Error Analysis: SPOT-Trankit on SSJ-UD test set

Model: SPOT-Trankit — CLARIN 11356/1997, run with trankit==1.1.2 using the Slovenian model archive trankit-sl-ssj+sst.zip (MD5 0ddfac8d7445f8fa300f59dde1a00352). CLASSLA-Stanza is included for reference and was run with classla==2.2.1 (classla.Pipeline('sl', pos_use_lexicon=True)). · Data: UD Slovenian SSJ test set, v2.17 (2025-10-22); predictions were run on pre-tokenised text (gold sentence and token boundaries supplied).

Overall Accuracy — SPOT-Trankit

LAS by Dependency Relation

Sorted by SPOT-Trankit LAS. Difference = SPOT-Trankit LAS − CLASSLA LAS.

Relation	Gold	SPOT-Trankit LAS	CLASSLA LAS	Diff

UPOS Accuracy by Tag

Sorted by SPOT-Trankit accuracy. Difference = SPOT-Trankit − CLASSLA.

UPOS	Gold	SPOT-Trankit	CLASSLA	Diff

Most Common Errors — Dependency Relations

SPOT-Trankit only. Click a row for examples. ↑↓ to navigate, Esc to close.

Most Common Errors — UPOS

SPOT-Trankit only. Click a row for examples.

Error pair	Count

Most Common Errors — XPOS (MSD)

SPOT-Trankit only. Hover or focus a row for MSD decode. Click for examples.

Error pair	Count

Most Common Errors — Lemma

SPOT-Trankit only. Click a row for examples.

Error pair	Count

Acknowledgement

This work was supported by the research projects SPOT (model development, ARIS grant no. Z6-4617) and MAPCASE (model evaluation, ARIS grant no. J6-70213). The interactive website was developed by Nives Hüll under the supervision of Kaja Dobrovoljc Zor.