System for modifying a source text into an output text, method for modifying a source text into an output text, and computer program product

Abstract

A system for modifying a source text into an output text is described, wherein a translation module is provided for converting the source text into a first intermediate text, which is connected to a content module, wherein the content module has a plurality of content agents, each content agent being trained to perform a specific modification task on the first intermediate text, wherein the content module is trained to generate a second intermediate text, wherein a verification module is provided which is trained to check the second intermediate text and, if necessary, modify it to create the output text, and wherein an output module is provided to output the output text.

Description

[0001] A system for modifying a source text into an output text, a method for modifying a source text into an output text and a computer program product are described.

[0002] Systems for modifying a source text into an output text, methods for modifying a source text into an output text and computer program products of the type mentioned above are known in the prior art.

[0003] Many texts—regardless of whether they are written by humans or generated automatically—are highly context-dependent. For example, the English word "plane" can be translated into German as "Flugzeug" or "Ebene," although these two terms have completely different meanings.

[0004] Translations or texts to be created or adapted for a specific subject matter must meet a wide range of different requirements, including accuracy, clarity, length, legal requirements, and, if applicable, corporate marketing guidelines. Some companies have precise guidelines regarding the use of certain terms.

[0005] When using large language models (LLMs), these texts currently translate context-independently, producing non-reproducible and unpredictable results that require extensive manual support for evaluation and, if necessary, adjustment or correction. This applies, for example, to translations of texts between different input and source languages, as well as the creation of specific texts such as instructions or marketing texts from existing text and, if necessary, image material in the same language.

[0006] Legal considerations such as product liability in connection with instructions are also not taken into account by such models. Specific language regulations cannot be easily integrated into common models. Political, company, or product-specific language regulations for customer addresses cannot be implemented by the existing language models.

[0007] The nature of large language models, neural networks or other machine translation techniques therefore does not achieve the desired or required results when translating specific texts, especially from certain very complex areas such as manuals, instructions or similar.

[0008] From EP 3 123 354 A1 a machine or computer-assisted translation system for translating a source text which transfers its meaning from one natural language to another is known, comprising software with a modular structure for organizing languages ​​and a transitory translation system which enables the creation of a multilingual system capable of performing translations in any direction between any of the included languages. The modular structure comprises a linguistic module of a dictionary of words and phrases, a linguistic module of a list of operational functions and parameters which control a conversion process required to perform a translation from one language to another, and an algorithm designed for rule-based machine or computer-assisted translation.

[0009] The task therefore arises of further developing systems for modifying a source text into an output text, methods for modifying a source text into an output text and computer program products of the type mentioned above in such a way that a better modification quality can be achieved when modifying texts.

[0010] The object is achieved by a system for modifying a source text into an output text according to claim 1, a method for modifying a source text into an output text according to the independent claim 6 and a computer program product according to the independent claim 10. Further embodiments and developments are the subject of the dependent claims.

[0011] A system for modifying a source text into an output text is described, wherein a translation module is provided for converting the source text into a first intermediate text, which translation module is connected to a content module, wherein the content module has a plurality of content agents, each content agent being designed to carry out a specific modification task on the first intermediate text, wherein the content module is designed to generate a second intermediate text, wherein a verification module is provided which is designed to verify the second intermediate text and, if necessary, to modify it in order to create the output text, wherein an output module is provided to output the output text.

[0012] The translation module can be a common, available translation module.

[0013] The content module can incorporate contextual knowledge and other requirements and modify the initial intermediate text accordingly. To do this, the content module can be configured to compare the source text with the initial intermediate text.

[0014] The system can be implemented in a computer, whereby the described modules translation module, content module, content agent and verification module can be implemented in one software and / or in several software modules.

[0015] The output module can be an acoustic and / or optical output module, for example, a loudspeaker or a screen. The output can be provided in parallel on multiple modules, for example, as speech from one or more loudspeakers and as text on one or more screens.

[0016] The modification of the source text into an output text can, for example, be a translation of a source text from one language to another and / or a conversion of the source text from one form to another, e.g., shortening a long text into a short text, converting a text requiring a high level of specialist knowledge into a text that is generally understandable, or converting a specialist text into a text suitable for marketing purposes. Such a conversion and a language translation can also occur simultaneously.

[0017] The content agents can have various specific tasks, e.g. a content agent can be provided which modifies the first intermediate text if necessary (i) with regard to orthography and syntactics, (ii) with regard to certain dictionaries or word ban lists, (iii) with regard to company-specific language regulations, (iv) with regard to market-specific requirements or (v) with regard to legal requirements for text forms and translations.

[0018] In addition, iteration loops can be avoided, which saves energy.

[0019] In a first further embodiment, it is provided that the verification module is followed by an evaluation module in which a user checks the output text, wherein the evaluation module is configured to start the content module to generate a modified second intermediate text again, taking into account the reasons for rejection, if the output text is rejected.

[0020] In this way, an improvement step can be incorporated that allows the content agents in the content module to be iteratively improved, so that over time, the modification results become increasingly better and more reliable, reducing the need for human revision. Furthermore, iteration loops in the system can be avoided, thus saving energy.

[0021] In a further refinement, it is provided that accepted or rejected output texts are stored in whole or in part in an example database with accepted or rejected example texts, wherein the content module is configured to regenerate a modified second intermediate text taking into account the example texts from the example database.

[0022] In this way, it is possible to train the content module and the content agents better over time and achieve better portability of the system.

[0023] In a further further embodiment, it is provided that the translation module and / or at least one of the content agents and / or the verification module have a large language model.

[0024] In this case, it is possible to describe the corresponding modification tasks in the form of prompts and to train the corresponding large language models more and more specifically over time, so that the output quality improves with increasing use.

[0025] In a further refinement, it is provided that at least one of the content agents can be selectively activated and deactivated.

[0026] This allows for more precise adaptation to the respective tasks, for example, if it is not necessary to replace certain generally understood terms with specific company-specific terms.

[0027] A first independent subject matter relates to a computer-implemented method for modifying a source text into an output text, wherein a translation module converts the source text into a first intermediate text and forwards it to a content module, wherein the content module has a plurality of content agents, wherein each content agent performs a specific modification task on the first intermediate text, wherein the content module generates a second intermediate text and forwards it to a verification module which verifies and optionally modifies the second intermediate text and creates the output text, wherein an output module outputs the output text.

[0028] In a first further embodiment, it is provided that the modification is a translation and / or reformulation.

[0029] In a further refinement, it is provided that the output text is checked before output, whereby if the output text is rejected, a modified second intermediate text is generated taking into account the reasons for rejection.

[0030] In a further further embodiment, it is provided that a context of the modification is specified to the translation module, wherein the translation module creates a first log file with the context and transmits it to the content module and / or that upon acceptance or rejection a second log file is created and transmitted to the content module.

[0031] When using large language models, the log file can be a prompt or be converted into a prompt.

[0032] A further independent subject matter relates to a computer program product comprising a computer-readable storage medium on which instructions are embedded which, when executed by at least one computing unit, cause the at least one computing unit to be configured to carry out the method of the aforementioned type.

[0033] The method can be carried out in a distributed manner on one or more computing units, so that certain method steps are carried out on one computing unit and other method steps are carried out on at least one further computing unit, whereby calculated data can be transmitted between the computing units if necessary.

[0034] Further advantages, features, and details will become apparent from the following description, which – where appropriate with reference to the drawings – describes at least one embodiment in detail. Identical, similar, and / or functionally equivalent parts are provided with the same reference numerals.

[0035] The single figure shows schematically a system 2 for modifying a source text T1 into an output text T4, T4'.

[0036] The system 2 comprises a translation module 4, a content module 6, a review module 8, an evaluation module 10, and an output module 12. A sample database 14 is also provided.

[0037] Translation module 4 can be a conventional translation module.

[0038] Content module 6 features specific content agents 6.1 to 6.6, each of which has specific tasks and is explained in more detail below. Other configurations may include more or fewer content agents.

[0039] The verification module 8 is designed to check the output of the content module 6 and modify it if necessary.

[0040] Evolution Module 10 allows interaction with a human.

[0041] The output module 12 may be an acoustic, visual or audiovisual output module.

[0042] A specific use case involves transmitting the source text T1 to the translation module 4. Upon input of the source text T1, a first log file 16.1 is created, which contains certain user specifications, e.g., the type of source text or the context and the task, for example, translation into another language and / or rewriting into a specific format.

[0043] Translation module 4 then generates a first intermediate text T2. This intermediate text T2 is a raw text that contains no specific context. If no translation from one language to another is required, the translation module prepares a basic adaptation of the text.

[0044] In the next step, a context-specific adaptation of the first intermediate text T2 takes place in content module 6 using content agents 6.1 to 6.6. These content agents 6.1 to 6.6 can each perform specific tasks, e.g., checking spelling and syntax, using specific dictionaries and word ban lists, implementing company-specific language regulations, market-specific text requirements, legal requirements, and the like. Depending on the respective task, content agents 6.1 to 6.6 can be activated and deactivated or modified using prompts based on the first log file 16.1.

[0045] The content module 6 thus generates a second intermediate text T3, which is passed to a verification module 8. The verification module 8 can check, based on the source text T1 and the first log file 16.1, whether the second intermediate text T3 meets the requirements or not, for example, whether it sufficiently corresponds to the source text T1 in terms of content. The verification module 8 can modify the second intermediate text T3 if necessary and generate the output text T4. If no modification is necessary, the output text T4 corresponds to the second intermediate text T3.

[0046] The second output text T4 is then passed to the evaluation module 10, where an end user receives the output text T4 along with explanations of the changes made to the first intermediate text T1. These explanations may include which of the content agents 6.1 to 6.6 made which changes. The end user then has the option of accepting or rejecting the output text T4 accordingly. Specific objections can be created, which are stored in a second log file 16.2.

[0047] Furthermore, all output texts T4 are stored in the example database 14 together with the respective status, in particular accepted or rejected.

[0048] If the output text T4 is rejected, the content module 6 is again tasked with modifying the first intermediate text T1, but this time taking into account the information from the second log file 16.2 and the now expanded sample database 14. The content module 6 creates a modified second intermediate text T3' from this, which is checked as previously described in the verification module 8 and, if necessary, converted into a modified output text T4'. The modified output text T4' is then reviewed in the evaluation module 10.

[0049] The end-user feedback loop can be repeated as often as necessary until the output text has been sufficiently modified to meet the requirements. Only after approval will the modified output text (T4') be used further.

[0050] Over time, the content agents 6.1 to 6.6 learn to perform the corresponding context-specific tasks better and better, so that the quality of the second intermediate texts T3 becomes increasingly better and the interaction of the end user is increasingly less necessary and can possibly be eliminated entirely.

[0051] Although the invention has been illustrated and explained in detail by means of preferred embodiments, the invention is not limited by the disclosed examples, and other variations may be derived therefrom by those skilled in the art without departing from the scope of the invention. It is therefore clear that a multitude of possible variations exist. It is also clear that the embodiments mentioned by way of example are truly only examples and should not be construed as limiting the scope, possible applications, or configuration of the invention in any way.Rather, the preceding description and the description of the figures enable the person skilled in the art to implement the exemplary embodiments in concrete terms, whereby the person skilled in the art, with knowledge of the disclosed inventive concept, can make various changes, for example with regard to the function or the arrangement of individual elements mentioned in an exemplary embodiment, without departing from the scope of protection defined by the claims and their legal equivalents, such as a further explanation in the description. List of reference symbols 2 systems 4 Translation module 6 Content module 6.1 - 6.6 Content agent 8 Verification module 10 Evaluation module 12 Output module 14 Example database 16.1 first log file 16.2 second log file T1 Source text T2 first intermediate text T3 second intermediate text T3' modified second intermediate text T4 output text T4' modified output text QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] EP 3 123 354 A1

[0008]

Claims

[1] System (2) for modifying a source text (T1) into an output text (T4, T4'), characterized by in that a translation module (4) is provided for converting the source text (T1) into a first intermediate text (T2), which is connected to a content module (6), wherein the content module (6) has a plurality of content agents (6.1 - 6.6), wherein each content agent (6.1 - 6.6) is designed to carry out a specific modification task on the first intermediate text (T2), wherein the content module (6) is designed to generate a second intermediate text (T3, T3'), wherein a checking module (8) is provided which is designed to check the second intermediate text (T3, T3') and, if necessary, to modify it in order to create the output text (T4, T4'), wherein an output module (12) is provided to output the output text (T4, T4'). [2] System (2) according to claim 1, characterized bythat the checking module (8) is followed by an evaluation module (10), in which a user checks the output text (T4, T4'), wherein the evaluation module (10) is configured to start the content module (6) to regenerate a modified second intermediate text (T3, T3') taking into account the reasons for rejection if the output text (T4, T4') is rejected. [3] System (2) according to claim 2, characterized by that accepted or rejected output texts (T4, T4') are stored in whole or in passages in an example database (16) with accepted or rejected example texts, wherein the content module (6) is set up to regenerate a modified second intermediate text (T3') taking into account the example texts from the example database (16). [4] System (2) according to one of the preceding claims, characterized bythat the translation module (4) and / or at least one of the content agents (6.1 - 6.6) and / or the verification module (8) have a large language model. [5] System (2) according to one of the preceding claims, characterized by that at least one of the content agents (6.1 - 6.6) can be selectively activated and deactivated. [6] Computer-implemented method for modifying a source text (T1) into an output text (T4, T4'), characterized byin that a translation module (4) converts the source text (T1) into a first intermediate text (T2) and passes it on to a content module (6), wherein the content module (6) has a plurality of content agents (6.1 - 6.6), wherein each content agent (6.1 - 6.6) carries out a specific modification task on the first intermediate text (T2), wherein the content module (6) generates a second intermediate text (T3, T3') and passes it on to a checking module (8) which checks the second intermediate text (T3, T3') and modifies it if necessary and creates the output text (T4, T4'), wherein an output module (12) outputs the output text (T4, T4'). [7] Computer-implemented method according to claim 6, characterized by that the modification is a translation and / or reformulation. [8] Computer-implemented method according to claim 6 or 7, characterized bythat the output text (T4, T4') is checked before output, whereby if the output text (T4, T4') is rejected, a modified second intermediate text is generated taking into account the reasons for rejection. [9] Computer-implemented method according to one of claims 6 to 8, characterized by that a context of the modification is specified to the translation module (4), wherein the translation module (4) creates a first log file (14.1) with the context and transmits it to the content module (6) and / or that upon acceptance or rejection a second log file (14.2) is created and transmitted to the content module (6). [10] Computer program product, comprising a computer-readable storage medium on which instructions are embedded which, when executed by at least one computing unit, cause the at least one computing unit to be adapted to carry out the method according to one of the preceding claims 6 to 9.

Images