JP6712973B2 - Sentence generation device, sentence generation learning device, sentence generation method, and program - Google Patents

Description

The present invention relates to a sentence generation device, a sentence generation learning device, a sentence generation method, and a program, and more particularly to a sentence generation device, a sentence generation learning device, a sentence generation method, and a program for automatically generating a natural language sentence. ..

In order for a computer to translate or summarize a document written in natural language, it is necessary not only to analyze a natural language sentence but also to have a technique of generating a sentence by a computer. In addition, it is necessary to automatically generate a natural language sentence when a machine and a person have a dialogue.

The computer generates sentences by inputting the first word, calculating the probability of the word that appears next, generating the word, and then using the output word as the input, calculating the probability of the word that appears next It is performed by repeating the process of.

In recent years, a method of calculating a word appearance probability based on all past word strings using a recurrent neural network (RNN) has become mainstream (Non-Patent Document 1).

In this method, a word is expressed by a fixed-length vector (called a fixed-length vector) in which each dimension element is a real value, and a combination between words is expressed by calculation of vectors. As a result, words having similar meanings become similar vectors, and there is an advantage that sparsity in combination calculation is reduced.

Specifically, the RNN first calculates a fixed length vector with the first word as an input, and outputs the next word. Next, a sentence is generated by repeating the process of calculating a fixed length vector using the output word as an input (FIG. 8).

Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals. Recurrent Neural Network Regularization.Proceedings of the 2nd International Conference on Learning Representations (ICLR 2014), 2014.

Since the sentence generation system using RNN can consider all the words output in the past, it can handle long-term dependency. On the other hand, since similar probabilities are given to words having similar meanings, there is a problem that a sentence including a grammatical error or a sentence having low readability is generated (FIG. 9).

The present invention has been made in view of this problem, and an object of the present invention is to provide a sentence generation device, a sentence generation method, and a program capable of generating a highly readable sentence with few grammatical errors. To do.

It is another object of the present invention to provide a sentence generation learning device and a program capable of learning a neural network for generating a highly readable sentence with few grammatical errors.

In order to achieve the above object, a sentence generation device according to the present invention includes an input conversion unit that converts an input word into a first fixed-length vector, and the first fixed-length vector obtained by the input conversion unit. And a second fixed-length vector generated in the past, using a pre-learned neural network for predicting the next word, a generation unit that generates a second fixed-length vector, and the input conversion. The second fixed length vector from the first fixed length vector obtained by the section using a pre-learned neural network for correcting the second fixed length vector generated by the generation section. Calculating a vector representing the appearance probability of each word from a correction unit that generates a correction vector for performing the correction, a second fixed-length vector generated by the generation unit, and the correction vector obtained by the correction unit. A word selected according to a vector representing the appearance probability of each word, and an output unit that outputs the next word, and the word output by the output unit is input to the input conversion unit, and the input conversion is performed. Unit, the generation unit, the correction unit, and the output unit to generate a word string in which the generated words are arranged by repeating each process as a candidate sentence, the generation unit, in the input conversion unit Corresponding to the obtained first fixed-length vector and the second fixed-length vector generated in the past, up to the word immediately preceding the word corresponding to the first fixed-length vector in the word string The second fixed length vector corresponding to the word string up to the word corresponding to the first fixed length vector in the word string is generated from the second fixed length vector.

Further, in the sentence generation method according to the present invention, the input conversion unit converts the input word into a first fixed-length vector, and the generation unit converts the input word into the first fixed-length vector obtained by the input conversion unit. , A second fixed-length vector generated in the past, using a pre-learned neural network for predicting the next word, generates a second fixed-length vector, and the correction unit performs the input conversion. The second fixed length vector from the first fixed length vector obtained by the section using a pre-learned neural network for correcting the second fixed length vector generated by the generation section. Generating a correction vector, and the output unit generates a vector representing the appearance probability of each word from the second fixed-length vector generated by the generation unit and the correction vector obtained by the correction unit. Calculating and outputting a word selected according to a vector representing the appearance probability of each word as the next word,
A word string in which words output by the output unit are input to the input conversion unit and words generated by repeating the processes of the input conversion unit, the generation unit, the correction unit, and the output unit are arranged. The candidate sentence is generated as a candidate sentence, and the generation unit includes the first fixed-length vector obtained by the input conversion unit and the second fixed-length vector generated in the past, the first fixed-length vector in the word string. From a second fixed-length vector corresponding to a word string up to the word immediately preceding the word corresponding to one fixed-length vector to a word string up to the word corresponding to the first fixed-length vector in the word string Generate a corresponding second fixed length vector.

According to the sentence generation device and the sentence generation method of the present invention, the input conversion unit converts the input word into a first fixed-length vector, and the generation unit acquires the first fixed length vector obtained by the input conversion unit. The second fixed-length vector is generated from the fixed-length vector and the second fixed-length vector generated in the past by using a pre-learned neural network for predicting the next word. The second fixed length vector is corrected using a pre-learned neural network for correcting the second fixed length vector generated by the generation unit from the first fixed length vector obtained by the input conversion unit. A correction vector for correcting the long vector is generated, and the output unit uses the second fixed-length vector generated by the generation unit and the correction vector obtained by the correction unit, and the appearance probability of each word. Is calculated, and the word selected according to the vector representing the appearance probability of each word is output as the next word.

Then, the words output by the output unit are input to the input conversion unit, and words generated by arranging the words generated by repeating the processes in the input conversion unit, the generation unit, the correction unit, and the output unit are arranged. In the word string, the sequence is generated as a candidate sentence, and the generation unit is the first fixed-length vector obtained by the input conversion unit and the second fixed-length vector generated in the past. Words from the second fixed-length vector corresponding to the word string up to the word immediately preceding the word corresponding to the first fixed-length vector to the word corresponding to the first fixed-length vector in the word string Generate a second fixed length vector corresponding to the column.

In this way, a correction vector for correcting the second fixed-length vector is generated, a word is output using the second fixed-length vector and the correction vector, and the output word is arranged. By generating a sequence as a candidate sentence, a sentence with few grammatical errors and high readability can be generated.

The sentence generation/learning apparatus according to the present invention includes a learning data conversion unit that converts a word of a correct sentence prepared in advance as learning data into a first fixed-length vector, and the first learning data conversion unit that obtains the first fixed-length vector. And a learning unit for generating a second fixed-length vector from the fixed-length vector of No. 1 and a second fixed-length vector generated in the past by using a neural network for predicting the next word. A neural network for correcting the second fixed-length vector generated by the generating unit from the first fixed-length vector obtained by the converting unit is used to correct the second fixed-length vector. A vector representing the appearance probability of each word is calculated from a correction unit that generates a correction vector, a second fixed-length vector generated by the generation unit, and the correction vector obtained by the correction unit. A word selected according to a vector representing the appearance probability of a word, an output unit that outputs the next word, and the word obtained by the output unit is a word next to the word of the correct sentence, A learning unit that learns a neural network for correcting the second fixed-length vector, and the learning data conversion unit, the generation unit, and the word output by the output unit as an input of the learning data conversion unit. The correction unit, the output unit, and the learning unit are configured to be repeated.

According to the sentence generation/learning apparatus of the present invention, the learning data conversion unit converts the word of the correct sentence prepared in advance as learning data into the first fixed-length vector, and the generation unit uses the learning data conversion unit. From the obtained first fixed-length vector and the second fixed-length vector generated in the past, a second fixed-length vector is generated and corrected using a neural network for predicting the next word. A second fixed length using a neural network for correcting the second fixed length vector generated by the generation unit from the first fixed length vector obtained by the learning data conversion unit. A correction vector for correcting the vector is generated.

Then, the output unit calculates a vector representing the appearance probability of each word from the second fixed-length vector generated by the generation unit and the correction vector obtained by the correction unit, and the appearance probability of each word is calculated. Is output as a next word, and the second fixed-length vector is set so that the word obtained by the output unit is the word next to the word of the correct sentence. A learning unit that learns a neural network for correction, and the learning data conversion unit, the generation unit, the correction unit, and the output unit using the word output by the output unit as an input of the learning data conversion unit. , And each processing in the learning unit is repeated.

As described above, a correction vector for correcting the second fixed-length vector is generated by using the neural network for correcting the second fixed-length vector, and the second fixed-length vector and the correction vector are generated. From, the vector representing the appearance probability of each word is calculated, the word selected according to the vector representing the appearance probability of each word is output as the next word, and the output word is next to the word of the correct sentence. Learning a neural network for correcting the second fixed-length vector so that the word becomes a word, and learning a sentence generation device for generating a highly readable sentence with few grammatical errors. You can

Further, in the sentence generation/learning apparatus according to the present invention, the learning unit is a neural network for predicting the next word so that the word obtained by the output unit is a word next to the word of the correct sentence. A network and a neural network for correcting the second fixed length vector may be learned.

A program according to the present invention is a program for causing a computer to function as each unit of the above sentence generation device or sentence generation learning device.

According to the sentence generation device, the sentence generation method, and the program of the present invention, it is possible to generate a sentence with few grammatical errors and high readability.

Further, according to the sentence generation learning device and the program of the present invention, it is possible to learn a neural network for generating a highly readable sentence with few grammatical errors.

It is a figure which shows the principle of the sentence generation apparatus which concerns on embodiment of this invention. It is a schematic diagram showing an example of composition of a sentence generation device concerning an embodiment of the invention. It is a figure showing the example of the vector showing the appearance probability of each word concerning the embodiment of the invention. It is a figure showing the example of the produced|generated output sentence which concerns on embodiment of this invention. 5 is a flowchart showing a sentence generation processing routine of the sentence generation device according to the exemplary embodiment of the present invention. It is a schematic diagram showing an example of composition of a sentence generation study device concerning an embodiment of the invention. It is a flowchart which shows the learning process routine of the sentence generation learning apparatus which concerns on embodiment of this invention. It is a figure showing the example of the conventional sentence generation method. FIG. 6 is a diagram illustrating an example in which a sentence including a grammatical error is generated.

Embodiments of the present invention will be described below with reference to the drawings.

<Outline of sentence generation device according to embodiment of the present invention>
In the above problem, all past words are represented by one vector, and it is not possible to explicitly handle where each word appears in the past, and the long-term dependency and the dependency between the most recent word Due to the inability to distinguish between.

In the present embodiment, in order to solve the above problem, the accuracy of the grammatical connection with the latest word is increased and the readability of the entire sentence is increased.

In the present embodiment, as in the past (Non-Patent Document 1), all past words are expressed by one vector using RNN. In addition to this, a correction unit for correcting the probability output by the RNN based on the input word is newly constructed (FIG. 1).

This correction unit is used to reduce the appearance probability of a word whose relation to the input word is grammatically incorrect. That is, erroneous words are removed from the output candidates and grammatically correct words are output.

Words are repeatedly generated using this system, and finally one sentence is output. However, it is assumed that whether or not there is a grammatical error (grammatical correctness) is determined based on learning data used when learning the correction unit.

In this respect, the method of reducing the distance between the input and the output on the network is close to Reference 1 below, in which the output of the lower layer is input to the upper layer separated by one or more.

[Reference 1] Deep Residual Learning for Image Recognition. Proceedings of the 29th IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2016), 2016.

However, in Reference Document 1, the output probability of a word is calculated by adding the output of the lower layer and the output of the upper layer, that is, all the words are expressed by one vector, and therefore, in which point each word is in the past We cannot solve the problem that we can't explicitly handle whether or not it appears in.

On the other hand, in the present embodiment, it is possible to explicitly consider the immediately preceding word by performing a two-step calculation in which the correction unit corrects the result calculated using the RNN.

According to this embodiment, it is possible to generate a word by explicitly considering the immediately preceding word. As a result, a word having a grammatically correct relationship with the immediately preceding word is generated, and a sentence with few grammatical errors and high readability can be output as the final output.

The present embodiment can be applied to all situations in which sentence generation by a computer is required, such as translation, summarization, and dialogue system. For example, in a dialogue system, the present invention can be used to generate more natural utterances with fewer errors when the system generates utterances.

In the following, in the interactive system, a task of generating an utterance (system utterance sentence Y) for which the system should return a response to a certain user utterance (user utterance sentence X) will be described as an example. First, an embodiment of a sentence generation device that generates a utterance sentence of a system utterance using a learned neural network will be described, and then a sentence generation learning device that learns a neural network will be described.

<Configuration of sentence generation device according to embodiment of the present invention>
The configuration of the sentence generation device according to the embodiment of the present invention will be described. FIG. 2 is a schematic diagram showing the configuration of the sentence generation device according to the embodiment of the present invention.

The sentence generation device 10 according to the embodiment of the present invention is a sentence generation device in a dialogue system. The sentence generation device 10 is configured by a computer including a CPU, a RAM, and a ROM that stores a program for executing a sentence generation processing routine described later, and is functionally configured as shown below. There is.

As shown in FIG. 2, the sentence generation device 10 includes an input unit 100, an input conversion unit 110, a generation unit 120, an RNN parameter storage unit 130, a correction unit 140, a correction parameter storage unit 150, and an output unit. And 160.

The sentence generation device 10 generates a sentence following the word from the input second fixed-length vector and the first word of the sentence.

The sentence generation device 10 generates one sentence with the highest probability by repeating the processes of the input conversion unit 110, the generation unit 120, the correction unit 140, and the output unit 160.

The input unit 100 is realized by an input device such as a known keyboard or storage device. The input unit 100 may also be a network interface that inputs observation data via a network such as the Internet.

The RNN parameter storage unit 130 and the correction parameter storage unit 150 each store a parameter corresponding to the structure of the neural network and the value of each parameter. The value of each parameter is updated by learning.

The output unit 160 is implemented by a display, a printer, a magnetic disk, or the like.

The input unit 100 receives, as an input, the first word of the sentence generated as the system utterance sentence Y, which is a response to the user utterance sentence X.

For example, as the first word of the generated sentence, “Yes” or “No” that is the beginning of the sentence in the reply utterance is received as an input.

The first word of the sentence to be input is determined by a rule base or a classifier. That is, when generating a subsequent sentence based on the input, the sentence generation device 10 according to the present embodiment is used. The first word of the sentence may be a "start" symbol indicating the start of the sentence.

The input conversion unit 110 converts a word into a first fixed-length vector representing the word.

Specifically, the input conversion unit 110 takes the first word of the sentence input by the input unit 100 or the word passed from the output unit 160 as an input, and sets the input word as a fixed-length vector.

(First fixed length vector).
More specifically, a fixed length vector is assigned to each word in advance, and the fixed length vector of the word corresponding to the input word is obtained. The number of dimensions of this fixed length vector is an arbitrary value N.

The input conversion unit 110 then converts the fixed-length vector

To the generation unit 120 and the correction unit 140.

The generation unit 120 predicts the next word from the first fixed-length vector obtained by converting the word by the input conversion unit 110 and the sequence of words up to the word or the second fixed-length vector representing the user's utterance. A second fixed length vector representing a series of words up to the word is generated using a pre-learned neural network for.

Specifically, first, the generation unit 120 generates a fixed-length vector representing a series of words up to the word.

(Second fixed length vector) is generated.

Here, when you enter the first word

, That is,

Is an arbitrary fixed length vector, and in the present embodiment, is a fixed length vector corresponding to the user utterance sentence X, but may be a zero vector or the like.

Further, the generation unit 120 is held in advance by the sentence generation device 10 when the first word is input.

Is input.

It is assumed that the fixed length vector corresponding to the user utterance sentence X is generated in advance by using the words forming the user utterance sentence X by the same procedure as that for generating the second fixed length vector of the system utterance sentence Y. .. Specifically, the fixed length vector corresponding to the user utterance sentence X is generated by repeating the same processing as the input conversion unit 110 and the generation unit 120. However, at that time, the input conversion unit 110 uses the words forming the user utterance sentence X, instead of the words output from the output unit 160.

Next, the generation unit 120 uses a pre-learned RNN parameter for predicting the next word from the RNN parameter storage unit 130.

,

,as well as

And a first fixed-length vector representing the word converted by the input conversion unit 110

, And the second fixed-length vector

Is input and the RNN is calculated. A second fixed-length vector representing the entire word string (sequence) from the first word to the current word (t-th word) by RNN

To get The second fixed-length vector is the first fixed-length vector obtained by the input conversion unit 110.

The same N dimension as

Also,

Is an N×N dimensional matrix,

Is an N×N dimensional matrix, and

Is an N-dimensional vector.

In this embodiment, there is no restriction on the RNN, and any RNN such as a long-term short-term memory (LSTM) can be used. Here, as an example, the calculation by the Ellman type RNN will be described. For a word string w[1:t] having a length t, a first fixed-length vector representing a word at the terminal point t (input word to RNN) is generated.

, The second fixed-length vector representing the word string w[1:t]

Is obtained by recursively calculating the following equation (1).

Here, tanh is a function that applies a hyperbolic tangent function to each dimension of a fixed length vector.

Then, the generation unit 120 generates the generated second fixed-length vector

To the output unit 160.

The RNN parameter storage unit 130 stores pre-learned RNN parameters for predicting the next word.
Specifically, the RNN parameter storage unit 130 is a parameter of the neural network that is learned in advance by the sentence generation/learning apparatus 20 described later.

,

,as well as

I remember.

Further, the structure of the RNN, which is a neural network for predicting the next word, and the initial value of each parameter can be set to arbitrary values before learning. However, at that time, an input layer (corresponding to a portion that receives an input to the input conversion unit 110) and an output layer (in the output unit 160,

Is the number of dimensions (M dimensions) corresponding to the number of vocabulary words (words used for generation) appearing in the learning data.

The correction unit 140 corrects the second fixed-length vector from the first fixed-length vector obtained by the input conversion unit 110 using a pre-learned neural network for correcting the second fixed-length vector. To generate a correction vector.

Specifically, the correction unit 140 first reads the first fixed length vector from the correction parameter storage unit 150.

Parameters of pre-learned neural network for correcting

,
as well as

To get

The structure of the neural network for correcting the second fixed length vector is determined according to the structure of RNN.

Also,

Is an M×N dimensional matrix,

Is an M-dimensional vector.

Next, the correction unit 140 uses the first fixed-length vector obtained by the input conversion unit 110.

Is input and the correction vector for correcting the second fixed-length vector that is the output of the RNN is calculated by the following equation (2).

here,

Is a function that applies a sigmoid function to each dimension of a fixed length vector. That is, the correction vector

Is a vector in which the value of each dimension takes a real value from 0 to 1, and is a correction vector

Each dimension of corresponds to a word that appears in the learning data.

That is, the correction unit 140 uses the first fixed-length vector representing the word obtained by the input conversion unit 110.

It is a forward-propagation type neural network with input as. However, other neural networks such as RNN and CNN may be used.

Then, the correction unit 140 uses the generated correction vector.

To the output unit 160.

The correction parameter storage unit 150 stores the parameters of the previously learned neural network for correcting the second fixed length vector.

Specifically, the first fixed-length vector

Parameter of the neural network, which is pre-learned by the sentence generation/learning device 20 described later, for correcting the second fixed-length vector

,
as well as

I remember.

The output unit 160 corrects the second fixed-length vector based on the second fixed-length vector and the correction vector and then calculates a vector representing the appearance probability of each word to calculate the appearance probability of each word. The word selected according to the vector representing is output as the next word.

Specifically, the output unit 160 outputs the second fixed-length vector that is the output of the RNN acquired from the generation unit 120.

And the correction vector acquired from the correction unit 140

Probability distribution showing the probability that each word will be generated using and

Is calculated by the following equation (3).

here,

Is the second fixed-length vector output by RNN

From, the matrix that is the parameter of RNN

And vector

A vector corresponding to the word in which each dimension appears in the training data, converted by

Is an operator that calculates the product of the dimensions of a vector, and softmax(·) is a softmax function that converts the value of each dimension of a given vector into a value from 0 to 1.

Also,

Is a matrix of M×N dimensions,

Is an M-dimensional vector.

That is, in this embodiment, the second fixed-length vector that is the output of the RNN is

And the correction vector generated by the correction unit 140

A vector representing the probability of each word occurrence, based on and

To calculate.

Then, the output unit 160 outputs a vector representing the appearance probability of each word.

, The word corresponding to the dimension with the highest probability is obtained as the (t+1)th output word (FIG. 3).

Then, the output unit 160 passes the output word to the input conversion unit 110 as the next word unless the output word is a symbol (end symbol) indicating the end of a sentence.

On the other hand, if the output word is the end symbol, the output unit 160 uses the words sequentially output by the output unit 160 to generate a sentence.

Specifically, the output unit 160 uses the first input word and the word string (w1 and w[2:t+1]) output by the output unit 160 so far to generate the system utterance sentence Y.

FIG. 4 shows an example of the generated output sentence when Yes is input as the first word of the sentence.

In addition, in the output unit 160, it is possible not only to select the word with the highest probability in each output and generate a sentence, but also to output some words with high probability as candidates for final output.

In this case, the number of candidate words may be determined in advance, or a threshold may be set for the high probability to generate a plurality of candidate words to be the final output and generate a plurality of sentences. The output unit 160 may output a plurality of highly probable sentences as final outputs from the generated plurality of sentences.

Further, in the dialogue system, a rule or a classifier for selecting whether or not it is consistent with the previous dialogue may be separately prepared, and one sentence may be selected from a plurality of generated sentences.

Then, the output unit 160 passes the generated sentence to the output unit 160.

The output unit 160 outputs the sentence generated by the output unit 160.

<Operation of the sentence generation device according to the embodiment of the present invention>
FIG. 5 is a flowchart showing a sentence generation processing routine according to the embodiment of the present invention.

When the first word of the sentence generated as the system utterance sentence Y, which is a response to the user utterance sentence X, is input to the input unit 100, the sentence generation device 10 executes the sentence generation processing routine shown in FIG.

First, in step S100, the input conversion unit 110 converts the input first word or the word selected in step S140 described below into a first fixed-length vector representing the word.

Next, in step S110, the generation unit 120 outputs the first fixed-length vector obtained by converting the word in step S100, the second fixed-length vector representing a series of words up to the word, or the user utterance sentence X. A second fixed-length vector representing a series of words up to the word is generated using a pre-learned neural network for predicting the next word from the second fixed-length vector represented.

In step S120, the correction unit 140 uses the pre-learned neural network for correcting the second fixed-length vector from the first fixed-length vector obtained in step S100 to obtain the second fixed-length vector. A correction vector for correcting the vector is generated.

In step S130, the output unit 160 corrects the second fixed-length vector based on the second fixed-length vector obtained in step S110 and the correction vector obtained in step S120. , Calculate a vector that represents the appearance probability of each word.

In step S140, the output unit 160 outputs the word selected according to the vector indicating the appearance probability of each word as the next word.

In step S150, the output unit 160 determines whether the word selected in step S140 is the end symbol.

When the selected word is not the end symbol (NO in step S150), the process returns to step S100, the selected word is input to the input conversion unit 110, converted by the input conversion unit 110, generated by the generation unit 120, and corrected. The generation by the unit 140 and the output by the output unit 160 (steps S100 to S140) are repeated.

On the other hand, when the selected word is the end symbol (YES in step S150), in step S160, the output unit 160 generates a sentence using each word selected in step S140, and the output unit 160 outputs the sentence. The generated sentence is output, and the sentence generation processing routine ends.

As described above, according to the sentence generation device of the present embodiment, a correction vector for correcting the second fixed length vector is generated, and based on the second fixed length vector and the correction vector. Then, by calculating a vector representing the appearance probability of each word and outputting the word selected according to the vector representing the appearance probability of each word as the next word, there are few grammatical errors and the readability is improved. Can generate high sentences.

<Outline of sentence generation learning device according to embodiment of the present invention>
Next, an outline of the sentence generation device according to the embodiment of the present invention will be described.

The parameters in the RNN and the correction unit are determined using learning data according to the task.

For example, in a dialogue system, dialogue history generated in a single language such as only English or Japanese is used as learning data. Note that document data such as newspapers may be used as learning data. The same data is used for parameter learning in the RNN and the correction unit.

After learning, the parameters are fixed. This parameter determines the accuracy of sentence generation.

The RNN and the parameter in the correction unit can be learned at the same time, or only one of them can be learned. For example, it is possible to fix one with a learned value and learn the other further.

The parameter is learned so that the sentence in the learning data can be generated.

<Configuration of sentence generation learning device according to embodiment of the present invention>
Next, the configuration of the sentence generation device according to the embodiment of the present invention will be described. FIG. 6 is a schematic diagram showing the configuration of the sentence generation/learning apparatus according to the embodiment of the present invention.

From the input learning data 30, the sentence generation/learning apparatus 20 calculates the probability that the word of the correct sentence corresponding to the next word in the vector representing the appearance probability of each word output by the neural network becomes higher. RNN, which is a neural network for predicting the word, and a neural network for correcting the second fixed-length vector are learned.

The sentence generation/learning apparatus 20 is configured by a computer including a CPU, a RAM, and a ROM that stores a program for executing a sentence generation/learning processing routine described later, and is functionally configured as shown below. ing.

As shown in FIG. 6, the sentence generation/learning apparatus 20 includes an input unit 200, a learning data conversion unit 210, a generation unit 220, an RNN parameter storage unit 230, a correction unit 240, and a correction parameter storage unit 250. The output unit 260 and the learning unit 270 are included.

The sentence generation/learning apparatus 20 is an RNN that is a neural network for predicting the next word by repeating the processes of the learning data conversion unit 210, the generation unit 220, the correction unit 240, the output unit 260, and the learning unit 270. , And a neural network for correcting the second fixed-length vector.

The input unit 200 is realized by an input device such as a known keyboard and storage device. The input unit 200 may be a network interface that inputs observation data via a network such as the Internet.

The input unit 200 receives the input of the learning data 30.

Here, the learning data 30 is configured to include at least one pair of a user utterance sentence and a system utterance sentence (correct answer sentence) for the user utterance sentence. However, the learning data does not need to be actually uttered by the user or the system, and one utterance in a dialogue between two users (or between systems) is used as a user utterance sentence and another utterance is uttered. May be used as the system utterance sentence.

The learning data conversion unit 210 converts a word of each sentence prepared in advance as learning data into a first fixed length vector representing the word.

Specifically, the learning data conversion unit 210 first receives the first word of the user utterance sentence of the learning data and converts the first word into a first fixed-length vector.

When the learning data conversion unit 210 receives a notification from the learning unit 270 that the processing of the current word is completed, the learning data conversion unit 210 takes the next word of the correct sentence as an input, similarly to the input conversion unit 110 in the sentence generation device 10. , The next word is converted to a first fixed length vector.

Then, the learning data converter 210 converts the converted first fixed-length vector

To the generation unit 220 and the correction unit 240.

Further, the learning data conversion unit 210 similarly inputs each word of the system utterance sentence of the learning data into a first fixed length vector.

The generation unit 220 uses a learning data conversion unit 210 to convert a word into a first fixed-length vector and a second fixed-length vector representing a series of words up to the word, and a neural network for predicting the next word. Using the network RNN, a second fixed-length vector representing a series of words up to the next word of the word is generated.

Specifically, first, the generation unit 220, like the generation unit 120 in the sentence generation device 10, outputs a second fixed-length vector representing a series of words up to the word.

Is input.

Further, the generation unit 220 is held in the sentence generation/learning apparatus 20 in advance when the first word is input.

Is input.

Next, the generation unit 220 uses the RNN parameter for predicting the next word from the RNN parameter storage unit 230.

,

,as well as

And the first fixed-length vector converted by the learning data conversion unit 210

, And the second fixed-length vector

Is input and the RNN is calculated. A second fixed-length vector representing the entire word string (series) from the first word to the next word (t-th word) by RNN

Is calculated using Equation (1) above.

The generation unit 220 then generates the generated second fixed-length vector.

To the output unit 260.

The RNN parameter storage unit 230 is a parameter of the RNN for predicting the next word.

When,

When,

I remember and.

The correction unit 240 uses the neural network for correcting the second fixed-length vector from the first fixed-length vector obtained by the learning data conversion unit and uses the neural network for correction to correct the second fixed-length vector. Generate a vector for use.

Specifically, the correction unit 240 first reads the first fixed-length vector from the correction parameter storage unit 250.

Parameters of neural network to correct

,
as well as

To get

Next, the correction unit 240, like the correction unit 140 in the sentence generation device 10, receives the first fixed-length vector obtained by the learning data conversion unit 210.

Is input, and a correction vector for correcting the second fixed-length vector that is the output of the RNN is calculated by the above-mentioned formula (2).

Then, the correction unit 240 uses the generated correction vector.

To the output unit 260.

The correction parameter storage unit 250 is a neural network parameter for correcting the second fixed-length vector.

When,

I remember and.

The output unit 260 corrects the second fixed-length vector based on the second fixed-length vector and the correction vector, and then calculates a vector representing the appearance probability of each word to calculate the appearance probability of each word. The word selected according to the vector representing is output as the next word.

Specifically, the output unit 260, like the output unit 160 in the sentence generation device 10, is the second fixed-length vector that is the output of the RNN acquired from the generation unit 220.

And the correction vector generated by the correction unit 240

A vector representing the probability of each word occurrence, based on and

Is calculated by the above equation (3).

Then, the output unit 160 outputs a vector representing the appearance probability of each word.

To the learning unit 270.

When the current word is included in the user's utterance, the learning unit 270 notifies the learning data conversion unit 210 that the processing of the current word is completed. When the current word is included in the system utterance sentence that is the correct sentence, in the vector representing the appearance probability of each word calculated by the output unit 160, the probability of the word of the correct sentence corresponding to the next word RNN, which is a neural network for predicting the next word, and a neural network for correcting the second fixed-length vector, are learned so that

Specifically, the learning unit 270 uses a vector representing the appearance probability of each word.

About the RNN and the first fixed-length vector using error backpropagation so that the probability of the word that appears next in the correct sentence of the learning data is high.

Learn a neural network to correct.

Then, the learning unit 270 stores the RNN parameter for predicting the next word stored in the RNN parameter storage unit 230.

When,

When,

When,

When,

And the first fixed-length vector stored in the correction parameter storage unit 250.

Parameters of neural network to correct

When,

And update.

Then, the learning unit 270 notifies the learning data conversion unit 210 that the processing of the current word is completed.

The learning data conversion unit 210 receives the second word, which is the next word, as an input. Further, as described above, by repeating the conversion by the learning data conversion unit 210, the generation by the generation unit 220, the generation by the correction unit 240, the output by the output unit 260, and the learning by the learning unit 270, the system utterance that is a correct sentence is uttered. For each word of the sentence, RNN, which is a neural network for predicting the next word, and a neural network for correcting the first fixed-length vector are learned so that the output probability of the word is high, and Update the parameters.

In addition, at the time when learning using a set of the user utterance sentence and the system utterance sentence of the learning data is completed, the second fixed-length vector

Is reset to a zero vector, and learning is performed using a set of the next user utterance sentence and system utterance sentence.

Further, when the learning using the last set of the user utterance sentence and the system utterance sentence in the learning data 30 is completed, the process returns to the first one set and the learning is performed again. Then, this procedure is repeated until the generation probability of each set in the learning data 30 is sufficiently high, and the neural network for correcting the first fixed-length vector and the RNN, which is a neural network for predicting the next word. Update network parameters.

<Operation of sentence generation learning device according to embodiment of the present invention>
FIG. 7 is a flowchart showing a sentence generation learning processing routine according to the embodiment of the present invention.

When the learning data is input to the input unit 200, the sentence generation/learning apparatus 20 executes the sentence generation/learning processing routine shown in FIG. 7.

First, in step S170, the learning data conversion unit 210 converts a word of the user utterance sentence of the learning data into a first fixed-length vector representing the word.

Next, in step S180, the generation unit 220 generates the first fixed-length vector obtained by converting the word in step S170 and the initialized second fixed-length vector, or the words up to the word of the user utterance sentence. A second fixed-length vector representing a sequence of words up to the next word is generated from the second fixed-length vector representing the sequence by using RNN, which is a neural network for predicting the next word. To do.

In step S190, it is determined whether or not the processing in steps S170 to S180 has been performed for all the words in the user's utterance, and there is a word for which the processing in steps S170 to S180 has not been performed (step S190: If NO, the process returns to step S170, and the word is converted into the first fixed-length vector.

On the other hand, when the processes of steps S170 to S180 are performed for all the words in the user's utterance (YES in step S190), the process proceeds to step S200.

In step S200, the learning data conversion unit 210 converts a word of the correct sentence prepared in advance as learning data into a first fixed-length vector representing the word.

In step S210, the generation unit 220 causes the first fixed-length vector obtained by converting the word in step S200 and the second fixed-length vector paired with the correct answer sentence included in the learning data, or the correct answer sentence of the correct answer sentence. A second fixed-length vector representing a series of words up to a word and a second series representing a series of words up to the next word of the word are used by using RNN which is a neural network for predicting the next word. Generate a fixed length vector.

In step S220, the correction unit 240 corrects the second fixed-length vector from the first fixed-length vector obtained in step S200 using a neural network for correcting the second fixed-length vector. A correction vector for is generated.

In step S230, the output unit 260 corrects the second fixed-length vector based on the second fixed-length vector obtained in step S210 and the correction vector obtained in step S220. , Calculate a vector that represents the appearance probability of each word.

In step S240, the learning unit 270 predicts the next word so that the probability of the correct word corresponding to the next word in the vector representing the appearance probability of each word calculated in step S230 is high. A neural network for correcting the second fixed-length vector is learned.

In step S250, it is determined whether or not the steps S200 to S240 have been performed on all the words in the correct sentence, and if there is a word that has not been subjected to the steps S200 to S240 (NO in step S250). ), the process returns to step S200, and the word is converted into the first fixed-length vector.

On the other hand, when the process of steps S200 to S240 is performed for all the words in the correct sentence (YES in step S250), the process proceeds to step S260.

In step S260, it is determined whether or not the processes of steps S170 to S250 have been performed for all learning data, and if there is learning data that has not been subjected to the processes of steps S200 to S250 (NO in step S260). ), the process returns to step S170, and the first word of the user utterance sentence of the learning data is converted into the first fixed-length vector.

On the other hand, when the processes of steps S170 to S250 have been performed on all learning data (YES in step S260), the process proceeds to step S270.

In step S270, the learning unit 270 determines whether learning has converged. Specifically, it is determined whether or not the generation probability of each correct answer sentence in the learning data 30 is sufficiently high, and when it is determined that the generation probability of each correct answer sentence in the learning data 30 is sufficiently high, When it is determined that the learning has converged, and it is determined that the generation probability of each correct sentence in the learning data 30 is not sufficiently high, it is determined that the learning has not converged.

If the learning has not converged (NO in step S270), the process returns to step S170, and the first word of the user utterance sentence of the first learning data is converted into the first fixed-length vector.

On the other hand, if the learning has converged (YES in step S270), the sentence generation learning processing routine ends.

As described above, according to the sentence generation/learning apparatus of the present embodiment, the correction vector for correcting the second fixed-length vector is corrected by using the neural network for correcting the second fixed-length vector. Is generated, a vector representing the appearance probability of each word is calculated based on the second fixed-length vector and the correction vector, and the word selected according to the vector representing the appearance probability of each word is By outputting the word as a word, and learning the neural network for correcting the second fixed-length vector so that the probability of the word of the correct sentence corresponding to the next word is high, there are few grammatical errors, Moreover, it is possible to learn a sentence generation device for generating a highly readable sentence.

The present invention is not limited to the above-described embodiments, and various modifications and applications are possible without departing from the scope of the present invention.

Although the present embodiment has been described using the example of the dialogue system, the present invention can be applied to all situations in which sentence generation by a computer is required, such as translation and abstract. In this case, a second fixed-length vector representing a sentence to be translated or a document to be summarized and the first word may be input. The second fixed-length vector that is input may be a zero vector (a vector whose value in each dimension is zero).

Further, in the present embodiment, the sentence generation learning device has described the case of learning the RNN and the parameter in the correction unit at the same time, but it is also possible to learn only one of them. For example, it is possible to fix one with a learned parameter and learn the other further.

Further, in the specification of the present application, the embodiment in which the program is pre-installed has been described, but the program can be stored in a computer-readable recording medium and provided.

10 sentence generation device 20 sentence generation learning device 30 learning data 100 input unit 110 input conversion unit 120 generation unit 130 parameter storage unit 140 correction unit 150 correction parameter storage unit 160 output unit 200 input unit 210 learning data conversion unit 220 generation unit 230 parameter Storage unit 240 Correction unit 250 Correction parameter storage unit 260 Output unit 270 Learning unit

Claims (5)

An input conversion unit for converting the input word into a first fixed-length vector,
The second fixed length vector obtained by the input conversion unit and the second fixed length vector generated in the past are used to generate a second word using a pre-learned neural network for predicting the next word. A generator for generating a fixed-length vector of
From the first fixed-length vector obtained by the input converter, a second learned fixed-length vector is generated by using a pre-learned neural network for correcting the second fixed-length vector generated by the generator. A correction unit that generates a correction vector for correcting the vector,
A vector representing the appearance probability of each word is calculated from the second fixed-length vector generated by the generation unit and the correction vector obtained by the correction unit, and according to the vector representing the appearance probability of each word. An output unit that outputs the selected word as the next word,
A word string in which the words output by the output unit are input to the input conversion unit and the words generated by repeating the processes of the input conversion unit, the generation unit, the correction unit, and the output unit are arranged. Is generated as a candidate sentence,
The generation unit is a first fixed-length vector obtained by the input conversion unit and a second fixed-length vector generated in the past, which is a word corresponding to the first fixed-length vector in the word string. From the second fixed-length vector corresponding to the word string up to the word immediately before the second fixed-length vector corresponding to the word string up to the word corresponding to the first fixed-length vector in the word string, Sentence generator to generate. A learning data conversion unit that converts a word of a correct sentence prepared in advance as learning data into a first fixed-length vector;
A second fixed length vector is obtained by using a neural network for predicting the next word from the first fixed length vector obtained by the learning data conversion section and the second fixed length vector generated in the past. A generator for generating a vector,
Correct the second fixed-length vector from the first fixed-length vector obtained by the learning data conversion unit using a neural network for correcting the second fixed-length vector generated by the generation unit. A correction unit for generating a correction vector for
A vector representing the appearance probability of each word is calculated from the second fixed-length vector generated by the generation unit and the correction vector obtained by the correction unit, and according to the vector representing the appearance probability of each word. An output unit that outputs the selected word as the next word,
A learning unit that learns a neural network for correcting the second fixed-length vector so that the word obtained by the output unit is a word next to the word of the correct sentence;
Including,
A sentence generation learning device that repeats each process in the learning data conversion unit, the generation unit, the correction unit, the output unit, and the learning unit by using the word output by the output unit as an input of the learning data conversion unit. The learning unit corrects the neural network for predicting the next word and the second fixed-length vector so that the word obtained by the output unit becomes the next word after the word of the correct sentence. The sentence generation/learning apparatus according to claim 2, wherein the neural network for learning is learned. The input conversion unit converts the input word into a first fixed-length vector,
A generation unit uses a pre-learned neural network for predicting the next word from the first fixed-length vector obtained by the input conversion unit and the second fixed-length vector generated in the past. To generate a second fixed-length vector,
The correction unit uses the pre-learned neural network for correcting the second fixed-length vector generated by the generation unit from the first fixed-length vector obtained by the input conversion unit, Generate a correction vector for correcting the fixed length vector of 2,
The output unit calculates a vector representing the appearance probability of each word from the second fixed-length vector generated by the generation unit and the correction vector obtained by the correction unit, and represents the appearance probability of each word. Including outputting the word selected according to the vector as the next word,
A word string in which words output by the output unit are input to the input conversion unit and words generated by repeating the processes of the input conversion unit, the generation unit, the correction unit, and the output unit are arranged. Is generated as a candidate sentence,
The generation unit is a first fixed-length vector obtained by the input conversion unit and a second fixed-length vector generated in the past, which is a word corresponding to the first fixed-length vector in the word string. From the second fixed-length vector corresponding to the word string up to the word immediately before the second fixed-length vector corresponding to the word string up to the word corresponding to the first fixed-length vector in the word string, The statement generation method to generate. A program for causing a computer to function as each unit of the sentence generation device according to claim 1.