Helpium

Inspiration

Everyday, hundreds of thousands of people ask questions about Canadian government services (benefits, immigration and citizenship, taxes etc). Even though the official website provides some information and instructions. We wanted to create a platform for helping low literacy people, non-native and second language learners, and those struggling with various types of reading comprehension problems to access these services.

In many cases, this work is often done by the english speaking children of first generation immigrants, or other family members and friends who may want to help but not know the language well themselves.

What it does

At the moment, Helpium can write drafts for emails intended to request information from government services. It cannot answer these questions directly, because it doesn't know - but hopefully it can help people ask for information they need from the right people.

How we built it

Powered by a large pre-trained text-to-text model offered by Cohere, we trained the model using few-shot learning based on examples we collected from people seeking support online (mostly on Canadian government websites and reddit posts). We read these posts, rephrased them as simple questions and used text generation tools to create letter templates that looked suitable for a small number of cases.

We then created a UI in R Shiny that dynamically constructs a prompt that instructs the Cohere model how to generate letters we thought looked useful and hooked it all together using their API.

Challenges we ran into

1. Training: As training examples (question-email template pairs specifically for Canadian government services) are hard to get(or scrape : D) within one day and require a lot of human annotations, fine-tuning an existing model on our tasks is not feasible. To address this problem, we have utilized the strategy of few-shot learning and have take advantage of the power of large language models.

2. Visualization: to build an interactive web app, we use Shiny(an R package) but since most of the code was written in Python for Cohere API, we need to adapt our code from Python to R.

3. Telling the truth! By default, these models make claims that are not always true. We removed such claims from the few-shot examples we accumulated but this offers no guarantees that the model will stay on topic. We hope that we can also teach users how to possibly navigate multiple drafts and pick out the relevant bits for their situation.

Accomplishments that we're proud of

We are proud of everything we have done :D

What we learned

• Use AI to help the community
• Prompt engineering
• API rendering
• Web app building

What's next for Helpium

• Add question-answering on this platform so that people can get their response fast. (Actually this feature is almost done but we don't have enough time to put in on the web app sadly. We will add it afterwards !)
• Improve the accuracy and factuality of the generated answer
• Include the options of customization of the email generation

Built With

• cohere
• openai
• python
• shiny