(17.01.2022 - 30.01.2022)

Further Fine-Tuning Optimisation

Using an agile development methodology incorporates a high degree of iteration. We prioritised getting a working pipeline complete before perfecting either side. Delivering a project this way means we need to test the working pipeline, find areas to improve and then adjust and adapt the code and strategy through each iteration. After several iterations of the project, this will lead to huge improvements in its functionality.

In our latest iteration, to improve the fine-tuning aspect of the project we have done 2 things:

1. Additions to Sequential Question Answering (SQA)
   • By testing our model we spotted several questions streams that the bot couldn’t produce a confident answer for.
   • To combat this we generate our own questions on a data set that mirror the ones the bot struggles with together to the correct answer.
   • We do this several times so the bot can learn with different variations how to answer certain types of questions given certain data.
2. Increasing data and variety of data
   • Secondly, we’ve increased the number of medical tables we input into the model for training. This should help the bot further familiarise itself with the language.

After making the following changes we then tested again to check for an improvement. We could clearly see a higher proportion of tests were passing and the bot could handle a larger variety of question streams.

Bot User Interface

We deployed the Rasa Bot to a webpage, which is currently only accessible from a localhost, in order to effectively allow users to interact with the model. The conversation begins with a greeting from the bot. A prompt to upload a report appears after the user's first communication. The client can attach their clinical report in a csv file to further investigate the details of the given document. In the future, clients will be able to use files in the txt and pdf formats to facilitate their interactions with the bot. After the material has been uploaded, the user can finally ask pertinent questions and receive accurate replies based on the test results data.