A harmonised testsuite for POS Tagging
of German social media data

Data sets

A harmonised POS testsuite of web data, CMC and Twitter microtext

• download the harmonised dataset (56,572 tokens)
• with word forms and STTS pos tags (+ some additional CMC-specific tags)
• UD pos tags have been automatically converted, based on the STTS pos tags
• no lemma information (or, for parts of the data, automatically predicted lemmas)

The original data comes from 3 different sources:

• a twitter dataset with 21,181 tokens (Rehbein 2013)
• two datasets from the Empirist shared task 2015
  • web data (12,718 tokens)
  • computer-mediated communication (10,505 tokens)
• tweeDe: a twitter dataset with 12,156 tokens (Rehbein, Ruppenhofer & Zimmermann 2018)

If you only want to download the new Twitter testsuite:

tweeDe (POS)

• the tweeDe dataset, split up into train/dev/test set
• 56,572 tokens

POS tagging models for German social media

HunPos model

• pre-trained model for the HunPos tagger (Halácsy et al. 2007)
  (including a readme)

Bi-LSTM-char-crf model

• pre-trained model for the biLSTM-CRF tagger (Reimers & Gurevych 2017)
  (including a readme)

Twitter SkipGram Embeddings

• download from here

Results

text type	HunPos	biLSTM-char-CRF
TiGer	96.48	97.56
web	93.73	93.93
cmc	89.49	91.44
twitter (GSCL)	91.23	92.20
tweeDe	93.10	92.31