that make up the search result sentences. The determination unit 114 determines one or more recommended words to be recommended for narrowing down the search result sentences to sentences desired by the user, based on the query sentence, the search result sentences, and the dependency relationships.

[0074] According to the above configuration, recommended words are determined taking into consideration the dependency relationships between words in the search result sentences, so that recommended words can be determined to be words that express the content that the user is actually asking, but that are not included in the query sentence, compared to recommended words selected based on the frequency of occurrence of each word in the search result sentences. By using such recommended words, interactive sentence searches can be performed more efficiently.

[0075] (Second embodiment) The determination unit 114 according to the second embodiment determines recommended words by combining the above-mentioned Examples 1 to 3 and Comparative Example. Second embodiment 2 will be described below. Second embodiment will be described below. In the following description, components having substantially the same functions as those in the first embodiment will be given the same reference numerals and will be described only when necessary. Furthermore, descriptions of parts common to the first embodiment, such as the processing procedure for dialogue search and the display processing of recommended words, will be omitted.

[0076] The determination unit 114 according to the second embodiment determines one or more first recommended word candidates based on the query sentence, the search result sentence, and the dependency relationships acquired by the analysis unit 113, determines one or more second recommended word candidates based on the query sentence, the search result sentence, and the occurrence frequency of each word constituting the search result sentence, and determines recommended words based on the first recommended word candidates and the second recommended word candidates. The first recommended word candidates are the recommended word candidates determined in accordance with the above-described Examples 1 to 3, as shown in FIG. 7, FIG. 9, and FIG. 11. The second recommended word candidates are the recommended word candidates determined in accordance with the above-described Comparative Example, as shown in FIG. 19.

[0077] There are various combination methods. For example, the determination unit 114 calculates a first score for the first recommended word candidate, calculates a second score for the second recommended word candidate, and calculates a total score of the first score and the second score for each recommended word candidate. Then, the determination unit 114 selects as recommended words those whose total scores meet the above selection criteria.

[0078] The total score TS can be calculated as a weighted sum of the first score S1 and the second score S2, as shown in the following formula (2). By setting the weights appropriately, it is possible to attach more importance to the recommended word candidates determined in Examples 1 to 3 than to the recommended word candidates determined in the comparative example. For example, it is preferable to set the weight w1 for the first score S1 higher than the weight w2 for the second score S2.

[0079] TS=w1·S1+w2·S2 (2)

[0080] Note that both the first score and the second score are not necessarily calculated for each word constituting the search result sentence. A word may be selected as both the first recommended word candidate and the second recommended word candidate, or may be selected as only one of the first recommended word candidate and the second recommended word candidate, or may not be selected as either the first recommended word candidate or the second recommended word candidate.

[0081] According to the second embodiment, recommended word candidates are determined based on the recommended word candidates determined in Examples 1 to 3 and the recommended word candidates determined in the comparative example. This makes it possible to determine recommended words by taking into account the number of occurrences of words constituting the search result sentences in the comparative example in addition to the dependency analysis in Examples 1 to 3.

[0082] (Third embodiment) The determination unit 114 according to the third embodiment deletes recommended words that are similar to other recommended words from among a plurality of recommended words. In other words, the determination unit 114 selects one representative recommended word from a group of recommended words that are similar to one another among a plurality of recommended words. The third embodiment will be described below. In the following description, components having substantially the same functions as those in the first embodiment will be assigned the same reference numerals and will be described only when necessary. In addition, descriptions of parts common to the first embodiment, such as the processing procedure for interactive search and the display processing of recommended words, will be omitted.

[0083] FIG. 13 is a diagram showing an example of operation according to the third embodiment. As shown in the table in the upper left of FIG. 13, five recommended word candidates are identified, and a score is calculated for each recommended word candidate. The refined search results indicate the ID numbers of the question-answer pairs (QA) to which the recommended word candidates belong. The ID numbers correspond to the ID numbers shown in FIGS. 7, 9, and 11. For example, the recommended word "patent proposal application approver" has a score of "0.3," and the refined search results are "1, 2, 3, 4, 5." In the example of FIG. 13, the number of hits in the refined search results is greater for recommended words with higher scores than for lower scores.

[0084] The recommended word candidates may contain the same word or spelling. In the third embodiment, recommended word candidates that partially contain the same word or spelling are considered to be similar to each other. If the top four scores of the five recommended words, "patent proposal application," "patent proposal application approver," "patent proposal application approver confirmation," and "patent proposal application approver change," are displayed as they are, only words containing the same word, "patent proposal application," will be presented as recommended words. As can be seen from the refined search results, these similarly related recommended word candidates belong to overlapping question-and-answer pairs. Therefore, presenting similarly related recommended word candidates as recommended words does not improve the effectiveness of narrowing down question-and-answer pairs.

[0085] For example, the refined search results for "patent proposal application" are "1, 2, 3, 4, 5, 6, 7," "patent proposal application approver" are "1, 2, 3, 4, 5," "patent proposal application approver confirmation" are "1, 3, 5," and "patent proposal application approver change" are "2, 4." Therefore, the range of question-answer pairs to which the recommended words belong remains the same whether only "patent proposal application" is presented as a recommended word or whether "patent proposal application," "patent proposal application approver," "patent proposal application approver confirmation," and "patent proposal application approver change" are presented as recommended words. In other words, "patent proposal application approver," "patent proposal application approver confirmation," and "patent proposal application approver change" have low contribution or value to narrowing down the search results. Instead, presenting the refined search result "8, 9," "application change deadline," as a recommended word would have a higher contribution to narrowing down the search results.

[0086] Therefore, the determination unit 114 according to the third embodiment deletes recommended word candidates that partially contain the same word or notation as a certain recommended word candidate. First, the determination unit 114 identifies two or more recommended word candidates that contain the same word or notation from among a plurality of recommended word candidates, and determines that the identified two or more recommended word candidates are similar to each other. Then, the determination unit 114 keeps one of the two or more recommended word candidates that are similar to each other and deletes the rest. Which recommended word candidate to keep may be determined based on any criteria. As an example, the determination unit 114 may keep the recommended word candidate with the highest score and delete the other recommended word candidates. This is because a higher score is considered to be more valuable as a recommended word. As another example, the determination unit 114 may keep the recommended word candidate that belongs to the largest number of question-and-answer pairs and delete the other recommended word candidates.

[0087] In the example of Figure 13, "patent proposal application," "patent proposal application approver," "patent proposal application approver confirmation," and "patent proposal application approver change" are all similar because they all contain the same word, "patent proposal application." Then, "patent proposal application," which has the largest number of question-and-answer pairs, is left, and the others, "patent proposal application approver," "patent proposal application approver confirmation," and "patent proposal application approver change," are deleted. As a result, "patent proposal application" and "application change deadline" are determined as recommended words and displayed.

[0088] FIG. 14 is a diagram illustrating another example of operation according to the third embodiment. FIG. 14 illustrates an example in which recommended words with lower scores produce more hits in the refined search results than recommended words with higher scores. As illustrated in FIG. 14, the top-ranked recommended word candidate is “patent proposal application approver,” and the second-ranked recommended word candidate is “patent proposal application.” Even if recommended words with lower scores produce more hits in the refined search results than recommended words with higher scores, the determination unit 114 may retain the recommended word candidate with the highest score and delete the other recommended word candidates. If only “patent proposal application approver” is retained, the question-and-answer pairs with ID numbers “6” and “7” obtained by filtering with “patent proposal application” cannot be included in the search results, but the other ID numbers “1” to “5” can be included in the search results. As a result, as illustrated in FIG. 14, “patent proposal application approver” and “application change deadline” are determined and displayed as recommended words.

[0089] FIG. 15 is a diagram illustrating another example of operation according to the third embodiment. FIG. 15 illustrates an example of the inclusion relationship of the notation strings of recommended word candidates, rather than the inclusion relationship of word elements in word strings as in FIGS. 13 and 14 . The notation of the extracted recommended word candidate, as in FIG. 15 , may be included in the notation of another word. In this case, the concept is similar to the example when elements of a word string are included. Because "patent proposal" and "proposal application" are included in "patent proposal application," which has the highest score, most of the search results obtained by narrowing down the recommended word candidates with scores of 22nd and 3rd will overlap. Therefore, "patent proposal" and "proposal application" can be considered to have low value. On the other hand, "proposal application approval" may produce different narrowing results than "patent proposal application" because some parts are not included in "patent proposal application." Therefore, it is determined to have value. As a result, "patent proposal application" and "proposal application approval" are determined to be recommended words and displayed, as shown in FIG. 15 .

[0090] As described above, according to the third embodiment, it is possible to prevent recommended word candidates that are similar to each other from being presented as recommended words. This reduces the possibility that words that contribute to narrowing down the search results in the same way are displayed, making it possible to efficiently present recommended words within the limited number of words that can be displayed.

[0091] (Fourth embodiment) The determination unit 114 according to the fourth embodiment estimates a category of interest, which is a category of interest for the user, from among a plurality of categories corresponding to a plurality of search result sentences, and determines recommended words based on the dependency relationship between the category of interest, the query sentence, and the search result sentence. The fourth embodiment will be described below. In the following description, components having substantially the same functions as those in the first embodiment are given the same reference numerals and will be described only when necessary. Furthermore, descriptions of parts common to the first embodiment, such as the processing procedure for dialogue search and the display processing of recommended words, will be omitted.

[0092] FIG. 16 is a diagram showing an example of the data structure of a collection of questions and answers stored in the sentence database 200 according to the fourth embodiment. As shown in FIG. 16, the collection of questions and answers according to the fourth embodiment is a collection of question and answer pairs, each of which is a combination of a question, an answer to the question, and a category of the question and answer pair. A category is attribute information assigned to classify the content of a question and answer pair. A category is information that indicates the genre or question category of a question and answer pair, and roughly classifies the content to which the question and answer pair relates. For example, a question and answer pair with ID "2" has a question "I would like to change my working hours," an answer "On the working hours management screen...," and a category of "Work approval." The question, answer, and category are usually entered manually, but may also be entered automatically by a computer.

[0093] Fig. 17 is a diagram showing an example of operation according to the fourth embodiment. As shown in the table in the upper left of Fig. 17, six recommended word candidates are identified, and a score is calculated for each recommended word candidate. The determination unit 114 identifies the category of question and answer pairs to which the recommended word candidates belong, based on the question and answer collection shown in Fig. 16. The determination unit 114 selects categories of question and answer pairs that should be displayed and categories of question and answer pairs that should not be displayed, based on the categories.

[0094] In the example of Figure 17, question and answer pairs are classified into three categories: "patent proposal," "work approval," and "visit." Typically, a user will search for a question and answer pair that belongs to one of the categories that classifies the user's problem. Therefore, it is pointless to display recommended words in a category other than the category in which the user is interested. Here, the category in which the user is interested is considered to be a category that accounts for a high proportion of search result sentences. Therefore, when recommended word candidates like those in the example of Figure 17 are obtained, the category in which the user is asking a question is first estimated. Figure 17 shows an example in which "patent proposal," the category in which question and answer pairs were most frequently searched, is used.

[0095] If no category is selected, "patent proposal application approver," "application deadline," "attendee addition," and "working hours change" will be displayed as recommended words, as shown in Figure 17. However, if a user is searching for a question-and-answer pair related to a patent proposal, "attendee addition" and "working hours change," which are in a different category, will never be selected. Therefore, the determination unit 114 estimates the category of interest of the user from the categories of question-and-answer pairs to which the recommended word candidates belong. Methods for estimating the category of interest include selecting the category to which the most question-and-answer pairs in the search result sentence belong, or selecting the category to which many of the top candidate question-and-answer pairs belong.

[0096] In the example of FIG. 17, of the six QA search results, four are in the category "patent proposal," one is in the category "work approval," and one is in the category "visit." From these results, it is highly likely that the user was looking for a question-and-answer pair related to the category "patent proposal." Therefore, the determination unit 114 presumes that "patent proposal," which is the category with the largest number of results, is the category of interest. The determination unit 114 selects the recommended word candidates associated with the category of interest "patent proposal" as recommended words. As a result, "patent proposal application approver," "application deadline," "file attachment," and "co-applicant addition" are selected and displayed as recommended words, as shown in FIG. 17.

[0097] According to the fourth embodiment, even when the search result text contains a mixture of question-and-answer pairs from multiple categories, it is possible to estimate the category that the user is likely to be inquiring about and display recommended words related to that category. By displaying only recommended words that are narrowed down to the category that the user is likely to be looking for, it is possible to make the display of recommended words more efficient.

[0098] (Fifth embodiment) The determination unit 114 according to the fifth embodiment determines the ranking of a plurality of search result sentences based on the degree of correspondence between the words constituting each of the search result sentences and the word string constituting the query sentence, and determines recommended words based on the determined ranking and the dependency relationship between the query sentence and the search result sentence. More specifically, the determination unit 114 calculates a score for each search result sentence based on the ranking of the search result sentence, and determines recommended words based on the calculated score as well as the dependency relationship between the query sentence and the search result sentence. This allows for prioritizing the selection of recommended words from words contained in the top search result sentences that are expected to be more likely to be correct, thereby improving the efficiency of conversational search. The fifth embodiment will be described below. In the following description, components having substantially the same functions as those in the first embodiment will be denoted by the same reference numerals and will be described only when necessary. Furthermore, descriptions of parts common to the first embodiment, such as the conversational search processing procedure and the display processing of recommended words, will be omitted.

[0099] Fig. 18 is a diagram showing an example of operation according to the fifth embodiment. In the example of Fig. 18, the query is "An error occurred in the approver and I cannot apply," the search words are "approver," "error," and "application," and eight search result questions are acquired based on the search words: "An error occurred when selecting the approver when proposing a patent and I cannot proceed," "An error occurs when changing the approver," "I would like to add multiple approvers, but an error occurs and I cannot add them," "An error occurred in the proposal department code," "An error occurred in the file attachment of the patent proposal," "An error is displayed when I press the submit button for the patent proposal," "An error is displayed on the screen for adding co-applicants," and "An error message stating an invalid date is displayed when I specify the deadline for the patent proposal."

[0100] When the search result sentences are acquired, the determination unit 114 determines the rank of each search result question sentence in terms of the degree of match with the query sentence. There are various methods for determining the rank, and no particular limitation is imposed. For example, the determination unit 114 counts the number of search words included in each search result question sentence and determines the rank based on this count. The number added to the left end of each search result question sentence in FIG. 18 indicates the rank calculated by the determination unit 114.

[0101] Since there are no question-and-answer pairs that match three of the search words, the question-and-answer pairs that match two of the search words rank first through third, and question-and-answer pairs that match only one of the search words rank fourth and third. The recommended words generated from this search result are shown in the right figure of 18. Here, the recommended word candidate "error display" appears four times, the most frequently, and is ranked first among the recommended word candidates. While the policy of determining the ranking of recommended word candidates based on the number of appearances is often adopted, there is a problem with the example in Figure 18. This is because the recommended word "error display" for the question-and-answer pairs ranked fourth and lower that match only one of the search words is given priority over the recommended words for the question-and-answer pairs ranked first through third that match two of the search words.

[0102] When a user specifies multiple search words, question-and-answer pairs that contain many of the search words are often the ones the user is looking for. To achieve this prioritization of recommended words, a score weight table is prepared for each ranking of question-and-answer pairs. The score weight table is a table that associates the ranking of search result questions with weight values. The higher the ranking, the larger the weight value should be set. For example, the weight value may be set as follows: ranking "1" is weighted to "10," ranking "2" is weighted to "8," ranking "3" is weighted to "6," ranking "4" is weighted to "4," ranking "5" is weighted to "2," and ranking "6" is weighted to "1." The score weight table is stored in the sentence database 200 or the storage device 12.

[0103] The determination unit 114 identifies the ranking of the search result question sentence to which each recommended word belongs, identifies a weight value corresponding to the identified ranking from the score weight table, and calculates a score based on the product of the ranking and the weight value. The ranking of recommended words is as shown in the right diagram of Figure 18. "Error Display," which appears four times, is ranked fourth, sixth, seventh, and eighth in the question-and-answer pair to which it belongs, so it receives 2 points, 1 point, 1 point, and 1 point, respectively, for a total of 5 points. Meanwhile, the recommended word "Select Approver" included in the first place in the question-and-answer pair receives 10 points, and the recommended word "Change Approver" included in the second place in the question-and-answer pair receives 8 points. As a result, when three recommended words are displayed as shown in Figure 18, "Select Approver," "Change Approver," and "Add Approver" are selected and displayed as recommended words.

[0104] As described above, the fifth embodiment makes it easier for recommended word candidates included in sentences with high-ranking search results to be selected as recommended words, which makes it possible to preferentially display recommended words related to question-and-answer pairs that are likely to be what the user is looking for, thereby improving the efficiency of displaying recommended words.

[0105] Thus, it becomes possible to select recommended words that are useful for narrowing down search results in an interactive text search.

[0106] Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be embodied in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, and are also included in the scope of the invention and its equivalents as defined in the claims. [Explanation of symbols]

[0107] 1...text search system, 11...processing circuit, 12...memory device, 13...input device, 14...communication device, 15...display device, 100...text search device, 111...reception unit, 112...search unit, 113...analysis unit, 114...determination unit, 115...display control unit, 116...overall control unit, 200...text database, 300...client terminal.

Claims

1. A reception unit for receiving an inquiry sentence related to a question; a search unit that searches through a plurality of stored answers to the question and obtains one or more answers to the question that match the query sentence as a search result sentence; an analysis unit that performs a sentence analysis on the search result sentence to obtain dependency relationships between words constituting the search result sentence; a determination unit that determines one or more recommended words based on the query sentence, the search result sentence, and the dependency relationships between the words; A text search device comprising:

2. each of the plurality of question and answer sentences has a combination of a question sentence and an answer sentence to the question sentence, The reception unit receives a query sentence in the form of a question as the query sentence, the search unit acquires the question sentence and / or the answer sentence as the search result sentence.

2. The text search device according to claim 1.

3. The text search device according to claim 1 , further comprising a display control unit that displays the recommended words on a display device.

4. The text search device according to claim 3 , wherein the display control unit displays the search result text together with the recommended words on the display device.

5. The text search device according to claim 4 , wherein the display control unit displays a relationship between the recommended words and words that constitute the search result text and are different from the recommended words.

6. The determination unit is determining one or more first recommended word candidates based on the received character string, the search result sentence, and the dependency relationship; determining one or more second recommended word candidates based on the query sentence, the search result sentence, and the frequency of occurrence of each word constituting the search result sentence; determining the recommended word based on the first recommended word candidate and the second recommended word candidate; 2. The text search device according to claim 1.

7. The text search device according to claim 1 , wherein the determining unit deletes, from the plurality of recommended words, recommended words that are similar to other recommended words.

8. The determination unit is estimating a category of interest that is a category in which the user is interested from among a plurality of categories corresponding to the plurality of answers to the questions, respectively; determining the recommended words based on the interest category, the query sentence, the search result sentence, and the dependency relationships; 2. The text search device according to claim 1.

9. The determination unit is determining a ranking of the plurality of answers to the question based on a degree of agreement between a word constituting each of the plurality of answers to the question and a word string constituting the query sentence; determining the recommended words based on the ranking, the query sentence, the search result sentence, and the dependency relationship; 2. The text search device according to claim 1.

10. The text search device according to claim 1 , wherein the analysis unit uses dependency analysis, syntactic analysis, and / or a graph-based ranking method as the text analysis.

11. The text search device of claim 1, wherein the determination unit determines as the recommended word one of a word that is in a modifying relationship with a word included in the query sentence among the words constituting the search result sentence, a subject or object of a base word among the words constituting the search result sentence, and an adverbial modifier clause of the base word among the words constituting the search result sentence.

12. A text search device comprising: Accepting inquiries related to questions, searching the stored plurality of answers to the question and obtaining one or more answers to the question that match the query sentence as a search result sentence; performing a sentence analysis on the search result sentence to obtain dependency relationships between words constituting the search result sentence; determining one or more recommended words based on the query sentence, the search result sentence, and the dependency relationships between the words; The text search method comprises:

13. On the computer, A function for accepting a query related to a question; a function of searching through a plurality of stored answers to the question and acquiring one or more answers to the question that match the query sentence as a search result sentence; A function of performing a sentence analysis on the search result sentence to obtain dependency relationships between words constituting the search result sentence; determining one or more recommended words based on the query sentence, the search result sentence, and the dependency relationships between the words; A text search program that achieves this.

Images