JP2017010201A - Procedure expression extraction method, procedure expression extraction device, and procedure expression extraction program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a procedure expression extraction method, a procedure expression extraction device, and a procedure expression extraction program for accurately extracting procedure expression from a text.SOLUTION: A parameter estimation part 20 estimates weight to each of predetermined feature quantities representing features regarding a tag showing that it is procedure expression to be extracted from a text as a parameter based on a plurality of pieces of training case data which is the text to which the tag is preliminarily given. A text input part 22 accepts input of the text to be considered as an object of extraction of the procedure expression. A procedure expression extraction part 24 extracts the procedure expression from the text based on the feature quantities to be extracted from the text accepted by the text input part 22, and the weight to each of the feature quantities.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a procedure expression extraction method, a procedure expression extraction apparatus, and a procedure expression extraction program, and more particularly to a procedure expression extraction method, a procedure expression extraction apparatus, and a procedure expression extraction program for extracting a procedure expression from text.

  A system in which a human makes a question to a machine and the machine that receives the question returns an appropriate response is called a question answering system. Since the user can quickly obtain desired information by using the question answering system, the importance of the question answering system is increasing day by day in the present age when various information is inundated.

  In many conventional question answering systems, the range of questions that the system can answer is limited to factoids. A factoid is a specific expression expressed by a person name, a place, a word, a noun phrase, or the like. For example, if you receive the question "Who is the first prime minister in Japan?" Or if you answer "Hirofumi Ito" or "Which prefecture is Japan with the highest population?" Or “Tokyo”.

  At present, however, systems have been developed that accept and answer questions that can only be answered in sentences, such as questions that ask reason. For example, when the question “why the sky is blue” is accepted, the answer is “caused by the wavelength of light”.

  In order for the question answering system to answer more various types of questions and be put to practical use, it is necessary to be able to answer even when a question asking for a procedure is accepted. For example, when a question “Please tell me how to make instant coffee” is accepted, it is necessary to answer in multiple sentences in order to explain a series of procedures. For example, to the above question, “First, open the lid of the instant coffee jar and put the instant coffee powder into a cup or other container. Next, pour hot water into the cup or other container. It is necessary to answer such as

  Non-patent document 1 is cited as a prior art document that generates an answer to such a question. In the technique disclosed in Non-Patent Document 1, first, a document search is performed using a given question, and then a keyword that has a high probability of being included in a place where a series of procedures described in advance is manually described. Using the set, generate an answer to the question that asks the procedure.

Satoshi Nakakura and Junichi Fukumoto, "Question Answering System beyond Factoid Type Questions", In Proceedings of the 23rd International Conference on Circuits / Systems, Computers and Communications (ITC-CSCC 2008), pp. 617-620, 2008.

  Here, the method disclosed in Non-Patent Document 1 has two problems. The first problem is that individual keywords must be set manually. Therefore, there is no method for verifying whether or not an appropriate set of keywords can be covered. The second problem is that, when generating an answer, only a plurality of sentences that are supposed to include a specific set of keywords are extracted. Therefore, even when a procedure is explained in order by a plurality of sentences, the structure of the order cannot be grasped.

  The present invention has been made in view of the above circumstances, and an object of the present invention is to provide an expression extraction method, a procedure expression extraction apparatus, and a procedure expression extraction program for accurately extracting procedure expressions from text.

  In order to achieve the above object, a procedural expression extraction method of the present invention is a procedural expression extraction method in a procedural expression extraction device comprising a parameter estimation unit, a text input unit, and a procedural expression extraction unit, wherein the parameter estimation unit Is based on a plurality of training case data in which a tag indicating that it is a procedural expression is a text given in advance. The step of estimating as, the step in which the text input unit accepts input of text to be extracted from the procedure expression, and the procedure expression extraction unit is extracted from the text received by the text input unit Based on the feature amount and the weight for each of the feature amount, Including a step of extracting more representation, a.

  The training example data text is preliminarily provided with a tag indicating that it is a procedure and a tag indicating that it is a procedure title, and the parameter estimation unit estimates a weight for the feature amount as a parameter. Includes a process in which the parameter estimation unit estimates, as the parameter, a weight for each feature amount representing a feature related to each of the tags extracted from a sentence based on the plurality of training case data, and the procedural expression The step of extracting the procedure expression by the extraction unit is based on the feature amount extracted from the text received by the text input unit by the procedure expression extraction unit and the weight for each of the feature amounts. Extract the title and the procedure from the text received by the text input unit That process may include a.

  The feature amount includes whether the sentence with the tag includes a predetermined word, whether the sentence with the tag includes a predetermined part of speech, and the tag It is also possible to include at least one of whether or not the beginning of the sentence is a number, and whether or not the similarity between the sentences to which the tag is attached is equal to or greater than a predetermined threshold.

  The procedural expression extraction device according to the present invention is a predetermined characteristic representing a characteristic relating to the tag, which is extracted from the text based on a plurality of training case data in which the tag indicating the procedural expression is a text given in advance. A parameter estimation unit that estimates a weight for each of the quantities as a parameter, a text input unit that receives input of text to be extracted from the procedural expression, and the feature amount extracted from the text received by the text input unit And a procedural expression extracting unit that extracts the procedural expression from the text based on the weights for each of the feature quantities.

  The procedure expression extraction program of the present invention is a program for causing a computer to execute each step of the procedure expression extraction method.

  According to the present invention, it is possible to make a response considering the order of procedures without manually describing individual procedure expressions.

It is a block diagram which shows the structure of the procedure expression extraction apparatus which concerns on embodiment. It is a flowchart which shows the flow of the procedure expression extraction process performed by the procedure expression extraction apparatus which concerns on embodiment. It is a schematic diagram which shows an example of the training example data which concern on embodiment. It is a schematic diagram which shows an example of the parameter data which concern on embodiment. It is a schematic diagram which shows an example of the text made into the object of extraction of the procedure expression which concerns on embodiment. It is a schematic diagram which shows an example of the text which was the object of extraction of the procedure expression which concerns on embodiment, and was annotated. It is a schematic diagram which shows an example of the procedure expression data which concern on embodiment.

  Hereinafter, a procedure expression extraction device according to the present embodiment will be described in detail with reference to the drawings.

  As shown in FIG. 1, the procedure expression extraction apparatus 10 according to the present embodiment is based on a plurality of training case data that is a text in which a tag indicating a procedure expression and a tag indicating a title are given in advance. The parameter estimation unit 20 estimates the weight for each of the predetermined feature amounts representing the features related to each tag as a parameter. The procedural expression extraction device 10 also includes a text input unit 22 that receives input of text to be extracted from procedural expressions. The procedural expression extraction device 10 extracts a procedural expression from the text received by the text input unit 22 based on the feature amount extracted from the text received by the text input unit 22 and the weight for each feature amount. A procedural expression extraction unit 24 is included.

  Furthermore, the procedural expression extraction device 10 includes a storage unit 18 that stores various data necessary for procedural expression extraction processing for extracting a procedural expression from text that is a target of procedural expression extraction. The storage unit 18 includes a training case data storage unit 18a that stores training case data described later, a parameter data storage unit 18b that stores parameter data described later, and a procedure expression storage unit 18c that stores procedure expressions described later. Have.

  The procedure expression extraction apparatus 10 according to the present embodiment includes, for example, a CPU (Central Processing Unit), a RAM (Random Access Memory), and a ROM (Read Only Memory) that stores various programs including a program for procedure expression extraction processing described later. It is comprised with the computer apparatus provided with. The computer constituting the procedural expression extraction device 10 may include a storage unit such as a hard disk drive or a nonvolatile memory. In the present embodiment, a case where a nonvolatile memory is provided instead of the ROM will be described. A program executed by the CPU may be stored in a storage unit such as a hard disk drive or a nonvolatile memory. When the CPU reads and executes a program stored in a storage unit such as a ROM or a hard disk, the hardware resource and the program cooperate to realize the function described above.

  Next, referring to FIG. 2, an operation of a procedure expression extraction process executed by the procedure expression extraction apparatus 10 according to the present embodiment when an execution instruction is input by a user will be described as an example.

  In step S101, the parameter estimation unit 20 reads the plurality of training case data described above from the training case data storage unit 18a. The training case data is data indicating a training case used when learning the procedural expression. As shown in FIG. 3 as an example, the training case number is annotated with a training case number for identifying the training case and a tag. This data is associated with training examples including procedural expressions. The training example includes a plurality of sentences that are procedural expressions, and is a part or all of a plurality of sentences representing a series of procedures, or the entire document posted on a single web page or the like.

  In the present embodiment, the training example includes a sentence that does not relate to a series of procedures, a sentence that represents a title of the series of procedures, and a sentence that represents a series of procedures. The sentence representing a series of procedures may be a sentence representing a part of the series of procedures or a sentence representing all of the series of procedures.

  In the example of training case number 1 in the training case data shown in FIG. 3, a sentence “There are various ways to make cup ramen” with no tag, and “How to make cup ramen in this way simply” The sentence "can do" corresponds to a sentence that is not related to a series of procedures. In addition, a sentence “an example of a procedure for creating a cup ramen” shown with a <title> tag as a title tag corresponds to a sentence representing a title. Furthermore, six sentences such as a sentence “1. Open the lid of the cup ramen” with a <procedure> tag as a procedure tag correspond to sentences representing a series of procedures.

  As described above, in this embodiment, the training example includes a sentence that does not relate to a series of procedures, a sentence that represents the title of a series of procedures, and a sentence that represents a series of procedures, but the present invention is not limited thereto. The training example may include only a sentence representing a series of procedures, or may include only a sentence representing a series of procedures and a sentence representing a series of procedures.

  In this embodiment, a sentence representing a series of procedures is enclosed in <procedure> tags and </ procedure> tags, and a sentence representing a series of procedures is enclosed in <title> tags and </ title> tags. However, it is not limited to this. A sentence representing a series of procedures is attached with a tag which can be determined to be a sentence representing a series of procedures, and a sentence representing the title of a series of procedures may be a sentence representing the title of a series of procedures. A tag that can be determined may be attached.

  In addition, when a sentence representing a more detailed procedure is incorporated in a part of a series of procedures, a hierarchical structure tag may be attached to the sentence representing the series of procedures. In this case, for example, a statement representing a series of procedures is enclosed in <main-procedure> tags and </ main-procedure> tags, and a statement representing further detailed procedures is enclosed in <sub-procedure> tags and </ sub-procedure tags. > Surround with tags. Thereby, a procedural expression having a hierarchical structure can be extracted.

  In step S103, the parameter estimation unit 20 estimates, as a parameter, a weight for each feature amount related to each tag from the plurality of read training case data. In the present embodiment, the parameters are estimated by the following formulas (1) and (2) using a conditional random field that is a known method.

In the above formulas (1) and (2), x = {x ₁ , x ₂ ,..., X _m } is a sequence of each sentence included in the training example. In the example shown in FIG. 3, x ₁ = “There are various ways to make cup ramen.”, X ₂ = “An example of a procedure for making a cup ramen is shown below.”, X ₃ = “1. Open the lid. ”

In the above formulas (1) and (2), y = {y ₁ , y ₂ ,..., Y _n } is a string of tags of each sentence included in the training example. In the example illustrated in FIG. 3, the tag of each sentence is “no tag”, “title”, or “procedure”.

In the above equations (1) and (2), the function f (x, y) is a function that returns a feature vector corresponding to the argument x and the argument y. In the present embodiment, the feature vector f (x, y) is a sequence of elements corresponding to a plurality of predetermined feature amounts representing the feature of the sentence related to each tag. In the present embodiment, these feature quantities are defined by a combination of an argument x and an argument y. In the example shown in FIG. 3, a sequence of feature quantities is represented by z = {z ₁ , z ₂ ,..., Z _k. }, Z ₁ = “procedure-sentence begins with a number”, z ₂ = “procedure—includes“ after ””, and z ₃ = “title—includes“ procedure ””. f (x, y) is represented by a vector in which an element corresponding to the combination x is 1 and an element that does not correspond is 0, depending on the combination of the values of the argument x and the argument y.

  The feature amount is, for example, whether a sentence with a <procedure> tag includes a predetermined word (for example, “after”, “procedure”, etc.), or a sentence with a <procedure> tag. Whether the sentence contains a specific part of speech (for example, adjectives, adverbs, etc.), whether the sentence beginning with a <procedure> tag is a number, and between sentences with a <procedure> tag It includes at least one of whether or not the sentence similarity is greater than or equal to a predetermined threshold.

  In the example of training case number 1 shown in FIG. 3, the feature quantity is “prodedure-sentence begins with a number” based on the description “<procedure> 1. Open the cup ramen lid. </ Procedure>” In this case, the element corresponding to the feature amount is set to 1. Also, based on the description of "<title> Cupramen making procedure shown below. </ Title>", if the feature is "title-" includes procedure "", it corresponds to the feature The element to be set is 1.

  In the above equations (1) and (2), w is a parameter vector representing the weight for each feature amount. In the above formulas (1) and (2), w · f (x, y) is an inner product of w and f (x, y). In the above equations (1) and (2), the probability P (y | x) represents the probability of being an argument y under the condition of the argument x.

  In the present embodiment, when the parameter estimation unit 20 estimates a parameter from the read training case data, when the feature vector f (x, y) is given to the above equation (1) for each training case, the probability P ( A parameter vector w that maximizes y | x) is obtained.

  As a method for obtaining the parameter vector w that maximizes the probability P (y | x), a known method can be applied. For example, the method disclosed in Reference Document 1 below can be applied. .

[Reference 1] John Lafferty, Andrew McCallum and Fernand CN Pereira, "Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data", In Proceedings of the 18th International Conference on Machine Learning (ICML 2001), pages 282-289, 2001.

  The parameter estimation unit 20 stores the obtained parameter vector w in the parameter data storage unit 18b as parameter data. In the present embodiment, as shown in FIG. 4 as an example, data in which each feature amount is associated with a weight for the feature amount is stored in the parameter data storage unit 18b as parameter data.

  In the example shown in FIG. 4, the weight corresponding to the feature quantity “procedure-sentence begins with a number” indicates how easily the <procedure> tag is attached when the beginning of the sentence is a number. ing.

  In step S <b> 105, the text input unit 22 receives an input of text to be extracted from the procedure expression and stores it in the storage unit 18. As shown in FIG. 5 as an example, the text includes a plurality of sentences such as “To make money with ATM, do as follows”.

  In step S <b> 107, the procedure expression extraction unit 24 extracts a procedure expression from the text received by the text input unit 22 based on the parameter data estimated by the parameter estimation unit 20. In this embodiment, the procedure expression is expressed by the following equation (3), and the y that is the sequence of tags of each sentence when the inner product w · f (x, y) is maximum is obtained using the Viterbi algorithm. Extract.

  For example, when the number of sentences included in the text is four and the extracted argument y is {title, procedure, procedure, procedure}, as illustrated in FIG. 6 as an example, a plurality of sentences included in the text are included. <Title> tag, <procedure> tag, <procedure> tag, and <procedure> tag can be added to each of the sentences in order.

  Then, by applying the extracted argument y to the above text, only a sentence with a <title> tag or a <procedure> tag is extracted as a procedure expression. Therefore, the sentence “No tag” is not extracted.

  In step S109, the procedure expression extraction unit 24 stores the extracted procedure expression in the procedure expression storage unit 18c, and ends the execution of the program for the procedure expression extraction process. In this embodiment, as shown in FIG. 7 as an example, the procedure expression is separated into a sentence (title) with a <title> tag and a sentence (procedure) with a <procedure> tag, Memorize it in correspondence. If there is no sentence (title) with the <title> tag in the text, only the sentence (procedure) with the <procedure> tag is stored.

  In the example shown in FIG. 7, in response to the title extracted as follows from the text shown in FIG. 5 "To make money with ATM, do as follows." A procedure expression associated with a plurality of procedures such as “Insert.” Is stored in the procedure expression storage unit 18c.

  As described above, the procedure expression extraction device 10 according to the present embodiment estimates, as a parameter, a weight for a predetermined feature amount representing a feature of the procedure expression from training example data including a plurality of procedure expressions, and the weight for the feature amount. Based on the procedural expression, the procedural expression is extracted from the text to be extracted. Thereby, it is possible to accurately extract the procedural expression from the text.

  In addition, each structure of the parameter estimation part 20, the text input part 22, and the procedure expression extraction part 24 with which the procedure expression extraction apparatus 10 which concerns on this embodiment is provided may be implement | achieved by dedicated hardware. Alternatively, it may be realized by a memory and a microprocessor. Each of these components is configured by a memory and a CPU (central processing unit), and a program for realizing the function of each component is loaded into the memory and executed to realize the function. Good.

  Further, a program for realizing the function of each processing unit of the procedure expression extraction device 10 according to the present embodiment is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into a computer system. The procedural expression extraction process may be performed by executing. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer system” includes a WWW system having a homepage providing environment (or display environment). The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Further, the “computer-readable recording medium” refers to a volatile memory (RAM) in a computer system that becomes a server or a client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. In addition, those holding programs for a certain period of time are also included.

  The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, what is called a difference file (difference program) may be sufficient.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes design and the like within the scope not departing from the gist of the present invention. .

DESCRIPTION OF SYMBOLS 10 Procedure expression extraction apparatus 18 Storage part 18a Training example data storage part 18b Parameter data storage part 18c Procedure expression storage part 20 Parameter estimation part 22 Text input part 24 Procedure expression extraction part

Claims (5)

A procedural expression extraction method in a procedural expression extraction device comprising a parameter estimation unit, a text input unit, and a procedural expression extraction unit,
Each of the predetermined feature amounts representing the feature related to the tag, which is extracted from the text based on a plurality of training case data in which the parameter estimation unit is a text given in advance as a tag indicating that it is a procedural expression Estimating a weight for as a parameter;
The text input unit accepting input of text to be extracted from the procedural expression;
The procedural expression extraction unit extracting the procedural expression from the text based on the feature amount extracted from the text received by the text input unit and the weight for each of the feature amount;
Procedural expression extraction method. In the text of the training example data, a tag indicating that it is a procedure and a tag indicating that it is a title of the procedure are given in advance,
The step in which the parameter estimation unit estimates the weight for the feature amount as a parameter includes a feature amount representing a feature related to each of the tags extracted from a sentence based on the plurality of training case data. Including estimating a weight for each as the parameter,
The step of extracting the procedural expression by the procedural expression extracting unit includes the step of extracting the procedural expression from the feature amount extracted from the text received by the text input unit and the weight for each of the feature amounts. The procedural expression extraction method according to claim 1, further comprising: extracting the title and the procedure from the text received by the text input unit. The feature amount includes whether the sentence with the tag includes a predetermined word, whether the sentence with the tag includes a predetermined part of speech, and the tag. 3. The method according to claim 1, comprising at least one of whether the sentence head is a number and whether the similarity between the sentences to which the tag is attached is equal to or greater than a predetermined threshold. Procedural expression extraction method. Based on a plurality of training case data in which a tag indicating procedural expression is a text given in advance, a weight for each of the predetermined feature amounts representing the feature related to the tag extracted from the text is estimated as a parameter. A parameter estimation unit to perform,
A text input unit that accepts input of text to be extracted from the procedural expression;
A procedural expression extracting unit that extracts the procedural expression from the text based on the feature amount extracted from the text received by the text input unit and the weight for each of the feature amount;
A procedural expression extraction device.   A procedural expression extraction program for causing a computer to execute each step of the procedural expression extraction method according to claim 1.

Images