JP2015087796A - Questioning field determination device and questioning field determination method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To determine accurately whether a question sentence is a question sentence in a specific field or not.SOLUTION: A question answering device 10 includes a question sentence input part 11 for inputting a question sentence, an entity extraction part 12 for extracting an entity which is a question object from an inputted question sentence, a perplexity calculation part 13 for calculating the value of perplexity based on a language model in which a question sentence in a specific field is assumed from the inputted question sentence, and a determination part 14 for determining whether the inputted question sentence is a question sentence in the specific field or not based on the extraction result of the entity and on the calculated perplexity value.

Description

  The present invention relates to a question field determination device and a question field determination method for determining a field of a question sentence.

  With the development of speech recognition technology and question answering technology, agent-type services that accept and answer questions in the natural sentence format of users are becoming widespread. The agent-type service can answer various questions such as weather, transfer guidance, chat, objective facts such as history and society. Questions can be conducted with natural sentences such as "tell me today's weather", and it is characterized by the fact that a personal secretary can be used as if it exists on a personal computer or mobile terminal.

  Patent Document 1 discloses a system configuration for typical question answering. The contents of the question from the user are analyzed, and the question type is classified according to the so-called 5W1H framework such as Who, Where, and When, and answers are extracted by different methods. Patent Document 2 discloses a typical framework for performing a chat conversation.

Japanese Patent Laid-Open No. 2002-132911 Japanese Patent No. 5150583

  The method for generating answers varies greatly depending on the field of questions the user asks, such as weather, transfer guidance, chat, etc. In the agent type service, in order to answer such questions, the process of actually generating the answer Before that, it is necessary to carry out a process of determining the target field of the question (question field determination). In particular, there is a case where a character visualizing an agent is displayed on the screen in order to obtain a feeling that the user is talking to someone, and a question sentence for the character (hereinafter referred to as chatting. For example, “What is your age?”) The objective structure (for example, “What is the age of the prime minister?”) Is very similar in composition, but the required actions are very different, so a method for selecting them is important.

  In the method disclosed in Patent Document 1, the question type is classified according to a simple feature of the sentence such as Who if the expression at the end of the sentence is “who”. Chatting and questions about objective facts are difficult to distinguish, especially with simple features such as sentence endings. For this reason, the method disclosed in Patent Document 1 cannot appropriately determine the question field, and a determination method specialized for it is necessary. In addition, since chats and questions about objective facts are mixed in the input performed by the agent-type service, a method of distinguishing them is required even in a system mainly having chat conversation as shown in Patent Document 2. .

  The present invention has been made in view of the above-mentioned problems. For example, a question sentence is in a specific field so that a user's question can be answered and received including a chat and a response to an objective fact. It is an object of the present invention to provide a question field determination device and a question field determination method for accurately determining whether or not a question sentence.

  To achieve the above object, a question field determination apparatus according to the present invention extracts a question sentence input means for inputting a question sentence, and an entity that is a question target from the question sentence input by the question sentence input means. An entity extraction means; a perplexity calculation means for calculating a perplexity value based on a language model assuming a question sentence in a specific field from a question sentence input by the question sentence input means; and Determination means for determining whether the question text input by the question text input means is a question text in a specific field based on the extraction result and the perplexity value calculated by the perplexity calculation means; .

  In the question field determination device according to the present invention, after considering both the entity extraction result from the question sentence and the perplexity value calculated from the question sentence, the question sentence is a question sentence in a specific field. It is determined whether or not there is. Thereby, according to the question field judgment device concerning the present invention, it becomes possible to judge correctly whether a question text is a question text of a specific field.

  The perplexity calculating means may calculate a perplexity value based on the context of words included in the question sentence. According to this configuration, the perplexity value can be calculated appropriately and reliably, and the present invention can be implemented appropriately and reliably.

  The determination means sets a threshold value according to the extraction result of the entity, compares the set threshold value with the perplexity value, and determines whether or not the question sentence is a question sentence in a specific field. Also good. According to this configuration, the determination can be performed appropriately and reliably, and the present invention can be implemented appropriately and reliably.

  The question field determination device includes a first answering unit that generates and outputs an answer to the question message by the first method when the determination unit determines that the question message is a question message in a specific field. When the determination means determines that the question sentence is a question sentence in a field other than the specific field, an answer to the question sentence is generated and output by a second method different from the first method. It is good also as providing a 2nd reply means. According to this configuration, an answer to the question sentence can be output.

  The determination means calculates the degree to which the question sentence is a question sentence in a specific field based on the entity extraction result and the perplexity value, and the question field determination device determines whether the question sentence is determined by the determination means. In addition, an information addition requesting unit that requests addition of information related to the entity extracted by the entity extracting unit for answering the question sentence may be further provided. According to this configuration, information necessary for grasping the question can be acquired.

  By the way, the present invention can be described as an invention of a question field determination method as described below, in addition to the invention of a question field determination device as described above. This is substantially the same invention only in different categories, and has the same operations and effects.

  That is, the question field determination method according to the present invention includes a question sentence input step for inputting a question sentence, an entity extraction step for extracting an entity to be questioned from the question sentence input in the question sentence input step, and a question sentence A perplexity calculation step for calculating a perplexity value based on a language model assuming a question text in a specific field from the question text input in the input step, an entity extraction result in the entity extraction step, and a perplexity A determination step of determining whether or not the question text input in the question text input step is a question text in a specific field based on the perplexity value calculated in the calculation step.

  In the present invention, it is determined whether or not the question sentence is a question sentence in a specific field after taking into account both the entity extraction result from the question sentence and the perplexity value calculated from the question sentence. Is done. Thereby, according to this invention, it becomes possible to determine correctly whether a question sentence is a question sentence of a specific field.

BRIEF DESCRIPTION OF THE DRAWINGS It is a figure which shows the question response system comprised including the structure of the question response apparatus which is a question field determination apparatus which concerns on 1st Embodiment of this invention, and the said question response apparatus. It is a table which shows the information for producing | generating the reply with respect to the question text regarding things other than objective fact. It is a figure which shows the hardware constitutions of the question response apparatus which is a question field determination apparatus which concerns on embodiment of this invention. It is a flowchart which shows the process (question field determination method) performed with the question response apparatus which is a question field determination apparatus which concerns on 1st Embodiment of this invention. It is a figure which shows the question response system comprised including the structure of the question response apparatus which is a question field | area determination apparatus which concerns on 2nd Embodiment of this invention, and the said question response apparatus. It is a table | surface which shows knowledge data of an entity property type. It is a table | surface which shows a user's past question history.

  Hereinafter, embodiments of a question field determination device and a question field determination method according to the present invention will be described in detail with reference to the drawings. In the description of the drawings, the same elements are denoted by the same reference numerals, and redundant description is omitted.

  FIG. 1 shows a question answering device 10 which is a question field determination device according to the first embodiment of the present invention, and a question answering system 1 including the question answering device 10. The question answering system 1 includes a question answering device 10 and a user terminal 20. The question answering system 1 is a system in which a question sentence is transmitted from the user terminal 20 to the question answering apparatus 10, and a response to the question sentence is generated in the question answering apparatus 10 and transmitted to the user terminal 20.

  The user terminal 20 is a general terminal (apparatus) including input means such as a keyboard and a voice input function for inputting a question sentence from a user, and output means for outputting an answer to the question sentence such as a monitor. Moreover, the user terminal 20 can communicate with the question answering apparatus 10 via a communication network. The communication network is, for example, a public communication network such as the Internet or a mobile communication network, or a dedicated line. Further, the connection between devices in the present embodiment is not limited to the communication network described above, and may be a communication bus in the same physical device, or a combination thereof. Specifically, the user terminal 20 corresponds to a personal computer or a mobile phone used by the user.

  The user terminal 20 inputs a question sentence by a user operation or the like using the above input means. The question sentence may be, for example, a keyboard used by the user or may be obtained by voice recognition by inputting the user's voice. The user terminal 20 transmits the input question sentence to the question answering apparatus 10 in the form of a text sentence (plain text). Examples of questions include "Your name", "The US President's name", "Who is the first astronaut?"

  Here, the question text has a plurality of fields. For example, the field includes a question sentence concerning objective facts and a question sentence concerning non-objective facts. The question sentence regarding objective facts is a question sentence regarding objective facts that are not related to the question answering apparatus 10 according to the present embodiment, such as guidance, chat, history, and society. In the example of the above question sentence, the question sentences “The name of the President of the United States” and “Who is the first astronaut?” Correspond to the question sentence concerning objective facts.

  The question sentence related to other than the objective fact is a question based on the existence of the question answering apparatus 10 according to the present embodiment. For example, it is a question with respect to a character (a character visualizing an agent) displayed on the user terminal 20 in order to perform a question response with the question answering device 10. The question sentence “Your name is” in the above example of the question sentence corresponds to a question sentence related to things other than objective facts.

  The user terminal 20 receives an answer to the question sentence transmitted from the question answering device 10 in response to the transmission of the input question sentence. The user terminal 20 performs output such as display output of the received answer. Thus, in the question answering system 1, it is possible to receive and answer questions including responses to chats and objective facts to the user's question.

  Subsequently, functions of the question answering apparatus 10 according to the present embodiment will be described. As shown in FIG. 1, the question answering apparatus 10 includes a question sentence input unit 11, an entity extraction unit 12, a perplexity calculation unit 13, a determination unit 14, an objective fact response unit 15, and other question response units. 16.

  The question sentence input unit 11 is a question sentence input means for inputting a question sentence. Specifically, the question sentence input unit 11 inputs a question sentence by receiving data of a text sentence (plain text) that is a question sentence from the user terminal 20. The question sentence input unit 11 outputs the input question sentence to the entity extraction unit 12, the perplexity calculation unit 13, and the perplexity calculation unit 13.

  The entity extraction unit 12 is an entity extraction unit that extracts an entity (Entity) as a question target from the question sentence input from the question sentence input unit 11. The entity that is the question target indicates what the question sentence asks, and is the subject of the question sentence. For example, in the question sentence “What is the name of the US President?”, “US President” corresponds to an entity, and in “Who is the first astronaut?” “First astronaut” corresponds to an entity. Specifically, for example, the entity extraction unit 12 extracts an entity as follows.

  For the entity extraction, a sequence labeling method, which is a method of machine learning, is applied. In the sequence labeling, a predetermined label is assigned to sequence data such as a word sequence. An operation example when applying sequence labeling to entity extraction in this embodiment will be described.

  The entity extraction unit 12 divides the text that is the input question sentence into morphemes that are the smallest meaningful units. For the division into morphemes, for example, a morphological analyzer that has been conventionally used can be used. If the question sentence is “What is the name of the US President?”, The result of the morphological analysis will be “US / President / NO / NAME / HA”. Here, “/” represents a morpheme break.

  Subsequently, the entity extraction unit 12 assigns three types of labels called BIO format to each morpheme. In the figure, B is the head of the section (entity here) to be detected (the first entity in the section), I is in the middle of the section, and O is a portion other than the section. In the above example, the labeling result is, for example, “America (B) / President (I) / (O) / Name (O) / Has (O)”. (B) (I) (O) attached after each morpheme represents that each label is given. Finally, all the morphemes to which the I label following the morpheme to which the B label is assigned are taken out and set as the entity as the final result. In the above example, “USA (B) / President (I)” is the entity.

  In the above example, since three types of labels are given to five morphemes, there is a possibility of labeling in the fifth power of 3. In the sequence labeling method, generally, a conditional probability of a label sequence when a sequence to be labeled (for example, a morpheme sequence) is given is calculated using a machine learning method. The relationship between the label target series and the label is learned using a character string to which a correct label is assigned in advance. At the stage of actual labeling, the probability for each labeling possibility (in the above example, 3 to the fifth power) is calculated using these relationships, and a plausible sequence is output.

  A conventional method can be used for sequence labeling. For example, the above operations can be implemented using any one of a number of commercially available tools such as CRF ++, which are available for free.

  Further, it is not always necessary to use a machine learning method for extracting an entity. For example, when an expression representing the subject of a thing such as “XX” appears at the beginning of the sentence, the entity is extracted by a simple method that regards the part of “XX” as an entity or a method that extracts a part corresponding to the main character based on syntax analysis. Also good.

  The entity extraction unit 12 outputs the extracted entity to the determination unit 14. Further, there are cases where the entity cannot be extracted because the question sentence does not form a decent question or the entity is not included in the question sentence. In that case, the entity extraction unit 12 notifies the determination unit 14 that the entity could not be extracted.

  The perplexity calculation unit 13 calculates a perplexity value (ppl) based on a language model assuming a question text in a specific field from the question text input from the question text input unit 11. It is a calculation means. The perplexity value based on the language model assuming a question text in a specific field is a value indicating the degree of conformity when it is assumed that the input question text is the above-mentioned specific type of question text. In the present embodiment, the question text in a specific field is a question text regarding objective facts. The language model is, for example, a model of a relationship between words and parts of speech when a certain language is assumed. In the present embodiment, the perplexity (objective fact perplexity) value is calculated using a language model based on the context of words included in the question sentence as follows.

There are various representation methods for the language model, but in this embodiment, an N-gram model is used. For example, in the case of N = 2, the probability (P (w _i | w _i−1 ), w _i is the i-th word in the sentence) related to a sequence of two words is expressed. As a result, the sentence appearance probability P (w ₁ ,..., W _n ) (n is the number of words in the sentence) is conditional on the assumption that the word appearance depends only on the previous word. It can be calculated as ΠP (w _i | w _i−1 ) multiplied by the probability.

  A language model is created by learning examples of sentences in an assumed application field. A conventional method can be used for generating the language model. For example, since there are general tools such as SRILM, a language model can be created by inputting an example sentence into these tools. In the present embodiment, a language model created from a sentence example of only a question sentence relating to an objective fact from which a question of other intentions such as chat is removed is used.

  The perplexity calculation unit 13 stores information related to the language model in advance, and calculates a perplexity value based on the information. The perplexity calculation unit 13 divides the text that is the input question sentence into morphemes that are the smallest meaningful units. The division into morphemes is performed in the same manner as the division into morphemes by the entity extraction unit 12. The perplexity calculation unit 13 may use the result of division into morphemes by the entity extraction unit 12. Alternatively, before the processing by the entity extraction unit 12 and the perplexity calculation unit 13, for example, the question sentence input unit 11 performs division into morphemes, and the result of the division is obtained by the entity extraction unit 12 and the perplexity calculation unit 13. You may enter

各単語位置において、Ｐ（ｗ_ｉ｜ｗ_ｉ−１，ｗ_ｉ−２，…，ｗ_１）はそれ以前までの単語の系列から見た場合の、その単語の出現確率（例えば、「アメリカ／大統領／の」と出現した場合の「名前」が出現する確率）となる。言語モデルを利用することで、この値を計算することができる。この（対数の）平均値をとるため、おおよそその文章が、想定する言語モデルを対象とした場合にどの程度珍しいものであるかを評価することができる。 The perplexity calculating unit 13 calculates the entropy H based on the following expression based on the positional relationship in the question sentence of each word that is a morpheme after the division.

At each word position, P (w _i | w _i−1 , w _i−2 ,..., W ₁ ) is the probability of appearance of the word (for example, “America / The probability of “name” appearing when “president / no” appears. This value can be calculated by using a language model. Since this (logarithmic) average value is taken, it is possible to evaluate how rare the sentence is when the intended language model is targeted.

The perplexity calculation unit 13 calculates 2 entropy power (2 ^H ) as a perplexity value. The perplexity value calculated as described above indicates that the lower the value, the higher the degree of conformity when it is assumed that the input question text is a question text regarding objective facts. The perplexity calculation unit 13 outputs the calculated perplexity value to the determination unit 14.

  Based on the entity extraction result by the entity extraction unit 12 and the perplexity value calculated by the perplexity calculation unit 13, the determination unit 14 determines that the question sentence input by the question sentence input unit 11 is an objective fact. It is the determination means which determines whether it is the question sentence regarding. The determination unit 14 sets a threshold value according to the entity extraction result (whether or not the entity is extracted) by the entity extraction unit 12, compares the set threshold value with the perplexity value, and performs the above determination. Do.

  Basically, when the perplexity value is low, the input question text is regarded as a question text regarding objective facts. However, information on whether or not an entity has been extracted is considered as information for ensuring that the question is a question about a more objective fact.

  Specifically, the determination unit 14 stores two threshold values thr1 and thr2 in advance. thr1 is a first threshold value used when no entity is extracted, and a value smaller than thr2 is set. thr2 is a first threshold value used when an entity has been extracted. For example, thr1 = 10 and thr2 = 100 can be set. The determination unit 14 sets thr1 or thr2 as a threshold according to the entity extraction result by the entity extraction unit 12. The determination unit 14 compares the set threshold value with the perplexity value.

  As a result of the comparison, the determination unit 14 determines that the perplexity value is less than or equal to the threshold value (or smaller than the threshold value) (that is, a question that is not uncommon by a language model that assumes a question sentence regarding objective facts). If it is a sentence), it is determined that the inputted question sentence is a question sentence regarding objective facts. In that case, the determination unit 14 outputs the question sentence to the objective fact response unit 15.

  As a result of the comparison, the determination unit 14 determines that the perplexity value exceeds the threshold value (or is greater than or equal to the threshold value) (that is, a question sentence that is rare by a language model that assumes a question sentence regarding objective facts). In the case of the above, it is determined that the inputted question sentence is a question sentence related to other than objective facts. In that case, the determination unit 14 outputs the question sentence to the other question answering unit 16.

  The objective fact response unit 15 generates and outputs an answer to the question text by the first method when the judgment unit 14 determines that the question text is a question text related to the objective fact. It is an answer means. Specifically, as a simple configuration, for example, the objective fact response unit 15 generates an answer as follows. First, when a question sentence is input from the determination unit 14, the objective fact response unit 15 divides the text that is the input question sentence into morphemes that are the smallest meaningful units. The division result into morphemes may be obtained by using the entity extraction unit 12 described above. The objective fact response unit 15 makes a keyword search request (inquiry) to the external search engine 30 using a noun among words that are morphemes after division as a keyword.

  The external search engine 30 is a device that receives a keyword search request and performs a search for a search target document (for example, a Web page) using a keyword included in the request. The external search engine 30 performs a search to extract a search target including the keyword, and information indicating the extracted search target or information on the search target itself (for example, a URL or a Web page that is information indicating a Web page) Itself) to the requester of the search. The question answering apparatus 10 and the external search engine 30 are connected via a communication network, and can communicate with each other.

  The objective fact response unit 15 outputs (generates) a response from the external search engine 30 to the request as an answer to the input question sentence. Specifically, the objective fact response unit 15 transmits a URL, which is information indicating a Web page, or the Web page itself to the user terminal 20 as an answer to the input question sentence. For example, if the input question is “Human first astronaut is”, the words (nouns) of “humanity”, “first”, and “astronaut” are extracted as search keywords and searched by them. The web page (web page including those search keywords) is transmitted from the objective fact response unit 15 (question answering device 10) to the user terminal 20 as an answer to the inputted question sentence.

  The objective fact response unit 15 may output a specific word as an answer based on the relationship with the search keyword from words included in the search result document as shown in Patent Document 1. Further, the objective fact response unit 15 may be realized by any function as long as it can perform a question response assuming an objective fact by inputting a text as a question sentence.

  The other question answering part 16 replies to the question sentence by the second method different from the first method when the decision part 14 determines that the question sentence is a question sentence related to other than the objective fact. Second answer means for generating and outputting. Specifically, the other question answering unit 16 generates an answer to a question sentence that is not an objective question, such as receiving and answering a chat based on a prestored rule. For example, the other question answering unit 16 stores information in which a user's question (question sentence) and an answer as shown in the table of FIG. 2 are associated, and generates an answer to the question sentence based on the information. .

  In the table shown in FIG. 2, the other question answering unit 16 obtains an answer that matches the question text input from the determination unit 14 among the user's questions (question text), and generates (generates) an answer to the input question text ) Transmit to the user terminal 20.

  Further, the other question answering unit 16 may be realized by any function as long as it is a function capable of answering a question assuming other than an objective fact by inputting a text as a question sentence. The functional configuration of the question answering apparatus 10 has been described above.

  FIG. 3 shows a hardware configuration of the question answering apparatus 10 according to the present embodiment. As shown in FIG. 3, the question answering apparatus 10 includes a central processing unit (CPU) 101, a random access memory (RAM) 102 and a read only memory (ROM) 103, which are main storage devices, a communication module 104 for performing communication, and The computer is configured to include a computer including hardware such as an auxiliary storage device 105 such as a hard disk. The functions of the question answering apparatus 10 described above are exhibited by operating these components by a program or the like. The above is the configuration of the question answering apparatus 10 according to the present embodiment.

  Subsequently, a question field determination method which is a process executed by the question answering apparatus 10 according to the present embodiment will be described using the flowchart of FIG. This process is started when a question sentence is transmitted from the user terminal 20 to the question answering apparatus 10. In this process, first, a question text is received and inputted by the question text input unit 11 (S01, question text input step). The input question text is output from the question text input section 11 to the entity extraction section 12, the perplexity calculation section 13, and the perplexity calculation section 13.

  Subsequently, the entity extraction unit 12 extracts entities from the question text input from the question text input unit 11 (S02, entity extraction step). The entity extraction result is notified from the entity extraction unit 12 to the determination unit 14.

  On the other hand, the perplexity calculation unit 13 calculates a perplexity value from the question text input from the question text input unit 11 (S03, perplexity calculation step). The calculated perplexity value is output from the perplexity calculation unit 13 to the determination unit 14. Note that the processing order of S02 and the processing of S03 are independent of each other, and therefore the processing order does not necessarily have to be as described above.

  Subsequently, the determination unit 14 determines whether or not the input question sentence is a question sentence regarding an objective fact based on the entity extraction result and the perplexity value as follows. First, it is determined whether an entity has been extracted (S04, determination step). When the entity has not been extracted (NO in S04), the determination unit 14 sets thr1 as the threshold value and compares the perplexity value with thr1 (S05, determination step). That is, it is determined whether the perplexity value is very low.

  If the entity has been extracted (YES in S04), the determination unit 14 sets thr2 as the threshold value and compares the perplexity value with thr2 (S06, determination step). That is, it is determined whether the perplexity value is low to some extent.

  As a result of the comparison, when the perplexity value is determined to be equal to or less than the threshold value (thr1 or thr2) (YES in S05 or S06), the question sentence input by the determination unit 14 is a question sentence regarding objective facts. And the question text is output to the objective fact response unit 15. Subsequently, an answer to the inputted question sentence is generated by the objective fact response unit 15 and output by being transmitted to the user terminal (S07, first answer step). The answer is generated by a keyword search request to the external search engine 30.

  As a result of the comparison in S05 or S06, if the perplexity value exceeds the threshold value (thr1 or thr2) (NO in S05 and S06), the question entered by the determination unit 14 relates to something other than objective facts. It is determined that it is a question sentence, and the question sentence is output to the other question answering unit 16. Subsequently, an answer to the inputted question sentence is generated by the other question answering unit 16 and transmitted to the user terminal to be output (S08, second answer step). The reply is generated by referring to the table shown in FIG. The answer transmitted from the question answering apparatus 10 in S07 or S08 is received by the user terminal 20 and output in a format that can be recognized by the user. The above is the question field determination method that is the process executed by the question answering apparatus 10 according to the present embodiment.

  As described above, in this embodiment, after considering both the entity extraction result from the question sentence and the perplexity value calculated from the question sentence, the question sentence is a question sentence regarding an objective fact. It is determined whether or not there is. In the extraction of an entity, an entity is extracted when the question sentence is in the form of a decent question and the question sentence includes a portion corresponding to the subject. In other words, in the present embodiment, it is evaluated whether or not it is possible to be sure that the question sentence is a more objective fact based on the perplexity value and the existence of the entity. In this way, not only the perplexity value but also the information on whether or not the entity has been extracted is used, so that the perplexity value can be evaluated more flexibly, and the question text can be accurately related to objective facts. It can be determined whether or not. This also makes it possible to accurately answer the user's question in the question answering system 1 that can respond to the user's question including a chat and a response to an objective fact.

  In the present embodiment, it is determined whether or not the question sentence is a question sentence regarding objective facts, but it is determined whether or not it is a question sentence regarding an arbitrary field other than the question sentence regarding objective facts. It is good as well. In this case, the perplexity value is calculated using a language model that assumes a question sentence in the field.

  Moreover, it is good also as calculating a perplexity value with the language model based on the context of the word contained in a question sentence like this embodiment. According to this configuration, the perplexity value can be calculated appropriately and reliably, and the present invention can be implemented appropriately and reliably. However, if the perplexity value can be calculated by another method (using a language model other than the above), that method may be used.

  Further, as in the present embodiment, a threshold value corresponding to the entity extraction result may be set, and the determination may be performed by comparing the set threshold value with the perplexity value. According to this configuration, the determination can be performed appropriately and reliably, and the present invention can be implemented appropriately and reliably.

  Moreover, it is good also as a means to produce | generate an answer with respect to question sentences, such as the objective fact response part 15 and the other question response part 16, like this embodiment. According to this configuration, an answer to the question sentence can be output. However, generating an answer to the question sentence is not necessarily essential for the present invention. That is, the present invention may determine whether or not the question sentence is a question sentence regarding an objective fact, and output the determination result.

  In addition, all the functions according to the present invention are provided in the question answering device 10 as the functions of the present embodiment. However, all the functions according to the present invention are not necessarily provided in the question answering apparatus 10. That is, the physical arrangement of functions according to the present invention is not limited to that of the above-described embodiment. For example, some of the functions according to the present invention may be provided in the user terminal 20. In that case, the question answering device 10 and the user terminal 20 constitute a question field determination device (question field determination system) according to the present invention. Thus, the question field determination device according to the present invention may be configured by a plurality of physical devices.

  Subsequently, a modification of the present embodiment will be described. In the embodiment described above, whether or not an entity can be extracted by the entity extraction unit 12, that is, whether an entity is “present” or “none” is used to determine whether or not the question is a question regarding an objective fact. However, in the entity extraction, not only the “existence” and “absence” of the entity, but also a probability value indicating the degree of certainty can be calculated, and this can be used for the determination. In many sequence labeling methods such as CRF, a probability value when a sequence is given is calculated for a sequence of labels of detection candidates, and the value that can be obtained is used to output a sequence with the highest probability. it can.

  In this case, the determination unit 14 applies a threshold value to a range of probability values instead of “present” and “none” of the entity. For example, when the probability value is 0.0 or more and less than 0.3, thr1 = 10, when 0.3 or more and less than 0.7, thr2 = 50, and when 0.7 or more and 1.0 or less, thr = 100. Compare the values of. This makes it possible to flexibly take into account the degree of conformity of the language model of objective facts, depending on the degree of confidence regarding the extraction of entities. For example, compared to the case where only two threshold values are used, when the entity is extracted but the probability that the sequence is extracted is low, the question sentence is objectively set unless the perplexity value is slightly higher than normal. It is possible to perform an operation such that the question is not considered to be a question sentence related to an actual fact.

  Subsequently, a second embodiment according to the present invention will be described. FIG. 5 shows a question answering device 10a which is a question field determination device according to the second embodiment of the present invention, and a question answering system 1a including the question answering device 10a. The second embodiment according to the present invention is the same as the first embodiment with respect to parts that are not particularly described. In the question answering apparatus 10a according to the present embodiment, the determination unit 14a and the objective fact response unit 15a have different functions from the determination unit 14 and the objective fact response unit 15 according to the first embodiment. Further, the question answering apparatus 10a according to the present embodiment includes an additional information prompting part 17a as a new functional part in addition to the question answering apparatus 10 according to the first embodiment.

  The determination unit 14a calculates the degree that the question sentence is a question sentence related to an objective fact based on the entity extraction result by the entity extraction unit 12 and the perplexity value calculated by the perplexity calculation unit 13. Based on the calculated degree, it is determined whether or not the question sentence is a question sentence regarding objective facts. As the degree, the determination unit 14a calculates the probability that the question sentence is a question sentence regarding an objective fact. That is, the determination unit 14a does not directly determine whether the question sentence is a question sentence related to the objective fact directly from the entity extraction result and the perplexity value, but once the question sentence relates to the objective fact. The probability of being a question sentence is calculated as a value between 0 and 1.

  The probability that the question sentence is a question sentence related to the objective fact is obtained by a class classification method such as a naive Bayes classifier. Here, the probability that the question sentence is a question sentence regarding an objective fact is expressed as P (t = 1 | ppl, e). When t = 1, it indicates that the question sentence is a question sentence regarding objective facts, and when t = 0, the other is indicated. ppl represents a perplexity value. e indicates the presence or absence of entity extraction.

In the naive Bayes classifier, P (t = 1 | ppl, e) is calculated by the following equation using a learning model in which a question sentence regarding objective facts and a question sentence that is not so are used as sentence examples.
P (t = 1 | ppl, e) = P (ppl, e | t = 1) × P (t = 1) / P (ppl, e)
In the above formula, the right-hand side indicates the perplexity value and the probability (P (ppl, e | t = 1)) regarding the set of presence / absence of entity extraction in the case of a question sentence regarding objective facts, and the question sentence regarding objective facts. Divided by the probability (P (ppl, e)) of the pair of perplexity value and entity extraction presence / absence (probability of being asked about objective fact) (P (t = 1)) Value. Each value can be estimated by obtaining sufficient example sentences.

  Specifically, P (ppl, e | t = 1) is a set of target perplexity value and entity extraction presence / absence of the question example regarding objective fact (calculated by the perplexity calculation unit 13). The perplexity value and the entity extraction result set by the entity extraction unit 12). P (t = 1) is estimated based on the ratio of the question sentence example regarding the objective fact to the full sentence example including the question sentence example regarding the objective fact and the question sentence example that is not so. P (ppl, e) is estimated based on the ratio of the target perplexity value and the presence / absence of entity extraction for all sentence examples.

  As described above, by calculating the probability of the question sentence relating to the objective fact, it is possible to accurately grasp the degree to which the question sentence is the question sentence relating to the objective fact. The determination unit 14a holds a new preset threshold value (Rthr), and determines that the calculated probability is equal to or greater than the threshold value when it is determined that the question sentence is not a question sentence regarding objective facts. To do. If it is determined that the calculated probability is greater than or equal to the threshold, it can be considered that the user has asked some objective fact as the intent of the question, but there is not enough information. In this case, the user is further prompted for additional information for answering the question text. If the determination unit 14a determines that the calculated probability is greater than or equal to the threshold, the determination unit 14a notifies the additional information prompting unit 17a that the user is prompted (requested) for additional information. When the entity extraction unit 12 extracts an entity, the determination unit 14a notifies the additional information prompting unit 17a together with the entity.

  The additional information prompting unit 17a is an information addition requesting unit that requests addition of information related to the entity extracted by the entity extracting unit 12 for answering the question sentence. Upon receiving the notification from the determination unit 14a, the additional information prompting unit 17a makes the request by transmitting (replying) information to the user terminal 20 to prompt (request) information addition. For example, the information (reply) to be transmitted to the user terminal 20 is “Please listen a little more clearly. Please include the subject, predicate, and interrogative. For example,“ What is the name of the American president? ”. . When the entity has already been extracted (that is, when the entity is notified from the determination unit 14a to the additional information prompting unit 17a), the additional information prompting unit 17a can reply using the character string of the entity. is there. For example, when the question sentence initially input from the user terminal 20 is “America President?”, The extracted entity “U.S. President” is used, You can also reply like "Do you want to?"

  The user terminal 20 receives the transmitted information and performs output such as display output. The user of the user terminal 20 recognizes the necessity of additional information from the output, and inputs the additional information corresponding to the output to the user terminal 20. The user terminal 20 transmits the input additional information to the question answering apparatus 10a. The question answering apparatus 10a receives the additional information. The additional information may be processed in the same manner as the original question sentence, or the original question sentence and the additional information are combined and input to the objective fact response unit 15a, and the objective fact response unit 15a performs the initial information. An answer may be generated from the question sentence and the additional information.

  Further, the additional information prompting unit 17a may hold entity-property knowledge data in advance and enumerate property values corresponding to the extracted entities to prompt the user to select. Entity-property-type knowledge data is one of the forms of data representation used when answering questions about objective facts. A fact is a property (property) that is an entity that is a subject and an attribute that the subject has. ) Value (Value). FIG. 6 shows knowledge data of an entity-property type.

  The additional information prompting unit 17a can include, in the reply, a property associated with the extracted entity in the entity-property type knowledge data. In the example of knowledge data shown in Fig. 6, if the entity "America" is extracted, for example, "What do you look for in the US? Can you answer about the President / Prime Minister, area, and population?" Prompt for additional information by information.

  Further, the additional information prompting unit 17a holds another threshold value (Pth), and if there are fewer property types associated with the extracted entity in the knowledge data, the additional information prompting unit 17a is prompted for additional information. Alternatively, all information on the corresponding entity may be returned as an answer to the question sentence using the knowledge data. In the example of the knowledge data shown in FIG. 6, the reply is “American president and prime minister is Mr. xxxx. The area is yyyyyy square meters and the population is zzzzz people”.

  Further, the additional information prompting unit 17a may hold a user's past question history that is a question sentence transmitted in the past from the user terminal 20, and return a question example therefrom. An example of the user's past question history is shown in FIG. In this question history, the question text, the entity extracted from the question text, and the perplexity value calculated from the question text are associated with each other. The entity associated with the question sentence is, for example, extracted by the entity extraction unit 12 in the past. The perplexity value associated with the question sentence is, for example, calculated by the perplexity calculating unit 13 in the past. As for past question histories, past question histories of all users can be used without distinguishing between users.

  The additional information prompting unit 17a searches for a question history in which the same entity as the question text to be processed is extracted. The additional information prompting unit 17a prompts information addition based on the question history. When “USA” is extracted as an entity, in the example of the question history shown in FIG. 7, there are three histories of “Who is the US President?” “Who is the US President?” Is extracted. Most simply, the additional information prompting unit 17a returns all of these question histories to the user as answer examples. For example, the answer sentence can be “What do you look for in the United States? For example,“ Who is the American president? ”“ Who is the US president?

  Since it may be possible to output a large amount of the above-mentioned history when operated for a long period of time, these may be narrowed down by the perplexity value. For example, only perplexity values that are equal to or less than a preset threshold value (eg, 3000) may be output. In this case, two question histories of “who is the US president” and “who is the US president” are presented to the user.

  In addition, not only the question text but also the quality of the answer may have some judgment criteria, and only the question history with high answer quality may be answered. For example, the number of search results acquired by the objective fact response unit 15a from the external search engine 30 may be used as an evaluation value, and only the question texts obtained with many search results may be used. What kind of index is actually used is determined depending on the mechanism used when the objective fact response unit 15a evaluates the answer.

  As described above, according to the present embodiment, even if it is not possible to sufficiently determine whether or not the input question sentence is a question sentence regarding objective facts, information necessary for grasping the question is acquired. be able to. This also enables an accurate response to the user's question.

  DESCRIPTION OF SYMBOLS 1, 1a ... Question answering system 10, 10a ... Question answering device, 11 ... Question text input part, 12 ... Entity extraction part, 13 ... Perplexity calculation part, 14, 14a ... Determination part, 15, 15a ... Objective Fact response unit, 16 ... other question response unit, 17a ... additional information prompting unit, 101 ... CPU, 102 ... RAM, 103 ... ROM, 104 ... communication module, 105 ... auxiliary storage device, 20 ... user terminal, 30 ... external search engine.

Claims (6)

A question sentence input means for inputting a question sentence;
Entity extraction means for extracting an entity that is a question target from the question text input by the question text input means;
Perplexity calculation means for calculating a perplexity value by a language model assuming a question text in a specific field from the question text input by the question text input means,
Based on the entity extraction result by the entity extraction means and the perplexity value calculated by the perplexity calculation means, the question sentence input by the question sentence input means is a question sentence in the specific field. Determination means for determining whether or not there is;
A question field judgment device comprising:   The question field determination device according to claim 1, wherein the perplexity calculating unit calculates the value of the perplexity based on a context of words included in the question sentence.   The determination unit sets a threshold value according to the extraction result of the entity, compares the set threshold value with the perplexity value, and determines whether the question message is a question message in the specific field. The question field judgment device according to claim 1 or 2 which judges. If the determination means determines that the question sentence is a question sentence in the specific field, a first answer means for generating and outputting an answer to the question sentence by a first method;
When the determination means determines that the question text is a question text in a field other than the specific field, an answer to the question text is generated by a second method different from the first method. A second answering means for outputting;
The question field judgment device according to any one of claims 1 to 3, further comprising: The determination means calculates the degree that the question sentence is a question sentence in the specific field based on the extraction result of the entity and the value of the perplexity,
The information addition request | requirement means which requests | requires the information addition regarding the entity extracted by the said entity extraction means for answering the said question sentence according to the said degree calculated by the said determination means is further provided. The question field judgment device given in any 1 paragraph of above. A question sentence input step for inputting a question sentence;
An entity extraction step of extracting an entity as a question target from the question sentence input in the question sentence input step;
A perplexity calculation step of calculating a perplexity value by a language model assuming a question sentence in a specific field from the question sentence input in the question sentence input step;
Based on the entity extraction result in the entity extraction step and the perplexity value calculated in the perplexity calculation step, the question sentence input in the question sentence input step is a question sentence in the specific field. A determination step of determining whether or not there is;
Question field judgment method including

Images