Sparv-Superlim: A plugin for classifying text using the models trained on tasks in Superlim

Sparv-Superlim is a Sparv plugin for classifying text using the Superlim baseline models. Superlim is a multi-task benchmark for Swedish, which includes baseline models.

How to use?

Install Sparv-Superlim by injecting it into the Sparv Pipeline:

pipx inject sparv-pipeline git@github.com:spraakbanken/sparv-sbx-superlim.git

See the Sparv documentation for more details on how to install plugins.

Then make a config file and choose the relevant annotations:

metadata:
  id: corpora
  name:
    eng: corpora
    swe: korpora
  language: swe
  description:
    eng: Swedish political manifestos with Superlim annotations
    swe: Svenska valmanifest med Superlimannoteringar
import:
  source_dir: source_small
  importer: text_import:parse
export:
  default:
    - xml_export:pretty
    - sbx_superlim:predictions
  annotations:
    - <token>
    - <sentence>
    - <sentence>:sbx_superlim.migration_stance
    - <sentence>:sbx_superlim.nuclear_stance
sbx_superlim:
  hf_model_path:
    absabank-imm: 'sbx/bert-base-swedish-cased_absabank-imm'
    argumentation: 'sbx/bert-base-swedish-cased-argumentation_sent'
  hf_inference_args:
    batch_size: 32

Plugin-specicific variables which start with hf are HuggingFace parameters. The most important one is the hf_model_path which tells which fine-tuned model to use for each task.

Full working examples can be found in the examples folder.

Available annotations

So far, Sparv-Superlim provides 10 different annotations. These are summarized in the table below:

Superlim task	Sparv-Superlim Annotation	Annotation	Label	Segment
absabank-imm	migration_stance	Attitude towards immigration	float between 1-5	sentence
argumentation- sentences	[topic]_stance	Stance to a given topic	pro, con or neutral	sentence
dalaj-ged	correct_swedish	Correct Swedish	correct or incorrect	sentence
swenli	previous_entailment	The logical relationship of two sentences	entailment, contradiction or neutral	sentence pair
sweparaphrase	similarity	Similarity between two sentences	float between 1-5	sentence pair

Wish to contribute?

Do you have new, innovative ways of incorporating models trained on Superlim into Sparv-Superlim? Make a feature request or even better a pull request!

How to cite?

Please cite the following technical report: Felix Morger. 2024. When Sparv met Superlim…A Sparv plugin for natural language understanding analysis of Swedish. Tech. rep. University of Gothenburg. You can also use the bibtex entry below.

@techreport{sparv-superlim,
  title =	 {When {S}parv met {S}uperlim\ldots {A} {S}parv Plugin for Natural Language Understanding Analysis of {S}wedish},
  author =	 {Morger, Felix},
  url = {https://hdl.handle.net/2077/83664},
  year =	 {2024},
  publisher = {Språkbanken Text},
  institution =	 {University of Gothenburg},
}