chatbot-tamil-rasa

This is a chatbot in Tamil language developed with Rasa

Introduction

Many users struggle to reload through their carrier’s respective online portals. The main reason being difficult to understand English and code words for reload etc. are not familiar. As a solution, I developed this chatbot to facilitate reload requirements for a user. The bot is developed with Rasa.

The bot can reload, cancel last reload, get a refund, get coverage for a particular area etc. in Tamil. Other than that, standard greetings, bot_question and other standard questions can now be asked in Tamil language.

Because the actual backend implementation is not possible (that is not a requirement), I have added outputs based on some randomness. So, for example, reload might be successful or unsuccessful depends on a random True/False result. The actions depend on random execution for most of such user queries so all the outcomes could be emulated.

Issues I faced during development

1. Rasa could not detect certain numbers in a statement.

   1. With default pipeline, Rasa could not detect phone number.
   2. But it detected account number and amount correctly.
   3. So, I played a little bot with the pipeline and switched to RegexEntityExtractor. It was worse. It detected numbers but did not correctly identify them as correct entities. For example, an account number is wrongly slotted into phone number.
   4. I have asked the problem in stack overflow, (https://stackoverflow.com/q/69626455/10582056), but no answer
   5. So as an ultimate solution, I have changed regex for phone, account, and amount (Initially there was no regex, I just gave some examples for Rasa to identify correct slots itself) and replaced RegexEntityExtractor with RegexFeaturizer.
   6. It works now, however in forums they said the inconsistencies in training is common, and sometimes it might not work well.

2. Sometimes slots are forgotten. a. The culprit is the MemoizationPolicy. I changed history to 7 and now it works. b. Also, for TEDPolicy, the epochs to train are increased to 200

Also, I have utilized most of Rasa’s core concepts such as rules, actions, NLU and domain. Some new things such as SlotSet took some time to work out, but eventually I got through it. I have added enough in-code comments to project as well.