JP5718405B2 - Utterance selection apparatus, method and program, dialogue apparatus and method - Google Patents

Description

  The present invention relates to an utterance selection device, method, and program, and an interaction device and method, and more particularly, to an utterance selection device, method, and program for selecting an utterance used in an interaction system, and the selected utterance. The present invention relates to an interactive apparatus and method for interacting with a user.

  It is important for a chat dialogue system that performs dialogues with users on various topics to grasp the topics in the current dialogue and respond to the topics. However, it is difficult in terms of cost to prepare a list of utterances that the system should perform in advance for all topics. Therefore, for words included in the user's utterance indicating the topic in the current dialogue, a related sentence is extracted from the Internet using a web search engine, and one or more sentences are extracted as utterances from the extracted sentence. An interactive system that performs the above has been proposed (see Non-Patent Document 1, for example).

  There is also an interactive system that uses Twitter data, which is a microblog service, as a search target instead of a general web document, acquires related tweets based on search keywords obtained from user utterances, and uses them as system utterances. It has been proposed (see, for example, Non-Patent Document 2).

Shibata, M., Nishiguchi, T., and Tomiura, Y. (2009). "Dialog system for open-ended conversation using web documents." Infomatica, 33 (3), pp. 277-284. F. Bessho, T. Harada, and Y. Kuniyoshi, "Dialog system using real-time crowdsourcing and Twitter large-scale corpus," in Proc. SIGDIAL, 2012, pp.227-231.

  In the techniques described in Non-Patent Document 1 and Non-Patent Document 2, search results obtained for a topic word are used as utterances, but search results obtained from web pages and Twitter include various topics. Therefore, it is not always possible to obtain appropriate search results for a given topic. Therefore, there is a possibility that a low quality utterance will be performed.

  The present invention can perform a high-quality response using an utterance selection device, method, and program capable of selecting a high-quality utterance at low cost as an utterance used in a dialogue system, and the selected utterance. An object of the present invention is to provide an interactive apparatus and method.

In order to achieve the above object, an utterance selection device according to a first aspect of the present invention provides a related word related to a word on the basis of a language pattern composed of a word and part-of-speech sequence for the word. A related word acquisition unit that acquires a related word related to at least one extracted word extracted from a user's utterance from a dictionary that is collected and correlated, and from the utterance database including a plurality of utterances, the extracted word and the An utterance search unit that searches for an utterance corresponding to a search query including a related word related to the extracted word acquired by the related word acquisition unit, and at least one utterance is selected from the utterances searched by the utterance search unit And an utterance selection unit.

According to the utterance selection device according to the present invention, the related word acquisition unit collects related words related to the word based on a language pattern including a word and a part-of-speech sequence that is predetermined together with the word. Then, a related word related to at least one extracted word extracted from the user's utterance is acquired from the associated dictionary. Then, the utterance search unit searches the utterance database corresponding to the search query including the extracted word and the related word related to the extracted word acquired by the related word acquiring unit from the utterance database including a plurality of utterances. Further, the utterance selection unit selects at least one utterance from the utterances searched by the utterance search unit.

  As described above, since the utterance is selected from the utterance searched by including the related word related to the extracted word in the search query, a high-quality utterance can be selected at low cost as the utterance used in the dialogue system.

  In addition, the related word acquisition unit acquires a related word related to the extracted word using a dictionary in which the related word is associated with a degree of association with the word as the dictionary, and the utterance search unit Among the related words related to the extracted word, related words having the relevance level equal to or higher than a predetermined value can be included in the search query. Further, the degree of association may be the degree of co-occurrence of the word and the related word, or the degree of appearance of the related word in the utterance subsequent to the utterance including the word. As a result, it is possible to select an utterance that is more relevant to the topic represented by the extracted word.

  Further, the utterance selection unit excludes utterances including words other than stop words that are predetermined as words that do not affect the extracted words, the related words, and utterance contents, from utterances searched by the utterance search unit. can do. Thereby, high quality speech can be selected.

  Further, the utterance selection unit can cluster the utterances searched by the utterance search unit based on the related word or the stop word included in the utterance, and select a representative utterance of each cluster. . Thereby, diversity can be given to the selected utterance.

  Further, the utterance selection unit clusters the utterances searched by the utterance search unit based on the related word or the stop word included in the utterance, and the number of utterances included in each cluster is a predetermined number or more. The representative utterance can be selected. As a result, it is possible to select an utterance that tends to become a topic for the extracted word.

  Further, the utterance selection unit can acquire at least the utterance immediately before the user, and can select an utterance including a word included in the immediately preceding utterance out of the utterances searched by the utterance search unit. Thereby, it is possible to select an utterance for realizing a lubricious dialogue.

  The dialogue apparatus according to the second aspect of the invention includes an utterance accepting unit that accepts a user's utterance and an utterance output unit that outputs one of the utterances selected by the utterance selection device. Thereby, a high-quality response can be performed.

The utterance selection method according to the third aspect of the invention relates to a related word related to a word based on a language pattern consisting of a word and a part of speech sequence that is previously determined by the related word acquisition unit together with the word. from speech databases from the collected association dictionary, comprising the steps of: obtaining related terms associated with at least one extracted words are extracted from the utterance of the user, the speech retrieval unit, a plurality of utterances, said extraction word And a step of searching for an utterance corresponding to a search query including a related word related to the extracted word acquired by the related word acquisition unit, and an utterance selection unit at least 1 from the utterance searched by the utterance search unit Selecting one utterance.

  The dialogue method according to the fourth invention includes a step in which the utterance receiving unit receives the user's utterance and a step in which the utterance output unit outputs one of the utterances selected by the utterance selection method. Is the method.

  An utterance selection program according to a fifth aspect of the invention is a program for causing a computer to function as each unit constituting the utterance selection device.

  As described above, according to the utterance selection device, method, and program of the present invention, an utterance used in a dialog system to select an utterance from utterances that are searched by including a related word related to an extracted word in a search query. As a result, it is possible to select a high quality speech at a low cost. In addition, according to the interactive apparatus and method of the present invention, an effect that a high-quality response can be performed using the utterance selected as described above can be obtained.

It is a block diagram which shows the functional structural example of the speech selection apparatus which concerns on 1st Embodiment. It is a figure which shows an example of the table for calculating an association degree. It is a figure which shows an example of the related word about "ramen." It is a figure which shows an example of the table for calculating the relevance using dialog data. It is a figure which shows an example of the related word about "ramen" which recalculated the relevance degree. It is a figure which shows an example of a search query. It is a figure which shows an example of a search result. It is a figure which shows an example of a clustering result. It is a flowchart which shows the speech selection process in 1st Embodiment. It is a block diagram which shows the functional structural example of the dialogue apparatus which concerns on 2nd Embodiment.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<First Embodiment>
The utterance selection device 10 according to the first embodiment is configured by a computer including a CPU, a RAM, and a ROM that stores a program for executing an utterance selection process described later. Functionally, the computer includes a word input unit 11, a related word acquisition unit 12, an utterance search unit 13, a filtering unit 14, a clustering unit 15, and an utterance candidate selection unit 16, as shown in FIG. Can be expressed as The filtering unit 14, the clustering unit 15, and the utterance candidate selection unit 16 are examples of the utterance selection unit of the present invention.

  The word input unit 11 receives, as an input, a word composed of one or more morphemes corresponding to the current topic extracted from the user's utterance. The word extracted from the user's utterance is referred to as “extracted word”. For example, a word whose part of speech included in the user's utterance is a noun can be extracted as an extracted word. In addition to the extracted word, the word input unit 11 receives a dialogue history including at least the utterance immediately before the user. The word input unit 11 passes the received extracted word to the related word acquisition unit 12 and stores the dialogue history in the dialogue history storage unit 21.

  The related word acquisition unit 12 acquires related words related to the extracted word delivered from the word input unit 11 with reference to the related word dictionary 22 in which related words related to each of the plurality of words are recorded. In the related word dictionary 22, the part of speech related to the heading word is recorded in association with the heading word as the related word related to the heading word, each of the noun, verb and adjective words. For example, the related terms can be collected from large-scale blog data using a predetermined language pattern. More specifically, the noun has the pattern “B (noun) of A (noun)”, the verb has the pattern “A (noun) B (verb)”, and the adjective “A (noun) B ( Pattern matching by regex on the blog data subjected to morphological analysis using the pattern “adjective” or “A (noun) is B (adjective)”, and the related part of speech for the heading word A is the noun, verb, and adjective Words B can be listed and collected as related words.

  In addition, with respect to a word that is a headword in the related word dictionary 22, in order to include only related words having a high degree of relevance in the related word dictionary 22, related words may be filtered using a log likelihood ratio. . For example, when there is a related word r for a certain heading word w, the degree of association indicating how much the related word r is related to the heading word w is calculated according to the table shown in FIG. In the table shown in FIG. 2, N is the total number of headword-related word pairs (total number of appearances). c12 is the total number of pairs in which the headword is w and the related word is r. c1 is the number of pairs whose headword is w, and c2 is the number of pairs whose related word is r. From here, how much the value of c12 is biased is calculated using the log likelihood ratio. The log likelihood ratio can be calculated by, for example, the following equation (1).

Here, O _i is the value of each cell in the table shown in FIG. 2, that is, c12, c1-c12, c2-c12, and N-c2-c1 + c12. E _i is the expected value of O _i . For example, the expected value of c12 can be c1 * (c2 / N). The log likelihood ratio obtained in this way is used as the relevance. Since the log-likelihood ratio follows a chi-square distribution, if the log-likelihood ratio is 10.83 or more, it is considered that the headword and the related word are statistically significant (p <0.001). It is done. Therefore, for each headword, only related words having a log likelihood ratio exceeding 10.83 can be recorded. The relevance may be a scale representing the co-occurrence of a word and a related word, and in addition to the log likelihood ratio, a chi-square value, a T score, a mutual information amount, and the like may be used. FIG. 3 shows an example of a related word for the heading word “ramen”. In FIG. 3, the top 20 items are displayed in the log likelihood ratio order as the degree of association.

  Further, the degree of co-occurrence between words in the dialog data may be used as the degree of association suitable for the dialog system. For example, if the word w1 appears in a certain utterance and the word w2 tends to appear in a subsequent utterance, it can be said that the word w1 and the word w2 are strongly related. Therefore, as described above, for the related words filtered using the log likelihood ratio calculated using the appearance frequency of the headword and related words, the relevance is recalculated using the table shown in FIG. . In the table shown in FIG. 4, M is the total number of adjacent pairs of utterances in the dialog data. c12 is the total number of pairs in which the word w1 appears in the previous utterance of the utterance pair and the word w2 appears in the subsequent utterance of the utterance pair. c1 is the total number of pairs in which the word w2 appears in the subsequent utterance of the utterance pair, and c2 is the total number of pairs in which the word w1 appears in the previous utterance of the utterance pair. Using the value of each cell in the table shown in FIG. 4, the log likelihood ratio calculated by equation (1) is recalculated as the relevance.

  Since the utterance selection device 10 in the first embodiment is for selecting an utterance to be used in the dialogue system, the case where the degree of association calculated from the dialogue data as described above is used will be described. Use Twitter data, which is a microblog service, as a corpus of dialogue data, collect tweets in a reply relationship as a pair, create a database, and calculate the log-likelihood ratio indicating the degree of co-occurrence as the relevance from this database be able to. FIG. 5 shows an example of a related word obtained by recalculating the relevance level for the headword “ramen”. In FIG. 5, the top 20 items are displayed in the log likelihood ratio order, which is the degree of association calculated from the conversation data.

  In the first embodiment, a related word dictionary 22 having about 2 million words extracted from blog data as headline words and having a log likelihood ratio as a relevance level for each related word for the headline word is used. . The related word acquisition unit 12 transfers the related word of the extracted word acquired with reference to the related word dictionary 22 together with the extracted word transferred from the word input unit 11 to the utterance search unit 13.

  The utterance search unit 13 searches the utterance database (DB) 23 using the extracted word and the related word delivered from the related word acquisition unit 12. For example, in a utterance that always includes an extracted word, if a related word is included, the logarithm of the log likelihood ratio of the related word is added to the score of the search result. Get utterances. For example, L can be 100. The use of the logarithm of the log likelihood ratio is a measure for the case where the log likelihood ratio may become very large. By searching in this way, an utterance including as many related words as possible can be acquired from the utterance DB 23.

  As the utterance DB 23, for example, it is possible to use data obtained by collecting Twitter data, dividing the collected data into sentences by punctuation marks and symbols, and registering one sentence as one record indicating one utterance. The utterance DB 23 can hold a transposed index so that a full text search is possible. The utterance DB 23 can be constructed by using a general procedure for constructing a full-text search engine database. For example, it can be constructed by using free software Lucene or Namazu.

  FIG. 6 shows an example of a search query using the extracted word “ramen” and its related words. The example of FIG. 6 uses a search query format by Lucene. Each blank delimiter is a search condition, and “content: noodle ^ 9.170959” contains the word “noodle” in the content of the utterance (content), the score of 9.170959 (“noodle” (Logarithm of log likelihood ratio) is given. A search condition that starts with a plus, such as “+ content:“ ramen ”” represents an indispensable condition, which represents a condition that the extracted word “ramen” is necessarily included.

  FIG. 7 shows an example of the top 10 search results when the utterance DB 23 is searched using the search query shown in FIG. It can be seen that only utterances particularly related to the extracted word “ramen” are retrieved. The “search result score” in FIG. 7 is the sum of the logarithm of the log likelihood ratio of the related words included in the acquired utterance. The utterance search unit 13 passes the acquired utterance set to the filtering unit 14.

  The filtering unit 14 filters the utterance set delivered from the utterance search unit 13. Each of the utterances included in the utterance set acquired by the utterance search unit 13 is an utterance having a relatively high relationship with the extracted word, but is a process for leaving only the utterance related to the extracted word.

  Specifically, the filtering unit 14 performs morphological analysis on each utterance included in the utterance set. For morphological analysis, NTT's JTAG, free software ChaSen, Mecab, or the like can be used as a morphological analyzer. The filtering unit 14 extracts content words (nouns, verbs, adjectives, and unknown words) from the morphological analysis result, and classifies the extracted content words into extracted words, related words, stop words, and other words. . Then, the filtering unit 14 excludes utterances including words classified as other words from the utterance set.

  Here, verbs and adjectives are used as words corresponding to stop words. Unlike prescriptive words (nouns), predicates are words that do not significantly affect the utterance content. In addition to the precautions, auxiliary nouns (such as “thing” and “thing”) that do not affect the utterance content may be included in the stop word. The filtering unit 14 passes the filtered utterance set to the clustering unit 15.

  When the clustering unit 15 includes each utterance included in the utterance set passed from the filtering unit 14, the part of speech included in each utterance is a noun related word, or the utterance does not include a noun related word Cluster by stop word. FIG. 8 shows an example of the result of clustering utterances acquired and filtered for the extracted word “ramen”. In the clustering result of FIG. 8, “eating” is included in the related word because the part of speech of “eating” is set to “noun” due to a morphological analysis error. Sa variant nouns (behavioral nouns) are treated as verbs.

  If an utterance does not contain a related word whose part of speech is a noun, clustering is performed using stop words included in the utterance. For example, a cluster composed of utterances whose semantic contents are not changed with respect to the extracted words, such as “good ramen” and “good ramen” can be obtained. These are categorized into clusters with the phrases “delicious” and “delicious” respectively. The clustering unit 15 passes the clustering result to the utterance candidate selection unit 16.

  The utterance candidate selection unit 16 creates three utterance sets based on the clustering result transferred from the clustering unit 15. First, in a cluster including related words, when a typical utterance of the cluster includes a content word included in the utterance immediately before the user stored in the dialogue history storage unit 21, the utterance Is an utterance set obtained by extracting. Hereinafter, this utterance set is referred to as “set 1”. The content words included in the utterance immediately before the user are extracted from the morpheme analysis result by reading the utterance immediately before the user stored in the dialogue history storage unit 21 and performing morphological analysis on the utterance immediately before the user. can do. The second is an utterance set obtained by extracting representative utterances one by one from a cluster having a plurality of utterances of two or more utterances. Hereinafter, this utterance set is referred to as “set 2”. The third is an utterance set obtained by extracting one representative utterance from each of all clusters. Hereinafter, this utterance set is referred to as “set 3”.

  Any method may be used as a representative utterance extraction method for each cluster. For example, the shortest utterance in the cluster can be set as the representative utterance of the cluster. This is because the shortest utterance is the simplest sentence and is considered to contain the contents of the cluster densely. In addition to using the shortest utterance, a representative utterance of each cluster may be extracted based on a standard such as TFIDF (TF: Term Frequency, IDF: Inverse Document Frequency) generally used for important sentence selection. Good.

  The utterance candidate selection unit 16 creates three utterance sets and, if the set 1 is not an empty set, selects the set 1 as an utterance candidate set. If set 1 is an empty set, set 2 is selected as the utterance candidate set. If neither set 1 nor set 2 is an empty set, set 3 is selected as the utterance candidate set. By selecting the utterance candidate set in this order, it is possible to return a response according to the immediately preceding user utterance, and even when the set 1 is empty, a plurality of utterances are included as in the set 2 By preferentially selecting an utterance from a certain cluster, it is possible to select an utterance having a content often spoken about the topic (extracted word). When set 1 and set 2 are empty, by using set 3, any utterance related to the extracted word can be selected.

  The utterance candidate selection unit 16 outputs the selected utterance candidate set. When all of the sets 1 to 3 are empty sets, the utterance candidate selection unit 16 outputs the empty set as a selection result.

  Next, the operation of the utterance selection device 10 according to the first embodiment will be described. When the extracted word and the conversation history are input to the utterance selection device 10 in a state where the related word dictionary 22 and the utterance DB 23 are prepared, the utterance selection processing shown in FIG. 9 is executed in the utterance selection device 10.

  In step 100 of the utterance selection process, the word input unit 11 receives the input extracted word and dialogue history, passes the received extracted word to the related word acquisition unit 12, and stores the dialogue history in the dialogue history storage unit 21. To do.

  Next, in step 102, the related word acquisition unit 12 refers to the related word dictionary 22 in which the related words for each of the plurality of words are recorded for the related words of the extracted words passed from the word input unit 11. get. The related word acquisition unit 12 transfers the acquired related word to the utterance search unit 13 together with the extracted word transferred from the word input unit 11.

  Next, in step 104, the utterance search unit 13 searches the utterance DB 23 using the extracted word and the search query based on the related word delivered from the related word acquisition unit 12, and includes the extracted word. L utterances are acquired in the order of scores according to the degree of association of related words included in the utterance. The utterance search unit 13 passes the acquired utterance set to the filtering unit 14.

  Next, in step 106, the filtering unit 14 performs morphological analysis on each utterance included in the utterance set passed from the utterance search unit 13, extracts content words from the morphological analysis results, and extracts the extracted content. The words are classified into extracted words, related words, stop words, and other words. Then, the filtering unit 14 excludes utterances including words classified as other words from the utterance set. The filtering unit 14 passes the filtered utterance set to the clustering unit 15.

  Next, in step 108, the clustering unit 15 selects each utterance included in the utterance set passed from the filtering unit 14, the part of speech included in each utterance is a noun related word, or the part of speech contains a noun related word. If is not included, clustering is performed using stop words. The clustering unit 15 passes the clustering result to the utterance candidate selection unit 16.

  Next, in step 110, the utterance candidate selection unit 16 is stored in the dialogue history storage unit 21 among representative utterances of clusters including related words based on the clustering result passed from the clustering unit 15. An utterance set (set 1) is created by extracting utterances including content words included in the utterance immediately before the user. Further, the utterance candidate selection unit 16 creates an utterance set (set 2) in which representative utterances are extracted one by one from a cluster having a plurality of utterances of two or more utterances. Further, the utterance candidate selection unit 16 creates an utterance set (set 3) in which representative utterances are extracted one by one from all the clusters.

  Next, in step 112, the utterance candidate selection unit 16 determines whether or not the set 1 is an empty set. If the set 1 is not an empty set, the process proceeds to step 114, and the utterance candidate selection unit 16 selects and outputs the set 1 as the utterance candidate set. When the set 1 is an empty set, the process proceeds to step 116, and the utterance candidate selection unit 16 determines whether or not the set 2 is an empty set. If the set 2 is not an empty set, the process proceeds to step 118, where the utterance candidate selection unit 16 selects the set 2 as an utterance candidate and outputs it. When the set 2 is an empty set, the process proceeds to step 120, and the utterance candidate selection unit 16 determines whether or not the set 3 is an empty set. If the set 3 is not an empty set, the process proceeds to step 122, where the utterance candidate selection unit 16 selects the set 3 as an utterance candidate and outputs it. When the set 3 is an empty set, the process proceeds to step 124, where the utterance candidate selection unit 16 outputs the empty set and ends the utterance selection process.

  As described above, according to the utterance selection device according to the first embodiment, an utterance is selected from an utterance set searched by using an extracted word and a related word of the extracted word as a search query. As an utterance, a high quality utterance can be selected at low cost.

<Second Embodiment>
Next, a second embodiment will be described. The dialogue apparatus 20 according to the second embodiment is configured by a computer including a CPU, a RAM, and a ROM that stores a program for executing dialogue processing. Functionally, this computer can be represented by a configuration including an utterance receiving unit 25, an utterance candidate set selecting unit 26, and an utterance output unit 27, as shown in FIG. The utterance candidate set selection unit 26 includes each unit constituting the utterance selection device 10 according to the first embodiment. That is, the dialogue device 20 according to the second embodiment is a dialogue device in which the utterance selection device 10 according to the first embodiment is incorporated.

  The utterance reception unit 25 receives a user's utterance as an input. The user's utterance may be input as voice data. When the user's utterance is input as voice data, the utterance reception unit 25 converts the user's utterance, which is the received voice data, into text by voice recognition. The utterance reception unit 25 performs morphological analysis on the text data indicating the user's utterance, and extracts words whose part of speech included in the utterance is a noun. If there are a plurality of words whose part of speech is a noun, one of them is selected. The method of selecting one can be selected according to a predetermined rule, for example, selecting a word whose part of speech that appears at the end of the received user's utterance is a noun. Then, the extracted word and the utterance of the previous user are transferred to the utterance candidate set selection unit 26.

  The utterance candidate set selection unit 26 selects and outputs an utterance candidate set in the same manner as the utterance selection apparatus 10 according to the first embodiment.

  The utterance output unit 27 randomly selects one utterance from the utterance candidate set output from the utterance candidate set selection unit 26, and outputs it to the user as a system utterance. When an empty set is returned from the utterance candidate set selection unit 26, a list of utterances when no utterance candidates are obtained is prepared separately, and one or more utterances are selected from the list. Output.

  For example, in the dialogue apparatus 20, the following dialogue can be realized between the user and the system. The first "Hello" is a system prompt (initial utterance). “Ramen” is input to the utterance candidate set selection unit 26 as an extracted word, and an utterance is selected from the utterance candidate set obtained as a result, thereby realizing a chat. “Korakuen's ramen is delicious” is preferentially considered as an utterance candidate because there is “delicious” in the utterance immediately before the user.

Dialogue system (system) and user (you) dialogue example:
[System]: Hello [you]: I ramen is love.
[System]: Ramen is miso ramen [you]: That's right.
[System]: I want to eat ramen [you]: It's delicious [system]: Koraku ramen is delicious [you]: Is it true?

  As described above, by incorporating the utterance selection device according to the first embodiment into a dialogue device, it is possible to realize a chat dialogue system that performs high-quality utterances on various topics at low cost. Chat is used as a lubricant for conversation between humans. By enabling high-quality chat, human interaction with the system becomes smoother and the efficiency of collaboration between computers and users increases.

  In the second embodiment, the configuration in which the utterance selection device is incorporated in the dialog device has been described. However, the dialog device including the utterance reception unit and the utterance output unit, and the utterance selection configured as an external device of the dialog device. You may comprise as an interactive system which communicates between apparatuses and outputs the response with respect to the speech received from the user.

  The present invention is not limited to the above-described embodiments, and various modifications and applications can be made without departing from the gist of the present invention.

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

DESCRIPTION OF SYMBOLS 10 utterance selection apparatus 11 word input part 12 related word acquisition part 13 utterance search part 14 filtering part 15 clustering part 16 utterance candidate selection part 20 dialogue apparatus 21 dialogue history storage part 22 related word dictionary 23 utterance database 25 utterance reception part 26 utterance candidate Set selection unit 27 Utterance output unit

Claims (12)

At least one extracted from a user's utterance from a dictionary in which related words related to the word are collected and associated with the word based on a language pattern composed of a word and part-of-speech sequence, together with the word. A related word acquisition unit that acquires related words related to one extracted word;
An utterance search unit for searching for an utterance corresponding to a search query including a related word related to the extracted word and the extracted word acquired from the extracted word and the related word acquiring unit from an utterance database including a plurality of utterances;
An utterance selection unit that selects at least one utterance from the utterances searched by the utterance search unit;
Utterance selection device including The related word acquisition unit acquires, as the dictionary, a related word related to the extracted word using a dictionary in which the related word is associated with a degree of association with the word,
2. The utterance selection device according to claim 1, wherein the utterance search unit includes, in the search query, a related word having a relevance level equal to or higher than a predetermined value among related words related to the extracted word.   The utterance selection apparatus according to claim 2, wherein the degree of association is a degree of co-occurrence between the word and the related word.   The utterance selection device according to claim 2, wherein the degree of relevance is a degree at which the related word appears in an utterance subsequent to an utterance including the word.   The utterance selection unit excludes utterances including words other than stop words predetermined as words that do not affect the extracted words, the related words, and utterance contents from utterances searched by the utterance search unit. The utterance selection device according to any one of claims 1 to 4.   The utterance selection unit clusters the utterances searched by the utterance search unit based on the related word or the stop word included in an utterance, and selects a representative utterance of each cluster. Utterance selection device.   The utterance selection unit clusters the utterances searched by the utterance search unit based on the related word or the stop word included in the utterance, and represents a cluster whose number of utterances included in each cluster is a predetermined number or more. The utterance selection device according to claim 5 which selects a typical utterance.   The utterance selection unit acquires at least the utterance immediately before the user, and selects an utterance including a word included in the immediately previous utterance from utterances searched by the utterance search unit. The utterance selection device according to claim 1. An utterance reception unit for receiving user utterances;
An utterance output unit that outputs any of the utterances selected by the utterance selection device according to any one of claims 1 to 8;
Interactive device including The related word acquisition unit collects related words related to the word based on a language pattern including a word and a part of speech that is predetermined together with the word, and the user's utterance Obtaining a related word related to at least one extracted word extracted from;
An utterance search unit searches an utterance corresponding to a search query including a related word related to the extracted word and the extracted word acquired by the related word acquiring unit from an utterance database including a plurality of utterances;
An utterance selection unit selecting at least one utterance from the utterances searched by the utterance search unit;
Utterance selection method including An utterance accepting unit accepting a user's utterance;
An utterance output unit outputting any of the utterances selected by the utterance selection method according to claim 10;
Interactive method including   The utterance selection program for functioning a computer as each part which comprises the utterance selection apparatus of any one of Claims 1-8.

Images