JSP3

Semester project 3 for Linguistics

Goals

• analyze some text
• learn techniques for analyzing text
• learn some software engineering skills
• learn git

2025 task: Word Sense Disambiguation

Take a text, sense-annotate open-class words with a wordnet.

In previous work, we have done this manually:

Teaching Through Tagging — Interactive Lexical Semantics (Bond et al., GWC 2021)

• Do this automatically
  • evaluate how well this works EVAL
    • WHO:
    • all the WSD tasks need this
    • look at common errors
    • fix some (e.g. se/si)
  • test different contexts WSD-C
    • Effective context engineering for AI agents
  • test different wordnet information WSD-W
    • watch Let the LLM Write the Prompts: An Intro to DSPy in Compound AI Pipelines
  • test different LLM models WSD-M
    • look at settings to make it more efficient
    • e.g., prompt caching
  • find translations and use as context ALIGN
    • find epub, extract text, split into para, align
      • use: https://github.com/averkij/lingtrain-aligner/releases/tag/0.1.0
    • then add to WSD-C
  • textual criticism --- compare versions TEXT
    • CNK version
    • https://cs.wikisource.org/wiki/V%C3%A1lka_s_mloky
  • look at sentiment over the story SENTI
    • use general tool
    • use senses
    • compare
    • visualize
  • improve Czech wordnet EXPAND
    • add new senses to existing concepts using aligned data
    • verify with LLM?
    • create candidate definitions/examples?
  • add new Czech concepts NEW
    • add new suggestions for concepts to the hierarchy difficult

Goals

• get something useful for each task
• combine to make a best-of-breed
• write, submit and publish a paper
• release at least one automatically tagged, aligned corpus

Approach

• work on tasks in pairs
• use github to coordinate
  • GitHub quickstart
  • git an Introduction

Fortnightly meeting

• 5-10 minutes progress

• longer discussion of issues as necessary

• small meetings pair+me or WSD+me, .... as necessary

Next tasks

• ALL make github account
  • send me accountname
  • I will add to github
  • then add your name to a task
  • WSD --- try to run ollama on a prompt
  • ALIGN --- try to run align on chapter 1 of VsM
  • TEXT --- look for existing information on Karel Capek and versions, maybe ask Bohemian studies
• FCB
  • prepare databases and data
  • meet to set up eval, ...