Data & code for textual features influence on human sentiment perception in literary texts
| No. texts | No. annotations | No. words | Period | |
|---|---|---|---|---|
| Fairy tales | 3 | 772 | 18,597 | 1837-1847 |
| Hymns | 65 | 2,026 | 12,798 | 1798-1873 |
| Prose | 1 | 1,923 | 30,279 | 1952 |
| Poetry | 40 | 1,579 | 11,576 | 1965 |
We present the Fiction4 corpus of literary texts, spanning 109 individual texts across 4 genres and two languages (English and Danish) in the 19th and 20th century. The corpus consists of 3 main authors, Sylvia Plath for poetry, Ernest Hemingway for prose and H.C. Andersen for fairytales. Hymns represent a heterogenous colleciton from Danish official church hymnbooks from 1798-1873. The corpus was annotated for valence on a sentence basis by at least 2 annotators/sentence.
Full Fiction4 corpus data in \data\fiction4_data.json
We compare this fiction corpus again nonfiction texts (across genres)
The nonlit considered is:
- EmoBank (from this paper https://aclanthology.org/E17-2092/), repo here. So these are multigenre sentences. (n=10,062 & range=(1 to 674 toks) & mean_length=87.8 toks)
- Facebook posts (from this paper https://aclanthology.org/W16-0404.pdf), repo here. So these are facebook posts (multiple sentences)(n=2,895 & range=(2 to 445 toks) & mean_length=86.7 toks)
All code for our study on human/model sentiment perception across these corpora is available in this repository, see primarily feature extraction (get_features.py) and analysis (analysis.py).
Annotator agreement calculation for each subcategory of the Fiction4 corpus is in /annotation/annotator_agreement.py