Samer Mohamad, Yasmina Regional Director for MENA at Yango, highlights the unique characteristics of Yasmina, an AI assistant designed specifically for the Middle East.
How is Yasmina’s AI technology different from other kinds of AI assistants?
Our intent from the outset was to treat the Middle East not as an afterthought, but as a region to invest in. We wanted to create something for the local people and build a legacy by utilising our GLOCAL (think global, act local) approach.
From day one, we were committed to building Yasmina as a human-like AI assistant uniquely designed for the Middle East. We did not want to use an existing international model with Arabic as an add-on.
What sets Yasmina apart from other assistants is the ability to maintain engaging, fun and consistent dialogues that feel like you are talking to a real person. Another remarkable distinction is that Yasmina understands major Arabic dialects and speaks Khaleeji fluently. Additionally, Yasmina excels at telling culturally relevant jokes that resonate with local humour.
A great example of our technology is that Yasmina is capable of recognising whether the user is male or female and puts on only age-appropriate content when interacting with children.
Another example is its ability to respond in real time to changing user preferences. For example, if the user switches between Arabic and English mid-conversation, Yasmina will respond accordingly.
Yasmina’s uniqueness is built on its technological foundation which consists of natural language understanding, speech recognition and speech synthesis.
Yasmina’s human-like conversational skills are based on a proprietary Large Language Model (LLM) developed from scratch using a refined dataset of Arabic web pages. It is further fine-tuned over an extended period of time in collaboration with a diverse group of Arabic-speaking experts selected for their varied dialects, pronunciation and vocabulary.
A large number of people from various locations – such as Riyadh, Jeddah, Dammam and more – took part in enhancing Yasmina’s speech recognition. These diverse voices, spanning different ages and genders, trained Yasmina to adapt to local preferences, grasp the subtleties of language and deliver authentic and contextually appropriate interactions.
We have improved Yasmina’s ability to generate natural language by exposing its LLM to diverse linguistic contexts, including colloquialisms and vernacular expressions. Yasmina’s LLM is trained on rich and varied datasets, which capture the intricacies of regional dialects, idioms and cultural references.
What technological innovations are required to develop an AI assistant capable of delivering hyper-relevant AI-driven conversations?
To generate content that is highly relevant to the user’s environment or query, an AI assistant would need a Large Language Model capable of analysing a vast corpora of language data, learning intricate patterns and nuances essential to understanding and generating natural language. This is the essence of a localised LLM.
LLMs enable AI assistants to understand and generate language that resonates with users from different regions. An LLM trained on large, diverse and locally relevant datasets enhances an AI assistant’s local capabilities. Similarly, leveraging insights from previous interactions or related information allows LLMs to engage in context-aware conversations.
Yasmina is learning every day – what does this mean and how do you achieve this?
It means that we are expanding the capabilities of our LLM model to continuously improve Yasmina’s vocabulary, comprehension and response quality.
We have established learning from human feedback (RLHF) loops so the feedback from our active beta user community is incorporated into the LLM. The result is an AI that gets better every day.
