Using AI for analysis and forecasting
Conversational and speech systems: past and present
Two decades ago there were already systems that conversed in natural language, but they were far from perfect and limited in numerous ways. Speech recognition technologies were available long before ChatGPT arrived, yet they wrestled with complex challenges and real limits on what they could do.
The challenge of training AI systems
Developing AI systems requires feeding in a massive volume of examples. For a system to answer questions correctly, it needs intensive training across varied scenarios. Ask it “what sound does a bird make?” and it is expected to answer “tweet tweet”. When the system gets it wrong, it is corrected — and over time it improves.
Learning from mistakes
For an AI system to identify its errors and correct itself, somebody has to know the right answer to every question. If the system answers incorrectly, the correct answer is supplied and it learns from it. In the past this process demanded supervised learning — enormous manual effort writing thousands of questions and their correct answers.
Ilya Sutskever and the OpenAI breakthrough
Ilya Sutskever, born in Russia and raised in Israel, became one of the leading figures in artificial intelligence. With a doctorate in computer science from Canada and a role as chief scientist at OpenAI, he chose to change tack and explore a new way of training chatbots: unsupervised learning.
What is unsupervised learning?
The approach is built on drilling the system with complete sentences taken from existing texts, with the final word removed, and asking the system to guess it. When the system guesses wrong, it is corrected and carries on learning. The result is a far more automatic and efficient training process.
Adding the Transformer
Unsupervised learning got a significant upgrade with the development of the Transformer by Google researchers in 2017. Combining the Transformer with the unsupervised approach produced GPT-1, which performed far better than anything before it.
The evolution of the GPT models
OpenAI kept improving and extending the model, reaching a peak with the launch of GPT-3.5 in 2022. It was regarded as one of the most advanced chatbots available, trained on vast quantities of information and offering genuinely groundbreaking capabilities.
Private language models on your own network
Many organisations avoid integrating language models such as ChatGPT for fear of leaking sensitive information. Others buy licences for commercial services but restrict their use, particularly where the content is shareable. More recently we have seen a revolution in open-source language models — models that excel on quality and sometimes surpass the commercial ones.
We propose running these open-source models on your internal network, with no internet connection (on-premise), which brings significant advantages. First, complete information security: all data stays inside the internal network with no risk of leakage. Second, a close fit to your organisation’s needs — full access to the model allows precise tuning to your own subject matter. Third, a substantial saving: the models themselves are free, so the only expenditure is compute, which is normally far lower than commercial licence fees.
Our team is glad to advise on the structure of the internal network and to build a tailored plan for deploying a private language model.
What language models can actually read
Language-based models and advanced linguistic analysis open the door to a deeper understanding of text and speech. With this technology, models can read and analyse texts of any length while extracting insights that serve a wide range of applications.
They make it possible to understand and evaluate texts in depth, including recognising the emotions reflected in the content — joy, frustration, anger — and even distinguishing urgent enquiries from less urgent ones. They can also assess the level of politeness in a conversation, or tell everyday language from formal register, which makes it possible to streamline services and improve the customer experience.
The technology, which includes models such as Llama 3 and Mixtral, supports Hebrew and can analyse texts and conversations accurately, including live voice recordings. These applications improve customer service, whether by identifying patterns and refining the sales process or by spotting business trends and emerging shifts. That in turn supports tailored training programmes, better marketing strategies and fresh insight extracted from professional material.
Other models used for language understanding
- BERT (Bidirectional Encoder Representations from Transformers) — developed by Google and used to understand the context and meaning of words in sentences through bidirectional analysis of text. It suits tasks such as question answering, named-entity recognition and text classification, and handles complex texts precisely.
- GPT-4 (Generative Pretrained Transformer) — a powerful generative model from OpenAI used for tailored text generation, translation, conversation analysis and more. It works across many languages and supports accurate automated content creation matched to the user’s needs.
- RoBERTa (Robustly Optimised BERT) — an improved version of BERT built specifically for deep text understanding, using additional training data and new methods that make it faster and more accurate.
- DistilBERT — smaller and lighter than BERT while retaining impressive performance on language-understanding tasks. Suited to applications that need faster processing with limited resources.
- XLNet — more advanced than BERT, developed by a Google team to overcome the limitations of learning word order. It analyses text across a range of possible orders and improves contextual accuracy.
- T5 (Text-to-Text Transfer Transformer) — a versatile model that treats every problem as a text-to-text conversion task, enabling a broad range of NLP work such as summarisation, translation and question answering.
- ALBERT (A Lite BERT) — a more efficient version of BERT that reduces the parameter count and streamlines processing while maintaining high accuracy, making it an accessible and effective tool for language analysis.
- Whisper — a generative model from OpenAI for analysing and understanding spoken language, providing speech recognition and conversation-to-text conversion, including language identification, transcription and translation.
Using these models helps organisations improve customer service, conversation analysis, comprehension of complex documents and the overall user experience, tuned precisely to business needs and specific language requirements.
Key areas
- Data analysis with AI: understanding data and user behaviour.
- Intelligent automation: building automated processes that save time.
- Business forecasting: using models to predict future trends.
Examples of application
- Analysing user data and behaviour to identify trends and opportunities for improvement.
- Creating automations that carry out repetitive tasks without human involvement.
- Developing models that forecast market trends and company performance.
Pace of impact
Technology companies see rapid impact — meaningful performance gains within months. Financial firms move more slowly, but with cumulative effect on trend forecasting and business performance.
Machine learning is impressive on paper — but in practice, without accessible, clean data there is nothing to teach it. And when deploying AI in an organisation, the first question should be about the infrastructure, not the technology.