Initially, LLMs were used at the design phase of NLU-based chatbots to help build intents and entities. Now, they have stepped out from the shadow of NLU and are starting to take centre stage with their almost magical abilities to generate understandable text.
While NLU focuses on finding meaning from a person’s message (intents), LLMs use their vast knowledge base to generate relevant and coherent responses.
A Word of Caution
LLMs and generative AI are not completely accurate and can produce wild content that is not factual. So, on its own without guardrails, it is not suitable for customer-facing enterprise use cases, especially where sensitive and private information is involved.
There is also the matter of compliance and not exposing personal information. Personal data should never be passed out of the confines of the enterprise and never used to train an LLM.
Hallucinations and security risks can be addressed by fine-tuning an LLM for a specific industry, and implementing Retrieval Augmented Generation (RAG) which provides the LLM with factual data from an external source.
On their own, LLMS are also slower and more expensive to run.
The jury is still out, but as technology develops, it appears that a good approach is a hybrid approach. By combining NLU and LLMs, chatbots can better understand queries, extract relevant information, and generate responses that are not only contextually appropriate but also linguistically natural, creating a more human-like conversational experience.
Here's a breakdown of the interplay between NLU and LLMs in chatbots:
Extracting Intent and Entities
The NLU component is responsible for extracting meaning from a person’s inputs. It involves tasks like intent recognition, entity extraction, and sentiment analysis.
NLU helps the chatbot understand what the user is asking or expressing. It breaks down the text into individual words or phrases, tagging them with grammatical roles like nouns, verbs, and adjectives. This information is then used to determine the overall intent of the message, such as asking a question, seeking information, or requesting a service. For example, a customer might message: "I'd like to pay my account" and the NLU will categorise that as an intent to pay.
What’s more, NLU identifies entities, which are specific pieces of information mentioned in a person’s conversation, such as numbers, post codes, or dates.
For example, if a customer asks, "I can pay a hundred towards my debt." NLU would identify the intent as "promise to pay" and extract the relevant entity, the amount "£100".
Control, Consistency and Reliability
Using NLU to power conversational AI is more reliable and predictable than using just LLMs, which are prone to hallucinations and are not as secure. To be on the safe side, many customer engagement bots are using NLU with user-verified responses.
Generating Contextual Responses
LLMs are powerful AI models, like OpenAI’s GPT, that have been trained on massive amounts of data to understand and generate human-like language (and they can also create images, write music and code). They possess a deep understanding of language nuances and context and are excellent at generating grammatically correct content and simulating conversations that are fit to the specific context.
Aid in Understanding language
In the example used above where the customer said, "I'd like to pay my account", the NLU categorised it as an intent to pay. However, it had to be trained on the many ways a person might express this. With an LLM, it can more fully grasp what a person is saying regardless what terms they use. They could have said "I'm going to clear what I owe" or "I'd like to pay off my debt" and the LLM would understand that as an intent to pay.
The Importance of Guardrails
For security reasons, LLMs need to be programmed with guardrails in place that narrow their responses to the data provided and strictly exclude anything that is not within the desired parametres.
Interactive Dialogue Flow
The interplay between NLU and LLMs helps chatbots to maintain a coherent dialogue flow. NLU provides the intent recognition within a context while the LLM accesses its knowledge base and responds appropriately. This back-and-forth exchange results in more engaging conversations, mimicking human-to-human interactions.
Context Establishment
The LLMs come into play to maintain context and generate responses.
Dynamic Conversations
LLMs contribute to the dynamic nature of conversations. They can generate diverse and relevant responses, giving interactions with a chatbot a more natural flavour.
Learning and Adaptation
Some chatbots leverage the learning capabilities of LLMs to adapt and improve over time. They can be fine-tuned based on user interactions and feedback and so continually improve their performance.
Together, NLU and LLMs empower chatbots to communicate with people in a more personalised, knowledgeable and accurate way. Their combined capabilities help customer engagement chatbots to fulfill their role in customer service, information retrieval, and task automation.
To learn more, see:
Subscribe to the Webio newsletter to stay up to date with the latest news in conversational AI for customer engagement.
Enjoy news on Webio podcasts, webinars, articles and more.