Manually annotated corpora

From Clarin K-Centre
Jump to navigation Jump to search

Manual corpora are collections of texts containing manually validated or manually assigned linguistic information, such as morphosyntactic tags, lemmas, syntactic parses, named entities etc. These corpora can be used to train new language annotation tools as well as to test the accuracy of existing annotation tools.

Corpus Gesproken Nederlands

This corpus is at manually transcribed and annotated for all its layers.

Eindhoven Corpus

The Eindhoven corpus (VU version) is a collection of Dutch written and transcribed spoken texts from the period from 1960 to 1976. The corpus contains approximately 768,000 tokens. It is the first collection of Dutch written and (transcribed) spoken texts.

Lassy Small

The Lassy Small Corpus is a corpus of approximately 1 million words with manually verified syntactical annotations. The lemmas and POS-tags were generated with Tadpole (now Frog) and the syntactical depency structures were generated with Alpino. The lemmas, POS-tags and syntactic tree structures were manually verified and corrected.

SoNaR Small Corpus

The SoNaR Small Corpus Commercial contains approx. 825.000 words which were manually semantically annotated. This is a subset of the much larger (approx. 500 million word) SoNaR corpus.

ACTER: Annotated Corpora for Term Extraction Research

The ACTER (Annotated Corpora for Term Extraction Research) is an annotated dataset for term extraction. Terms and Named Entities have been manually annotated in specialised comparable corpora covering 3 languages (English, French, and Dutch), and 4 domains (corruption, dressage, heart failure, and wind energy).

  • CLARIN page
  • Github page
  • Rigouts Terryn, Ayla, 2019, ACTER (Annotated Corpora for Term Extraction Research) v1.3, Eurac Research CLARIN Centre

Dutch Archaeology NER Training Dataset

A manually annotated NER dataset, consisting of Dutch archaeological excavation reports. The following entity types are labelled: Artefacts, Time periods, Materials, Places (geographical locations), Archaeological contexts and Species. The dataset is provided in the BIO format, with each token on 1 line and empty lines denoting sentence boundaries. On each line you can find the token, PoS tag, morphological segmentation and finally the label, separated by spaces. The PoS tag and morphological segmentation are assigned by Frog.


This data repository contains an English-Dutch machine translation (MT) of the detective novel "The Mysterious Affair at Styles" by Agatha Christie. The translation was generated in May 2019 by Google's neural MT system. The complete novel (5,276 sentence pairs) has been manually annotated with translation errors using an adapted version of the SCATE MT error taxonomy (Tezcan et al. 2017).

  • Download page
  • Tezcan, A., Hoste, V., & Macken, L. (2017). SCATE taxonomy and corpus of machine translation errors. In G. C. Pastor & I. Durán-Muñoz (Eds.), Trends in E-tools and resources for translators and interpreters (Vol. 45, pp. 219–244). Brill | Rodopi.

Pilot Annotation Corpus PLATOS NL-EN

This annotated corpus consists of 100 user-comments sourced on Facebook in the Summer of 2020, on the Facebook page of a popular flemish newspaper (Het Laatste Nieuws, corpus contains both the original comments in Dutch and a parallel English translation (Google Translate output lightly edited for correctness by a native speaker). The annotation covers topic and aspect labelling, stance labelling, argumentativeness detection and claim identification.

  • Download link
  • Bauwelinck, N., & Lefever, E. (2020, December). Annotating Topics, Stance, Argumentativeness and Claims in Dutch Social Media Comments: A Pilot Study. In Proceedings of the 7th Workshop on Argument Mining (pp. 8-18).


The EventDNA corpus is a Dutch-language corpus comprising 1,773 news documents in which news events, entities, IPTC Media Topic codes and coreference links have been manually annotated.


The EmotioNl corpus is an emotion-annotated corpus of Dutch texts. It consists of 1,000 Dutch tweets and 1,000 captions from reality TV-shows, annotated with the categories anger, fear, joy, love, sadness and neutral, and scores (real-valued, between 0 and 1) for the emotional dimensions valence, arousal and dominance.