Prototyping a map interface for Tamil Watch Dog

• The TWD database is a collection of news articles, tweets, and other sources of information about the activities of Tamil Nationalism.

Prototype

• The database is scrapes from websites and stored in a CSV
• SpaCy is used to extract named locations (NER) from the text
• The lowest denominator location is selected and geocoded. This might be a cause for concern as the geocoding is not always accurate. Choices have to be weighted better
• It is then mapped using Leaflet,
• Svelte and Svelvet are used to create interactions and a narrative maker.

Roadmap

• Refine geocoding to be more accurate.
• Persistent state and sharing for narratives created on the narrative maker
• Research on event classification and other narrative building elements.

LLM models

• Alpaca, AlpacaLora, LLaMa etc... (https://huggingface.co/Pi3141/alpaca-native-13B-ggml/tree/main)
  • Both the 7B and 13B are highly dependant on the prompts, they need to be explored

• Instructor Embedder (https://huggingface.co/hkunlp/instructor-large)

• Electra (https://huggingface.co/spaces/brijw/transformers_ner/blob/main/app.py)