Why You Need a Large Language Model for Intent Recognition

Identifying customer intent and routing to the right place is key to a positive support experience. In this post, we show how it's done with LLMs.

Why You Need a Large Language Model for Intent Recognition

Identifying a customer’s intent correctly and routing to the correct place is key to a positive customer support experience. When designing your customer support solution, there are three key components to consider:

  1. Correctly identifying the customer’s problem and desired outcome (their “intent”): 63% of customers expect businesses to know their unique needs and expectations, with the number increasing to 76% for B2B buyers (Salesforce Research, 2020)
  2. Achieving the customer’s desired outcome the first time: Nearly 70% of customers are irritated when their call is transferred from department to department because they need to repeat the same information (Zendesk, 2020)
  3. Decreasing time to resolution: 90% of customers rate an “immediate” response as integral when they have a customer service question (Hubspot, 2021) and 60% say the worst experience involves a long wait time (Zendesk, 2020)

Conversational assistants have been widely deployed to solve this problem in 2022, experts estimate 80% of banking and healthcare queries will be answered by a conversational assistant (CNBC, 2017). In addition, 87% of customers report neutral or positive experiences with conversational assistants (Startup Bonsai, 2022) which indicates they have been successful in providing fast and scalable support.

However, today’s conversational self-service assistants still do not have the best reputation for being able to understand a customer’s query, especially if their query is long, complex, or doesn’t contain certain keywords.

Today's conversational AI is primarily rule-based

Userlike surveyed chatbot users for the negative aspects of their experience and 38% of respondents stated that the assistant could not understand them (Userlike, 2021). Anecdotal quotes indicated that most assistants struggle because they “depend on the user to follow a strict flow that results in rigid interactions and conversational dead ends” (Userlike, 2020).

IT departments have a long way to go before they develop a chatbot that can capably handle the nuances of language and complex questions, which are challenging even for a human service person. – Userlike, 2020

To get the best accuracy, consider using a large language model to perform intent recognition.

Large language models are neural networks trained on terabytes of human language text to achieve breakthrough results on language comprehension and reasoning tasks. In the last few years, large language models have greatly closed the gap between AI and humans in terms of performance.

Large language models maintain the benefit of being cheaper and faster at recognizing intent than humans. However, they have traditionally remained inaccessible to conversational AI developers because of the prohibitive cost to train and serve these large neural networks. Only large companies like Google, Facebook, and Microsoft were able to produce these models, and they did so primarily for internal usage: Google to power their translation and intelligent search and Facebook to perform content moderation on their platform.

Cohere was founded to make these large language models available to the public to democratize NLP for developers. We enable users to “finetune” these large language models to a specific classification task, such as intent recognition, for the best natural language understanding ability available. To demonstrate our performance on intent recognition even against other large language models, we used a public dataset called Banking77 that contains 13,000+ banking queries. Each query is labeled into 1 of 77 intents.

Our Large model outperforms the best natural language processing (NLP) models in the world:

Performance on Banking77 using Cohere finetuned large embedding model

Getting started with Cohere Classify

Start by reading our Guide to Text Classification. After signing up, navigate to the Classify playground and try our “Restaurant customer inquiries” preset, which classifies frequently asked questions a restaurant would receive into intents. You’ll note that our models only require 5 examples per intent label to be able to classify new intents due to strong baseline performance.  Simply click Export Code when you’re ready to add a classifier to your application.

Large language models can give you the best chance of determining your customers’ intent accurately the first time they ask without a human in the loop. Interested in using Cohere for your customer support or intent recognition solution? Click below to speak to an expert.