Project overview
Large Language Models, such as ChatGPT, are typically trained to handle numerous general-purpose, cognitively challenging tasks on thousands of state-of-the-art specialized processors. However, there are significant shortcomings when using general LLMs in specialized domains such as engineering or in specific industries like telecom.
Nokia Language Model & Generative AI tool provides a completely new way of querying, managing, operating and interacting with a telecom network, providing a new paradigm in automation.
Today’s environment
Large Language Models (LLMs) are arguably the fastest growing segment of AI today. ChatGPT has captivated the expert and the novice in its remarkable language fluency and general encyclopedic knowledge. In the past year, there has been a proliferation of LLMs with hundreds of pre-trained models, ranging in size from a few to hundreds of billion parameters. These models are typically trained on hundreds or thousands of state-of-the-art specialized processors (GPUs) for days and weeks on a multiplicity of tasks such as Q&A, translation, summarization, classification, and more. Pre-trained LLMs are able to handle numerous general-purpose, cognitively challenging tasks remarkably well.
However, with many now using ChatGPT and LLMs, users have observed significant shortcomings in LLM skills in specialized domains such as engineering, medicine, law, among others. It is now generally agreed that LLMs trained or fine-tuned on specific domains of knowledge will be needed for AI deployment for the enterprise beyond the general-purpose skills of ChatGPT. This is particularly true of domains such as communication technology where proliferation of specialized terminology, phrases and acronyms are particularly high.
Through our research, we have learned that AI tools for specialized knowledge domains are different in many respects, as they need 1) less effort in training, 2) provide higher accuracy, and 3) require modest computation during a conversation. As an example, a telecom engineer may need to inquire about an external interface to a Nokia baseband unit. When using ChatGPT for this inquiry, the generated response would not deliver useful information. However, our Nokia Language Model (NLM) tool has the specialized knowledge to provide an actionable response.
Our solution
Nokia’s Language Model & Generative AI is a completely new way of querying, managing, operating and interacting with a telecom network, providing a new paradigm in automation at an unprecedented scale. Our solution makes knowledge extraction simple by training language models on specialized knowledge domains.
In response to a query, the language model retrieves relevant information from the archive, which is subsequently summarized by generative AI. Our research shows that this approach is not only less costly to develop but is also more accurate compared to finetuning LLMs.
We trained the Nokia Language Model (NLM) on an archive of tens of thousands of Nokia product documents with over one-third of a billion words covering descriptions, releases, features, installation, deployment, and troubleshooting. The top retrievals are then used as context to Generative AI for a summary. Nokia Language Model is the first such technology in telecommunications, as announced as part of the evolution of the Nokia Digital Assistant earlier last year.
Nokia’s Digital Assistant is an existing chatbot product provided to customer engineers for troubleshooting and knowledge discovery. It incorporates customized language models with generative AI capabilities and requires modest compute resources. Many networks include multi-vendor infrastructure. Our NLM tool can also be extended to include other network vendors’ documentation.
Our differentiators
Proven technology which has been commercially implemented
Provides more accurate and useful query results by matching all sentences and focusing on semantics, saving time and improving efficiencies
Industry-first expert chatbot delivers both answers to technical questions and engages in multiple follow up Q&As
Our vision of AI in language modeling goes beyond chatbots with specialized domain knowledge and extends to extracting dynamic network information and its synthesis for high-level snapshots and summarization for real-time status and diagnostics. Imagine an AI agent which can answer questions about tens of thousands of network elements and highlight those which have a high chance of failing in the next days or weeks.
As network technology and network state become amenable to AI analysis through text and speech, we begin to see a new way of managing and operating a telecommunication network, mostly in an autonomous way but answerable to humans - a new paradigm in automation at an unprecedented scale.
APA style publications
-
J. Wen, C.-N. Yu, R. Greiner. Robust Learning under Uncertain Test Distributions: Relating Covariate Shift to Model Misspecification. International Conference on Machine Learning (ICML), 2014.
-
S. Ravanbakhsh, C.-N. Yu, R. Greiner. A Generalized Loop Correction Method for Approximate Inference in Graphical Models. International Conference on Machine Learning (ICML), 2012.
-
C.-N. Yu. Transductive Learning of Structural SVMs via Prior Knowledge Constraints. International Conference on Artificial Intelligence and Statistics (AISTATS), 2012.
-
I. Saniee, L. Zhang, B. Magnetta, Truncated Lottery Ticket for Deep Pruning. Proceedings of ICIP, Bordeaux, Oct 2022.
-
D. Kushnir, S. Jalali, I. Saniee.Towards clustering high-dimensional Gaussian Mixture clouds in linear running time. Proceedings of 21st International Conference on Artificial Intelligence and Statistics (AISTATS), PMLR 89: 1379-1387, 2019.
