A method of connecting multilingual translation module / s with large language model / s for optimized output generation and output translation of user inputs in user preferred language

Abstract

The invention discloses a method of connecting multilingual translation module / s with large language model / s for optimized output generation and output translation of user inputs in the user preferred language. This method is carried out by a worker module which is connected to large language model / s and third-party translation module / s to provide seamless translation and text generation capabilities for a user input. Existing machine translation systems and text generation models have limitations in handling multiple languages and integrating translation with complex text generation. The proposed invention addresses these limitations by connecting and integrating user input translation, text generation, and subsequent translation of the output to the user's original language. The invention offers a more coherent and contextually relevant interaction for users who prefer to communicate in their native language with generative text AI by bridging the gap between translation and text generation.

Description

[0001] A method of connecting multilingual translation module / s with large language model / s for optimized output generation and output translation of user inputs in user preferred languageTechnical Field

[0002] The invention relates to artificial Intelligence, more particularly language processing and translations using neural networksBackground Art

[0003] Machine translation systems have long been used to translate text from one language to another. Existing machine translation systems produce translations in a variety of languages. There are also systems that utilize neural machine translation models, which are trained on large bilingual text corpora, to generate translations. However, these systems typically focus solely on translation and do not integrate other language processing tasks such as text generation. For instance, Google's neural machine translation system and Facebook's systems and methods for neural machine translation are examples of standalone translation technologies that use neural networks. Yet, they are primarily dedicated to translation and do not integrate language generation in their scope.

[0004] On the other hand, the field of natural language processing has seen significant advancements in text generation models. Large language models like GPT, BERT, and others use large-scale machine learning to generate text that is contextually relevant and grammatically correct based on an input prompt. However, these systems have generally been designed to operate in a single language and do not include integrated translation capabilities. The translation capabilities of these models are sub-par as the translation capabilities are restricted to language processing alone.

[0005] They are primarily been designed to operate in a single language, usually English, and do not include integrated translation capabilities. Issues arise when these two distinct systems - translation and text generation - interact. Forinstance, in the case of GPT, the original model was trained primarily on English text, limiting its effectiveness in understanding, and generating text in other languages as the GPT model is not trained to effectively handle various other languages.

[0006] Furthermore, some systems have attempted to integrate translation and text generation. However, these systems typically involve distinct, separate steps - first translating the input, then generating text based on the translation. Such systems do not include a mechanism for converting the generated text back into the user's original language. Users who don't understand English would not be able to comprehend the output.

[0007] The present invention aims to overcome these limitations by integrating a worker module. It distinguishes itself from the prior art by providing a seamless integration of translation and text generation by connecting the user input translation - connecting inputs to a neural network for text generation and translation of the output to user’s language. It is capable of handling a corpus of texts more than 100 languages. It utilizes a unique worker module to prepare translated inputs for the language model and to convert the model's output back into the user's original language, providing a more coherent and contextually relevant interaction for users who wish to interface with the system in their native language.

[0008] On the other hand, Open Al's GPT represents an advanced language model focusing on text generation based on an input prompt. However, it operates primarily in a single language and lacks an integrated translation module.

[0009] Microsoft's multilingual language model processes multiple languages using a single language model but doesn't specify a workflow that involves translation of input, text generation, and subsequent translation of output, as does our patent application.

[0010] In the realm of chat applications, IBM's dynamic language translation carries out real-time translation, but this patent is more about facilitating multilingual chat and does not concern itself with Al-driven text generation.

[0011] The current invention distinguishes itself from the prior art by introducing an integrated system that links a language translation module with a transformer-based language model, supplemented by a translation worker module. This setup facilitates text translation from the user's source language to a target language, contextually relevant text generation based on the translated input, and subsequent translation of the generated text back to the user's source language. This results in a more coherent and user-friendly interaction.

[0012] In contrast, machine translation technologies, such as Google's neural machine translation system and Facebook's systems and methods for neural machine translation, translate text between languages. These systems usually utilize Sequence-to-Sequence (Seq2Seq) models that are trained on large bilingual corpora. While highly effective at translation, these models don't exhibit the creative text generation capability of models like GPT. Furthermore, these patents do not articulate the integration of machine translation with complex text generation.

[0013] IBM's interactive multi-lingual conversation agent and Microsoft's multilingual language model are attempts to bridge this gap. IBM's patent proposes an agent capable of conversing in multiple languages, and Microsoft's application describes a single model processing multiple language. However, these approaches have certain limitations. IBM's patent is more focused on maintaining a conversation in various languages rather than generating creative text in response to a prompt. Meanwhile, Microsoft's patent application doesn't specify a mechanism for back-translating generated text into the user's original language.

[0014] IBM's dynamic language translation in a chat application introduces real-time translation in a chat context, which is a step towards bridging communication gaps across languages. Yet, this patent focuses on translating existing messages rather than generating novel text and does not incorporate Al-driven text generation.

[0015] The proposed invention departs from these existing technologies by integrating the functionalities of language translation and creative text generation into a single, streamlined system. It not only includes a language translation module and a large language model but also introduces a unique mark down system which token izes and formats text which creates a more coherence output. This design ensures efficient text translation from the user's source language to atarget language, context-aware text generation based on the translated input, and subsequent translation of the generated text back to the user's original language.

[0016] It is crucial to note that a system like this would have the potential to greatly improve user interaction and accessibility for non-English speakers, addressing a significant gap in current Al-driven communication technologies.

[0017] Furthermore, the transformer based neural network models have been trained on large amounts of text data and are able to generate coherent and fluent text in response to a given prompt. This has applications in a wide range of areas, including chatbots, language translation, and text summarization. The use of transformer based neural network models has improved the quality of automated text generation and made it possible to generate more accurate and useful explanations. It has the capability to carry out complex language processing based on the user requirements. It can generate human-like responses to textbased inputs. These models use a deep neural network architecture and has been trained on large amounts of text data to generate coherent and contextually appropriate responses.

[0018] There have been numerous prior art references related to neural networking, including the foundational work of McCulloch and Pitts, the introduction of perceptrons by Rosenblatt, the development of backpropagation by LeCun et al. These scholarly works have materialized into GPT models and other neural networks today. However, these works focus primarily on transformer-based architecture and not multilingual accessibility of such architecture.

[0019] These neural networking patterns that exist are focused most only the data processing and output generation aspects and are not focused on the making the outputs accessible to non-native English speakers. Most of the data bases used for the pre-training is based off English data bases and non-English data sets used for analysis purposes are not as effectively run by the training models due to issues with language structure of the non-native languages. This problem is significantly more prevalent in languages that are not spoken much in the world such as Sinhala, Hindu, Tamil and other languages which are used by a small minority of countries, which are not mainstream languages in the global arena unlike languages such as English, Chinese, French and German.

[0020] In the instant method the translation and text generation are connected by a ‘Worker’, in which a third-party translation module. The system is set up such a way where the system can connect to multiple third-party translation modules for the best outcome. The worker module can direct the translated text to multiple transformer-based neural networks. Currently, the ‘worker’ has been connected to the Open Al GPT architecture and the LLaMA language model (Large Language Model Meta). These databases are trained in extremely large corpus of data which can be fine-tuned to specific applications. This database contains a vast amount of text data from different sources, such as web pages, books, and news articles, which have been cleaned and preprocessed to ensure high quality. The text data is then used to train language models that can be used for various natural language processing tasksSummary of Invention

[0021] The claimed invention is an Al supported translation module, which gives answers in the preferred language. The gpt interface receives the query from the user in the user preferred language and translated it to the common language before answering the question. Then it finds most appropriate answer before translating them in the user preferred language. The selected answer is translated to the user preferred language at last.Technical Problem

[0022] The field of natural language processing has seen a convergence of two main areas: machine translation and text generation, with large language models at the heart of this. Open Ai’s and similar large language models are powerful text generators but have been primarily trained and optimized for English. As such, they may not effectively understand or generate text in other languages.

[0023] As highlighted above there is a lack of efficient multilingual language processing applications that use neural networking or any artificial intelligence data processing and output generators. Current language processing applications such as GPT, Bard and LLaMA has the ability to respond to a different language, but the responses are quite ineffective (figure 5- 9) because data search is done in the language the input is entered into. As most of the data bases are currently available and is trained are in English, answers generated are of very poor inquality as the answers are limited to the data base that is available in that language alone. The training models only search for data available in that particular language, if the data is not available in that particular language, then the training models are unable to respond effectively to the inputs.Solution to Problem

[0024] This method seamlessly integrates translation and text generation. This integration involves connecting the translated user input to a neural network for text generation and then translating the output back into the user's language. It can handle a corpus of texts in more than 100 languages, employing a unique translation worker module to prepare translated inputs for the language model and convert the model's output back into the user's original language and format using tokenization and via mark down process to provide coherent output to the user. This approach ensures a more coherent and contextually relevant interaction for users desiring to interface with the system in their native languageAdvantageous Effects of Invention

[0025] Most important advantage of this method is that regardless of the input language, the output will remain consistent, which makes Artificial intelligence language processing available throughout the world and not restricted to English speaking countriesBrief Description of Drawings

[0026] The present patent application is further elucidated by the accompanying figures / drawings, which are provided to enhance the understanding of the disclosed subject matter. These drawings are an integral part of this specification and showcase practical implementations of the disclosed subject matter. Together with the detailed description, the drawings serve to illustrate the underlying principles and embodiments of the disclosed subject matter

[0027] [Fig.1 ] Discloses an overview of the process of receiving user input by this method of connecting multilingual translation module with a transformer based neural network for optimized output generation and the output generated by the transformer based neural network is given in the user’s preferred language.

[0028] [Fig.2] Discloses the details of the worker module which is the key aspect of this patent application

[0029] [Fig.3] Discloses an overview of the key aspects of this method, including input prompts by user being translated into common language, and the output generated by the transformer based neural network being translated back into user’s preferred language

[0030] [Fig.4] Illustrates a comparison between the response (output) generated by Chat GPT, large language model for an input prompt in a different language (Sinhala) and the response generated by this model. Explanation: In figure 4 the response by the Chat GPT model is nonsensical as the current transformerbased is incapable of understanding inputs in other languages (non-English) as the large language model is not properly trained in data from that particular language

[0031] [Fig.5] Illustrates a comparison between the response generated by Chat GPT, large language model for an input prompt in a different language (Sinhala) and the response generated by this model. It further illustrates the response generated by the Chat GPT language model, where the input prompt is in English, which is the main language the language model is trained in. Explanation: In figure 5, as in the above figure 4, the response by the Chat GPT Open Al model does not answer the input prompt of the user as the current transformer-based large language models are incapable of understanding inputs in other languages (non-English) as the transformer based language model is not properly trained in data from that particular language. This is evident from the Chat GPT Open Al model’s response to the same question in English being vastly better and more similar to the response generated by this model in the user’s preferred language

[0032] [Fig.6] Illustrates a comparison between the response generated by Open Al GPT, large language model for an input prompt in a different language (Hindi) and the response generated by this model. Explanation: In figure 6, the user’s prompt is ‘write a brief summary of the cold war’ in HindiHowever, as clearly visible from the translated answers (using Google translate) shown in the figure 6, the answer / output generated by Open Al does not respondto the user’s input at all due to Open Al’s inability to effectively understand prompts in different languages

[0033] [Fig.6.2] Illustrates a comparison between the response generated by Bert Google Al, large language model for an input prompt in a different language (Hindi) and the response generated by this model. In the above image, it clearly shows that Bard Al has not been trained to be multilingual.

[0034] [Fig.7]lllustrates a comparison between the response generated by Open Al, large language model for mathematical input prompt / input prompt which require a mathematical formula as a response. Explanation: In Figure 7, the Chat GPT Open Al responds to a user prompt which require a mathematical formular as an answer by providing the mathematical formula in a text format. In contrast, the current method, uses a mark down process which is integrated into the worker module, to format the mathematical equations into math Jax formula for a much more accurate and interactive response for a user’s prompt.Advantageous Effects

[0035] Most important advantage of this method is that regardless of the input language, the output will remain consistent, which makes Artificial intelligence language processing available throughout the world and not restricted to English speaking countries.Mode for invention

[0036] In this invention, the Large Language model has been integrated into a translation model connected by a worker module which ensure seamless connection between the two models. The invention provides a simple interface for user input prompts in any language to be translated into the mostly commonly used language by the large language model, processed and translated back to user’s language in a seamless coherent manner.

[0037] The translation models connected to the worker allows this method of connecting multilingual translation module with a transformer based neural network for optimized output generation in user’s preferred language to accept user prompts in over 100 different languages.

[0038] Users are allowed to choose the language of input out of 100 different languages and submit the input to this method using the interface provided’. Input by the user is translated into a common language (English or any other such language) by a Worker module which is connected to a third-party translation module which uses neural networking for translation for the best translations. (Input by the user can be translated into several languages which have been used to train the data in the Large Language Models. However, currently it is only being translated into English as English is the mostly commonly used language in the data sets available in large language models).

[0039] Worker converts the translated user inputs to enhance quality and compatibility with large language models. Translated prompt is directed to a large language model / s using an application programming interface (API call). The worker calls the API of a third-party large language model . The parameters and context of the conversation are submitted in the request by the way of headers and a request body.

[0040] The worker module then translates the response (concatenated response if required) by the large language model model to the user’s preferred language.

[0041] The worker module uses a markdown process to format the application and to convert the output from the large language model using tokenizing and formatting parameters to improve the final output that is accessed by the users.

[0042] Prior to translating the response, the worker caches the response in the common language (English), which can be directly accessed by the worker if similar / same question is asked by a user, which by passes the need to access large language models / s

[0043] he unique feature of this method of language processing is that it streamlined the manner in which inputs and outputs in different user preferred languages interact. As set out in figure 1 , this application has a ‘Worker’ which is a training model inbuilt into the operational structure to translate inputs from any language into a common language (at the moment English) and access trained data bases from that particular language. This worker is trained based off multilingual language database, which makes translations more accurate, and the database is consistently being updated.

[0044] The training models included in the method also has the ability to translate all available data that is used to train the model into a specific language (English) to ensure that the entire neural network, regardless on the language of the data base is stored in is converted into English. This allows answers to be generated regardless of the input language in English. Afterwards, as explained in the 2ndpart of Figure 2, the output is re-translated into the language of the user / input by a worker, then presented to the user. This enables all outputs generated to be consistent across all languages as the output is generated in the common language (English).

[0045] Furthermore, the creators can feed data at large and small scale and change the search priority of the data to the creator inserted data, which allows data processing to be improved based on the user’s requirements.

[0046] This method of connecting multilingual translation module with a transformer based neural network for optimized output generation makes Al more accessible to all people across the world.

[0047] 1ststep: the input: the input prompt by the user, which can be entered in whatever language user prefers is translated to a common language (English) by a ‘Worker’ module, refer to figure 1 , part (i). This allows the user to ask questions in whatever language and for the language model to understand and process the question in the common language (presently English). Translated input prompt is directed towards a transformer based neural network using an (Application programming interface (API call), which connects the said method to the transformer based neural network which initiates a response from the neural network.

[0048] 2ndStep: the output: After accessing the transformer based neural networks in English (which is the language vast majority of currently available data is available in - also referred to as the common language) the output generated for the user’s input prompt is directed back to the worker module. The translated output is directed back into the user’s preferred language by a ‘Worker’ module (refer to figure 1 , part(ii)) and produced to the user.

[0049] 3rdStep: If the user has followed up questions in his native language, the output which was initially translated into the user’s language is cached in thecommon language to improve processing ability. User’s new input prompt will be translated by a ‘Worker’ as explained in step one above. Afterwards, initial response saved in the cache in the common language (English) and the translated new prompt is processed via the data base and an output is generated. This process is repeated every time and input, or a follow-up input is submitted by the user.

[0050] Every time a different user inputs the same / similar prompts in a different language, the worker will translate the input prompt into the common language (English) and first search the cached responses, if the cached responses provide an answer, the response will be directly sent to the user bypassing the need to direct the translated user input through a transformer based neural network.Industrial ApplicabilityThis invention has the capability to make current large language model / s and Al systems more accessible to user from various countries who do not speak / use mainstream languages. This can be used in industries such as e-commerce, finance, telecommunications, and others, to provide instant customer support in multitude of languages. This method also can be used by content creators to generate text for various purposes such as social media posts, product descriptions, and news articles. This can help in reducing the time and effort required for content creation, especially for businesses that require a large amount of content on a regular basis in the native language of the creator. This model can be used in research to analyze large volumes of text data and extract insights in the English language and generate an efficient output in the language chosen by the user, regardless of the input language.

Claims

Claims1. A method for connecting multilingual translation module / s with large language model / s for optimized output generation and output translation of user inputs in the user preferred language, into user preferred language comprising: a forward processing of user inputs using a worker module connected to a transformer based language model / s and a third party translation module ; a backward processing of output translations using the worker module connected to a transformer based language model / s and the third-party translation module and caching and retrieving the user inputs and translated outputs processed using the forward processing of inputs and backward processing of output translations.

2. The method for connecting multilingual translation modules with large language models as claimed in claim 01 wherein, the forward processing of user inputs using a worker module connected to a transformer based language model / s further comprises the following steps: receiving input in a user-selected language; directing the input into the forward processing worker module box using connecting procedure; translating the input into a common language using a third-party translation module ;Pre-processing the translated input for compatibility with a large language model / s based on pre-programmed parameters; directing the translated input to a large language model / s which is trained in a large corpus of data for language processing and generating contextually relevant output based on the translated input prompt using a language generation model.

3. The method for connecting multilingual translation modules with large language models as claimed in claim 01 wherein, the backward processing of output translations using the worker module connected to a large language model further comprises the steps of:directing the output generated by the large language model / s to the worker module via connecting procedure; combining each of the several translated outputs via a concatenate formula with integrated parameters incorporated into the worker module for the best possible answer; tokenizing the output generated by the large language model / s or concatenated output to format the response by segmenting the output based on the formatting parameters pre-programmed onto the worker module, wherein the output generated by the large language model is connected to the worker module, which acts as a tokenizer, where in the output is received as tokens using a tokenizer (figure 2, 205), transforming the tokens into a string array ; translating the output generated by the large language model / s or concatenated output back to the user-selected language using a third- party translation module which is integrated into the worker module using connecting procedure, wherein the tokenizer ensures that the output is appropriately segmented for processing and the tokenized output is transformed into a string array and the string array is sent through a translation module by the worker wherein the output in the common language is translated into user’s selected language and delivering the translated output text to the user in the user preferred language.

4. The method for connecting multilingual translation modules with large language models as claimed in claim 1 wherein, Caching and retrieving of user inputs and output translations comprises the steps of :Caching of the output generated by the large language model / s in the common language by the worker module prior to translating the output;Caching of User inputs at the forward processing worker module via integrated connecting procedure and retrieving of user inputs and output translations by comparing the user input and relevant output translations.

5. A method of translation with user input with a forward processing worker module comprising: receiving input prompt in a user-selected language;directing the input into a forward processing worker module using connecting procedure; translating the input text into multiple languages using a third-party translation module ; preprocessing the translated input text for compatibility with a large language model based on pre-programmed parameters; directing the each of the separate translated text inputs to the worker module, whereas, translated texts are directed to the large language model / s trained in a large corpus of data for language processing and generating contextually relevant output text based on the translated input prompt using the language generation model.

6. A method of translation with output translations using backward processing worker module comprising; directing each of the outputs generated by the large language model to the worker module via connecting procedure; tokenizing each of the outputs generated by the large language model in each respective language to format the response by segmenting the output based on the formatting parameters preprogrammed transforming the tokens into a string array; translating the output generated by the large language model back into the user-selected language wherein the tokenizer ensures that the output is appropriately segmented for processing and the tokenized output is transformed into a string array and the string array is sent through a translation module by the worker where in the output in the common language / s is translated into user’s selected language; combining each of the several translated outputs via a concatenate formula with integrated parameters incorporated into the worker module for the best possible answer and delivering the translated and concatenated output text to the user in the user preferred language.

7. A system for translating user inputs into a common language for processing such inputs for output generation via a large language model and translating the output into user preferred language comprising : a forward processing worker module connected to one or more largelanguage model / s and a translation module, which receives inputs and processes the inputs by translating them into a common language and directs the translated inputs to a large language model; a backward processing worker module with an integrated tokenizer connected to the said translation module and to one or more large language models; a user interface for receiving inputs and delivering translated outputs.

8. The system as claimed in claim 7 wherein the forward processing worker module comprises of: a input receiving prompt to receive inputs in a preferred language; a forward processing worker module connected to the said input receiving prompt; an integrated third-party translation module; a transformer based large language model connected to the said translation module and the worker module and a language generation model which redirects the translated input to transformer-based language model / s , to generate contextually relevant output text based on the translated inputs.

9. The system as claimed in claim 7, wherein the backward processing worker module comprising of: a transformer based language model and language generation model to receive outputs, combine such outputs; integrated parameters for generating the best possible responses; a tokeniser which tokenizes the output generated by said transformer-based language model; an integrated third-party translation module and a user interface which delivers the translated output text to the user in their preferred language.

10. The system as claimed in claim 7 wherein the backward processing worker module further comprising a caching storage which incorporates with translation module and outputs , enabling faster response generation for similar or identical inputs

Images