INTELLIGENCE ARTIFICIELLE ET LANGAGES
Innovation
Our technological breakthrough, and why it matters
The entire OWI project is based on a groundbreaking discovery: a unique language modeling approach that allows for the seamless connection of all signs (words) just as humans naturally do. This enables the extraction of all information contained within a message.
This discovery, developed between 1992 and 2013, brings an unparalleled level of precision and reliability compared to current approaches.
When it comes to automatically analyzing emails or providing telephone assistance, this precision and reliability are not just advantages; they fundamentally change what can be automated, thus impacting ROI and enhancing the quality of the customer experience.
Language processing by OWI
Interestingly, almost all approaches to natural language processing, including GPT, have overlooked a fundamental principle of linguistics (Ferdinand de Saussure): the understanding of language involves two distinct steps:
- The production of signs (recognition of “signifiers” in linguistic terms), which depends solely on the language.
- The interpretation of these signs (moving from “signifiers” to “signifieds”), which depends on the context of the message.
The future of human language processing necessarily lies in AI that closely resembles how humans process language. This means utilizing two artificial intelligences that collaborate with each other:
- Linguistic AI capable of producing, from a message, all potentially meaningful signs.
- The principles of the linguistic AI model are the same for all languages, but each language requires a specific linguistic model.
- This model is referred to as the “dictionary” in OWI.
- Real-world knowledge AI (referred to as business AI in OWI) capable of interpreting the signs produced by the linguistic AI into actionable information.
- The principles of the business AI model are the same for all sectors and companies, but each specific context requires a specific model.
- This model is referred to as the “analysis plan” in OWI.
Our learning model
L'apprentissage
For the first AI, the linguistic one, the work is already done using an LLM approach. It is trained on large volumes of data for each language, including the use of Transformer algorithms, similar to GPT. As a result, it performs immediately well for each new client and use case.
The second AI, the domain-specific one, requires specific training for each sector and each company:
- We offer sector-specific packages that allow for quick implementation.
- Because this domain-specific AI builds upon an already proficient linguistic AI, it learns quickly and requires low volumes of data (between 10 and 100 times less data compared to other “semantic” AIs).
- Finally, we have developed a technology called UTML (User to Machine Learning) that transforms available data within the company into directly usable training datasets.
This fast learning capability and low data volume requirement ensure very short projects (maximum 1 to 2 months) at a low cost. Additionally, this provides each OWI client company with:
- Complete data protection since the available data is sufficient.
- Explanations for all decisions made by our AI system.
Innovations in usage
Our R&D teams are also dedicated to supporting our clients’ innovation efforts:
"Semantic Extractors," or SEMEX, for completing appointment requests, retrieving customer numbers, etc.
"Revocalization," or SPELLING, to overcome the difficulties of the best Speech To Text solutions in transcribing identifiers (since these identifiers are not common words).
"Attachment Analysis," using the best OCR technologies, for emails and chatbots.
"Advisor Assistance," meaning the integration into their workstation of response models, relevant links to handle the request, XLS files prepared with all useful information already extracted from different systems, etc.
"Quality Control," with automatic checks to avoid the biases inherent in any Machine Learning.
"Detailed Analysis" for a message, a conversation, providing necessary explanations and justifications for each result produced by our AI.
