Optimization system for technical documents

Abstract

The invention relates to a method of providing a technical document, comprising: a) Integrating an initial technical document into a computer implemented document optimization system by an integration unit of the computer implemented document optimization system, b) Preprocessing the technical document to generate a plurality of different text representations of the technical document by a preprocessing unit of the document optimization system, c) Indexing the plurality of different text representations using an indexing algorithm by an indexing unit of the document optimization system, d) Providing the indexed plurality of different text representations to generative artificial intelligence models to perform a text processing, e) Benchmarking the text processing of the generative artificial intelligence models by assessing a quality of answers of the generative artificial intelligence models by a benchmarking unit of the computer implemented document optimization system, f) Merging the plurality of different text representations into one single technical document by a merging unit of the computer implemented document optimization system.

Description

[0001] Description

[0002] Optimization system for technical documents

[0003] Users in any industrial domains have multiple challenges for which they can state questions, and the relevant information is available in form of multiple documents that need to be combined. Searching all these documents is cumbersome and inefficient.

[0004] As today’s documents are intended for human consumptions, they are not necessarily written in a form that automated artificial systems can interpret and correctly understand their content. Therefore, there is a need for augmented documents which are better processible, interpretable, and understandable to artificial intelligence-based systems.

[0005] Large Language Models (LLMs) are trained with manifold available text data. For specialized use cases either a finetuning with documents from relevant domains is executed in addition or only the relevant documents are defined with a prior retrieval search and used via prompt engineering (retrieval augmented generation). Applying these LLMs in a Generative Al application, correct answers based on the documents are expected. But if the technical content is not “understood” correctly by the LLM, the answers are incorrect. A typical example for a technical document comprises various structures (like tables and diagrams) as well as nomenclatures (like the FeatureBit explanation which requires to combine letters from a table in a specific way). All these aspects cannot be interpreted out-of-the-box by todays text or image embedding algorithms or generative Al models.

[0006] It is an object of the present invention to provide a method of providing a technical document and an according computer implemented document optimization system which increase the reliability and quality of an Al-based processing of the technical document in the field of a technical plant.

[0007] The problem is solved by a method of providing a technical document, comprising:

[0008] a) Integrating an initial technical document into a computer implemented document optimization system by an integration unit of the computer implemented document optimization system,

[0009] b) Preprocessing the technical document to generate a plurality of different text representations of the technical document by a preprocessing unit of the document optimization system, c) Indexing the plurality of different text representations using an indexing algorithm by an indexing unit of the document optimization system,

[0010] d) Providing the indexed plurality of different text representations to generative artificial intelligence models to perform a text processing,

[0011] e) Benchmarking the text processing of the generative artificial intelligence models by assessing a quality of answers of the generative artificial intelligence models by a benchmarking unit of the computer implemented document optimization system,

[0012] f) Merging the plurality of different text representations into one single technical document by a merging unit of the computer implemented document optimization system.

[0013] The technical document has a typical data format which is used in the course of industrial documentation. This means that the document has a defined structure and / or grammar and / or defined relations between the single elements of the document. An example for such a technical document is a technical documentation of a process control system like PCS neo of SIEMENS AG.

[0014] In a first step a typical technical document is integrated into a computer-implemented document optimization system which is done by integration unit if the optimization system. The optimization system is preferably implemented on a server of a process control system.

[0015] Afterwards, there is done a preprocessing of the technical document. Thereby, the technical document is processed to generate a plurality of different text representations of the technical document. Thereby, there can be done a chunking of the content of the technical document using different parameter settings.

[0016] In a following step, the plurality of different text representations is indexed using an indexing algorithm. Thereby, different text and image embeddings can be calculated. Index Algorithms can be ElasticSearch or vector databases.

[0017] Afterwards, the accordingly preprocessed document representations are provided to generative artificial intelligence models to perform a text processing. Thereby, a generative artificial intelligence (Al) model is a type of artificial intelligence designed to create new content. This can include text, images, music, and videos. The generative artificial intelligence models learn patterns and structures from existing data and use this knowledge to generate new, similar data. The generative artificial intelligence models in the course of the invention do text pro- cessing like text summarization, keyword generation or image caption generation. The generative artificial intelligence models preferably are multi-modal generative artificial intelligence models to better cope with the plurality of different text representations of the technical document.

[0018] In a next step, there is done a benchmarking of the text processing of the generative artificial intelligence models by the benchmarking unit of the computer implemented document optimization system. Thereby, a quality of the answers of the generative artificial intelligence models is assessed. The benchmark can be a predefined benchmark for technical questions answering. Different large language models can be used to compare expected and provided answers.

[0019] Finally, the plurality of different text representations is merged into one single technical document by a merging unit of the computer implemented document optimization system. The technical document is now ready for a subsequent usage by technical systems like a large language model based assistant.

[0020] According to one aspect of the invention, in the course of the benchmarking a Top-K / Bottom-K evaluation is done to integrate feedback of a human user into the benchmarking. During this evaluation, the K best and worst results from the benchmark are selected and provided to the human user I expert for a manual assessment. With the Top-K evaluation, the focus is on the top K items in a ranked list. This method is commonly used in recommender systems to measure how well the system predicts the most relevant items for a user. With the Bottom-K evaluation, on the other hand, there is looked at the bottom K items in a ranked list. This can be useful for identifying items that are consistently ranked poorly and may need improvement. The human expert can provide additional labels (like providing additional questions / answer-pairs using the document-specific technical terms) or explanation of certain terms with definitions. The human feedback can afterwards be integrated into the benchmark for future document optimizations.

[0021] Preferably, after the step of benchmarking there is added a glossary of terms for explaining technical terms of the technical document. Thereby, the glossary is a list of terms and their definitions, as it can usually be found at the end of a book or technical document. It is designed to help to understand specialized or technical language used in the text. Additionally, an ambiguity of vocabulary can be reduced. Thereby, an ambiguity of vocabulary can occur when a word or phrase has multiple meanings, making it unclear which interpretation is intended in the context of the technical documentation. This can lead to misunderstandings or confusion in understanding the content of the technical document. Most preferably, after the step of benchmarking there is added an annotation of rows of tables and / or an annotation of images of the technical document.

[0022] The technical document can be translated into another language during the processing by the computer implemented document optimization system.

[0023] According to one aspect of the invention, the technical document, after having been processed by the computer implemented document optimization system, is used by another generative artificial intelligence model to get a technical command for a technical system. Such a technical system can be e.g., a control system for a technical plant, preferably a process control system for a process plant.

[0024] Preferably, the technical command is executed by the technical system, most preferably by a process control system.

[0025] The problem as described above is also solved by a computer implemented document optimization system, comprising:

[0026] a) An integration unit, designed for integrating an initial technical document into the computer implemented document optimization system,

[0027] b) A preprocessing unit, designed for generating a plurality of different text representations of the technical document,

[0028] c) An indexing unit, designed for indexing the plurality of different text representations using an indexing algorithm,

[0029] d) A benchmarking unit, designed for benchmarking the text processing of the generative artificial intelligence models by assessing a quality of answers of the generative artificial intelligence models,

[0030] e) A merging unit, designed for merging the plurality of different text representations into one single technical document.

[0031] The previously described properties, features, and advantages of this invention, as well as the manner in which they are achieved, become clearer and more distinctly comprehensible in the context of the following description of embodiments, which will be further explained in connection with the drawings. It shows:

[0032] FIG 1 a schematic view of a method according to the invention; and FIG 2 a schematic view of a process control system.

[0033] In FIG 1, there is illustrated a method according to the invention. In a first step I, an initial technical document 1 like a manual of a process control system is integrated into a computer implemented document optimization system 2. In a subsequent step II, the integrated document 3 is preprocessed to generate a plurality of different text representations 4 of the integrated technical document 3. In a further step III, the plurality of different text representations 4 is indexed using an indexing algorithm. The indexed plurality of different text representations 5 is, in a subsequent step IV, is provided to multi-modal generative artificial intelligence models to perform a text processing on the indexed plurality of different text representations 5.

[0034] In a further step V, the text processing 6 of the generative artificial intelligence models is bench-marked by assessing a quality of answers of the generative artificial intelligence models. In the course of the benchmark a Top-K / Bottom-K evaluation is done. Thereby, feedback of a human user 7 is integrated into the benchmarking. To the benchmarked text processing 8 of the generative artificial intelligence models, there is added an annotation of rows and tables and / or an annotation of images of the plurality of different representations of the technical document 1.

[0035] In a further step VI, the plurality of different representations of the technical document 1 is merged into one single document 9. The merged technical document 9 can be translated into an arbitrary language.

[0036] After having been processed by the computer implemented document optimization system 2, the technical document 9 can be used by another generative artificial intelligence model to get a technical command for a technical system like a process control system for a process plant. This technical command can be executed by the technical system afterwards.

[0037] FIG 2 shows a process control system 10 for a process plant. The process control system 10 comprises an operator station server 11 and an operator station client 12. The operator station server 11 is connected to the operator station client via a bus infrastructure 13 and transmits data for control and command the process plant to the operator station client 12, and vice versa. The operator station server 11 is connected to an automatization device 14 which is connected to a plurality of process devices 15a, 15b, 15c, 15d, 15e like sensors or actors. The operator station server 11 comprises a computer implemented document optimization system 2 as it is used according to FIG 1. The computer implemented document optimization system 2 comprises an integration unit 16, designed for integrating an initial technical document 1 into the computer implemented document optimization system 2. Furthermore, the computer implemented document optimization system 2 comprises a preprocessing unit 17, designed for generating a plurality of different text representations of the technical document 1. Furthermore, the computer implemented document optimization system 2 comprises an indexing unit 18, designed for indexing the plurality of different text representations using an indexing algorithm.

[0038] Furthermore, the computer implemented document optimization system 2 comprises a generative artificial intelligence model unit 19, designed for performing a text processing on the different text representations. Furthermore, the computer implemented document optimization system 2 comprises a benchmarking unit 20, designed for benchmarking the text processing of the generative artificial intelligence models by assessing a quality of answers of the generative artificial intelligence models. Furthermore, the computer implemented document optimization system 2 comprises a merging unit 21 , designed for merging the plurality of different text representations into one single technical document 9. List of Reference

[0039] 1 Initial technical document

[0040] 2 Computer implemented document optimization system 3 Integrated document

[0041] 4 Plurality of different text representations

[0042] 5 Indexed plurality of different text representations

[0043] 6 Text processing

[0044] 7 Human user

[0045] 8 Benchmarked text processing

[0046] 9 Merged technical document

[0047] 10 Process control system

[0048] 11 Operator Station Server

[0049] 12 Operator Station Client

[0050] 13 Bus infrastructure

[0051] 14 Automation device

[0052] 15a Sensor or Actor

[0053] 15b Sensor or Actor

[0054] 15c Sensor or Actor

[0055] 15d Sensor or Actor

[0056] 15e Sensor or Actor

[0057] 16 Integration unit

[0058] 17 Preprocessing unit

[0059] 18 Indexing unit

[0060] 19 Generative artificial intelligence model unit

[0061] 20 Benchmarking unit

[0062] 21 Merging unit

Claims

Patent claims1. Method of providing a technical document, comprising:a) Integrating an initial technical document into a computer implemented document optimization system by an integration unit of the computer implemented document optimization system,b) Preprocessing the technical document to generate a plurality of different text representations of the technical document by a preprocessing unit of the document optimization system, c) Indexing the plurality of different text representations using an indexing algorithm by an indexing unit of the document optimization system,d) Providing the indexed plurality of different text representations to generative artificial intelligence models to perform a text processing,e) Benchmarking the text processing of the generative artificial intelligence models by assessing a quality of answers of the generative artificial intelligence models by a benchmarking unit of the computer implemented document optimization system,f) Merging the plurality of different text representations into one single technical document by a merging unit of the computer implemented document optimization system.

2. Method according to claim 1, wherein in the course of the benchmarking a Top-K / Bottom-K evaluation is done to integrate feedback of a human user into the benchmarking.

3. Method according to claim 1 or 2, wherein after the step of benchmarking there is added a glossary of terms for explaining technical terms of the technical document.

4. Method according to any of the preceding claims, wherein after the step of benchmarking there is added an annotation of rows of tables and / or an annotation of images of the plurality of different representations of the technical document.

5. Method according to any of the preceding claims, wherein the technical document is translated into another language during the processing by the computer implemented document optimization system.

6. Method according to any of the preceding claims, wherein the generative artificial intelligence models are multi-modal generative artificial intelligence models.

7. Method according to any of the preceding claims, wherein the technical document, after having been processed by the computer implemented document optimization system, is used by another generative artificial intelligence model to get a technical command for a technical system.

8. Method according to claim 7, wherein the technical command is executed by the technical system.

9. Method according to claim 7 or 8, wherein the technical system is a control system of a technical plant.

10. Computer implemented document optimization system, comprisinga) An integration unit, designed for integrating an initial technical document into the computer implemented document optimization system,b) A preprocessing unit, designed for generating a plurality of different text representations of the technical document,c) An indexing unit, designed for indexing the plurality of different text representations using an indexing algorithm,d) A generative artificial intelligence model unit, designed for performing a text processing on the different text representations,e) A benchmarking unit, designed for benchmarking the text processing of the generative artificial intelligence models by assessing a quality of answers of the generative artificial intelligence models,f) A merging unit, designed for merging the plurality of different text representations into one single technical document.

Images