Introduction to conversational flow
Humans interact using voice, written text, gesture and tonal modes. These interactions can be short, long, carry over from a previous episode, be completely new or stop for now to be completed later types. Often Human intelligence factor these aspects in consideration whenever a person has an interaction with another sentient being.
With the advent of technology in day to day life and previously science fiction topics like Artificial Intelligence, Machine learning, Natural language processing becoming mainstream there is a race to create chat bots which mimic humans and are tasked to take care of the trivial task for their users for now. The goal is for these chat bot to grow intelligent over time as we study how humans interact with them and they learn to become more effective.
A typical conversational transaction may go through multiple states with each state belonging to a certain type of meta states depicting the stage of the conversation. The conversation may jump across states either in the forward or backward direction. What finally matters is each of the conversation transactions passes through one or more of the meta states, whilst exchanging and collecting data as the conversation flows through the steps.
Meta states within a conversation
A conversation can be analysed to belong to following states
- Initiate / Start
- Context / Discover
- Input /Information
- Action / Goal
- Close / Summarize
Typically a conversation will follow the above meta-steps in a sequential manner with each meta-state having one of many steps within the state to proceed to the next state. For a successful conversation, the flow should ideally go past the Goal state at a minimum. The wrap-up states are primarily summaries due to which drop out at that stage in the conversation flow does not lead to a missed opportunity for a successful conversation transaction.
For a typical conversation transaction to reach the goal step it requires to go through the context step where the intent is discovered and the input state where all the information required for processing the intent is captured and confirmed with the other party.
To take an example to better describe the flow, say a person walk up to a store selling mobile phones with the intention of buying a phone the conversation could be modeled along following lines
Typical sales flow
A human being keeps track of the end goal for the conversation and the current state so as to course correct if the conversation veers off track.
A conversational bot understands Intents and Entities but it needs to understand a meta state flow which govern human interaction so as to be as effective as a human it seeks to mimic.
These meta states are also an effective measure for monitoring the bots effectiveness and figure out which stages need better training. Say for example data shows that conversations are jumping states back from the actionable stage to the information stage showing the information being shared is not scoped enough for the goal in mind to be met. An ideal conversation should move across states only in the forward direction with any backward movements pointing to improper scripting of the conversation in context.
Modelling a conversation in states also gives the engine to add cognitive capabilities where one can predict how a conversation is going to go and take corrective actions if the path of the conversation tree being followed doesn’t lead to an end goal one has in mind.
Conclusion
To summarize, conversation bots currently are an virtual implementation of erstwhile IVR systems with some learning built into it where humans do the majority of the training. For conversation services to mature to the next generation, they need to be programmed to first understand how a human conversation can be modeled in machine terminology as close to the real version as possible and then build self learning tracks into it.
