Presupposed Taxonomies: Evaluating Neural Network Semantics (PreTENS)

Original Paper: https://aclanthology.org/2022.semeval-1.29.pdf

This dataset comes from SemEVAL-2022 shared tasks.

The PreTENS task aims at focusing on semantic competence with specific attention on the evaluation of language models with respect to the recognition of appropriate taxonomic relations between two nominal arguments.

We collected the Italian part of the original dataset, and more specifically only the first sub-task: acceptability sentence classification.

Example

Here you can see the structure of the single sample in the present dataset.

{
  "text": string, # sample's text
  "label": int, # 0: non ha senso, 1: ha senso
}

Statitics

PRETENS	0	1
Training	3029	2808
Test	7707	6853

Proposed Prompts

Here we will describe the prompt given to the model over which we will compute the perplexity score, as model's answer we will chose the prompt with lower perplexity. Moreover, for each subtask, we define a description that is prepended to the prompts, needed by the model to understand the task.

Description of the task: "Indica se le seguenti frasi hanno senso.\n\n"

Cloze Style:

Label (non ha senso): "{{text}}\nLa frase precedente non ha senso"

Label (ha senso): "{{text}}\nLa frase precedente ha senso"

MCQA Style:

{{text}}\nDomanda: La frase precedente ha senso senso? Rispondi sì o no:

Results

The following results are given by the Cloze-style prompting over some english and italian-adapted LLMs.

PRETENS	ACCURACY (15-shots)
Gemma-2B	53.5
QWEN2-1.5B	56.47
Mistral-7B	66.5
ZEFIRO	62
Llama-3-8B	72.34
Llama-3-8B-IT	65.58
ANITA	66.1

Aknowledgement

We would like to thank the authors of this resource for publicly releasing such an intriguing benchmark.

Additionally, we extend our gratitude to the students of the MNLP-2024 course, whose first homework explored various interesting prompting strategies.

The original dataset is freely available for download here.

License

The data come under MIT license.