Deep learning for chatbots remains a hot topic as more and more companies look for different approaches to develop their chatbots. Deep learning techniques for chatbots are only one of several different approaches that use Artificial Intelligence (AI) to simulate human conversations. When a chatbot has to answer complex questions and/or understand with good accuracy a wide range of different intents (e.g. more than 100+ user intents), a more sophisticated approach is required.
Pros and cons of deep learning chatbots
In a previous post, we talked about how chatbots simplify the interaction between people and machines (What Is a Chatbot?) Today, we will explore pros and cons affecting chatbots based on deep learning or machine learning.
Machine learning and deep learning for chatbots are on the rise especially thanks to virtual assistants like Siri, Alexa or Cortana that allow consumers to interact using their voice via smartphone or even appliances and devices used in the home.
There are basically two different techniques at the core of these kinds of systems:
- Supervised learning. In this case, the chatbot software is trained by a large set of requests. Each request is correlated to a specific “tag”, which represents a specific user intent. Pros: Deep learning or machine learning based on a supervised approach can enable a good level of accuracy. Cons: However, this approach needs a really large set of good, correctly tagged examples, otherwise the training process won’t be successful.
- Unsupervised learning. In this case, the chatbot software relies on a very high number of examples to independently identify the requests and corresponding user intents. Pros: The system potentially does not require human supervision and a set of explicitly tagged examples. Cons: A fully independent approach is not possible, as the training set of examples should be not only extremely wide and varied but also of high quality to ensure that the chatbot is properly trained.
Both approaches may leverage some basic Natural Language Processing (NLP) capabilities, but at the same time require a long time to be trained and a vast amount of good, appropriate data. A system like Siri receives more than 1 billion requests on a weekly basis, while data volumes necessary for training a deep learning system in today’s enterprise world are considerably lower.
Poor data means poor results. A lack of good examples with which to appropriately train the system will lead to approximative results or, even worse, completely wrong answers or no answers, unless the training phase is continuously repeated.
Thanks to consumer application like Siri, chatbots have been receiving a lot of attention from private and public organizations, and they are expected to continue their rise in the Artificial Intelligence solutions market. According to Gartner, “by 2020 customers will manage 85% of their relationship with the enterprise without interacting with a human,” and “by 2020, over 50% of medium to large enterprises will have deployed product chatbots.”
There are several reasons why consumers love chatbots (and they include curiosity and entertainment); and there are as many reasons why companies benefit from chatbots as well, such as the opportunities to improve routine tasks, to simultaneously process multiple requests from users, to reduce customer care efforts and costs, to acquire a better understanding of customer issues and expectations, and finally, to enhance their loyalty.
The problem is that the typical enterprise data world scenario is completely different from that of the consumer, especially if we look at how chatbots can be implemented.
Deep learning for chatbots, for example, may be appropriate in a consumer context, while in the enterprise world it’s usually difficult to easily gather large volumes of good examples that are ready to be used to train the system.
For this reason, more and more companies are looking for more sophisticated approaches such as hybrid approaches based on different AI algorithms, machine learning / deep learning and advanced NLP. In the end, don’t forget that chatbots, as well as other forms of AI, are not magic, and any AI solution is always based on programming.