How to make your chatbot more human-like

Chatbot’s potential is nothing new, and at Bitext we have been talking about them for a while. We emphasize the importance of Natural Language Processing to overcome the current limitations users and developers of bots are facing when trying to create human-like chatbots.

Over the lasts months, we have been researching bots and bot developing platforms like wit.ai, api.ai, and LUIS. And we detected some issues that have been around for a while and seem not to be solved yet. The issues found are fundamental for accurate human language understanding and linguistic technology is a perfect match to solve them.

We decided to put our resources to work with the objective of improving the bot user experience. We introduce to you our Chatbot interactive infographic so you can see some examples of what is behind our NLP platform and what it can do.

The demo proposes that you follow different sentences within a conversation. The sentences contain several issues we have detected while testing bots and platforms:

• Negation: we realized that many bots don’t understand negation in a phrase because they have been built based on a keyword approach. That makes it difficult for users to ask for something as simple as “I want a barbecue pizza with no pork”. Let’s see some examples:

o “I want a barbecue pizza with no pork” (only negates pork).

o “We don’t want any drinks” (negates the whole event).

o “I’m not sure… I’ll take a beer (It doesn’t negate the main event)”

• Coordination: it is one of the most used elements in how humans talk, and after our research we found out that most relevant platforms do not support a request where elements are joined by a coordinator. Our linguistic knowledge makes us capable of solving this issue.

o “[[I want a Hawaiian pizza] and [my wife will have a Margherita]]” (two main events).

o “I’ll have a Hawaiian [with [extra cheese] and [onion]] (two changes in ingredients)”.

o “I’ll take [[a Hawaiian [with [extra cheese] and [onion]]] and [a Margherita]]” (two pizzas, the first one with two ingredients)

• The connection between different phrases: Most of the chatbots have been designed following a tree model, so it’s not possible for a user to change his request and that forces him to start over. As a solution we propose the usage of connectors as in the following examples:

o “I want a Margherita with onion… Moreover, add extra cheese” (adds info to the first sentence: adds an ingredient).

o “I want a Hawaiian with extra pineapple. However, I prefer it with no ham” (also adds info to the previous one).

During the demo, you will see not only the conversation flow but also the text structure and linguistic knowledge that we extract from raw text. Text structure is key to understand requests and answer properly. That linguistic knowledge is shown in the demo as a parse tree and a JSON with the semantic information that can be derived from it. Our parsing technology is what allows us to extract relevant information that can be used by the Machine Learning powered bot engines to improve their accuracy extracting intents.

If you want to start trying our interactive chatbot infographic demo click here