Our chatbot,“α-ßΩτ”, turns 1 year old, becomes smarter

|

This page in:

If you’ve visited the World Integrated Trade Solution (WITS) website, perhaps you’ve already met and interacted with our chatbot known as α-ßΩτ who lives in the top right corner of  the page. We are happy to announce that α-ßΩτ recently turned one year old, and has become smarter. Last year, α-ßΩτ could only help users find data, but now it can search data and metadata, as well as help users with questions on methodology and definitions.  

chatbot image

Why build a chatbot?

The TradeStats module of WITS provides pre-aggregated trade statistics and allows users to browse data like top exporting countries and US imports and tariffs by country, all the way to more complex data like Germany Revealed Comparative Advantage (RCA) for Chemical exports. The pre-aggregated data contains over 500 million records spanning few million pages. Navigating the site is not always easy as it requires some understanding to adjust the 6 dimensions of reporter and partner country, flow, product, indicator, and time. In addition, many users find WITS through search engines, but they do not always land on the correct page.

What can you use it for?

The chatbot was launched last year with a Data Search function, as well as a search wizard to walk users step by step on how to obtain search results. Since the launch, users have asked over 6,500 questions, with α-ßΩτ having answered over 75% of the questions. The chatbot now supports search of knowledge articles and using knowledge search. Using the bot, users can search for information on these pages: Glossary, FAQ, Did you know, NTM Measures (Feb 2014 Version), Trade Metadata, and select pages from WITS help. The bot was developed for users to find data, and not pass the Turing test; hence greetings/questions like, “Hi” and “Who are you?” initially didn't return any answer. The bot now gives a courteous response when greeted, "Hello" or "Good Morning", and graciously replies when thanked.

chatbot thanks

How do you use the bot?

You can launch the chatbot by clicking on the α-ßΩτ Chat bot Search icon on the top of the page and choosing from the following options: 1) Data Search with or without Metadata, 2) Narrow Search and 3) Knowledge Search. You can use Data Search to find links to the correct page. If you need help in searching, you can select Narrow Search, which will walk you through the selection. If you want to ask questions on methodology, definitions or metadata you can use the Knowledge Search. You can also view the type of questions users have been asking in the recently asked questions page and refer to the help page on how to use the chatbot.

How did we build α-ßΩτ ?

We used a combination of Microsoft LUIS and QNA Maker. LUIS is used for Data Search while QNA Maker is used for Knowledge Search. The site itself has around 15 to 20 main pages driven by different parameters. These parameters contain one or more combinations of (a) reporter country (b) partner country (c) trade flow (d) indicator (e) product and (f) time. Using the combination of these parameters we can form URLs to the corresponding data page. For example, Brazil exports to Argentina has the following parameters:

  • Reporter: Brazil
  • Partner: Argentina
  • Flow: export
  • Indicator: trade value
  • Product: all/total; and
  • Time: all years

and the corresponding URL will be: https://wits.worldbank.org/CountryProfile/en/Country/BRA/StartYear/LTST/EndYear/LTST/TradeFlow/Export/Partner/ARG/Indicator/XPRT-TRD-VL

For any user search we had to identify the above 6 parameters and the purpose/intent of the search. It made a lot of sense to use natural language processing (NLP) to identify the parameters and intent of the search. We used Microsoft Language Understanding Intelligent Service (LUIS) to parse the search terms and identify the intent of the search and entities/parameters.

We trained QNA Maker with a combination of metadata, help, and methodology pages to answer user questions.

Finally, we used the Bot Framework to build the chat interface.

 

Let us know in the comments below if you would like a blog on “How to train a chat bot”.

Authors

Join the Conversation