Latest news

How We Upgraded an Already Great Chatbot: Techcrunch’s Case

No comments

If you have been reading about chatbots for some time, you have probably heard of TechCrunch’s Messenger bot, which was awarded the Best News Bot prize by ChatBottle, as it delivered news from their site and handled subscriptions to different topics, authors and sections efficiently.

Now, together with Chatfuel and TechCrunch teams, Bitext has developed a new and improved version of this bot. We integrated our NLP middleware into the existing bot architecture to make it benefit from our rewriting service for chatbots. The result is an enhanced version that handles conversational interaction, improved natural language understanding, including double intents, and unique features like negation understanding.

Bitext’s approaches

There are two main ways Bitext technology can make bots smarter. They are both focused in simplifying the training process:

In TechCrunch’s case, the Query Rewriting service was the one applied. As it was an existing bot already working, the solution had to be flexible enough to integrate into the current architecture. The rewriting technology was perfect for that because it acts as a NLP layer that allows the bot to understand any kind of query, therefore boosting its NLU performance.

Why are we so obsessed with simplifying the training process?

There are deploying times that cannot be reduced: the designing of the bot, i.e., deciding its purpose, its personality, or the intents and entities it is going to handle is crucial and shouldn’t be be done in a hurry. Then, it has to be built, which depends entirely on the objectives set, but isn’t usually a problem.

But if there is a clear black hole of time in bots developing projects, it is the training phase. It can take months to get users to interact enough, in quality and quantity, for the bot to learn from them. That’s why everyone who is thinking of developing a bot for their businesses is running to start their projects, just to be the first of the race.

To us, this race seems insane, even more when there’s no guarantee the bot will become fully functional at the end. We go instead for a different approach: a short, controllable training process that is aimed to improve the “brains”, and not the “looks”, of the chatbot. That way businesses can benefit from implementing their bot with complete confidence in the technology.

So what does the upgrade exactly consist on?

TechCrunch launched their bot to deliver news and for users to be able to ask for new stories on a certain topic. It was able to attend basic requests, like asking for the main menu, managing subscriptions or giving feedback. Also, users could write down the name of any news section in the web to get the latest stories featured in it.

But all these interactions with the bot were limited to certain strings: you had to type exactly “main menu”, “about”, “manage subscriptions”, etc. So users had to remember the precise phrase, its words and order, and couldn’t get out of that.

With Bitext technology, some advanced features were added to the bot’s Natural Language Understanding (NLU) system, allowing it to correctly interpret three new types of queries:

  • simple intent queries expressed in natural language
  • double intent queries
  • negated queries

This operation is seamless and doesn’t interfere with previous features of the bot, as can be seen in the table below.

Sample benchmark of TechCrunch’s bot before and after Bitext’s upgrade

How was this transformation into a more conversational bot possible?

The current pipeline works like this:

  1. First, any query sent by a user goes directly through the Query Rewriting middleware service.
  2. The query gets simplified, the non-important parts being reduced, and a simpler, canonical version of the original query is generated. This rewritten sentence is sent to the intent detection engine so the chatbot can figure out what the user is asking for. That way, the bot had to be trained only with a small amount of simplified queries like this one to be able to understand all kinds of natural language queries (NLQ).
  3. Once the intent is detected, Bitext once again rewrites the sentence into a boolean string, adding commands such as “AND” or “OR”. This step is key, so when the basic phrase including booleans goes to the search engine, the retrieval answer is the correct one.
  4. Finally, the search engine delivers the results, the ones that match the boolean query received from the previous step, to the bot user.

Query Rewriting service simplifies users’ queries and transforms them into a simpler version

As you can see, the architecture is completely modular, and Bitext middleware just acts as a convenient NLP layer to boost the chatbot performance.

Now, users can express themselves in natural language to get their stories and we think that’s fantastic. It’s great to be part of this revolution that is enabling everyone to communicate with machines in a more and more comfortable way. For years, we had to learn how to exchange communication with every single device, and from now on the task consists on simply getting used to not making that effort.

(This story was previously published in Bitext’ blog in four parts: 1, 2, 3, 4)

Clap to recommend this article to others so that small business can grow!

magnoliaHow We Upgraded an Already Great Chatbot: Techcrunch’s Case

Related Posts

Leave a Reply