Aroma Rodrigues: The limits of imagination

Historically we have all used computers mathematically, but now we get to use them linguistically. I will tell some stories of my work over time.

PyCon India: Terms and Conditions summarizer. You have an unreadable (for you) legal document. You get an email saying your rights have changed, but hardly anyone ever reads that. You could hire a lawyer, but I wanted to use a computer. Problem: no existing dataset. Approach: build a labeled dataset from scratch. Outcome: an NLP pipeline to extract obligations, permissions, and risks from such a doc. But the dataset was not really usable.

PyCon US 2024: Only bad demos in the building. With all the LLMs and AI, I was wondering what I would do if I could not keep my job? Maybe I could become a forensic portrait artist, so creating pictures of suspects of a robbery? I started experienting with giving a description of a man, and using an LLM to generate a picture. Kind of worked for some descriptions.

I also let an AI make a travelogue based on my photos. You run into things like the AI saying "I saw a car coming down the road" when I made the photo not for the car, but for the buildings.

Clothes matching: can AI do it? So picking out pieces of clothes that match nicely together. I got something together.

So basically, I can do anything!

PyCon-ZA 2019: NLP Fake News detector. There is improvement: today, LLMs offer richer context better understanding of blame dynamics, and scalable solutions. My source: 22M conversations from BuzzFeed on top 50 fake stories. Compare articles from spoof website and mainstream media. Use of fact-checking platforms like Alt News and SMHoaxSlayer. Looking for syntactic patterns, tagging statements to detect blame assignment, praise, event causality, active/passive voice patterning using nltk.RegExpParser.

PyCon Estonia 2023: If your friends are bullshitting, using SNLI. Goal: use LNP to spot contradictions in statements, proving when your friends are contradicting themselves. Check for two sentences: Entailment (they are roughly the same), contradition (they contradict), or neutrality (they are about something else).

I have lots more to tell, but time is up.