readme: add initial version

README.md

@@ -1,3 +1,36 @@
----
-license: unknown
----
+---
+license: unknown
+language:
+- de
+tags:
+- historical
+- newspapers
+---
+
+# SentiAnno: A Sentiment-Annotated Corpus of Austrian Historical Newspapers
+
+This repository hosts training, development and test splits for the recently introduced "SentiAnno" dataset from the ["Constructing a Sentiment-Annotated Corpus of Austrian Historical
+Newspapers: Challenges, Tools, and Annotator Experience"](https://aclanthology.org/2024.nlp4dh-1.6/) paper by [Lucija Krušic](https://huggingface.co/lukru).
+
+More from the paper:
+
+> This study presents the development of a sentiment-annotated corpus of historical newspaper texts in Austrian German, addressing a gap in annotated corpora for Natural Language Processing in the field of Digital Humanities. Three annotators categorised 1005 sentences from two 19th-century periodicals into four sentiment categories: positive, negative, neutral, and mixed. The annotators, Masters and PhD students in Linguistics and Digital Humanities, are considered semi-experts and have received substantial training during this annotation study.
+
+# Dataset Stats
+
+We create a 80/10/10 dataset split from the gold standard annotations, using the [`0ecb222`](https://github.com/lucijakrusic/SentiAnno/tree/0ecb2228e6c290dd22836024f32e559cc9b9711e) revision.
+
+Label negative has 447 sentences
+Label mixed has 56 sentences
+Label positive has 81 sentences
+Label neutral has 345 sentences
+
+For each label category (`positive`, `negative`, `neutral`, `mixed`) this dataset split ratio is performed leading to:
+
+* 741 training examples
+* 93 development examples
+* 95 test examples
+
+# License
+
+License is still be to cleared out, for now it is "unknown".