JP2017091368A - Paraphrase device, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a paraphrase device capable of replacing with an expression according to knowledge amount of interlocutor.SOLUTION: A paraphrase candidate enumeration section 32 refers to a predetermined thesaurus with respect to an input word of a noun and enumerates a plurality of paraphrase candidates. A paraphrase target selection section 34 selects a paraphrase expression from the plurality of paraphrase candidates based on a predetermined knowledge amount with respect to the input word. A paraphrase sentence generation section 36 generates a paraphrase sentence in which an input word of the input sentence is replaced with a paraphrase expression.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a paraphrase device, method, and program, and more particularly, to a paraphrase device, method, and program for paraphrasing an input sentence.

  Non-Patent Document 1 describes a method for paraphrasing words with high difficulty into plain words (basic learning vocabulary) in order to create a readable document even for children with a small number of vocabularies that can be understood compared to adults. ing.

Tomoyuki Sugawara, Kazuhide Yamamoto, Simplification and evaluation of vocabulary using multiple paraphrasing knowledge to support reading comprehension for elementary school students, Proc. Of the 19th Annual Conference of the Language Processing Society, pp.272-275, 2013. Kyoko Kanzaki, Francis Bond, Noriko Tomuro, Hitoshi Isahara, Extraction of Attribute Concepts from Japanese Adjectives, In LREC-2008, 2008.

  However, the method of Non-Patent Document 1 is intended to improve readability for children, and does not deal with differences in knowledge amount regarding specific fields and topics of adults.

  In addition, there has never been a technology that tries to give the impression that the knowledge of speakers and writers is abundant by rephrasing difficult, uncommon expressions.

  The present invention has been made to solve the above problems, and an object of the present invention is to provide a paraphrasing device, a method, and a program that can be paraphrased into expressions according to the amount of knowledge of a conversation person.

  To achieve the above object, the paraphrase device according to the first aspect of the present invention enumerates a plurality of paraphrase candidates with reference to a predetermined thesaurus for input words composed of nouns included in the input sentence. A paraphrase candidate enumeration unit; a paraphrase destination selection unit that selects a paraphrase expression from a plurality of paraphrase candidates enumerated by the paraphrase candidate enumeration unit based on a predetermined knowledge amount corresponding to the input word; and A paraphrase sentence generation unit that generates a paraphrase sentence in which the input word of the input sentence is paraphrased using the paraphrase expression selected by the paraphrase destination selection unit.

  Further, in the paraphrase device according to the first aspect, the paraphrase destination selection unit selects the paraphrase expression having a lower appearance frequency as the knowledge amount corresponding to the input word is larger, and the knowledge amount corresponding to the input word You may make it select the said paraphrase expression with high appearance frequency, so that there are few.

  Further, in the paraphrase device according to the first aspect, the paraphrase destination selecting unit is, for each of the paraphrase candidates, a document frequency that is the knowledge amount and the number of documents in which the paraphrase candidate appears in a predetermined corpus. The paraphrase expression may be selected by calculating a paraphrase cost based on the distance between the paraphrase candidate and the input word on the thesaurus.

  The paraphrase method according to the second invention is a step in which the paraphrase candidate enumeration unit enumerates a plurality of paraphrase candidates with reference to a predetermined thesaurus for an input word including nouns included in the input sentence. A paraphrase destination selecting unit selecting a paraphrase expression from a plurality of paraphrase candidates enumerated by the paraphrase candidate enumeration unit based on a predetermined knowledge amount corresponding to the input word; and paraphrase sentence generation And a step of generating a paraphrase sentence obtained by paraphrasing the input word of the input sentence using the paraphrase expression selected by the paraphrase destination selection unit.

  In the paraphrase method according to the second invention, the step of selecting by the paraphrase destination selecting unit selects the paraphrase expression having a lower appearance frequency as the knowledge amount corresponding to the input word increases, and selects the paraphrase expression as the input word. The smaller the corresponding knowledge amount is, the higher the appearance frequency is selected.

  In the paraphrasing method according to the second invention, the step of selecting by the paraphrase destination selecting unit includes, for each of the paraphrase candidates, the knowledge amount and the number of documents in which the paraphrase candidates appear in a predetermined corpus. The paraphrase expression is selected by calculating a paraphrase cost based on the document frequency and the distance on the thesaurus between the paraphrase candidate and the input word.

  A program according to the third invention is a program for causing a computer to function as each part of the paraphrasing device according to the first invention.

  According to the paraphrase device, method, and program of the present invention, with respect to an input word composed of nouns, a plurality of paraphrase candidates are listed with reference to a predetermined thesaurus, and a plurality of paraphrase candidates correspond to the input word. The paraphrased expression is selected based on the predetermined knowledge amount, and the paraphrase expression is used to rephrase the input word of the input sentence. The effect that it can be obtained.

It is a block diagram which shows the structure of the paraphrase apparatus which concerns on embodiment of this invention. 5 is a diagram illustrating an example of a Japanese thesaurus in Example 1. FIG. 6 is a diagram illustrating an example of a Japanese thesaurus in Example 2. FIG. It is a figure which shows an example of a knowledge amount setting file. It is a figure which shows an example of calculation of a paraphrase cost. It is a flowchart which shows the paraphrase process routine in the paraphrase apparatus which concerns on embodiment of this invention.

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

<Outline according to Embodiment of the Present Invention>

  First, an outline of the embodiment of the present invention will be described.

  In the embodiment of the present invention, broad words, synonyms, and instances (examples) that can be acquired from a thesaurus (systematic synonym dictionary) prepared in advance are listed as candidates for paraphrase destination expressions. Next, among the listed candidates, (1) the set knowledge amount, (2) the generality of the expression (document frequency in a Japanese corpus prepared in advance), and (3) the paraphrase source expression The paraphrasing cost is calculated using the distance on the thesaurus. Then, one paraphrase expression as a paraphrase destination having the lowest calculated paraphrase cost is selected, and a paraphrase sentence is created using this expression.

  As an example of paraphrasing, for example, when the sentence “I like Beaujolais” is entered, using the wine's proper name like “I like Joseph Drouin” gives a more knowledgeable impression. Can be given. On the other hand, if you rephrase using a broader word like “I like red wine”, even if a speaker (writer) who is assumed to be inexperienced in wine speaks (even if written), there is no sense of incongruity, and wine It is possible to generate a sentence that can be easily communicated to listeners (readers) who are not familiar with.

  As an example of a different nature from the above, if you enter the phrase “Please search using a search engine,” use a specific unique name such as “Search by google (R)” or “ "Please search with a search engine such as (R)", it is possible to change to a sentence that is easy to communicate, suitable for people with little knowledge, contrary to the previous example.

  In addition, it is determined based on the appearance frequency in the corpus what expressions (higher terms, synonyms, instances (examples such as proper names)) are more easily transmitted.

<Configuration of Paraphrase Device According to Embodiment of the Present Invention>

  Next, the configuration of the paraphrasing device according to the embodiment of the present invention will be described. As shown in FIG. 1, a paraphrase device 100 according to an embodiment of the present invention is a computer including a CPU, a RAM, and a ROM that stores a program and various data for executing a paraphrase processing routine described later. Can be configured. Functionally, the paraphrasing device 100 includes an input unit 10, a calculation unit 20, and an output unit 50 as shown in FIG.

  The input unit 10 receives an input sentence to be paraphrased. The input sentence input in Example 1 is “I like Beaujolais”. In addition, the input sentence to be input in Example 2 is “Please search with a search engine”. Hereinafter, for the convenience of explanation, the processing based on the input sentence of Example 1 is described as Example 1 and the processing based on the input sentence of Example 2 is described as Example 2.

  The calculation unit 20 includes a noun string extraction unit 30, a paraphrase candidate enumeration unit 32, a paraphrase destination selection unit 34, a paraphrase sentence generation unit 36, a Japanese thesaurus 40, a knowledge amount setting file 42, and a Japanese corpus 44. It is comprised including.

  The noun string extraction unit 30 extracts a noun string including nouns included in the input sentence from the input sentence received by the input unit 10 as an input word. Specifically, the input sentence is subjected to morphological analysis and divided into morphemes, and at the same time, the part of speech of each morpheme is acquired. Then, from the morphological analysis result, all noun strings composed of one or more nouns or continuous nouns are extracted as input words.

  When the input sentence is Example 1, it is divided into morphemes such as “bojolay / ga / like / is”, and “noun / particle / adjective / auxiliary verb” is acquired as a part of speech. Next, “Beaujolais” (noun) is extracted as an input word to be paraphrased.

  When the input sentence is Example 2, it is divided into morphemes such as “search / engine / de / search / do / te / please”, and the part of speech is “noun / noun / particle / noun-sa-variant / verb / particle / verb”. Is obtained. Next, “search” (noun) and “search engine” (noun, noun) are extracted as input words to be paraphrased.

  In the present embodiment, an operational noun (sa-changing noun) such as “search (do)” is excluded from the paraphrase target.

  The paraphrase candidate enumeration unit 32 enumerates a plurality of paraphrase candidates with reference to the Japanese thesaurus 40 for the input words extracted by the noun string extraction unit 30.

  As the Japanese thesaurus 40, for example, a Japanese thesaurus created by adding an instance (proprietary name) to a part of the thesaurus of Non-Patent Document 2 is used, but in a systematic Japanese thesaurus Any one can be used as long as it exists. A partial example of the Japanese thesaurus 40 corresponding to Example 1 and Example 2 is shown in FIGS.

  Specifically, the paraphrase candidate enumeration unit 32 collates all the extracted input words with the recorded words of the Japanese thesaurus 40, and among the input words that match the recorded words, synonymous with the input word having the largest number of constituent characters. A word, a broader word (a word on the 1 to N hierarchy and its synonym), and an instance (an example of a proper name) are acquired as paraphrase candidates. As shown in FIG. 3, in the case of Example 2, “search engine” having more constituent characters than “search” is used as a paraphrase target, and the paraphrase candidates are acquired from words recorded in the Japanese thesaurus 40.

  In the present embodiment, an input word that is an original noun sequence included in the input sentence (in the case of Example 1, “Beaujolais”, in the case of Example 2 “Search Engine”) is also included in the paraphrase candidates. You may not include it. When an input word is included in a paraphrase candidate, as a result, a paraphrase sentence may be generated with the original noun string.

  In the case of Example 1, “alcohol”, “sake”, “wine”, “grape wine”, “red wine”, “bojolet”, “Joseph Drouin” are listed from the Japanese thesaurus 40 as the paraphrase candidates. .

  In the case of Example 2, as paraphrase candidates, “code”, “software”, “program”, “search engine”, “search engine”, “yahoo (R)”, “google ( R) ".

  In this embodiment, N = 3, and the broader term is acquired retroactively up to three levels, but any number of levels may be used.

  Further, when applied to a dialogue system, the paraphrase candidates to be enumerated may be limited by using conversation topics (topic words). For example, if the conversation about “Wine” has already been exchanged, it is unnatural for the system to say “I like wine” again, so the hierarchy of the thesaurus is the topic word “ Words above “wine” (“wine”, “vinegar”, “alcohol”, “alcohol”) may be excluded from the paraphrase candidates.

  The paraphrase destination selection unit 34 appears from the plurality of paraphrase candidates listed by the paraphrase candidate enumeration unit 32 as the knowledge amount corresponding to the input word increases based on the knowledge amount predetermined corresponding to the input word. A paraphrase expression with a lower frequency is selected, and a paraphrase expression with a higher appearance frequency is selected as the knowledge amount corresponding to the input word is smaller. Specifically, as described below, the paraphrase destination selection unit 34 sets the knowledge amount set in the knowledge amount setting file 42 and the document in which the paraphrase candidate appears in the Japanese corpus 44 for each of the paraphrase candidates. A paraphrase expression is selected by calculating a paraphrase cost based on a document frequency that is a number and a distance on the thesaurus between the paraphrase candidate and the input word.

  An example of the knowledge amount setting file 42 is shown in FIG. As the knowledge amount for each of the topic categories, “high knowledge amount” or “low knowledge amount” is set for each character. The setting of the knowledge amount is a setting corresponding to the character to be a speaker or a listener. For example, in FIG. 4, it is assumed that each character has a setting a and a setting b. As for the knowledge amount, the recorded words of the Japanese thesaurus 40 are classified into rough topics in advance, and the knowledge amount for the rough topic classification is set. For example, in the present embodiment, the recorded word “red wine” is classified in advance into the topic classification “gourmet”, and the recorded word “search engine” is classified in advance into the topic classification “computer”, and the topic classification “ Set the amount of knowledge for “Gourmet” and “Computer”. Instead of determining the knowledge amount for topic classification, for example, the knowledge amount may be determined for each recorded word of the Japanese thesaurus 40, and the knowledge amount of the recorded word corresponding to the input word may be used. Good.

  The Japanese corpus 44 stores a plurality of documents. In the present embodiment, “document” refers to one posting on SNS, one blog article, one news article, and the like.

  The paraphrase destination selection unit 34 performs the following first to third processes to select a paraphrase expression.

  As a first process, the paraphrase destination selection unit 34 counts the document frequency (the number of documents in which the paraphrase candidate appears) in the Japanese corpus 44 for each paraphrase candidate. An expression with a high document frequency is an expression that appears in many documents, and can be said to be a more general expression. On the other hand, an expression with a low document frequency is an unusual expression that is not commonly used.

  Next, as the second process, the paraphrase destination selection unit 34, for each paraphrase candidate, (1) a set knowledge amount, (2) document frequency, and (3) a thesaurus with an input word that is a paraphrase source expression The paraphrasing cost is calculated using the above distance. In the case of the “high knowledge amount” setting, the cost is designed so that the general expression has a higher paraphrase cost. In other words, unusual expressions are easily selected. In the present embodiment, the paraphrase cost c (w) to the expression w is as follows when the document frequency is df (w) and the distance on the thesaurus from the input word that is the paraphrase source notation is d (w): 1) Calculated according to the equation.

  In the case of the setting of “small amount of knowledge”, the cost is designed so that the paraphrasing cost becomes higher for the rare expression. That is, general expressions are easily selected. In the present embodiment, when the total number of documents included in the Japanese corpus is D (for example, D = 1200,000), the paraphrasing cost for the expression w is calculated according to the following equation (2) using the inverse document frequency idf (w). calculate.

  Then, the paraphrase destination selecting unit 34 selects the candidate having the lowest calculated cost as the paraphrase destination by the third process. FIG. 5 shows an example of cost calculation. As shown in FIG. 5, the movement distance from the input word is reduced to a paraphrasing cost in order to make it easier to select a paraphrase as an expression with a closer positional relationship on the Japanese thesaurus 40 and as a farther expression is less likely to be selected. Calculate to include. In the present embodiment, the movement distance uses a simple number of movements. For example, in the case of FIG. 3, “search engine” and “search engine” are synonyms and are located at the same node. Therefore, in FIG. 5, the movement distance (number of movements) from “search engine” to “search engine”. ) Is 0, and the movement distance is 1 because one movement is required from the “search engine” to the “program”.

  For the paraphrase candidate in the case of Example 1, the amount of knowledge determined for the topic classification “gourmet” is used. For the paraphrase candidates in the case of Example 2, the amount of knowledge determined for the topic classification “computer” is used.

  The case where setting a and setting b are applied to Example 1 and Example 2 will be described below.

  When the setting a in FIG. 4 is applied, the knowledge amount determined for the topic classification “gourmet” is “high knowledge amount”, and the knowledge amount determined for the topic classification “computer” is “small knowledge amount”. Is. In this case, the paraphrase expression that becomes the paraphrase destination for the paraphrase candidate (topic classification “gourmet”) of Example 1 is “Joseph Drouin”, and the paraphrase that becomes the paraphrase destination for the paraphrase candidate (topic classification “computer”) of Example 2 The expression is “google (R)”.

  When the setting b in FIG. 4 is applied, the knowledge amount determined for the topic classification “gourmet” is “low knowledge amount”, and the knowledge amount determined for the topic classification “computer” is “high knowledge amount”. Is. In this case, the paraphrase expression that is the paraphrase destination of Example 1 (topic classification “gourmet”) is “red wine”, and the paraphrase expression that is the paraphrase destination of Example 2 (topic classification “computer”) is “search engine” (paraphrase). As in the previous expression).

  In addition, the formulas for calculating the paraphrasing costs in the above formulas (1) and (2) are only examples, and it is easy to select an unusual expression in the case of a setting with a large amount of knowledge, and it is common in the case of a setting with a small amount of knowledge. Any calculation formula can be used as long as the cost can be designed so that a simple expression can be easily selected.

  In this embodiment, the movement distance uses a simple number of movements, but is not limited to this. The method of measuring the movement distance is not limited to this method, and any method may be used as long as the scale is such that a low value is attached to a representation that is close in meaning and concept and a high value is attached to a distant representation.

  The paraphrase sentence generation unit 36 generates a paraphrase sentence obtained by paraphrasing the input word of the input sentence using the paraphrase expression selected by the paraphrase destination selection unit 34. Specifically, as will be described below for Examples 1 and 2, the noun string included in the input sentence is replaced with the paraphrase destination expression.

  In the case of Example 1, if the paraphrase expression selected in the case of “small amount of knowledge” is “red wine”, the input sentence “I like Beaujolais” is replaced with “red wine” and “red wine” Is generated. If the paraphrase selected in the case of “Knowledge” is “Joseph Drouin”, replace “Beaujolais” with “Joseph Drouin” and generate the paraphrase sentence “I like Joseph Drouin” .

  In the case of Example 2, if the paraphrase expression selected in the case of “small amount of knowledge” is “google (R)”, “search engine” of the input sentence “search with search engine” is set to “google (R) ”is replaced, and a paraphrase sentence“ Please search with google (R) ”is generated. If the paraphrase expression selected in the case of “some knowledge amount” is “search engine”, a paraphrase sentence “please search with search engine” is generated with the expression before paraphrase being maintained.

  In this embodiment, the expression of the paraphrase source and the paraphrase destination is simply replaced. However, when an instance (an example such as a unique name) is the paraphrase destination, “Beautiful such as Joseph Druen” As in “search engine such as google (R)”, it may be “<paraphrase source expression> such as <instance>”.

<Operation of the paraphrasing device according to the embodiment of the present invention>

  Next, the operation of the paraphrasing device 100 according to the embodiment of the present invention will be described. When the input unit 10 receives an input sentence, the paraphrasing device 100 executes a paraphrase processing routine shown in FIG.

  First, in step S100, a noun string included in the input sentence received by the input unit 10 is extracted as an input word.

  Next, in step S102, with respect to the input word extracted in step S100, a plurality of paraphrase candidates are listed with reference to the Japanese thesaurus 40.

  In step S104, a candidate for processing is selected from the paraphrase candidates listed in step S102.

  In step S106, for the paraphrase candidates selected in step S104, the document frequency that is the number of documents in which the paraphrase candidates appear in the Japanese corpus 44 is counted.

  In step S108, based on the knowledge amount set in the knowledge amount setting file 42, the document frequency counted in step S106, and the distance on the thesaurus between the paraphrase candidate and the input word, the above equation (1) or ( 2) The paraphrase cost is calculated according to the equation.

  In step S110, it is determined whether or not the processing in steps S106 to S108 has been completed for all the paraphrase candidates. If completed, the process proceeds to step S112. If not completed, the process returns to step S104 to select the next paraphrase expression. And repeat the process.

  In step S112, the paraphrase candidate with the lowest calculated cost is selected as the paraphrase expression from among the plurality of paraphrase candidates whose costs are calculated in steps S106 to S108.

  In step S114, the paraphrase text which paraphrased the input word of the input sentence using the paraphrase expression selected by step S104 is produced | generated, and it outputs to the output part 50, and complete | finishes a process.

  As described above, according to the paraphrase device according to the embodiment of the present invention, with respect to an input word composed of nouns, a plurality of paraphrase candidates are listed with reference to the Japanese thesaurus 40, and By selecting a paraphrase expression based on the knowledge amount of the knowledge amount setting file 42 determined in advance corresponding to the input word, and generating a paraphrase sentence in which the input word of the input sentence is rephrased using the paraphrase expression, It can be paraphrased as an expression according to the amount of knowledge of the interlocutor.

  Further, according to the technique according to the embodiment of the present invention, it is possible to rephrase the expression according to the knowledge amount regarding a specific topic possessed by an assumed speaker (writer) or listener (reader). When this is applied to a computer (dialogue system) that interacts with humans, the expression included in the utterance of the system can be changed according to the character setting of the dialogue system. For example, in the case of a child character dialogue system, there should be little knowledge about alcohol, so changing `` Dad loves Beaujolais '' to `` Dad likes red wine '' makes it a more natural utterance Can do. On the other hand, it is possible to create a dialogue system that speaks like a wine in detail by rephrasing it with uncommon expressions (such as unique product names and producer names). Moreover, it can contribute to implement | achieving a user friendly dialog system by paraphrasing the expression which is easy to understand according to the knowledge amount which the other party (user) has.

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

DESCRIPTION OF SYMBOLS 10 Input part 20 Calculation part 30 Noun string extraction part 32 Paraphrase candidate enumeration part 34 Paraphrase destination selection part 36 Paraphrase sentence generation part 40 Japanese thesaurus 42 Knowledge amount setting file 44 Japanese corpus 50 Output part 100 Paraphrase apparatus

Claims (7)

A paraphrase candidate enumeration unit that enumerates a plurality of paraphrase candidates with reference to a predetermined thesaurus for input words consisting of nouns included in the input sentence,
A paraphrase destination selection unit that selects a paraphrase expression from a plurality of paraphrase candidates listed by the paraphrase candidate enumeration unit based on a predetermined knowledge amount corresponding to the input word;
Rephrasing the input word of the input sentence using the paraphrase expression selected by the paraphrase destination selection unit;
Including paraphrasing device.   The paraphrase destination selection unit selects the paraphrase expression having a lower appearance frequency as the knowledge amount corresponding to the input word is larger, and the paraphrase expression having a higher appearance frequency as the knowledge amount corresponding to the input word is smaller. The paraphrasing device according to claim 1 to select.   The paraphrase destination selection unit, for each of the paraphrase candidates, the knowledge amount, the document frequency that is the number of documents in which the paraphrase candidates appear in a predetermined corpus, and the paraphrase candidates and the input words The paraphrase device according to claim 1, wherein the paraphrase expression is selected by calculating a paraphrase cost based on a distance on a thesaurus. The paraphrase candidate enumeration unit enumerates a plurality of paraphrase candidates with reference to a predetermined thesaurus for input words composed of nouns included in the input sentence,
A paraphrase destination selecting unit selecting a paraphrase expression from a plurality of paraphrase candidates enumerated by the paraphrase candidate enumeration unit based on a knowledge amount determined in advance corresponding to the input word;
A step of generating a paraphrase sentence in which the paraphrase sentence generation unit paraphrases the input word of the input sentence using the paraphrase expression selected by the paraphrase destination selection unit;
Including paraphrasing methods.   The step of selecting by the paraphrase destination selection unit selects the paraphrase expression having a lower appearance frequency as the knowledge amount corresponding to the input word is larger, and the appearance frequency is higher as the knowledge amount corresponding to the input word is smaller. The paraphrase method according to claim 4, wherein the paraphrase expression is selected.   The step of selecting by the paraphrase destination selecting unit includes, for each of the paraphrase candidates, the knowledge amount, a document frequency that is the number of documents in which the paraphrase candidates appear in a predetermined corpus, the paraphrase candidates, and the input The paraphrase method according to claim 4, wherein the paraphrase expression is selected by calculating a paraphrase cost based on a distance on the thesaurus with a word.   The program for functioning a computer as each part of the paraphrasing apparatus of any one of Claims 1-3.