KR102109866B1 - System and Method for Expansion Chatting Corpus Based on Similarity Measure Using Utterance Embedding by CNN - Google Patents

Abstract

The present invention relates to a chat system, specifically using word embedding and convolutional neural networks to effectively generate utterance unit representations and utterances for short utterances. A device and a method for expanding a chat corpus based on similarity measurement using speech embedding by a convolutional neural network that can be expressed, the chat pair extractor extracting a random chat pair from speech data using a window size; A chat unit representation generation unit that generates a utterance unit representation so that the machine can understand; a chat similarity calculation unit that calculates a chat similarity of a chat pair and a pre-built chat corpus randomly generated by the machine; a chat similarity threshold If it is higher than the value (Threshold), the random chat pair is the matching chat pair. And a chat corpus construction unit that determines that the chat corpus is expanded.

Description

System and Method for Expansion Chatting Corpus Based on Similarity Measure Using Utterance Embedding by CNN}

The present invention relates to a chat system, specifically using word embedding and convolutional neural networks to effectively generate utterance unit representations and utterances for short utterances. The present invention relates to an apparatus and method for expanding chat corpus based on similarity measurement using speech embedding by a convolutional neural network.

A chat system is a system that communicates between people and machines. The communication method used in the chat system is characterized by using the natural language used only between people.

The chat system is largely divided into a similarity-based chat system and a generation-based chat system.

First, the similarity-based chat system is a system that builds a database of a large number of user utterances and a system utterance response pair, finds the most similar user utterances in the database when user utterances are received as input, and outputs system utterances.

It is characterized by being able to output a robust response speech when it finds the most accurate speech.

An example of the prior art is a method of finding a semantically similar utterance using a three-step sentence search method. This method uses different morphemes and heuristic techniques for each step to increase coverage, and uses various qualities such as sentence patterns and positive / negative.

Another method is to improve the performance of the similarity-based chat system without effort of special feature selection using LSTM, which is one of deep learning models.

In addition, the generation-based chat system is a system that generates an appropriate response when a user's speech is input using the sequence to sequence model used in neural machine translation.

The sequence-to-sequence model consists of an encoder that summarizes the user's utterance and a decoder that generates the system utterance. In the step of generating the system utterance, it must be generated according to the sentence structure. It is characterized by requiring more corpus than similarity-based chat system, which is a previous study, for generation.

As a method according to an example of the prior art, a system capable of daily conversation and simple question-and-answer is proposed using a general sequence-to-sequence model, and another method recognizes a user's emotion in a general sequence-to-sequence model, and the user's emotion A method for generating a response utterance suitable for the proposed method is proposed.

In order to create such a chat system, a large number of chat corpuses in which a user's speech and a system's speech are bundled in a pair are essential.

However, since the chat corpus is currently uncommon, many studies have constructed and used the chat corpus through a lot of efforts, such as manually refining the log of the unrefined utterance.

The utterance data refers to all data uttered by a person, and refers to data in which only utterances exist, such as movies and manuscripts, but are not paired.

To create a chat pair using the utterance data, a random chat pair is constructed by grabbing a window size.

In addition, the similarity between a random chat pair and a pre-built chat standard corpus is calculated to determine whether the system speech is an appropriate response in response to a user's speech.

Representation of the unit of utterance means to express the utterance as a vector so that the machine can understand it.

TF (Term Frequency) and IDF (Inverted Document Frequency) are frequently used as methods of the prior art for generating a speech unit representation.

However, since chat utterance is composed of very short lengths, TF (Term Frequency) and TF * IDF (Inverted Documents Frequency), which are commonly used to express sentences and documents, are used as very sparsity vectors. There is a problem that expressive and semantic information is not included.

Republic of Korea Registered Patent No. 10-1814958 Republic of Korea Registered Patent No. 10-1741248

The present invention is to solve the problems of the chat system (Chatting system) of the prior art, by using a word-level embedding vector (Word embedding) and convolutional neural networks (Convolutional Neural Networks) effective speech representation for short speech An object of the present invention is to provide an apparatus and method for expanding a chat corpus based on similarity measurement using speech embedding by a convolutional neural network that enables to generate and express speech.

The present invention is for the expansion of the chat corpus based on similarity measurement using the speech embedding by the convolutional neural network that determines the machine as the correct chat pair and the chat corpus represented by the machine understandable vector as 0 and 1 The purpose is to provide an apparatus and method.

The present invention uses a word-level embedding and a convolutional neural network model to effectively express short-length chatty speech, a speech unit that reflects low dimensions and semantic information well. An object of the present invention is to provide an apparatus and method for expanding chat corpus based on similarity measurement using speech embedding by a convolutional neural network that generates a representation (Utterance Representation) and calculates similarity between speeches using the same.

The present invention creates a random pair from a large amount of speech data such as movie subtitles and manuscripts, calculates the chat similarity with a pre-built chat standard, and calculates the calculated chat similarity through an experiment threshold If it is greater than (Threshold), it is an object of the present invention to provide an apparatus and method for expanding a chat corpus based on similarity measurement using utterance embedding by a convolutional neural network that effectively determines a correct chat pair and expands the chat corpus. .

The present invention is semi-automatic because a person can modify a corpus constructed by a machine (Machine Labeled Chatting corpus), and because the machine judges primarily, speech embedding by a convolutional neural network has the effect of reducing the cost of human effort. An object of the present invention is to provide an apparatus and method for expanding a chat corpus based on similarity measurement.

Other objects of the present invention are not limited to those mentioned above, and other objects not mentioned will be clearly understood by those skilled in the art from the following description.

The apparatus for expanding the chat corpus based on similarity measurement using utterance embedding by the convolutional neural network according to the present invention for achieving the above object extracts a chat pair extracting a random chat pair from utterance data using a window size Wealth; a utterance unit representation generation unit that generates utterance unit representations so that the machine can understand utterances; a chat similarity calculation unit that calculates chat similarity between randomly created chat pairs and pre-built chat corpus; If the similarity is higher than the threshold (Threshold), the random chat pair is determined to be a chat pair with a correct response relationship, the chat corpus construction unit to expand the chat corpus; characterized in that it comprises a.

Here, the chat similarity calculator calculates the chat similarity when the i-th random pair comes in as an input,

으로 구하고, i번째 쌍은 길이가 n인 미리 구축된 채팅 말뭉치의 전체 쌍과 각각 유사도를 계산하여, 구해진 유사도 중 가장 큰 값을 i번째 쌍의 채팅 유사도라 하고, 이 채팅 유사도가 미리 정의된 임계값보다 크다면 올바른 쌍이라고 판단하는 것을 특징으로 한다.

The i-th pair is calculated by calculating the similarity with the entire pair of pre-built chat corpus having a length of n, and the largest value among the obtained similarities is called the chat similarity of the i-th pair, and this chat similarity is a predefined threshold If it is greater than the value, it is characterized by determining that it is the correct pair.

) 이용으로, 선형 결합(Linear combination)하여 하나의 채팅 유사도로 표현하는 것을 특징으로 한다.And, to calculate chat similarity, cosine similarity is used, and since the randomly extracted pair and the pre-built chat corpus are composed of pairs of user speech and system speech, each similarity is calculated, and the two similarities are similar. Gamma (reflection ratio of

), It is characterized by expressing as a single chat similarity by linear combination.

In addition, the utterance unit representation generation unit selects only morphemes that can effectively express short-length utterances by using morphological DF (Document Frequency) in order to express them as vectors containing semantic information of low dimensions, and average embedding. The morpheme used to generate a vector is characterized by selectively including a common noun, a proper noun, a rhetoric, a verb, an adjective, and a general adverb.

In addition, the speech unit representation generation unit generates a speech unit representation using a convolutional neural network model and word unit embedding.

Then, the utterance is expressed as a morphological unit embedding vector through the projection layer, and is derived as a deep feature representation using a convolution layer and max polling, and a final output vector using the derived deep feature representation. It is characterized by inducing) and learning by calculating the difference with the answer vector.

And for learning, the correct answer vector of the convolutional neural network model is generated using Latent Semantic Analysis (LSA) and TF * IDF, and LSA that expresses using TF * IDF for utterance, reduces dimension, and performs potential semantic analysis. Use to separate the matrix, use a low-dimensional dense vector as the correct answer vector, derive an output vector using the convolutional neural network model, and reduce the correct answer vector and cosine distance. Characterized by the progress of learning.

And, the output vector is used for learning, and the vector used as the actual speech unit representation is characterized by using the deep feature representation of the learning-composite neural network model as the speech unit representation.

A method for expanding a chat corpus based on similarity measurement using speech embedding by a convolutional neural network according to the present invention for achieving another object includes a chat pair extraction step of extracting a random chat pair from speech data using a window size; A utterance unit representation generation step of generating a utterance unit representation so that the machine can understand the utterance; a chat similarity calculation step of calculating a chat similarity of a chat pair and a pre-built chat corpus randomly generated by the machine; chat similarity If it is higher than the threshold (Threshold), the random chat pair is determined as a matching chat pair, the chat corpus expansion step to expand the chat corpus; characterized in that it comprises a.

Here, in the chat similarity calculation step, when the i-th random pair (pair) comes into the input, the chat similarity,

으로 구하고, i번째 쌍은 길이가 n인 미리 구축된 채팅 말뭉치의 전체 쌍과 각각 유사도를 계산하여, 구해진 유사도 중 가장 큰 값을 i번째 쌍의 채팅 유사도라 하고, 이 채팅 유사도가 미리 정의된 임계값보다 크다면 올바른 쌍이라고 판단하는 것을 특징으로 한다.

The i-th pair is calculated by calculating the similarity with the entire pair of pre-built chat corpus having a length of n, and the largest value among the obtained similarities is called the chat similarity of the i-th pair, and this chat similarity is a predefined threshold If it is greater than the value, it is characterized by determining that it is the correct pair.

) 이용으로, 선형 결합(Linear combination)하여 하나의 채팅 유사도로 표현하는 것을 특징으로 한다.And, to calculate chat similarity, cosine similarity is used, and since the randomly extracted pair and the pre-built chat corpus are composed of pairs of user speech and system speech, each similarity is calculated, and the two similarities are similar. Gamma (reflection ratio of

), It is characterized by expressing as a single chat similarity by linear combination.

Also, in the step of generating the representation of the unit of speech, only a morpheme capable of effectively expressing a short utterance is selected and averaged by using a document frequency (DF) in order to express a vector containing semantic information of low dimensions. The embedding vector is generated and the morpheme used is characterized by selectively including a common noun, a proper noun, a rhetoric, a verb, an adjective, and a general adverb.

Also, in the step of generating a speech unit representation, a speech unit representation is generated using a convolutional neural network model and word unit embedding.

Then, the utterance is expressed as a morphological unit embedding vector through the projection layer, and is derived as a deep feature representation using a convolution layer and max polling, and a final output vector using the derived deep feature representation. It is characterized by inducing) and learning by calculating the difference with the answer vector.

And for learning, the correct answer vector of the convolutional neural network model is generated using Latent Semantic Analysis (LSA) and TF * IDF, and LSA that expresses using TF * IDF for utterance, reduces dimension, and performs potential semantic analysis. Use to separate the matrix, use a low-dimensional dense vector as the correct answer vector, derive an output vector using the convolutional neural network model, and reduce the correct answer vector and cosine distance. Characterized by the progress of learning.

And, the output vector is used for learning, and the vector used as the actual speech unit representation is characterized by using the deep feature representation of the learning-composite neural network model as the speech unit representation.

The apparatus and method for expanding the chat corpus based on similarity measurement using speech embedding by the convolutional neural network according to the present invention as described above has the following effects.

First, using word embedding and convolutional neural networks, it is possible to effectively generate utterance unit representations and express utterances for short utterances.

Second, it is possible to effectively expand the chat corpus by determining whether the machine is the correct chat pair as 0 and 1 as the speech pair represented by the machine understandable vector.

Third, by using word embedding and convolutional neural networks models, a low-dimensional, semantic information well reflected Utterance Representation is generated and spoken using it. By calculating liver similarity, short-length chat utterances can be effectively expressed.

Fourth, a random pair is created from a large amount of speech data, such as movie subtitles and manuscripts, the pre-constructed chat standard (Golden standard chatting corpus) and the chat similarity are calculated, and the calculated chat similarity is the threshold value obtained through the experiment ( If it is greater than Threshold), it is possible to expand the chat corpus effectively by determining that any pair is a correct chat pair.

Fifth, it is semi-automatic because a machine can modify a corpus (Machine Labeled Chatting corpus) built by a machine, and it has an effect of reducing the cost of human effort because the machine judges it primarily.

1 is a block diagram of an apparatus for expanding chat corpus based on similarity measurement using speech embedding by a convolutional neural network according to the present invention;
2 is a block diagram showing an example of a movie subtitle that is one of utterance data
3 is an overall configuration diagram of a chat corpus semi-automatic construction model according to the present invention
4 is a block diagram showing an example of an average embedding vector
5 is a block diagram showing the generation of speech unit representations using a convolutional neural network model.
6 is a flow chart illustrating a method for expanding chat corpus based on similarity measurement using speech embedding by a convolutional neural network according to the present invention.

Hereinafter, a preferred embodiment of an apparatus and method for expanding chat corpus based on similarity measurement using speech embedding by a convolutional neural network according to the present invention will be described in detail as follows.

The features and advantages of the apparatus and method for expanding the chat corpus based on similarity measurement using speech embedding by the convolutional neural network according to the present invention will become apparent through detailed description of each embodiment below.

1 is a block diagram of a device for expanding chat corpus based on similarity measurement using speech embedding by a convolutional neural network according to the present invention.

The apparatus and method for expanding chat corpus based on similarity measurement using speech embedding by the convolutional neural network according to the present invention extracts chat pairs using a large amount of speech data and expands chat corpus to reduce the difficulty of constructing chat corpus Is to do.

In order to reduce the difficulty of constructing a chat corpus using such a large amount of speech data, the present invention defines a chat corpus extension system and uses a morpheme unit embedding vector and a convolutional neural network to effectively express short-length chattability speech. Creates a representation of deep qualities that express the utterance well.

Using this, chat corpus can be easily expanded, and chat corpus having various expressions can be constructed.

The reason why it is difficult to secure a good quality and large amount of chat corpus is that the user talk and the system talk must first be paired, and the system talk must be appropriate in response to the user talk.

Since such a chat corpus is a public corpus, it is generally used by a person who judges and builds a conversation log between real people.

However, there are a lot of utterance data that are simply arranged in chronological order, such as movie subtitles and scripts. In the present invention, it is possible to reduce the cost of establishing a chat corpus by extracting the correct chat pair from such a large amount of speech data.

The apparatus for expanding the chat corpus based on similarity measurement using utterance embedding by the convolutional neural network according to the present invention is a chat pair extractor extracting a random chat pair from utterance data using a window size as shown in FIG. 10), a utterance unit representation generation unit 20 for generating utterance unit representations so that the machine can understand utterances, and a chat pairing randomly made in the machine and a chat corpus of a pre-built chat corpus that are calculated The chat similarity calculator 30 includes a chat corpus construction unit 40 that determines if a chat similarity is higher than a threshold and a random chat pair is a matching chat pair and expands the chat corpus.

The present invention having such a configuration is to extract a correct chat pair from a large amount of speech data using a similarity technique and a pre-built chat corpus for semi-automatic expansion of the chat corpus.

The utterance data means data in which utterances are simply arranged in chronological order, such as movie subtitles and manuscripts. There is a large amount of such utterance data, but in order to construct a chat corpus, it must be extracted as a pair having a correct response relationship.

2 is a block diagram showing an example of a movie subtitle that is one of utterance data.

Since the utterance data is organized in chronological order, in the present invention, first, it is composed of arbitrary pairs.

To build a pair, we assume that the t-th utterance is a chat pair with the t + 1, t + 2, ..., t + window size utterance with a window size.

Table 1 is an example of a randomly created chat pair using a window size.

The number is an index of the talk pair, and the chat similarity is to determine whether the response relationship between the user talk and the system talk is correct.

The number 1 is the correct real answer pair in the movie, and the number 4 is the real answer in the movie that is not the real answer of 'What time?'

And numbers 2, 3, and 5 are incorrect response pairs. By simply constructing the pair using the window size as described above, it can be seen that the pair of the actual response in the movie and the pair that is not the actual response in the movie but which is considered to be correct are mixed.

Therefore, in the present invention, a model for semi-automatic construction of a chat corpus is constructed as follows.

3 is an overall configuration diagram of a chat corpus semi-automatic construction model according to the present invention.

Calculates the chat similarity between a randomly created chat pair on a machine named Machine and a pre-built chat corpus, and if the chat similarity is higher than the threshold, the random chat pair is a matching chat pair I judge that.

It is called a Machine Labeled Chatting corpus built by a machine, and it is semi-automated because a human can modify it, and because the machine judges it primarily, it can reduce the cost of human effort.

In order to obtain the chat similarity, it is necessary to first generate a vector or a unit of speech representation so that the machine can understand the speech.

In the basic model of the present invention, a utterance unit representation is generated using TF and TF * IDF, which are commonly used to express utterances as vectors.

Equation 1 is a formula for calculating the chat similarity when the i-th random pair comes in as an input.

The i-th pair calculates the similarity with the entire pair of pre-built chat corpus of length n.

The largest value among the obtained similarities is referred to as the i-th pair's chat similarity, and if the chat similarity is greater than a predefined threshold, it is determined that the pair is correct.

In the present invention, cosine similarity is used to calculate the similarity.

) 이용으로, 선형 결합(Linear combination)하여 하나의 채팅 유사도로 표현한다.And since both randomly extracted pairs and pre-built chat corpus are composed of pairs of user utterances and system utterances, each similarity is calculated, and gamma (reflection ratio of the two similarities)

), It is expressed as a single chat similarity by linear combination.

In the basic model according to the present invention, TF and TF * IDF, which are commonly used for generating utterance unit representation, were used.

However, since the chat utterance is very short in length, when a speech unit representation is generated using TF and TF * IDF, it is expressed as a very rare vector, and there is a problem that the utterance cannot be expressed well.

Therefore, in the present invention, a speech unit representation is generated using a word unit embedding vector.

The word unit embedding vector used uses a morpheme unit embedding vector pre-trained using a large amount of corpus and word2vec.

First, the first method is to use the average embedding vector of morphemes appearing in the utterance as a representation of the utterance unit.

The method of making the average embedding vector is as in FIG. 4.

FIG. 4 shows an example of an average embedding vector, and is an example in which the utterance 'I love my father' is expressed as an average of morpheme unit embedding vectors.

If the average embedding vector is generated in this way, it can be expressed as a vector containing low dimensions and semantic information.

However, if an average embedding vector is generated using the entire morpheme as described above, a lot of morphemes commonly found in other utterances, such as irradiation, are included.

In particular, in the case of very short utterances such as chat utterances, it is difficult to distinguish similar utterances from similar utterances by reflecting a lot of values for investigation.

In the present invention, an average embedding vector is generated by selecting only morphemes capable of effectively expressing short utterances using document morphology (DF).

The morphemes used are general nouns, proper nouns, rhetoric, verbs, adjectives, and general adverbs.

The second method is to generate a speech unit representation using a convolutional neural network model and word unit embedding.

5 is a block diagram showing the generation of utterance unit representation using a convolutional neural network model.

FIG. 5 shows the structure of a convolutional neural network model that generates an utterance unit representation for a short utterance, and short-length utterance is used as an input to the convolutional neural network model.

The short-length utterance is expressed as a morphological unit embedding vector through the projection layer, and is derived as a deep feature representation using the convolution layer and max polling.

This is a general convolutional neural network model that derives the final output vector using the derived deep feature representation and learns by calculating the difference with the answer vector.

However, in FIG. 5, since the purpose is to create an utterance unit representation rather than a general task of matching a correct answer tag using a convolutional neural network, learning must proceed well by making a correct answer to the input.

In the present invention, for smooth learning, the correct answer vector of the convolutional neural network model is generated using LSA (Latent Semantic Analysis) and TF * IDF.

First, a short utterance is expressed using TF * IDF, the matrix is separated using an LSA that performs a dimension reduction effect and potential semantic analysis, and a low-dimensional dense vector is used as a correct answer vector.

Then, an output vector is derived using a convolutional neural network model, and learning is performed so that a correct answer vector and a cosine distance are reduced.

The output vector is used for learning, and the vector used as the actual speech unit representation uses the deep feature representation of the learning-converged neural network model as the speech unit representation.

When learning is performed in this way, important qualities can be well extracted from the morphological unit embedding vector, and the short-length utterance can be expressed more effectively than using the average embedding vector as the utterance unit representation.

A method for expanding the chat corpus based on similarity measurement using speech embedding by the convolutional neural network according to the present invention will be described in detail as follows.

6 is a flow chart illustrating a method for expanding chat corpus based on similarity measurement using speech embedding by a convolutional neural network according to the present invention.

First, an arbitrary chat pair is extracted from the utterance data (S601), and a utterance unit representation generation step of generating a utterance unit representation so that the machine can understand the utterance is performed (S602).

Subsequently, the chat similarity of the chat pair pre-built and the chat pair randomly made by the machine is calculated. (S603)

Then, if the calculated chat similarity is higher than the threshold, the random chat pair is determined to be a chat pair with a correct response relationship (S604), and a chat corpus is constructed (S605).

A description of the chat corpus confirmation performance of the apparatus and method for expanding the chat corpus based on similarity measurement using speech embedding by the convolutional neural network according to the present invention described above is as follows.

Random pairs were made from a large amount of utterance data, and whether a person is available as a chat corpus or a chat similar or a chat not similar was directly attached. And experiments were configured to increase the performance of predicting chat similarity.

In addition, since the length of chat utterance is very short, a comparative experiment was conducted on a method of expressing utterance.

First, a large amount of utterance data used in the chat corpus semi-automatic construction experiment used subtitles from movies and foreign dramas collected from Opensubtitle.

For threshold and quality experiments and evaluation, development data and evaluation data were extracted from 6 films and foreign dramas respectively, and 3 corpus construction personnel attached presence or absence of chat similarity.

Table 2 shows development data and Table 3 shows statistics of evaluation data.

The kappa coefficient was measured for data reliability, and the kappa coefficient was 0.8114.

And about 400,000 pairs of chat corpus used in similarity calculation were used. The morpheme analyzer used for speech data and chat corpus [Hyemin Kim, Jungmin Min, Jaehyun Ahn, Baemanman, Koyoungjoong, "Synchronous morpheme analyzer using part-of-speech distribution and bidirectional LSTM-CRFs", 28th Korean and Korean Information Processing Conference , pp.3-8, 2016.], and used 100-dimensional morpheme unit embedding vector was learned using a large amount of corpus and word2vec.

Table 4 is a performance comparison table of the baseline system that generated the ignition unit representation using a typical TF.

)를 설정하기 위해 진행하였다.Gamma, which is the ratio reflecting the similarity between user speech and system speech in the basic baseline system

) To proceed.

F1 was used as the evaluation standard.

As a result of the experiment in Table 4, the gamma, the reflection ratio, showed the highest performance when 0.5, and the gamma used in all experiments was set to 0.5 to conduct the experiment.

In addition, the threshold value when the highest F1 determined through the development data was used was used for all experimental performances shown hereinafter.

Table 5 is a comparison table of basic TF, IDF, and TF * IDF.

As shown in Table 5, it can be seen that performance is improved when TF * IDF is used than when only TF and IDF are used.

Table 6 is a performance comparison table when the utterance unit representation is generated with TF * IDF and average embedding vector.

It can be seen that the performance of the average embedding vector is improved compared to the use of TF * IDF when generating the utterance unit representation.

Table 7 is a performance comparison table when a speech unit representation is generated using the final model, the convolutional neural network model and the morphological unit embedding vector.

Through the results in Table 7, it can be seen that the utterance can be more effectively expressed by using the morphological unit embedding vector and the convolutional neural network model.

The apparatus and method for expanding chat corpus based on similarity measurement using speech embedding by the convolutional neural network according to the present invention described above extracts chat pairs from a large amount of speech data and reduces correct chat pairs to reduce the difficulty of constructing chat corpus In order to extract well, the ignition unit representation is generated.

As a result, the method that can effectively express short-length speech by increasing the accuracy rate, reproducibility, and F1 by 5.16% p, 6.09% p, and 5.73% p, respectively, is better than using the morphological unit embedding vector and convolutional neural network model. It can be seen that it can be improved in terms of.

It will be understood that the present invention is implemented in a modified form without departing from the essential characteristics of the present invention as described above.

Therefore, the specified embodiments should be considered in terms of explanation rather than limitation, and the scope of the present invention is shown in the claims rather than the foregoing description, and all differences within the equivalent range are included in the present invention. Should be interpreted.

10. Chat Pair Extraction Unit
20. Ignition unit representation generation unit
30. Chat similarity calculator
40. Chat Corps Construction Department

Claims (16)

A chat pair extractor that extracts a random chat pair from the utterance data using the window size;
An utterance unit representation generation unit for generating an ignition unit representation so that the machine can understand the ignition;
A chat similarity calculation unit that calculates a chat similarity between a chat pair randomly made in a machine and a pre-built chat corpus;
If the chat similarity is higher than the threshold (Threshold), the random chat pair determines that the response is a matching chat pair, the chat corpus construction unit to expand the chat corpus; includes
The utterance unit representation generation unit selects only morphemes that can effectively express short-length utterances using document morphology DF (Document Frequency) in order to express them as vectors containing semantic information of low dimensions, and averages the embedding vector. The device for generating and using the morpheme based on similarity measurement based on speech similarity using speech embedding by a convolutional neural network, characterized in that it includes a general noun, a proper noun, a rhetoric, a verb, an adjective, and a general adverb. The method of claim 1, wherein the chat similarity calculator calculates the chat similarity when an i-th random pair (pair) is input.

With
The i-th pair calculates the similarity with each pair of pre-built chat corpus of length n, and the largest value among the obtained similarities is called the chat similarity of the i-th pair, and the chat similarity is greater than a predefined threshold. A device for expanding chat corpus based on similarity measurement using speech embedding by a convolutional neural network, characterized in that it is determined to be the correct pair of faces. The method of claim 2, wherein a cosine similarity is used to calculate the chat similarity,
Since the randomly extracted pair and the pre-built chat corpus are composed of a pair of user speech and system speech, each similarity is calculated, and gamma (reflection ratio of the two similarities)

), A device for expanding chat corpus based on similarity measurement using speech embedding by a convolutional neural network, characterized in that it is expressed as a single chat similarity by linear combination. delete According to claim 1, The utterance unit representation generation unit,
A device for expanding chat corpus based on similarity measurement using speech embedding by a convolutional neural network, characterized in that a speech unit representation is generated using a convolutional neural network model and word unit embedding. The method of claim 5, wherein the utterance is expressed as a morpheme unit embedding vector through a projection layer,
Derived into a deep feature representation using the convolution layer and max polling, derived final output vector using the derived deep feature representation, and calculated the difference with the answer vector (Answer vector) A device for expanding a chat corpus based on similarity measurement using speech embedding by a convolutional neural network, characterized by learning by using a convolution. The method of claim 6, for learning, the correct answer vector of the convolutional neural network model is generated using Latent Semantic Analysis (LSA) and TF * IDF.
I use TF * IDF for utterance, separate the matrix using LSA to reduce the dimension and perform potential semantic analysis, and use the low-dimensional dense vector as the correct answer vector,
Based on similarity measurement using speech embedding by a convolutional neural network, characterized in that an output vector is derived using a convolutional neural network model, and learning is performed so that the correct answer vector and the cosine distance are reduced. Device for expanding chat corpus. The output vector is used for learning, and a vector used as a real speech unit representation is used for a speech representation of a convolutional neural network, characterized by using the deep feature representation of the learning completed convolutional neural network model as the speech unit representation. Device for expanding chat corpus based on similarity measurement using speech embedding. A chat pair extraction step of extracting a random chat pair from the utterance data by using a window size in the chat pair extraction unit;
An utterance unit representation generation step of generating an utterance unit representation so that the machine can understand the utterance in the utterance unit representation generation unit;
Chat similarity calculation step of calculating the chat similarity (Chatting similarity) of the chat pairs and pre-built chat pairs randomly made by the machine in the chat similarity calculation unit;
If the chat similarity is higher than the threshold (Threshold), the chat corpus construction unit determines that any chat pair is a chat pair having a correct response relationship and constructs a chat corpus to expand the chat corpus;
In order to express as a vector containing semantic information of low dimensions in the utterance unit representation generation step, an average embedding vector is selected by selecting only morphemes that can effectively express short-length utterances using DF (Document Frequency) Method for generating a chat corpus expansion based on similarity measurement using utterance embedding by a convolutional neural network, characterized in that the morpheme used includes a general noun, a proper noun, a rhetoric, a verb, an adjective, and a general adverb. 10. The method of claim 9, In the chat similarity calculation step,
Chat similarity when the i-th random pair comes in as input,

With
The i-th pair calculates the similarity with each pair of pre-built chat corpus of length n, and the largest value among the obtained similarities is called the chat similarity of the i-th pair, and the chat similarity is greater than a predefined threshold. A method for expanding chat corpus based on similarity measurement using speech embedding by a convolutional neural network, characterized by determining that the face is a correct pair. 11. The method of claim 10, Cosine similarity (Cosine similarity) is used to calculate the chat similarity,
Since the randomly extracted pair and the pre-built chat corpus are composed of a pair of user speech and system speech, each similarity is calculated, and gamma (reflection ratio of the two similarities)

), A method for similarity measurement based chat corpus expansion using speech embedding by a convolutional neural network, characterized in that it is expressed as a single chat similarity by linear combination. delete 10. The method of claim 9, In the ignition unit representation generation step,
A method for expanding chat corpus based on similarity measurement using speech embedding by a convolutional neural network, characterized by generating speech unit representations using a convolutional neural network model and word unit embedding. The method of claim 13, wherein the utterance is expressed as a morpheme unit embedding vector through a projection layer,
Derived into a deep feature representation using the convolution layer and max polling, derived final output vector using the derived deep feature representation, and calculated the difference with the answer vector (Answer vector) A method for expanding a chat corpus based on similarity measurement using speech embedding by a convolutional neural network, characterized by learning by using the method. 15. The method of claim 14, For learning, the correct answer vector of the convolutional neural network model is generated using Latent Semantic Analysis (LSA) and TF * IDF.
I use TF * IDF for utterance, separate the matrix using LSA to reduce the dimension and perform potential semantic analysis, and use the low-dimensional dense vector as the correct answer vector,
Based on similarity measurement using speech embedding by a convolutional neural network, characterized in that an output vector is derived using a convolutional neural network model, and learning is performed so that the correct answer vector and the cosine distance are reduced. Ways to expand chat corpus. 16. The method according to claim 15, wherein the output vector is used for learning, and the vector used as the actual speech unit representation is used for the speech recognition unit representation using the deep feature representation of the completed learning product neural network model as the speech unit representation. Method for expanding chat corpus based on similarity measurement using speech embedding.

Images