How Are IBM Researchers Improving Chatbot Dialogue Flow?

Key takeaways

Chatbots have significantly transformed customer service across various industries, providing 24/7 assistance.
Current chatbot frameworks are easy to use, yet many fail to ensure natural and intuitive conversation flows.
Effective chatbots must anticipate user queries and respond appropriately to maintain a seamless dialogue.
IBM Research is leveraging deep learning and natural language processing to enhance chatbot interactions by analysing human conversation logs.
Understanding past human interactions helps designers create chatbots that better understand user intent and expression.

Chatbots have revolutionised the customer services space with most industries resorting to this ingenious piece of innovation. Be it booking a flight, ordering food, playing music or anything else, we can now refer to these familiar friends who are there for us 24x7. Building chatbots today is fairly easy as there are various different frameworks available today for building them. However, not many of them meet expectations. While today’s frameworks are easy to model and deploy, most of the chatbots lack the ability to perform a simple task and the conversation flow. It can be quite intimidating to induce natural dialogue flows and richness of natural conversation into these bots. It should be able to answer what are my users going to ask, how are users going to ask them, how should it respond to them, how is the conversation going to proceed and more.

The team at IBM Research is using past human to human logs and analysing them using deep learning and state of the art natural language processing techniques to get answers to these questions. The analysis of human to human logs provides suggestions to a dialogue designer on what are the key things users are going to ask as well as the dominant ways in which users are going to express themselves. To understand more about how IBM is overcoming the challenges of designing a chatbot, the evolution of chatbot industry in India and more, Analytics India Magazine got in touch with Gargi Dasgupta, Sr. Manager at IBM Research India, who leads the cognitive tech support area. She shares some interesting insights on the subject in this detailed interview.

Analytics India Magazine: While we see a lot of chatbots coming up in the market, their practical implementations aren’t quite successful. What is the reason for it? How can we overcome these?

Gargi Dasgupta: A lot of the failures with current chatbots in the market are because of two main factors:

Over-dependence on the manual effort for creating chatbots which makes them both time consuming and brittle
Lack of an end-to-end fall-back plan with continuous learning and improvement.

We at IBM Research are working towards training these applications with data by infusing AI and related technologies to help organisations get business value. We are working on scalable domain modelling from unstructured data like chatlogs and documents so that domain knowledge can be created leveraging the existing data. Chatbots can then learn from the existing domain models. We are also working on techniques where continuous feedback can be learned by the chatbot to improve the conversations for the next time.

AIM: What are the major challenges in building a fully-functional and conversant chatbot? Why do chatbots fail to have the richness of human conversations despite several attempts?

GD: The biggest challenges in building chatbots are:

Handling nuances of the user language
Handling domain and context understanding
Making them affordable and useful

Human conversations are very rich and diverse in the sense of usage of words, sentences, languages, emotions and contextual references. It is natural for humans to use domain-specific references in their conversation. For example, when asked the question: “Do you enjoy Apple?” in the middle of a workplace discussion, humans are likely to infer the domain and interpret that the question relates to the company Apple and not the fruit. However, this contextualisation is hard for a machine and one of the reasons why chatbots may fail to impress. Another challenge is natural human conversations contain a lot of cross-references. For example, a follow-up question is, “How does that compare with your previous one at Facebook?” is actually asking for a comparison of the stints at Apple and Facebook. Co-relating “that” to reference of the stint at Apple is intuitive for the human to understand but harder for the machine. Thus, the lack of richness in chat conversations can be attributed to technology not yet being able to handle the whole diverse range of human language variations, domain contextualisation, cross-references, colloquial expressions, and emotions.

AIM: How can companies make use of the continuous user inputs that are fed in a chatbot? How can intelligence be derived from consumer insights?

GD: Companies can drive insights from user inputs and strive to continuously improve the quality of conversations. Users often give explicit feedback (e.g. thumbs up or down) or can give implicit feedback (e.g. saying thank you or abruptly closing the session and a few tries). All these are valuable inputs and are being used to retrain the current dialogue models. Using AI models, user feedback is fed back into the dialogue box helping to learn new things leading to answers improve over time.

AIM: How have been the evolution of chatbots industry in India?

GD: The enterprise chatbot evolution in India started with everyone being wowed by that fact that they can have a 24x7 channel of communication with their customers. In addition, they can delight their customers with an instantaneous response to their queries. For businesses, it opened up yet another channel of communication with their customers. India is a key player in the chatbot market today where chatbots have been introduced by many banking and insurance companies. We have seen banks and insurance providers to be one of the early adopters of Chatbots in India helping customers for bill payments, mobile recharges, booking travel, so on and so forth. Further many are looking at offering more services like register a claim, get a quote and help customers understand their various policies better. We realise in all of this, one big challenge for chatbots in India is the support for local languages. 90% of the new internet users are expected to be non-English speakers in the next few years. Research is trying to work with linguists alongside technologists developing verticalised vernacular models for handling local languages.

AIM: What are the efforts carried out by IBM in the area of chatbots? How is the company using deep technologies to continuously improve chatbots?

GD: Enterprises today have existing customer care support channels where end users can either call or chat with human agents to resolve their problems. These human-to-human conversation logs (h2h logs) are a goldmine of customer insights such as "what are the problems being faced by customers" and "how agents are handling and resolving customers' problems". IBM is using AI techniques to analyse these logs and generate customer insights automatically. These insights are used by enterprises for modelling of the domains for which they need the bots. The other big challenge that IBM is focussing on is to help chatbot technology to advance for taking seamless actions in the enterprise. 50% of chat users feel today chatbots cannot execute tasks where the real benefits lie. However, with the help of integration into enterprise APIs and evolved reasoning and understanding IBM is looking forward to a world of seamless question-answering, taking actions and falling back to live agents whenever required.

AIM: What are the key points to keep in mind while designing a chatbot? What are the areas the IBM researchers, in particular, give a lot of importance to?

GD: While designing chatbots it is crucial to keep in mind the affordability of creating chatbots and their usability. Chatbots are only useful when they help a particular business function to improve. Enterprise chatbots today need to address two fundamental challenges:

Robust modelling of business function(s) for which questions are to be answered
Seamless handling of answering questions, taking actions, falling back to live agents

IBM Research is focussed on spearheading the adoption of chatbots in the enterprise by working on the above research problems.

AIM: How analytics play a crucial role in building smart chatbots?

GD: Analytics can give us pointers to how the chatbot will be used and also tell us how the current chatbot is performing. It also provides pointers to areas of improvement in the form of commonly misunderstood questions and intents.

AIM: How IBM makes use of customer insights to drive automation?

GD: By 2020, 30% of the enterprise tasks will be executed by automation bots. Conversation provides a natural and intuitive interface for executing the automation of simple, repetitive tasks. The way to drive the success of these virtual assistants in the enterprises lies in combining automation through conversational interfaces. For e.g. the ask “Can you help me transfer $100 to account Y” requires integration with backend enterprise APIs as well as entails an understanding of the transfer action, the amount of money and the payee account. IBM is focusing on driving automation through conversational interfaces by focusing on developing the backend integrators as well as the front-end NLP or ML techniques for user question understanding.

AIM: What is the cost involved in building a chatbot? Do these costs compensate for the benefits?

GD: The cost involved in building a chatbot includes:

A one-time investment in understanding the domain, the target audience and the use case for which the chatbot is being built. This involves effort from domain experts as well as a data scientist. Currently, it takes 3-6 months for a data science team to build a chatbot for a domain.
Once built, frequent re-training or “care and feeding” of the chatbot is needed. Depending on the amount of re-training to be done, it can take from a few hours to a few days.

Factoring these two, costs and depending on the complexity of the domain it can take 5K to 40K to build an enterprise chatbot. The target is to make chatbots much more affordable by reducing the chatbot creation time to days instead of months.

AIM: How do you intend to revamp customer services through AI? Would you like to highlight use cases?

GD: Today, customer expectations are very high, and they would like to have a 24x7 connection with high levels of speed and accessibility. Industry-wide there are about 265 billion customer calls every year of which 50% go unresolved. At the same time, support agents struggle with the deluge of updated product documentation and new releases. AI can help transform this support industry in 3 ways:

Create customer-facing technologies (like a virtual assistant) that can understand customer intents and personalise and contextualise responses to customer
Integrate with backend information systems which access key customer information as well as enterprise capabilities so that useful advice and actions can be offered whenever possible
Act as assistants to the thousands of support agents as they handle complex customer cases by presenting insights and recommendations

Here’s How IBM Researchers Are Making Chatbots More Intuitive With Natural Dialogue Flow