Many businesses have found ways to use simple automated response bots to improve call centre productivity and reduce service response times. The usage of chatbots to engage customers is fairly limited, as businesses are only beginning to grow their repository of data and fine-tuning these bots to serve customers better.
The troves of messaging conversations collected by businesses over time, can make chatbot systems smarter and more human-like. Messaging has become a major way of interaction among customers, and companies want more sophisticated chatbots that can support day to day business and improve productivity. Virtual assistant tools such as Apple Siri, Google Home and Amazon Alexa have also grown smarter with each use, handling follow-up conversations in context.
There are generally two types of chatbots:
- Retrieval-based Chatbots
- Generative-based Chatbots
What are Retrieval-based Chatbots?
Retrieval-based chatbots answer based on a fixed set of pre-defined responses. Its learning method is premised on static pre-fed information and rote learning. The sophistication of retrieval-based bots depends on a composite of machine learning classifiers, and largely adopt supervised learning techniques.
The simplest of chatbots include basic rule-based techniques like:
- Decision tree learning
- Questions to Answer formats
More sophisticated chatbots include natural language processing (NLP) techniques like:
- Text classification
- Language modelling
- Speech recognition
- Machine translation
- Documentation summarisation
More complex chatbots incorporate additional machine learning classification techniques to enhance model precision. Some examples include:
- Bayesian networks
- Clustering
- Support vector machines
What are Generative-based Chatbots?
Generative-based chatbot models are beginning to emerge but are tougher to implement.
Generative-based chatbots can provide original answers that are not based on pre-defined responses. Their learning is based on the concept of Recurrent Neural Networks (RNN), which allow them to learn dynamically.
RNN involves a deep learning model dedicated to the handling of sequences. This enables the chatbot to begin putting together a structured sentence, based on its understanding of the language principles. Because of this, generative-based chatbots require large amounts of training data and are harder to optimize. Given that these generative-based bots have relatively more freedom in how they can respond, they are prone to grammatical mistakes and may produce irrelevant, generic or inconsistent responses.
From Retrieval-based chatbots to Generative-based chatbots – how do we build effective chatbots?
Chatbots with strong business and user value today are built based on a combination of natural language processing (NLP) techniques and machine learning (ML) classifiers for a closed domain. These are mainly retrieval-based bots that limit the scope of the conversation to specific goals and solve problems for a specific context.
With more people using Artificial Intelligence (AI) personal assistants, more data will power neural network models and provide a better use case for generative bots. There are generally four key enablers that will improve the effectiveness of chatbots:
- Volume – large amounts of training data
- Structure – sophistication and robustness of taxonomy and ontology
- Model – robust learning techniques that comprise ML classifiers and RNN modelling
- Context – how scoped is the domain and how specific are the problem statements and business goals
In summary, the use of retrieval-based chatbots has grown exponentially over the past five years with more companies implementing them in day-to-day operations at scale. How soon we can move towards the adoption of generative-based chatbots will depend on the comprehensiveness of training data as well as the ability to build more refined and robust NLP and ML classifiers. It will be an uphill task ahead, but the horizon for generative-based chatbots is closer than we think, given its strong business potential.
This article, written by Johnson Poh, was first published on Learn@IBF.
Facebook Comments