The 2023 edition of Wikimedia Hackathon has just concluded, and I was one among the lucky participants to join on-site. It was a great feeling to re-connect with fellow Wikimedians after 4 long years of no-onsite events, and to be part of a community that thrives for free knowledge and code.
tl;dr: We built a user-script to summarise article sections on wiki pages. Check the code and demo.
Pre-hackathon statelite event in Stockholm
We had the idea of organizing a satellite event in Stockholm, and thanks to the support of WMSE (Wikimedia Sweden) at the opportune moment, we were able to secure the necessary space and time to host it. I would like to express my gratitude to Axel Petterson for assisting in the setup process. During this time, we also had a productive brainstorming session to determine the activities for the main hackathon.
Personally, I was eager to reconnect with the community after a period of relative inactivity, so it was important for me to gauge their interests. Lokal_Profil and Sebastian_Berlin graciously attended the event, and together we explored potential ideas for hacking during the main hackathon. Their expertise and insights into Wikispeech (text-to-speech functionality on wiki pages) sparked discussions related to improving accessibility and readability.
Towards the end of the event, with the help of Axel’s contributions and great refreshments, we came up with a feasible hackathon idea. Here are some refined iterations we discussed:
- Summarising info-boxes on wiki pages: We explored the possibility of enabling Wikispeech to read out loud the summarised content from info-boxes.
- Summarizing entire article pages: We considered developing a system to generate concise summaries of complete articles.
- Summarizing article sub-sections: We discussed using user-scripts and Wikimedia infrastructure to create summaries of specific sections within articles.
During our discussions, we made initial plans to leverage ChatGPT chat-completion endpoints for generating sub-section summaries. We also created an architecture diagram to visualize the implementation. To organize our efforts, we promptly created a Phabricator task, and we were fortunate to have Tgr join our team. Tgr suggested expanding the summarization feature to include Talk pages. Additionally, Tgr proposed the idea of using cookies or local storage to securely store OpenAI secrets, aiming to enhance user satisfaction and trust in the system.
Main hackathon at Technopolis
Athens is a beautiful place, and there is something historic on any corner you look. That was my impression of the city. We were accommodated and taken care of with a lot of planning and precision, and I thank the organising team for that. Technopolis was a great location as well, and we started the hackathon on time on the 19th. I was also part of the Trust and Safety team, and thanks to nothing going sideways, we had a relaxed hackathon.
During the idea pitching phase (strictly controlled to 30 seconds by Siebrand), I managed to pitch the idea in-front of the audience. It was not the clearest pitch, but straight after that, I got 1-2 people interested in the idea. Alexey had a lot of years under his belt in summarising wiki pages (mostly books, if I remember correctly), and then we had a couple of engineers who where working for the foundation (Wikimedia) along the lines of Machine Learning/ AI.
We found a table, and started working on how it would look like, and thanks to Alexey’s frontend/UI expertise, we got a very nice looking design fast.
I cannot mention enough the great fresh food, feta-cheese, tomatoes and drinks those were served continuously during the hacking days. Thank you.
By mid second day, we got Tgr too to sit with us, and he wrote a script that would extract well formatted text data from “Talk” sections on a wiki page. This was crucial, as until that point, we were relying on scraping text heading by heading to generate the text that has to be summarised.
On the third day, we realised that we have code in 3 repositories which were doing more or less the same thing, but better.
- Alexey had the one with the best UI (thanks to his expertise). He had also figured out how to stream ChatGPT responses right onto the page.
- Tgr had the best implementation to read and parse talk section pages (he used parsoid-html via API).
- I had maybe a bit improved version of scraping data under sections. Also, maybe a good way to continuously deliver updated scripts (via Github pages).
I got onto the task of merging the best of all these 3 together, and we managed to get https://github.com/tonythomas01/wikipedia-section-summaries/ before the demo. We created a small 30 second video (linked above), and managed to present it during the showcase. Since I do not like to see myself talking, I would not point to the time – but its somewhere in here 🙂
Future of the tool/script
We gathered a lot of feedback during the hackathon, and I can think about the following:
- Make the currently used ChatGPT prompt editable for users so that they can ask for further clarifications, or have a section explained in a different way.
- Send the script for evaluation, and maybe make it into a gadget.
- Explore if we can use any other free LLM models instead. There were a couple of talks about it during the hackathon.
- Refactoring of the script (in general), and rely on APIs for data, and not text scraping.
- Open up the script, so that it can be used for the Village Pump.
- Explore how to deliver the script in a better way, and not rely on Github pages.
Learnings
I would think about the following:
- Readability and accessibility is an area where generative AI can help Wikipedia a lot. By not necessarily dropping a lot of quality.
- From a hackathon point of view: Maybe we should have figured out a way to work together a bit better so that we did not end up with 3 versions on day 2. This is a challenge when multiple people are working on 1 user script. Hopefully, we can figure this out soon.
Other
I managed to explore the city around a bit. Managed to also buy a couple of olive oil bottles which we are using. Miss Athens, and hope to be back soon. Managed to make a lot more connections, and talk to a lot of hackers, students, and staff from the Ionian university as well. Cannot thank the organising team enough for taking care of us during the days.
Open and Free Source! 