Extended commodity ID module to cover the ~400 most common terms in the Girsu corpus. Sample outputs suggest we get ~85% accuracy for the commodity labeling task, but precision is closer to 95% because of focus on rules which are not overly general. Implemented labeling of modifiers and vessels, allowing distinction between different subtypes of items (male vs female animals, for example).
| # | Day | Date | A short description of the work done |
|---|---|---|---|
| 1 | Monday | 2020/06/01 | Finish checking wordnet performance: commodity code now covers 400 common terms (26454 total tokens). Testing suggests a need to refactor to jointly classify all words in an entry, rather than iteratively classifying one at a time. |
| 2 | Tuesday | 2020/06/02 | Restructured classification code to allow considering all words in an entry. Helps with cases where e.g. mun “salt” can be counted on its own, or can occur as a modifier describing fish. It is only the counted object in the former case. |
| 3 | Wednesday | 2020/06/03 | Improved ability to distinguish adjectives/modifiers from commodities. Tests show need to consider full tablet to disambiguate some implied objects (eg ration texts). Fixed handling of some vessels and metals. |
| 4 | Thursday | 2020/06/04 | Meeting. Implement handling of vessels in commodity ID module. |
| 5 | Friday | 2020/06/05 | Clean, document, and refactor dev branch of commodity ID module. Merge with master branch. |
| 6 | Saturday | 2020/06/06 | |
| 7 | Sunday | 2020/06/07 |