KR20220073475A - System and method for natural language understanding and method for processing learning data thereof - Google Patents

Abstract

A natural language understanding system and a method for processing learning data in the system are disclosed. The natural language understanding system according to an embodiment includes one or more intents and one or more learning sentences corresponding to the respective intentions, and the intention and the learning sentences are learning data configured in one of a first language or a second language. a learning data analysis module for receiving an input and determining whether there is a mixture of language between intentions of the learning data or within the same intention; and a weight adjustment module configured to adjust weights of at least a portion of one or more first language tokens and one or more second language tokens generated from the learning data based on the determination result.

Description

SYSTEM AND METHOD FOR NATURAL LANGUAGE UNDERSTANDING AND METHOD FOR PROCESSING LEARNING DATA THEREOF

Disclosed embodiments relate to a technology of natural language processing/understanding using machine learning.

In computer science, natural language understanding (NLU) means that a computer receives a sentence composed of a natural language such as Korean, Japanese, or English, which is commonly used by humans for communication, and infers the intent of the inputted sentence. means to do There are various technologies for understanding natural language on a computer, but recently, natural language understanding technology based on machine learning has been mainly researched.

Most of the NLU models currently on the market are modeled by selecting only one specific language, such as Korean or English. However, in this case, it is difficult for a service provider (eg, a chatbot service, etc.) to provide a multilingual service. This is because it is necessary to separately create different NLU models for each language type provided by the service provider. Also, from the service user's point of view, there was the inconvenience of having to separately select an appropriate NLU model according to the language they use before using the service. Accordingly, a means for configuring two or more different languages into one NLU model is needed.

Republic of Korea Patent Publication No. 10-2018-0109624 A (2018.10.08)

The disclosed embodiments are intended to provide technical means for implementing an NLU model capable of simultaneously processing two or more languages having different characteristics.

According to an exemplary embodiment, input learning data including one or more intents and one or more learning sentences corresponding to the respective intentions, wherein the intention and the learning sentences are configured in one of a first language or a second language a learning data analysis module for receiving and judging whether there is a mixture of language between intentions of the learning data or within the same intention; and a weight adjustment module configured to adjust weights of at least a portion of one or more first language tokens and one or more second language tokens generated from the learning data based on the determination result.

The weight may include one or more of a word frequency (TF) and an inverse document frequency (IDF).

As a result of the determination, if there is a mixture of language between intentions of the learning data, and the share of the first language intent in the learning data is higher than the share of the second language intent, the weight adjustment module is configured to: The frequency of the inverse document of the first language token generated from the learning sentence may be adjusted.

The system further comprises a balance analysis module for determining whether a ratio of the second language token to the first language token in the learning data is less than or equal to a preset threshold, the weight adjustment module, the balance When the analysis module determines that the ratio of the second language token to the first language token in the learning data is less than or equal to a preset threshold, the frequency of inverse documents of the first language token may be adjusted.

The weight adjustment module may adjust the inverse document frequency of the first language token based on a ratio of the first language intent to the second language intent.

If, as a result of the determination, there is a mixture of languages in the specific intention of the learning data, and the share of the first language learning sentences in the specific intent is higher than the share of the second language learning sentences, the weight adjustment module is configured to learn the second language A word frequency of the second language token generated from a sentence may be adjusted.

The system further includes a balance analysis module for determining whether a ratio of the second language learning sentence to the first language learning sentence in the specific intention in the learning data is less than or equal to a preset threshold, the weight The adjustment module may adjust the word frequency of the second language token when the ratio of the second language learning sentence to the first language learning sentence in the specific intention is less than or equal to a preset threshold as a result of the determination of the balance analysis module have.

The weight adjustment module may adjust the word frequency of the second language token based on a ratio of the first language learning sentence to the second language learning sentence.

According to another exemplary embodiment, it includes one or more intents and one or more learning sentences corresponding to the respective intentions, wherein the intention and the learning sentences include learning data composed of one of a first language or a second language. receiving input; determining whether there is a language mixture between intentions of the learning data or within the same intention; and adjusting a weight of at least a portion of one or more first language tokens and one or more second language tokens generated from the learning data based on the determination result.

The weight may include one or more of a word frequency (TF) and an inverse document frequency (IDF).

If, as a result of the determination, there exists a mixed language between intentions of the learning data, and the share of the first language intent in the learning data is higher than the share of the second language intent, the step of adjusting the weight may include: and adjust an inverse document frequency of the first language token generated from a corresponding learning sentence.

The adjusting of the weight may be configured to adjust the inverse document frequency of the first language token when a ratio of the second language token to the first language token in the learning data is less than or equal to a preset threshold value. .

The adjusting the weight may be configured to adjust the inverse document frequency of the first language token based on a ratio of the first language intent to the second language intent.

If, as a result of the determination, there is a mixture of languages in the specific intention of the learning data, and the occupancy of the first language learning sentences in the specific intent is higher than the occupancy of the second language learning sentences, the step of adjusting the weight may include: and adjust the word frequency of the second language token generated from a language learning sentence.

The adjusting of the weight may be configured to adjust the word frequency of the second language token when the ratio of the second language learning sentence to the first language learning sentence in the specific intention is less than or equal to a preset threshold value. have.

The adjusting the weight may be configured to adjust the word frequency of the second language token based on a ratio of the first language intent to the second language intent.

According to the disclosed embodiments, two or more languages having different characteristics may be processed through one NLU model. Therefore, it is possible to replace multiple NLU models for each language, particularly in chatbots or messengers that require multilingual services.

1 is a block diagram illustrating an NLU system 100 according to an embodiment.
2 is an exemplary diagram for explaining the configuration of learning data according to an embodiment;
3 is an exemplary diagram for explaining the configuration of learning data according to another embodiment;
4 is a block diagram illustrating in detail the feature extractor 104 of the NLU system 100 according to an embodiment.
5 is a flowchart illustrating a learning data processing process 500 according to an embodiment.
6 is a block diagram illustrating and describing a computing environment including a computing device suitable for use in example embodiments;

Hereinafter, specific embodiments of the present invention will be described with reference to the drawings. The following detailed description is provided to provide a comprehensive understanding of the methods, devices, and/or systems described herein. However, this is merely an example, and the present invention is not limited thereto.

In describing the embodiments of the present invention, if it is determined that the detailed description of the known technology related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description thereof will be omitted. And, the terms to be described later are terms defined in consideration of functions in the present invention, which may vary according to intentions or customs of users and operators. Therefore, the definition should be made based on the content throughout this specification. The terminology used in the detailed description is for the purpose of describing embodiments of the present invention only, and should in no way be limiting. Unless explicitly used otherwise, expressions in the singular include the meaning of the plural. In this description, expressions such as “comprising” or “comprising” are intended to indicate certain features, numbers, steps, acts, elements, some or a combination thereof, one or more other than those described. It should not be construed to exclude the presence or possibility of other features, numbers, steps, acts, elements, or any part or combination thereof.

1 is a block diagram illustrating a natural language understanding (NLU) system 100 according to an embodiment. The NLU system 100 according to an embodiment is a system for learning the learning data composed of a plurality of natural language sentences to configure an NLU model, and classifying intentions for the input natural language sentences using the NLU model. As shown, the NLU system 100 according to one embodiment includes a token generator 102 , a feature extractor 104 , and an intent classifier 106 .

The token generator 102 (Tokenizer) receives learning data and generates a token from learning sentences included in the learning data. In the disclosed embodiments, the learning data includes one or more intents and one or more learning sentences corresponding to the respective intentions, and the intent and the learning sentences may be configured in one of a first language or a second language. have. For example, when the first language is Korean and the second language is English, the learning data may include a plurality of learning sentences in English or Korean, respectively, and each of the learning sentences is a plurality of intentions in English or Korean, respectively. can be classified as either one. In the following description, for convenience, it is assumed that the first language is Korean and the second language is English. However, the disclosed embodiments are not limited to a specific type of language.

2 is an exemplary diagram for explaining the configuration of learning data according to an embodiment. In the illustrated embodiment, the learning data may include a plurality of intentions having any one of the first language and the second language. That is, in the present embodiment, there is a mixture of languages between a plurality of intentions in the learning data. However, in this case, language mixing does not occur between learning sentences with the same intention. For example, in the case of the first language intent, only learning sentences in the first language are included, and in the case of the second language intent, only learning sentences in the second language are included.

<Example 1> below is an example of a learning sentence in which language mixture between intentions exists. In <Example 1>, intentions 1 to 4 are Korean intentions, and intentions 5 and 6 are English intentions. Also, it can be seen that the Korean intention includes Korean learning sentences and the English intention includes only English learning sentences.

<Example 1>

Intent 1: Sound is not clear

- Learning sentence 1: The sound quality of the TV is not good. What should I do?

- Study Sentence 2: My TV sounds bad

Intent 2: TV won't turn on

- Learning sentence 1: How to fix a TV that won't turn on?

- Study Sentence 2: TV power problem.

Intent 3: How to adjust the tint?

- Learning Sentence 1: Please tell me the steps to correct the color tone.

- Learning sentence 2: How to change the color level?

Intent 4: My TV turns off automatically.

- Study Sentence 1: What can be the solution if the TV turns off automatically?

- Study Sentence 2: Help the TV turn off automatically?

Intent 5: How to Install an App?

- Learning sentence 1: how application can be downloaded in tv?

- Learning Sentence 2: What would be the method of installing app?

Intent 6: hat is Easy Sharing?

- Learning Sentence 1: Need to understand about easy sharing.

- Study Sentence 2: What is use of easy sharing?

3 is an exemplary diagram for explaining the configuration of learning data according to another embodiment. In the illustrated embodiment, the intentions included in the learning data may include a plurality of learning sentences each having one of a first language and a second language. That is, in the present embodiment, there is a mixture of languages within the same intention.

<Example 2> below exemplifies a learning sentence in which language mixture exists within the intention. In <Example 2>, among the learning sentences included in the same intention (intent 1), 1 to 14 are Korean sentences, and 15 to 18 are English sentences.

<Example 2>

Intent 1: How to Install an App?

- Learning Sentence 1: I want to install netflix.

- Learning sentence 2: How to download Facebook on TV?

- Learning Sentence 3: I want to install an application on my TV.

- Learning Sentence 4: No Color app on your Samsung Smart TV?

- Learning Sentence 5: How to Download Netflix?

- Learning Sentence 6: Please tell me the fastest way to install the app.

- Learning sentence 7: Help find the hot star app on TV.

- Learning Sentence 8: How can I add more apps on my smart TV?

- Learning Sentence 9: The procedure for downloading the app on the television.

- Learning Sentence 10: Need to know how to install an application on your TV?

- Learning Sentence 11: Applications from the App Store. How to install HOOQ

- Learning Sentence 12: How do I install the color app?

- Learning Sentence 13: Where is the Hotstar app on the TV?

- Learning Sentence 14: The Hot Star app does not appear on the TV.

- Learning sentence 15: how application can be downloaded in tv?

- Learning sentence 16: What would be the method of installing app?

- Learning sentence 17: I am not able to view zee5 app in my samsung smart TV

- Learning Sentence 18: Want to install applications in my tv.

Returning to FIG. 1 again, the feature extractor 104 (featurizer) generates a feature vector corresponding to each of the one or more tokens generated by the token generator 102 . In an embodiment, the feature vector may include at least a weight of each token. In this case, the weight may include at least one of a word frequency (TF) or an inverse document frequency (IDF) of each token.

An intent classifier (106) constructs an NLU model by learning the feature vector generated by the feature extractor (104). Thereafter, the intent classifier 106 estimates the intent of the input natural language sentence using the NLU model.

4 is a block diagram for explaining the detailed configuration of the feature extractor 104 according to an embodiment in more detail. In the disclosed embodiments, the NLU system 100 performs learning using learning data composed of two different languages. If the proportion of the first language and the second language in the learning data is similar, there is no problem, but if a specific language occupies an excessively large proportion, an imbalance problem between languages may occur.

Accordingly, the feature extractor 104 according to an embodiment includes a configuration for resolving such a language imbalance problem. Specifically, the feature extractor 104 includes a learning data analysis module 402 , a balance analysis module 404 , and a weight adjustment module 406 .

The learning data analysis module 402 receives the learning data, and determines whether there is a mixed language between intentions of the learning data or within the same intention.

The balance analysis module 404 analyzes the balance between the first language and the second language in the learning data, that is, the ratio of the first language and the second language.

In one embodiment, the balance analysis module 404 may analyze the ratio of the first language to the second language for the entire learning data. Specifically, the balance analysis module 404 may determine whether the ratio of the second language token to the first language token in the learning data is less than or equal to a preset threshold.

In one embodiment, the balance analysis module 404 may analyze the ratio of the first language and the second language for a specific intention in the learning data. Specifically, the balance analysis module 404 may determine whether the ratio of the second language learning sentence to the first language learning sentence within a specific intention in the learning data is less than or equal to a preset threshold value.

The weight adjustment module 406 adjusts the weights of at least some of the one or more first language tokens and the one or more second language tokens generated from the training data based on the determination result of the training data analysis module 402 . In one or more embodiments, the weight adjustment module 406 may adjust the weight in consideration of the language-specific ratio analysis result of the balance analysis module 404 .

5 is a flowchart illustrating a learning data processing process 500 according to an embodiment. The illustrated method is performed in a computing device 12 having one or more processors and a memory storing one or more programs to be executed by the one or more processors, for example in the feature extractor 104 of the NLU system 100 described above. can be performed. In the illustrated flowchart, the method or process is divided into a plurality of steps, but at least some of the steps are performed in a different order, are performed in combination with other steps, are omitted, are performed in separate steps, or are shown. One or more steps not included may be added and performed.

In step 502, the learning data analysis module 402 receives the learning data, and determines whether there is a mixed language between intentions of the learning data or within the same intention.

Then, the weight adjustment module 406 adjusts the weights of at least one of the one or more first language tokens and the one or more second language tokens generated from the learning data based on the determination result. This will be described in more detail as follows.

As a result of the determination in step 502, if there is a mixture of language between intentions of the learning data, in step 504, the balance analysis module 404 determines whether the ratio of tokens for each language extracted from the learning data is less than or equal to a threshold value.

For example, when the number of first language tokens among the tokens for each language extracted from the learning data is greater than the number of second language tokens, the balance analysis module 404 is a second language token for the first language token in the learning data It is determined whether the ratio of is equal to or less than a preset threshold value. In one embodiment, the threshold value may be set to a value between 60% and 90%, for example. If the ratio of languages with a small number of tokens in the training data is less than the threshold value, it is because an imbalance problem may occur in the future intention classification.

If it is determined in step 504 that the token ratio for each language is less than or equal to the threshold, in step 506, the weight adjustment module 406 calculates an intention ratio between languages in the learning data. For example, in the training data such as <Example 1>, the number of intentions for Korean as the first language is 4 and the number of intentions for English as the second language is 2. Accordingly, in this case, the intention ratio between the first language and the second language may be 2:1.

In step 508 , the weight adjustment module 406 adjusts a weight for an intention having a high language occupancy in the training data. In this case, the weight may be an inverse document frequency (IDF). For example, in the case of training data such as <Example 1>, the frequency of inverse documents of tokens corresponding to Korean intentions with a high language share is adjusted. The high share of the language of Korean intent means that the frequency of inverse documents of Korean tokens is lower than that of English tokens (that is, the scarcity of English tokens is high). Accordingly, in order to correct this, the weight adjustment module 406 corrects the frequency of inverse documents of Korean tokens having a high occupancy.

Specifically, the weight adjustment module 406 adjusts the inverse document frequency of the first language token based on the ratio of the first language intent to the second language intent. For example, in <Example 1>, the intention ratio between Korean (first language) and English (second language) is 2:1. Accordingly, by reflecting this, the weight adjustment module 406 multiplies the frequency of inverse documents of tokens generated from learning sentences with Korean intent by a value of 2 corresponding to the intent ratio.

On the other hand, if it is determined in step 502 that there is a mixture of languages within the intent of the learning data, the balance analysis module 404 in step 510 determines whether the ratio of tokens for each language extracted within a specific intent among the learning data is less than or equal to a threshold value to judge

For example, if the number of first language tokens among the tokens for each language extracted within a specific intent of the learning data is greater than the number of second language tokens, the balance analysis module 404 for the first language token in the corresponding intent It is determined whether the ratio of the second language tokens is equal to or less than a preset threshold. In an embodiment, the threshold value may be set to a value between 60% and 90%. If the ratio of languages with a small number of tokens within the same intent is less than the threshold, an imbalance problem may occur in the future intent classification.

If it is determined in step 510 that the token ratio for each language within the same intent is less than or equal to the threshold, in step 512 the weight adjustment module 406 calculates the ratio of learning sentences between languages within the specific intent. For example, in the learning data such as <Example 2>, the number of sentences for learning Korean as the first language included in intention 1 is 14 and the number of sentences for learning English as the second language is 4. Therefore, in this case, the learning sentence ratio of the first language and the second language may be 1:0.28.

In step 514 , the weight adjustment module 406 adjusts a weight for the learning sentence having a low language occupancy within the specific intent. In this case, the weight may be a word frequency (TF). For example, in the case of learning data such as <Example 2>, the word frequency of tokens extracted from English learning sentences with a low language occupancy is adjusted. The low share of English sentences means that the word frequency of English tokens is lower than that of Korean tokens. Therefore, in order to correct this, the weight adjustment module 406 corrects the word frequency of the English token having a low occupancy.

Specifically, the weight adjustment module 406 adjusts the word frequency of the first language token based on the ratio of the first language learning sentence to the second language learning sentence. For example, in <Example 2>, the learning sentence ratio of Korean (first language) and English (second language) is 1:0.28. Accordingly, by reflecting this, the weight adjustment module 406 multiplies the inverse document frequency of the token generated in the English learning sentence by a value of 3.57 (=1 ÷ 0.28) corresponding to the intention ratio.

6 is a block diagram illustrating and describing a computing environment including a computing device suitable for use in example embodiments. In the illustrated embodiment, each component may have different functions and capabilities other than those described below, and may include additional components other than those described below.

The illustrated computing environment 10 includes a computing device 12 . In one embodiment, computing device 12 may be NLU system 100 according to one embodiment. Computing device 12 includes at least one processor 14 , computer readable storage medium 16 , and communication bus 18 . The processor 14 may cause the computing device 12 to operate in accordance with the exemplary embodiments discussed above. For example, the processor 14 may execute one or more programs stored in the computer-readable storage medium 16 . The one or more programs may include one or more computer-executable instructions that, when executed by the processor 14, configure the computing device 12 to perform operations in accordance with the exemplary embodiment. can be

Computer-readable storage medium 16 is configured to store computer-executable instructions or program code, program data, and/or other suitable form of information. The program 20 stored in the computer readable storage medium 16 includes a set of instructions executable by the processor 14 . In one embodiment, computer-readable storage medium 16 includes memory (volatile memory, such as random access memory, non-volatile memory, or a suitable combination thereof), one or more magnetic disk storage devices, optical disk storage devices, flash It may be memory devices, other forms of storage medium accessed by computing device 12 and capable of storing desired information, or a suitable combination thereof.

Communication bus 18 interconnects various other components of computing device 12 , including processor 14 and computer readable storage medium 16 .

Computing device 12 may also include one or more input/output interfaces 22 and one or more network communication interfaces 26 that provide interfaces for one or more input/output devices 24 . The input/output interface 22 and the network communication interface 26 are coupled to the communication bus 18 . Input/output device 24 may be coupled to other components of computing device 12 via input/output interface 22 . Exemplary input/output device 24 may include a pointing device (such as a mouse or trackpad), a keyboard, a touch input device (such as a touchpad or touchscreen), a voice or sound input device, various types of sensor devices, and/or imaging devices. input devices and/or output devices such as display devices, printers, speakers and/or network cards. The exemplary input/output device 24 may be included in the computing device 12 as a component constituting the computing device 12 , and may be connected to the computing device 102 as a separate device distinct from the computing device 12 . may be

Meanwhile, an embodiment of the present invention may include a program for performing the methods described in this specification on a computer, and a computer-readable recording medium including the program. The computer-readable recording medium may include program instructions, local data files, local data structures, etc. alone or in combination. The media may be specially designed and configured for the present invention, or may be commonly used in the field of computer software. Examples of computer-readable recording media include hard disks, magnetic media such as floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and program instructions specially configured to store and execute program instructions such as ROMs, RAMs, flash memories, etc. Hardware devices are included. Examples of the program may include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

Although the present invention has been described in detail through representative embodiments above, those of ordinary skill in the art to which the present invention pertains can make various modifications to the above-described embodiments without departing from the scope of the present invention. will understand Therefore, the scope of the present invention should not be limited to the described embodiments, but should be defined by the claims described below as well as the claims and equivalents.

100: NLU system
102: Token Generator
104: feature extractor
106: intent classifier
402: training data analysis module
404: balance analysis module
406: weight adjustment module

Claims (16)

one or more intentions and one or more learning sentences corresponding to the respective intentions, wherein the intention and the learning sentences receive learning data composed of one of a first language or a second language, and the intention of the learning data a learning data analysis module for determining whether there is a mixture of languages between or within the same intention; and
and a weight adjustment module configured to adjust weights of at least a portion of one or more first language tokens and one or more second language tokens generated from the learning data based on the determination result. The method according to claim 1,
wherein the weight includes at least one of a Term Frequency (TF) or an Inverse Document Frequency (IDF). The method according to claim 1,
As a result of the determination, if there is a mixture of language between intentions of the learning data, and the share of the first language intent in the learning data is higher than the share of the second language intent,
The weight adjustment module is configured to adjust a frequency of an inverse document of the first language token generated from a learning sentence corresponding to the first language intent. 4. The method according to claim 3,
Further comprising a balance analysis module for determining whether the ratio of the second language token to the first language token in the learning data is less than or equal to a preset threshold,
The weight adjustment module is configured to adjust the inverse document frequency of the first language token when the ratio of the second language token to the first language token in the learning data is less than or equal to a preset threshold as a result of the determination of the balance analysis module a natural language understanding system. 4. The method according to claim 3,
The weight adjustment module,
and adjust the inverse document frequency of the first language token based on a ratio of the first language intent to the second language intent. The method according to claim 1,
If, as a result of the determination, there is a mixture of languages within the specific intention of the learning data, and the share of the first language learning sentences in the specific intent is higher than the share of the second language learning sentences,
and the weight adjustment module adjusts the word frequency of the second language token generated from the second language learning sentence. 7. The method of claim 6,
Further comprising a balance analysis module for determining whether the ratio of the second language learning sentence to the first language learning sentence in the specific intention in the learning data is less than or equal to a preset threshold,
The weight adjustment module is, as a result of the determination of the balance analysis module, when the ratio of the second language learning sentence to the first language learning sentence in the specific intention is less than or equal to a preset threshold, the word frequency of the second language token Coordinated, natural language understanding system. 7. The method of claim 6,
The weight adjustment module,
and adjusting the word frequency of the second language token based on the ratio of the first language learning sentence to the second language learning sentence. receiving learning data including one or more intentions and one or more learning sentences corresponding to the respective intentions, wherein the intention and the learning sentences are configured in one of a first language or a second language;
determining whether there is a language mixture between intentions of the learning data or within the same intention; and
Comprising the step of adjusting the weight of at least one of the one or more first language tokens and one or more second language tokens generated from the learning data based on the determination result, the learning data processing method. 10. The method of claim 9,
Wherein the weight includes at least one of a word frequency (TF, Term Frequency) and an inverse document frequency (IDF). 10. The method of claim 9,
As a result of the determination, if there is a mixture of language between intentions of the learning data, and the share of the first language intent in the learning data is higher than the share of the second language intent,
The adjusting of the weight is configured to adjust a frequency of an inverse document of the first language token generated from a learning sentence corresponding to the first language intent. 12. The method of claim 11,
The adjusting of the weight is configured to adjust the inverse document frequency of the first language token when a ratio of the second language token to the first language token in the learning data is less than or equal to a preset threshold value. How data is processed. 12. The method of claim 11,
The step of adjusting the weight is
and adjust the inverse document frequency of the first language token based on a ratio of the first language intent to the second language intent. 10. The method of claim 9,
If, as a result of the determination, there is a mixture of languages within the specific intention of the learning data, and the share of the first language learning sentences in the specific intent is higher than the share of the second language learning sentences,
The adjusting the weight is configured to adjust the word frequency of the second language token generated from the second language learning sentence. 15. The method of claim 14,
The adjusting the weight is configured to adjust the word frequency of the second language token when the ratio of the second language learning sentence to the first language learning sentence in the specific intention is less than or equal to a preset threshold value, How to process training data. 15. The method of claim 14,
and the adjusting the weight is configured to adjust the word frequency of the second language token based on a ratio of the first language intent to the second language intent.

Images