tupa
tupa copied to clipboard
danielhers
Transition-based UCCA Parser
Since both ucca and semstr require spaCy 2.2, which is incompatible with 2.1, all pre-trained TUPA models (trained with spaCy 2.1) will be unusable and need to be re-trained. See...
@OfirArviv, this is an issue we've been talking about a bit. I guess it could be called a DyNet issue but I'll try working around it. In the [mrp](https://github.com/danielhers/tupa/tree/mrp) branch,...
Since punctuation has a specific location it has to appear in (according to UCCA normalization rules, it has to be a child of the lowest common ancestor of its preceding...
Looks like new sub-configs and sub-namespaces keep getting added. Getting this when training on multiple formats (multitask) with many epochs, during dev evaluation: ``` ... File "tupa/config.py", line 344, in...
Constraints: - Top level must be F, L, H, G or U - L must also be LR - H must have P or S - F must not have...
TUPA often predicts a node with `A` (participant) children but no `P`/`S` (process/state). There are several ways to tackle this problem: 1. Add constraints on `Reduce` to force creating this...
TUPA often attached elements, such as Ls, Rs, and Ds, too high --- right to the top scene node instead of to the correct sub-unit. This is because when a...
The parser currently works in greedy mode only. In order to improve its accuracy, beam search should be helpful. This will require adding the `IDLE` transition (as in [Zhu et...
The distinction between P (process) and S (state) is unreliable: there is a low agreement on it and TUPA tends to make errors on them too. For example, in German,...