Feedback based Continuous Learning
It was first published on Zensar Blogs
Continuous learning is built on an idea of learning continuously and adaptively from the environment, retaining knowledge, skills, and using these for more complex and different tasks. Continuous learning is a key capability of human beings in an interactive environment. Its techniques promote the discovery of new features and unknown information from the data as compared to the traditional supervised learning methods that require a lot of labeled data and hinder the exploration of new features.
Continuous learning can be supervised, semi-supervised, or even unsupervised. The world is too complex to be only learned in a supervised way. In fact, we humans learn through our interactions with the environment and other humans from explicit and implicit feedback. This process of learning is called self-supervised as it does not require labeled training data. Similarly, for machines, one of the methods to continuously learn is self-supervised learning where unknown data is predicted based on initial data and a subset of information. This information is validated and corrected through feedback. In the context of conversational AI and natural language processing (NLP), continuous learning is critical for a conversational AI platform to be truly intelligent in any conversation to improve itself, understand and get to know each of its conversation partners. Further in this blog, we will talk about chatbots to refer to all kinds of conversational agents such as dialogue systems, question-answering systems, unstructured text retrieval, and generative models.
Early chatbot systems were mostly built using markup languages like artificial intelligence markup language (AIML), hand-crafted conversation generation rules and information retrieval techniques. They did not use explicit knowledge bases (KBs) which often resulted in dull responses. A major weakness of the existing chat systems is that they are exposed to new knowledge during conversations but the knowledge base is not expanded or updated during the conversation process. It limits the scope of their applications. Another thing to improve the learning is capturing and incorporating feedback in the chatbot systems. Without a feedback loop, you are limiting the intelligence that you get from the users and missing the opportunity to improve by adapting to user needs. A smart way to capture feedback is using multiple forms like a simple like/dislike button, a rating or even a comment.
Many research papers suggest that a simple like/dislike button is the best way to that as people don’t want to spend time deciding the rating for each question/answer pair. Comment based feedback is hard to interpret as the NLP engines need to find where and what is wrong.
To explain this, here’s a simple demonstration of the use of like/dislike buttons for taking feedback on each question-answer pair from the users.
Let’s start with a supervised approach in which we will get the actual feedback from the user in the form of like/dislike. After incorporating the feedback, the model which initially answered the questions incorrectly gave better answers for the same questions, and the questions which were answered correctly initially were answered with more confidence. Have you wondered what if there’s is no feedback mechanism? This is where the unsupervised approach can help. To begin with, the last layer of the above model will be retrained by incorporating the feedback, keeping the weight of the previous layer intact. Once it is done, implicit feedback specific to users can be utilized to train and update the underlying machine learning models. This task involves understanding the question asking the pattern of the user, time spent on questions, and similarity between successive questions, etc.
Now let’s share some thoughts on how continuous learning impacts business. According to Michel Falcon, regardless of your service, product, size of the business, or industry, customer experience has proven to be a key differentiating point for companies around the world. While there are a bunch of things you can do, chatbots are a must in customer experience strategy. The difference between a good chatbot and a great chatbot is a robust feedback mechanism. This functionality significantly improves the customer experience which in turn reduces the customer churn. Customers feel valued when they are probed for feedback and know that their feedback will contribute to improvements in the chatbot.
Gartner proved that customer experience is a game-changer in all domains. Feedback based continuous learning is beneficial for various domains like retail, BFSI, and manufacturing. In the BFSI sector, financial institutions with a digital-savvy customer base can enhance customer experience using this model, decreasing customer churn and overall cost. Continuous learning is also used for extracting better answers from large unstructured insurance documents.
In the retail context, AI and chatbots are an integral part of the online customer journey. This supports self-service and leads to faster problem resolution. For manufacturing, continuous learning can help chatbots to update managers about the demand for raw materials, the status of current inventory and help them order parts and materials. It can be used to improve the accuracy of the chatbots in giving recommendations about the suppliers of raw materials as well.
According to popular research, 85% of all customer interactions will be conducted without humans by 2021. Knowing when to be proactive and reactive in interactions is part of continuous learning. If you have any more questions about continuous learning, please leave a comment below.
