JP7155758B2 - Information processing device, information processing method and program - Google Patents

Description

The present invention relates to an information processing device, an information processing method, and a program.

There is a sentence generation technology in which a computer automatically generates sentences. For example, Patent Document 1 discloses a descriptive text generation method for generating descriptive text summarizing a document input to a computer using a recurrent neural network (RNN; hereinafter referred to as "RNN"). It is

JP 2018-28866 A

However, the invention according to Patent Literature 1 merely summarizes a document, and does not automatically generate sentences containing the text from the text specified by the user.

An object of one aspect of the present invention is to provide an information processing apparatus and the like capable of appropriately generating sentences including text specified by a user.

In one aspect, an information processing device stores a language model obtained by learning, in the order of the sentence, the elements that appear next to each element in a sentence composed of elements that are characters or character strings of a predetermined unit. a storage unit, a reception unit that receives input for specifying the elements to be included in a sentence to be generated , and an element that appears next to the element by inputting one of the specified elements to the language model, and specifying the element a generation unit that repeats the process of inputting the specified element and the obtained element into the language model and obtaining the element that appears next, and generates a sentence containing the specified element and the obtained elements ; and an output unit that outputs the sentence that has been written, and the generation unit inputs an element that exists at the beginning of the learning source sentence used for learning the language model to the language model, and an element that appears next to the element and generate a head phrase that includes the head element and the element that appears next to the obtained element, or input the element that exists at the head of the learning source sentence into the language model and next to the element Obtaining an appearing element, inputting the first element and the obtained element to the language model, and repeating the process of obtaining the next appearing element, and obtaining the first phrase including the first element and each obtained element Generate a sentence containing the generated head phrase and the specified element .

In one aspect, it is possible to appropriately generate sentences that include user-specified text.

It is a schematic diagram which shows the structural example of a sentence production|generation system. It is a block diagram which shows the structural example of a server. FIG. 4 is an explanatory diagram showing an example of a record layout of a keyword DB and a replacement dictionary; FIG. 10 is an explanatory diagram relating to sentence learning processing; FIG. 4 is an explanatory diagram of a language model; It is explanatory drawing regarding sentence production|generation processing. FIG. 3 is a screen image diagram displayed by a terminal; It is a flowchart which shows an example of the processing procedure of sentence learning processing. It is a flow chart which shows an example of a processing procedure of sentence generation processing. FIG. 10 is an explanatory diagram showing an overview of Embodiment 2; 10 is a flow chart showing an example of a processing procedure of sentence generation processing according to Embodiment 2; FIG. 11 is an explanatory diagram showing an overview of Embodiment 3; 14 is a flow chart showing an example of a processing procedure of sentence generation processing according to Embodiment 3; FIG. 12 is an explanatory diagram showing an outline of a fourth embodiment; FIG. FIG. 12 is a flow chart showing an example of a processing procedure of sentence generation processing according to Embodiment 4; FIG. FIG. 21 is an explanatory diagram of sentence learning processing according to Embodiment 5; FIG. 21 is an explanatory diagram of sentence generation processing according to Embodiment 5; FIG. 14 is a flow chart showing an example of a processing procedure of sentence learning processing according to Embodiment 5. FIG. FIG. 13 is a flow chart showing an example of a processing procedure of sentence generation processing according to Embodiment 5. FIG. It is a functional block diagram which shows operation|movement of the server of the form mentioned above.

Hereinafter, the present invention will be described in detail based on the drawings showing its embodiments.
(Embodiment 1)
FIG. 1 is a schematic diagram showing a configuration example of a sentence generation system. In this embodiment, a text generation system that automatically generates a text of a contract based on words specified by a user will be described. The text generation system includes an information processing device 1, terminals 2, 2, 2, . . . Each device is connected for communication via a network N such as the Internet.

The information processing device 1 is an information processing device capable of various types of information processing and transmission/reception of information, and is, for example, a server device, a personal computer, or the like. In the present embodiment, the information processing device 1 is assumed to be a server device, and is simply replaced with the server 1 in the following description. The server 1 uses a group of contracts for learning to perform machine learning to learn words and phrases that appear in the contract, and can estimate words and phrases that appear at any point in the contract from the words and phrases that appear immediately before. generate a language model. A language model is a model of the probability of generating natural language sentences, and includes, for example, an N-gram model and a hidden Markov model. As will be described later, in the present embodiment, the server 1 constructs RNN as a language model from a group of contracts given for learning, and uses the RNN to generate sentences for contracts.

The terminal 2 is a terminal device used by each user who creates a contract, and is, for example, a personal computer, a multifunctional terminal, or the like. The terminal 2 receives a specification input for specifying a phrase to be included in the text of the contract to be created, and requests the server 1 to generate a sentence including the specified phrase (hereinafter referred to as "specified phrase" as appropriate). The terminal 2 receives the generated sentence from the server 1 and presents it to the user.

FIG. 2 is a block diagram showing a configuration example of the server 1. As shown in FIG. The server 1 includes a control section 11 , a main storage section 12 , a communication section 13 and an auxiliary storage section 14 .
The control unit 11 has an arithmetic processing unit such as one or more CPU (Central Processing Unit), MPU (Micro-Processing Unit), GPU (Graphics Processing Unit), etc., and executes the program P stored in the auxiliary storage unit 14. By reading and executing, various information processing, control processing, etc. related to the server 1 are performed. The main storage unit 12 is a temporary storage area such as SRAM (Static Random Access Memory), DRAM (Dynamic Random Access Memory), flash memory, etc., and temporarily stores data necessary for the control unit 11 to perform arithmetic processing. Remember. The communication unit 13 includes a processing circuit and the like for performing processing related to communication, and transmits and receives information to and from the outside.

The auxiliary storage unit 14 is a large-capacity memory, a hard disk, or the like, and stores programs P and other data necessary for the control unit 11 to execute processing. The auxiliary storage unit 14 also stores a language model 141 , a keyword DB 142 and a replacement dictionary 143 . The language model 141 is, as described above, language model data generated from a learning contract group, and is model data of each clause included in various contracts. As will be described later, in this embodiment, the server 1 generates a language model 141 for each category of sentences included in a contract (type of contract and clauses in the contract) and stores the language model 141 in the auxiliary storage unit 14. I have

The keyword DB 142 is a database that stores keywords that frequently appear in each clause of various contracts. For example, when the server 1 performs machine learning to generate the language model 141 from a group of contracts for learning, it extracts the keywords of each clause in the contract according to the appearance frequency and stores them in the keyword DB 142. .

The replacement dictionary 143 is a table that associates and stores a plurality of related terms such as synonyms and synonyms. As will be described later, when a sentence is generated by receiving designation of a word from the user, the server 1 refers to the replacement dictionary 143 and replaces the word specified by the user to generate the sentence.

Incidentally, the auxiliary storage unit 14 may be an external storage device connected to the server 1 . Moreover, the server 1 may be a multicomputer consisting of a plurality of computers, or may be a virtual machine virtually constructed by software.

Further, in the present embodiment, the server 1 is not limited to the above configuration, and may include, for example, a reading unit for reading information stored in a portable storage medium, an input unit for receiving operation input, a display unit for displaying images, and the like. .

FIG. 3 is an explanatory diagram showing an example of the record layout of the keyword DB 142 and replacement dictionary 143. As shown in FIG. The keyword DB 142 includes a type column, a clause column, and a keyword column. The type column stores the types of contracts from which keywords are extracted. The clause column stores each clause in the contract from which keywords are extracted, in association with the contract type. The keyword column stores one or more keywords extracted from each clause of each type of contract in association with the type and clause of the contract.

The replacement dictionary 143 includes a target word string and a replacement word string. The target word string stores target words to be replaced using the replacement dictionary 143 for the words specified by the user. The replacement word string stores words and phrases that replace the target word in association with the target word.

FIG. 4 is an explanatory diagram relating to sentence learning processing. FIG. 4 conceptually shows how the server 1 performs machine learning for learning the sentences of a learning contract and generates a plurality of language models 141, 141, 141, . The outline of the processing executed by the server 1 will be described below.

The server 1 acquires a group of contracts for learning from the outside, and performs machine learning to learn words and phrases appearing in sentences (learning source sentences) of each contract. For example, the server 1 performs learning using an existing manually created contract as a learning contract. First, the server 1 determines the type of the contract and each clause included in the contract from each contract, and structuring the document by classifying the sentences in the contract for each category (type and clause). conduct. The server 1 then learns each sentence of the structured contract for each category to generate separate language models 141 .

The type of contract is a classification that broadly categorizes the content of the contract, such as a nondisclosure agreement, outsourcing agreement, joint research agreement, and the like. All of the above are examples, and the type of contract is not particularly limited. As an example on the left side of FIG. 4, a contract relating to a confidentiality agreement is conceptually illustrated. As shown in the thick line frame in FIG. 4, a general contract is divided into a plurality of clauses and the contents of the contract are described. The server 1 learns the text of each clause for each clause of the contract. In addition, when the term "sentence" is used in this specification, the sentence is not limited to one sentence, but may include a sentence consisting of a plurality of sentences or a sentence shorter than one sentence. As will be described later, the server 1 learns words and phrases that appear in each clause of each type of contract according to the order of the words and phrases in the contract. As a result, the server 1 generates a language model 141 corresponding to each clause.

First, the server 1 classifies a plurality of learning contracts according to the type of the contract. For example, the server 1 determines the type of contract based on the document name of the contract. In the contract shown in FIG. 4, the server 1 determines from the title "confidentiality agreement" that the contract is related to the nondisclosure agreement.

Further, the server 1 identifies text portions corresponding to each clause in the contract, and classifies the text in the contract for each clause. For example, the server 1 identifies a clause based on a subheading (subtitle) corresponding to the title of each clause. In the example of FIG. 4, the server 1 determines from the subheading "Purpose" of "Article 1" that the text portion is a clause relating to the purpose of the contract.

In the above description, the server 1 determines the category (type and clause) of the contract based on the rule based on the title name, subheading, etc. of the contract, but the present embodiment is not limited to this. For example, the server 1 uses the tf-idf method or the like to calculate the degree of similarity between sentences in each contract according to the appearance frequency of words contained in the sentences, and calculates the degree of similarity between sentences to determine the structure of the contract. may be modified.

The server 1 repeats the above-described processing for each contract, discriminates the type of each of the plurality of learning contracts, and discriminates each clause described in each contract. That is, the server 1 structures the contract. The server 1 generates each language model 141 based on sentences having a common category (type and clause) in each structured contract. For example, as shown on the right side of FIG. 4, the server 1 generates one language model 141 based on the text of each contract whose type is "confidentiality agreement" and whose clause is "compensation for damages" in common. do. Similarly for the other language models 141, the server 1 generates the language models 141 according to the categories of sentences.

FIG. 5 is an explanatory diagram of the language model 141. As shown in FIG. In this embodiment, the server 1 constructs RNN as the language model 141 . For example, the server 1 constructs an LSTM (Long Short-Term Memory), which is a type of RNN.

The server 1 inputs the sentences classified for each category as described above to the RNN, and learns the words appearing in the sentences of each category. Here, the server 1 first performs natural language processing such as morphological analysis on the text input to the RNN, and divides the text into words and phrases (elements) that are characters or character strings of predetermined units. This division unit is, for example, a unit such as a word or a clause, but is not particularly limited. For example, the server 1 stores in advance a dictionary (not shown) storing a plurality of words and phrases, and divides sentences according to the words and phrases stored in the dictionary.

It should be noted that the server 1 may divide the text in units other than words, phrases, and the like. As another division unit, a unit called a subword (partial word) can be assumed. A subword is a unit of words and phrases into which a sentence is divided according to the frequency of appearance in the sentence, unlike the usual spaced words. "Word", which is generally used as the minimum structural unit of sentences, is a unit that minimizes characters or character strings in sentences from the viewpoint of meaning, grammar, etc., but subwords are not units based on meaning, grammar, etc. It is a unit minimized according to the frequency used in a sentence. When the document is divided into subwords, the server 1 can also divide technical terms specific to the contract, so that the document can be divided more appropriately. In this manner, the server 1 may divide the text of the contract into elements, which are characters or character strings in predetermined units, and the elements to be divided are not limited to units such as words.

The server 1 inputs each divided word/phrase to the input layer related to the RNN, and performs machine learning. FIG. 5 conceptually illustrates the configuration of the RNN. As shown in FIG. 5, the RNN has an input layer, an intermediate layer (hidden layer), and an output layer. The input layer has a plurality of neurons that receive inputs of words that appear in order from the beginning to the end of a sentence. The output layer has a plurality of neurons for estimating and outputting words appearing next to words input to each neuron corresponding to each neuron in the input layer. The intermediate layer has a plurality of neurons for computing output values (words) in each neuron in the output layer with respect to input values (words) in each neuron in the input layer. Each neuron in the intermediate layer estimates the next word/phrase from one or a plurality of words/phrases appearing immediately before by performing an operation regarding the next input value using the operation result in the intermediate layer regarding the past input value.

Note that the configuration of the RNN shown in FIG. 5 is an example, and the present embodiment is not limited to this. For example, the intermediate layer is not limited to one layer, and may be two or more layers. Also, the number of neurons in the input layer and the number of output layers are not limited to the same number, and for example, the number of outputs may be at least as many as the inputs.

In addition, although the server 1 performs learning according to the RNN algorithm in the present embodiment, learning can be performed according to other algorithms such as other deep learning, N-gram model, SVM (Support Vector Machine), Bayesian network, decision tree, etc. may be performed to generate the language model 141 .

The server 1 inputs each word of the learning sentence to each neuron of the input layer according to the order of arrangement in the sentence, and obtains an output value from each neuron of the output layer. In the example of FIG. 5, the server 1 inputs the words "A", "And", "Otsu", "GA", etc. of the learning sentence to each neuron of the corresponding input layer according to the order in the sentence. . The server 1 performs calculations in each neuron of the output layer via the intermediate layer, calculates the occurrence probability of a word appearing at an arbitrary position (order) in the sentence based on the words appearing immediately before, infer words that appear in In the example of FIG. 5, the server 1 performs estimation by calculating the occurrence probability of the second appearing word based on the first word "ko". The server 1 also calculates and estimates the occurrence probability of the third word based on the first and second words "ko" and "and". Similarly, the server 1 estimates each word.

The server 1 compares the estimated word/phrase with the actual word/phrase (correct value), adjusts the parameters of each neuron so that the output value from each neuron in the output layer approximates the correct value, and constructs an RNN. For example, the server 1 adjusts the weight and the like in each neuron so that the word/phrase estimated as the word/phrase following "ko" becomes the actual word/phrase "and". As a result, the server 1 generates a language model 141 that has learned the order of words in the correct order of the learning sentences.

The server 1 performs the above-described processing for sentences in each category, and generates a language model 141 related to RNN for each category of contract.

In addition, the server 1 extracts keywords that frequently appear in sentences of each category and stores them in the keyword DB 142 at the time of learning. For example, the server 1 calculates the frequency of appearance of each word/phrase in a sentence, and extracts words/phrases whose frequency of appearance is greater than or equal to a threshold value as keywords. The server 1 stores the words extracted as keywords in the keyword DB 142 in association with the category (type and clause) of the text that is the extraction source. As will be described later, when the user designates words and phrases to be included in the text of the contract to be generated, the server 1 presents the words and phrases extracted as keywords to the user and accepts designation input.

FIG. 6 is an explanatory diagram regarding the text generation process. FIG. 7 is a screen image diagram displayed by the terminal 2. As shown in FIG. FIG. 6 conceptually illustrates how a contract sentence is generated from a plurality of words specified by the user. FIG. 7 schematically shows the display screen of the terminal 2 during sentence generation. A text generation process using the language model 141 will be described with reference to FIGS. 6 and 7. FIG.

First, the server 1 receives, from the terminal 2, a designation input designating a contract category for which the user wishes to create a text and a plurality of words to be included in the text. For example, the server 1 receives various information designation inputs from the terminal 2 via the input screen illustrated in FIG.

The input screen includes, for example, a clause selection field 71, a keyword selection field 72, and a text input field 73. The clause selection column 71 is a selection column for displaying the clauses of the type of contract already displayed in the upper part of the input screen as a pull-down menu, and allowing the user to select the clause of the text that the user desires to create. As shown in FIG. 7, in the clause selection field 71, the terminal 2 displays various clauses included in the contract type "confidentiality agreement" displayed at the top of the input screen in a pull-down, and selects one of the clauses. Accept input. As a result, the terminal 2 accepts the specification input of the type and clause of the contract that the user desires to create, that is, the category.

The keyword selection column 72 is a selection column for displaying the keyword of the clause selected in the clause selection column 71 as a pull-down menu and allowing the user to select the keyword to be included in the sentence to be created. The server 1 reads the keyword of the clause selected in the clause selection column 71 from the keyword DB 142 and outputs it to the terminal 2, and the terminal 2 displays the keyword in the keyword selection column 72 by pull-down, and selects and inputs one of the keywords. accept.

The text input field 73 is an input field for the user to input any keyword (word/phrase) desired to be included in the sentence to be created. The terminal 2 not only accepts a selection input from the keywords presented in the keyword selection field 72 , but also accepts input of any keyword in the text input field 73 .

When the user inputs arbitrary text as a keyword in the text input field 73, the user can freely input text. may be difficult to generate. Therefore, the server 1 may convert the text input by the user using the replacement dictionary 143 and use the converted text as a keyword. In the example of FIG. 7, the server 1 converts the words "document" and "permission" included in the text "permission by document" input by the user into "document" and "approval", and converts the converted text to keywords used as This prevents a situation in which sentence generation by the language model 141 becomes difficult.

As described above, the terminal 2 accepts designation input of the category of the contract that the user desires to create and a plurality of words and phrases (keywords) to be included in the text of the clause. When an operation input to the generation button 74 is accepted, the terminal 2 outputs the input content to the server 1 and requests generation of a sentence including the specified phrase.

Returning to FIG. 6 , the server 1 uses the language model 141 of the category specified by the user to generate a sentence containing a plurality of specified words and phrases, and outputs the sentence to the terminal 2 . FIG. 6 illustrates a case where the user specifies the clause "intellectual property rights" of the type "confidentiality agreement" as a category, and specifies "invention" and "sharing" as words to be included in the text. . In this case, the server 1 generates a sentence using the language model 141 in which the contract type is "confidentiality agreement" and the clause is "intellectual property rights".

First, the server 1 generates a plurality of sentence candidates (hereinafter referred to as “candidate sentences”) to be finally output from any of the specified words and phrases. FIG. 6 illustrates how a plurality of candidate sentences are generated from the word "invention". The server 1 inputs the word/phrase into the input layer related to the RNN, calculates the occurrence probability of the next word/phrase based on the word/phrase, and obtains one or more words/phrases with the highest occurrence probability as the second word/phrase as an output. Next, the server 1 inputs the second word/phrase to the input layer, calculates the probability of occurrence of the next word/phrase based on the first word/phrase "invention" and the second word/phrase, and obtains the third word/phrase as One or more words with a high probability of occurrence are obtained as an output.

In this way, the server 1 calculates the occurrence probability of the next word/phrase from the previous word/phrase, and estimates the next word/phrase. The server 1 repeats the above process to generate a plurality of sentences from one phrase. In this case, the server 1 calculates a score for evaluating the validity of each generated sentence based on the occurrence probability of each word calculated when generating the sentence, and selects a sentence with a high score as a candidate sentence. . For example, the server 1 uses the average probability of occurrence of each word included in the sentence as the score of the entire sentence. The server 1 regards sentences with an average value of occurrence probability equal to or greater than a threshold as candidate sentences.

Although the server 1 determines the candidate sentences by using the average value of the occurrence probabilities as the score, the candidate sentences may be determined based on other criteria. For example, the server 1 may determine candidate sentences by excluding sentences containing words and phrases whose occurrence probability is equal to or less than a threshold value, that is, sentences containing inappropriate words and phrases.

Also, in the above description, candidate sentences are generated from a single word out of a plurality of words specified by the user, but the present embodiment is not limited to this. For example, the server 1 may generate a plurality of candidate sentences from two or more of the three or more words specified by the user, and select candidate sentences containing the remaining specified words. That is, the server 1 only needs to be able to generate candidate sentences using some of the multiple words specified by the user, and the part of the words is not limited to a single word.

The server 1 selects candidate sentences including other specified words and phrases not used for sentence generation from among the plurality of words and phrases specified by the user, as sentences to be output to the terminal 2 from the generated candidate sentences. In the example of FIG. 6, the server 1 selects candidate sentences surrounded by dotted-line rectangular frames that include the word "share" other than "invention" as output sentences.

As described above, the server 1 selects candidate sentences that include all of the specified words and phrases and that have a high overall score from the generated candidate sentences. A single candidate sentence may be selected, or a plurality of candidate sentences may be selected. The server 1 outputs the selected candidate sentences to the terminal 2 . The terminal 2 displays the candidate sentences output from the server 1 as shown in FIG. 7 and presents them to the user.

FIG. 8 is a flow chart showing an example of the procedure of sentence learning processing. Based on FIG. 8, the contents of the sentence learning process will be described.
The control unit 11 of the server 1 acquires a contract group for learning (step S11). The control unit 11 performs a process of structuring the obtained text of each contract according to categories such as the type of contract and clauses in the contract, and classifies the text in each contract by category ( step S12).

The control unit 11 divides each sentence classified into each category into a plurality of words (elements), which are characters or character strings of a predetermined unit (step S13). For example, the control unit 11 may divide the sentence into semantic units such as words and clauses, or may divide the sentence into units according to the appearance frequency of subwords and the like.

The control unit 11 performs machine learning processing for learning each word of the divided sentence according to the order of arrangement in the sentence, and generates the language model 141 for each category (step S14). For example, the control unit 11 performs machine learning based on the RNN (LSTM) algorithm, and from one or more words that appear in order from the beginning of the sentence, a language model that estimates words that appear following the one or more words. 141 is generated. The control unit 11 generates separate language models 141 for each category. The control unit 11 stores the generated language model 141 for each category in the auxiliary storage unit 14 .

Further, the control unit 11 extracts the keywords of the sentences of each category based on the appearance frequency in the sentences of each category and stores them in the keyword DB 142 (step S15). The control unit 11 ends the series of processes.

FIG. 9 is a flow chart showing an example of a processing procedure of sentence generation processing. Based on FIG. 9, the processing contents of the text generation processing will be described.
The control unit 11 of the server 1 receives, from the terminal 2, a designation input that designates the category of the text to be generated via the input screen illustrated in FIG. 7 (step S31). As described above, the text category is the contract type, clause, and the like. The control unit 11 also receives, from the terminal 2, a designation input for designating a plurality of words to be included in the sentence to be generated via the input screen illustrated in FIG. 7 (step S32). For example, the control unit 11 refers to the keyword DB 142, presents a plurality of keywords of sentences of the category specified in step S31, and receives selection input by the user. Further, the control unit 11 may accept text input of arbitrary phrases from the user. Further, when receiving a text input of an arbitrary word/phrase, the control unit 11 may use the replacement dictionary 143 to replace the input word/phrase.

Using the language model 141 of the category specified in step S31, the control unit 11 generates a plurality of candidate sentences from any of the words specified in step S32 (step S33). For example, the control unit 11 inputs one of the specified words into the language model 141, which is the RNN, and obtains the words following the specified words as output. Furthermore, the control unit 11 obtains the following phrase as an output based on the specified phrase and the outputted phrase. The control unit 11 repeats the processing to generate a plurality of sentences. The control unit 11 calculates a score (for example, an average value of occurrence probabilities) of the entire sentence of each pattern based on the occurrence probability of each word/phrase calculated when generating the sentence. The control unit 11 determines candidate sentences according to the calculated scores. For example, the control unit 11 determines sentences with scores equal to or higher than a threshold as candidate sentences.

The control unit 11 selects, from among the plurality of generated candidate sentences, candidate sentences containing designated words other than the words used when generating the candidate sentences in step S33 as sentences to be finally output (step S34). The control unit 11 outputs the selected text to the terminal 2 (step S35), and ends the series of processes.

In the above description, a sentence is generated using a plurality of words specified by the user, but a sentence may be generated only from a single specified word.

In addition, although the contract is given as an example of the text to be generated above, the text to be generated is not limited to the contract, and may relate to other documents.

In the above description, the server 1 generates the language model 141 by learning each word appearing in the text of the learning contract from the beginning to the end in the correct order of the text, but the present embodiment is limited to this. not to be For example, the server 1 replaces the beginning and end of a sentence as in Embodiment 5, which will be described later, and generates a language model 141 that has learned each word in reverse order from the end to the beginning of the sentence. You may make it generate|occur|produce a sentence from. Further, for example, the server 1 may construct a bidirectional RNN that has been learned in both the forward order and the reverse order as the language model 141, and use the language model 141 to generate sentences. In this way, the server 1 only needs to be able to generate a sentence from the language model 141 that has learned the word order in the sentence, regardless of whether the word order is forward or reverse. The language model 141 may be constructed by learning bi-directionally and in reverse order.

As described above, according to the first embodiment, by using the language model 141 that learns each phrase appearing in the sentence of the learning contract in the order of the sentence, the sentence including the phrase specified by the user can be appropriately reproduced. can be generated to

Moreover, according to the first embodiment, by generating a sentence from a plurality of words specified by the user, it is possible to generate a more appropriate sentence.

Further, according to the first embodiment, by calculating the score of a sentence generated from the language model 141 from the probability of occurrence of each word, the validity of the sentence is evaluated, and an appropriate sentence is presented to the user. can be done.

Further, according to the first embodiment, a language model 141 is prepared for each category of sentences, and a sentence is generated from the language model 141 corresponding to the category specified by the user, thereby presenting a more appropriate sentence to the user. can do.

Further, according to the first embodiment, when a sentence is generated from a word arbitrarily input by the user, the sentence is generated using the language model 141 by replacing the input word using the replacement dictionary. It is possible to prevent situations in which generation is difficult.

(Embodiment 2)
In this embodiment, a form is described in which phrases that frequently appear at the beginning of sentences are generated based on the language model 141, and sentences including words specified by the user are generated from the phrases. In addition, the same code|symbol is attached|subjected about the content which overlaps with Embodiment 1, and description is abbreviate|omitted.
FIG. 10 is an explanatory diagram showing an overview of the second embodiment. Based on FIG. 10, the sentence generation process which the server 1 performs in this Embodiment is demonstrated.

The server 1 according to the present embodiment first uses the language model 141 to generate a plurality of words and phrases that are presumed to appear at the beginning of the text of the contract, ie, frequently appearing first phrases. For example, the server 1 uses the language model 141 to generate the first phrase from the words and phrases present at the beginning of the text of the learning contract used when generating the language model 141, that is, the learning source text. In the example of FIG. 10, "ko" corresponds to the word at the beginning of the learning source sentence. While referring to the language model 141, the server 1 selects words and phrases having occurrence probabilities equal to or greater than a certain value as words and phrases following the first word and phrase of the learning source sentence. Furthermore, the server 1 selects words and phrases having occurrence probabilities equal to or greater than a certain value as words and phrases following the selected words and phrases. The server 1 repeats this process, sequentially selects words and phrases having occurrence probabilities equal to or greater than a certain value as words and phrases following the head word, and generates sentences.

As described above, the server 1 sequentially selects words and phrases having occurrence probabilities equal to or higher than a certain value from the beginning of the sentence, and generates a plurality of words and phrases as the first phrase. For example, the server 1 generates sentences until one of the phrases specified by the user (in FIG. 10, "invention" or "shared") appears, and when the specified phrase appears, the word preceding the specified phrase appears. Alternatively, a plurality of words may be specified as the first phrase. Alternatively, the server 1 may generate a predetermined number of words and phrases from the beginning of the sentence, and from the predetermined number of words and phrases, specify a phrase portion having a cohesive meaning, grammar, and the like. It does not matter how far the phrase from the beginning of the sentence is specified as the first phrase.

The server 1 generates a sentence including the phrase specified by the user from the generated first phrase. Specifically, the server 1 arranges one of the plural words specified by the user as the word following the first phrase. Then, the server 1 inputs the initial phrase and the designated word/phrase to the RNN (language model 141), estimates the word/phrase following the designated word/phrase, and generates a plurality of candidate sentences.

The subsequent processing is the same as in the first embodiment, and the server 1 selects candidate sentences containing other specified words and phrases not used for sentence generation as sentences to be finally output, and outputs them to the terminal 2 .

FIG. 11 is a flowchart illustrating an example of a processing procedure of sentence generation processing according to the second embodiment.
After accepting input specifying the category of the sentence to be generated (step S31) and accepting input specifying multiple words to be included in the sentence (step S32), the control unit 11 of the server 1 executes the following processes. Based on the language model 141 of the category specified in step S31, the control unit 11 generates a phrase (a plurality of words) appearing at the beginning of sentences in the category (step S201). Control unit 11 generates a plurality of candidate sentences based on the generated phrase and one of the plurality of words specified in step S31 (step S202). Specifically, the control unit 11 arranges one of the designated phrases next to the first phrase, inputs it to the RNN, estimates the subsequent phrases, and generates a sentence. The control unit 11 shifts the process to step S34.

As described above, according to the second embodiment, the first phrase is generated based on the language model 141, and the subsequent phrases are generated by arranging the specified phrase next to the phrase, thereby making the sentence more suitable. can be generated.

(Embodiment 3)
In this embodiment, a form in which a sentence is generated in which specified words appear in the order specified by the user will be described.
FIG. 12 is an explanatory diagram showing an overview of the third embodiment. The text generation process executed by the server 1 according to the present embodiment will be described with reference to FIG. 12 .

In the present embodiment, the server 1 accepts not only input specifying a plurality of words on an input screen similar to the screen illustrated in FIG. . Then, the server 1 generates sentences in which each word appears in the designated order.

For example, as shown in FIG. 12, consider a case where "invention" is specified as the first word, "attribution" as the second word, and "consultation" as the third word. In this case, the server 1 inputs "invention" specified as the first word/phrase to the RNN (language model 141), estimates subsequent words/phrases, and generates a plurality of candidate sentences.

The server 1 selects, from among the plurality of generated candidate sentences, sentences in which other specified words and phrases not used for sentence generation appear in the order specified by the user. In the example of FIG. 12, the specified words other than "invention" and "attribution" and "consultation" are included, and the second specified word "attribution" precedes the third specified word "consultation". Select the text that appears. The server 1 outputs the selected text to the terminal 2.

FIG. 13 is a flowchart illustrating an example of a processing procedure of sentence generation processing according to the third embodiment.
After receiving the designation input of the category of the text to be generated (step S31), the control section 11 of the server 1 executes the following processing. The control unit 11 receives a designation input that designates a plurality of words and phrases to be included in the sentence to be generated and the order in which the words and phrases appear in the sentence (step S301). The control unit 11 generates a plurality of candidate sentences from one of the specified words (step S302). Then, the control unit 11 selects a candidate sentence in which each specified word/phrase appears in the order specified in step S301 from a plurality of candidate sentences as a sentence to be finally output (step S303). The control unit 11 shifts the process to step S35.

As described above, according to the third embodiment, a more accurate sentence can be generated by generating a sentence in which specified words appear in a specified order.

(Embodiment 4)
In this embodiment, when a candidate sentence containing all of a plurality of words specified by the user is not generated, the words included in the candidate sentence are replaced with the specified words to generate a sentence including all the specified words.
FIG. 14 is an explanatory diagram showing an overview of the fourth embodiment. The text generation process executed by the server 1 according to this embodiment will be described with reference to FIG. 14 .

Server 1 according to the present embodiment generates a plurality of candidate sentences based on one of the plurality of phrases specified by the user, as in the first embodiment. In this case, depending on the phrase specified by the user, sentences containing other specified phrases not used for candidate sentence generation may not be generated as candidate sentences. For example, as schematically shown in the upper part of FIG. ” may not be included.

In this case, the server 1 selects sentences to be output to the terminal 2 according to the score of the sentences while replacing the words included in the candidate sentences with the specified words. For example, the server 1 first selects the candidate sentence with the highest score from among the plurality of candidate sentences generated as described above. Then, the server 1 replaces each word/phrase included in the selected candidate sentence with the remaining designated word/phrase to generate a plurality of new candidate sentences. In the example of FIG. 14, regarding the sentences surrounded by the dotted rectangular frame with the highest score, the second, third, fourth, and so on following "invention" are replaced with the remaining designated words "shared". generate candidate sentences.

In the above description, all words other than the specified word "invention" are replaced with the specified word "shared" in round robin, but the present embodiment is not limited to this. For example, the server 1 may narrow down words having the same part of speech as the rest of the specified words "shared" out of the specified words "invention" and replace them with the specified words "shared" to generate candidate sentences. That is, the server 1 generates a candidate sentence by converting a word having the same part of speech as another specified word/phrase (element) not used for candidate sentence generation into the other specified word/phrase. By converting between the same parts of speech, more appropriate candidate sentences can be created and the processing load associated with sentence generation can be reduced.

The server 1 inputs the generated word/phrase of each candidate sentence to the RNN (language model 141), calculates the probability of occurrence, and calculates the score of each candidate sentence. Then, the server 1 selects a candidate sentence with a high score. For example, the server 1 selects candidate sentences whose scores are equal to or greater than a threshold. In this way, the server 1 generates candidate sentences that include all of the specified words and phrases and have scores equal to or higher than the threshold. The server 1 outputs the generated candidate sentences to the terminal 2 and presents them to the user.

FIG. 15 is a flowchart illustrating an example of a processing procedure of sentence generation processing according to the fourth embodiment.
After generating a plurality of candidate sentences using any one of the plurality of phrases specified in step S32 (step S33), the control section 11 of the server 1 executes the following processing. The control unit 11 determines whether or not there is a candidate sentence including specified words other than the words used when generating the sentence in step S33 among the plurality of words specified in step S32 (step S401). ). If it is determined that there is a candidate sentence containing the specified word (S401: YES), the control section 11 shifts the process to step S34.

When it is determined that there is no candidate sentence containing the designated word (S401: NO), the control unit 11 generates a new candidate sentence while replacing each word included in the candidate sentence generated in step S33 with the designated word. (Step S402). For example, the control unit 11 selects a candidate sentence with the highest score from among the plurality of candidate sentences generated in step S33, replaces each word included in the candidate sentence with a specified word not used for generating a sentence, and selects a plurality of candidate sentences. New candidate sentences for are generated. In this case, the control unit 11 generates a new candidate sentence by, for example, replacing words and phrases of the same part of speech as specified words and phrases (other elements) not used for sentence generation among the words and phrases included in the candidate sentences. Generate.

The control unit 11 uses the language model 141 to calculate the score of each of the newly generated candidate sentences, and selects the candidate sentence with the highest score as the sentence to be finally output (step S403). For example, the control unit 11 selects candidate sentences whose scores are equal to or greater than a threshold. The control unit 11 shifts the process to step S35.

As described above, according to the fourth embodiment, even when a sentence including all specified words is not generated, a sentence including the specified words can be appropriately generated by replacing the words in the sentence.

Further, according to the fourth embodiment, by performing the above replacement between the same part of speech, it is possible to more appropriately generate a sentence including the designated word/phrase.

(Embodiment 5)
In this embodiment, not only the forward order language model 141 (first language model) but also the reverse order language model 141 (second language model) are generated from the learning contract, and both language models 141 are generated. A form of generating a sentence by combining will be described.
FIG. 16 is an explanatory diagram of sentence learning processing according to the fifth embodiment. FIG. 16 conceptually illustrates how words appearing in a learning contract are learned in reverse order from the end of the sentence.

As in the first embodiment, the server 1 divides the text of the learning contract into predetermined units of words and phrases, learns the words and phrases from the beginning to the end of the text in positive order, and generates the language model 141. . Further, in this embodiment, the server 1 rearranges the words in the sentence in reverse order, learns the words in reverse order from the end of the sentence, and generates the language model 141 in reverse order.

For example, after completing the learning process in the forward order, the server 1 rearranges the learning sentences from the beginning to the end, and rearranges them in reverse order with the last word at the beginning. That is, as shown in FIG. 16, the server 1 converts the phrases "K", "and", "B", "ga", ... "notice", "to do" to "do", "notify", " Arrange them in the order of "to", "counterparty", ... "and", and "Party A". The server 1 inputs each rearranged word/phrase to the input layer of the RNN for reverse order learning having the same configuration as the RNN shown in FIG. Then, the server 1 constructs an RNN for estimating a word that appears immediately before one or more words, based on one or more words that appear in reverse order from the end of the sentence. In other words, the server 1 constructs an RNN that estimates a word/phrase at an arbitrary position in a sentence from words/phrases that follow the word/phrase.

FIG. 17 is an explanatory diagram of sentence generation processing according to the fifth embodiment. When generating sentences from words specified by the user, the server 1 combines the occurrence probabilities of each word/phrase calculated from both the forward-order and reverse-order language models 141 to generate sentences.

For example, the server 1 first generates candidate sentences using the forward order language model 141 . Specifically, as in the first embodiment, the server 1 generates a plurality of candidate sentences from any of a plurality of specified words. The server 1 then selects candidate sentences containing the rest of the specified words. Here, consider a case where the candidate sentences illustrated in FIG. 17 are selected.

The server 1 reverses the arrangement order of the words in the candidate sentence, inputs it to the language model 141, and calculates the probability of occurrence of each word. Then, the server 1 determines whether or not the candidate sentence is an appropriate sentence based on the probability of occurrence calculated from the reverse order language model 141 . For example, the server 1 compares the probability of occurrence of each word/phrase with a predetermined threshold value, and determines whether or not there is a word/phrase that is equal to or less than the threshold value. As a result, the server 1 checks whether there is an inappropriate word or phrase from the reverse order language model 141 .

When determining that there is a word whose occurrence probability is equal to or lower than the threshold, the server 1 replaces the relevant word with another word using the reverse order language model 141, for example. That is, the server 1 proofreads the relevant words. For example, the server 1 determines the position of the relevant phrase from one or more phrases that appear immediately after the relevant phrase (in FIG. 17, the phrases “opponent”, “to”, “no”, etc. , and replace the words. As a result, the server 1 corrects the candidate sentences into more appropriate sentences. The server 1 outputs the proofread sentence to the terminal 2 .

18 is a flowchart illustrating an example of a processing procedure of sentence learning processing according to Embodiment 5. FIG.
After dividing the text of the contract into a plurality of words (step S13), the control section 11 of the server 1 executes the following processes. The control unit 11 performs machine learning processing for learning each word of the divided sentence according to the order in which the sentence is arranged, and generates the forward order language model 141 (step S501). Next, the control unit 11 rearranges the sentences from the beginning to the end, and reverses the arrangement order of each word (step S502). Then, the control unit 11 generates the language model 141 by learning each word in reverse order (step S503). The control unit 11 stores the forward-order language model 141 generated in step S501 and the reverse-order language model 141 generated in step S503 in the auxiliary storage unit 14, and shifts the process to step S15.

19 is a flowchart illustrating an example of a processing procedure of sentence generation processing according to Embodiment 5. FIG.
After receiving the designation input of a plurality of words to be included in the sentence to be generated (step S32), the control section 11 of the server 1 executes the following processing. Based on the normal order language model 141, the control unit 11 generates a plurality of candidate sentences from any of the plurality of words specified in step S32 (step S521). The control unit 11 selects a candidate sentence including the specified word/phrase not used in step S521 from the generated candidate sentences (step S522).

The control unit 11 reverses the arrangement order of the words included in the candidate sentence selected in step S522 (step S523). Then, the control unit 11 calculates the occurrence probability of each word included in the candidate sentence based on the reverse order language model 141 (step S524). The control unit 11 determines the final sentence to be output according to the occurrence probability calculated from the reverse order language model 141 (step S525). For example, the control unit 11 uses the reverse order language model 141 to compare the probability of occurrence of each word included in the candidate sentence with a predetermined threshold, and determines whether there is a word below the threshold. When determining that there is no word or phrase below the threshold, the control unit 11 determines the candidate sentence as the output sentence. Further, for example, when the control unit 11 determines that there is a word or phrase below the threshold, it performs a process of replacing (proofreading) the corresponding word or phrase with another word or phrase using the reverse order language model 141, and generates a candidate sentence after replacement. The text to be finally output may be determined. The control unit 11 shifts the process to step S35.

Alternatively, candidate sentences may be generated from the language model 141 in reverse order and checked by the language model 141 in normal order.

As described above, according to the fifth embodiment, a more appropriate sentence can be presented to the user by generating a sentence by combining the forward-order and reverse-order language models 141 .

Further, according to the fifth embodiment, candidate sentences are generated from either the forward-order or reverse-order language model 141, and the other language model 141 is used to check the candidate sentences, thereby making the sentences more appropriate. can be generated.

(Embodiment 6)
FIG. 20 is a functional block diagram showing the operation of the server 1 of the form described above. When the control unit 11 executes the program P, the server 1 operates as follows.
The storage unit 201 stores a language model obtained by learning, in order of the sentence, the elements that appear next to each element in a sentence composed of each element that is a character or character string of a predetermined unit. The receiving unit 202 receives an input specifying the elements to be included in the text to be generated. The generating unit 203 refers to the language model and generates sentences based on the designated elements. The output unit 204 outputs the generated text.

The sixth embodiment is as described above, and the other parts are the same as those of the first to fifth embodiments.

The embodiments disclosed this time are illustrative in all respects and should not be considered restrictive. The scope of the present invention is indicated by the scope of the claims rather than the above-described meaning, and is intended to include all modifications within the scope and meaning equivalent to the scope of the claims.

1 server (information processing device)
11 Control Unit 12 Main Storage Unit 13 Communication Unit 14 Auxiliary Storage Unit P Program 141 Language Model 142 Keyword DB
143 replacement dictionary 2 terminal

Claims (12)

a storage unit that stores a language model obtained by learning, in the order of the sentence, the elements appearing next to the elements in a sentence composed of elements that are characters or character strings of a predetermined unit;
a receiving unit that receives a specified input of the element to be included in the generated text;
Input any of the specified elements into the language model to obtain an element appearing next to the element, input the specified element and the obtained element into the language model, and then appear next. a generation unit that repeats the process of acquiring elements to generate sentences that include the specified elements and the acquired elements ;
and an output unit that outputs the generated text ,
The generating unit
An element existing at the beginning of a learning source sentence used for learning the language model is input to the language model, an element appearing next to the element is acquired, and the element appearing at the beginning and the element appearing next to the acquired element are obtained. or input an element existing at the beginning of the learning source sentence into the language model to obtain an element that appears next to the element, and convert the first element and the obtained element to the Repeat the process of inputting to the language model and acquiring the next appearing element to generate a leading phrase including the leading element and each acquired element,
Generate a sentence containing the generated first phrase and the specified element
Information processing equipment. The language model calculates the probability of occurrence of the element appearing next based on the input leading element , and outputs the element for which the occurrence probability of a certain value or more is calculated,
2. The information processing apparatus according to claim 1 , wherein the head phrase including the element output by the language model is generated. a storage unit that stores a language model obtained by learning, in the order of the sentence, the elements appearing next to the elements in a sentence composed of elements that are characters or character strings of a predetermined unit;
a receiving unit that receives a specified input of the element to be included in the generated text;
Input any of the specified elements into the language model to obtain an element appearing next to the element, input the specified element and the obtained element into the language model, and then appear next. a generation unit that repeats the process of acquiring elements to generate sentences that include the specified elements and the acquired elements;
an output unit that outputs the generated text;
with
The reception unit receives designation input of a plurality of the elements,
The generation unit inputs any one of the specified plurality of elements to the language model to obtain a plurality of elements that appear next, and for each of the plurality of elements that appear next, Repeating the process of inputting the specified element input to the language model and the next appearing element into the language model and obtaining the next appearing element, and generating a plurality of said sentences containing an element to
A selection unit that selects the sentences by determining whether the sentences contain other elements specified in addition to the elements input to the language model from the plurality of sentences,
The information processing apparatus, wherein the output unit outputs the selected text. The reception unit receives an input specifying the order in which the plurality of elements appear in the sentence,
The information processing apparatus according to claim 3 , wherein the selection unit selects the sentence in which the plurality of elements appear in a designated order. If the sentence including the other element is not generated, the generating unit sequentially replaces each element included in the generated sentence with the other element to generate a plurality of sentences ;
Using the language model, the probability of occurrence of each element contained in each of the plurality of generated sentences is calculated, and based on the probability of occurrence of each element, a score indicating the validity of each of the plurality of sentences is calculated. a calculator,
5. The information processing apparatus according to claim 3 , wherein the selection unit selects sentences whose score is equal to or greater than a predetermined value . 6. The information processing apparatus according to claim 5 , wherein among the elements included in the generated sentence, the generation unit replaces an element having the same part of speech as the other element with the other element. The generation unit generates the sentence by calculating the probability of occurrence of each element following the specified element based on the language model,
A calculation unit that calculates a score indicating the validity of the generated sentence based on the probability of occurrence of each element,
The information processing apparatus according to any one of claims 1 to 6 , wherein the output unit outputs the text whose calculated score is equal to or greater than a predetermined value . The storage unit stores a first language model that has learned the elements in forward order from the beginning to the end of the sentence, and a second language model that has learned the elements by rearranging the elements from the beginning to the end of the sentence in reverse order. store the language model and
The generating unit
generating the sentence using the first or second language model ;
Converting the arrangement order of the elements included in the generated sentence to reverse order,
Using the second or first language model, calculate the occurrence probability of each element contained in the converted sentence,
The information processing apparatus according to any one of claims 1 to 7 , wherein the output unit outputs the text when the calculated probability of occurrence of each element is equal to or greater than a predetermined value . The storage unit stores a plurality of the language models that differ according to categories of sentences,
The receiving unit receives a specified input of the category of the sentence to be generated,
The information processing apparatus according to any one of claims 1 to 8 , wherein the generation unit generates the sentence based on the language model corresponding to the designated category. the storage unit stores a replacement dictionary for replacing the element;
The receiving unit receives a specified input of the element by text input,
a replacement unit that refers to the replacement dictionary and replaces the input element;
The information processing apparatus according to any one of claims 1 to 9 , wherein the generation unit generates the sentence based on the replaced element. Receiving a specified input of an element that is a character or a character string of a predetermined unit to be included in the generated sentence,
Input any of the specified elements to a language model that has learned the elements that appear next to each element in the sentence in the order of the sentence, acquire the element that appears next to the element, and obtain the element that appears next to the specified element. Repeating the process of inputting the element and the obtained element into the language model and obtaining the element that appears next, generating a sentence containing the specified element and each obtained element ,
A process of outputting the generated text,
An element existing at the beginning of a learning source sentence used for learning the language model is input to the language model, an element appearing next to the element is acquired, and the element appearing at the beginning and the element appearing next to the acquired element are obtained. or input an element existing at the beginning of the learning source sentence into the language model to obtain an element that appears next to the element, and convert the first element and the obtained element to the Repeat the process of inputting to the language model and acquiring the next appearing element to generate a leading phrase including the leading element and each acquired element,
Generate a sentence containing the generated first phrase and the specified element
An information processing method characterized in that a computer executes processing. Receiving a specified input of an element that is a character or a character string of a predetermined unit to be included in the generated sentence,
Input any of the specified elements to a language model that has learned the elements that appear next to each element in the sentence in the order of the sentence, acquire the element that appears next to the element, and obtain the element that appears next to the specified element. Repeating the process of inputting the element and the obtained element into the language model and obtaining the element that appears next, generating a sentence containing the specified element and each obtained element ,
A process of outputting the generated text,
An element existing at the beginning of a learning source sentence used for learning the language model is input to the language model, an element appearing next to the element is acquired, and the element appearing at the beginning and the element appearing next to the acquired element are obtained. or input an element existing at the beginning of the learning source sentence into the language model to obtain an element that appears next to the element, and convert the first element and the obtained element to the Repeat the process of inputting to the language model and acquiring the next appearing element to generate a leading phrase including the leading element and each acquired element,
Generate a sentence containing the generated first phrase and the specified element
A program characterized by causing a computer to execute processing .

Images