README.md

metadata

language:
  - it
license: mit
size_categories:
  - 1K<n<10K
task_categories:
  - text-classification
  - feature-extraction
  - text-generation
pretty_name: Dataset Metafore
configs:
  - config_name: analisi
    data_files: analysis/*.csv
  - config_name: analysis/min_1b_logprobs
    data_files:
      - split: train
        path: analysis/min_1b_logprobs/train-*
  - config_name: analysis/min_3b_logprobs
    data_files:
      - split: train
        path: analysis/min_3b_logprobs/train-*
  - config_name: clustering
    data_files: clustering/*.csv
  - config_name: default
    data_files:
      - metafore/*.csv
      - results/interpretazioni_esperimento.csv
  - config_name: metafore
    data_files: data/metafore.csv
  - config_name: partecipanti
    data_files: data/partecipanti_dati_anagrafici.csv
  - config_name: risultati
    data_files: results/*.csv
tags:
  - metaphor
  - psychology
  - italian
  - linguistics
dataset_info:
  - config_name: analysis/min_1b_logprobs
    features:
      - name: m_id
        dtype: string
      - name: metafora
        dtype: string
      - name: interpretazione
        dtype: string
      - name: model_score
        dtype: float64
      - name: probability_normalized
        dtype: float64
      - name: __index_level_0__
        dtype: int64
    splits:
      - name: train
        num_bytes: 541701
        num_examples: 2540
    download_size: 120341
    dataset_size: 541701
  - config_name: analysis/min_3b_logprobs
    features:
      - name: m_id
        dtype: string
      - name: metafora
        dtype: string
      - name: interpretazione
        dtype: string
      - name: model_score
        dtype: float64
      - name: probability_normalized
        dtype: float64
      - name: __index_level_0__
        dtype: int64
    splits:
      - name: train
        num_bytes: 541701
        num_examples: 2540
    download_size: 120370
    dataset_size: 541701

📊 Dataset Metafore

Dataset title: Dataset Metafore
Related paper: Language Models and the Magic of Metaphor: A Comparative Evaluation with Human Judgments — S. Mazzoli, A. Suozzi, G. E. Lebani (CLiC-it 2025).

This repository contains CSV files used in the study described above. The data were collected from human participants who provided metaphor interpretations and ratings, and from several Italian-trained language models whose normalized log-probabilities were computed on human interpretations and systematically constructed distractors.

---

🗂 Overview

The dataset contains 140 metaphors extracted from Italian parliamentary transcripts (Chamber of Deputies, 2008–2022). Human participants provided free-form interpretations for each metaphor and rated (1–5) the conventionality of the expression and the adequacy of the sentence context. For each metaphor there are two distractors created programmatically: a Literal Distractor (LD) and an Opposite Metaphorical Distractor (OMD). Several autoregressive LLMs (GePpeTto, Minerva family, LLaMAntino) were evaluated by computing log-likelihoods for each human interpretation and distractor; normalized log probabilities and ranks are provided.

Refer to the paper for full methodology and analyses.

---

Organized variable list (by file)

dataset-metafore.csv — master / metadata for each metaphor

Description: one row per metaphor (n = 140). This file contains the main experimental items, their balancing information, prompts shown to participants, and aggregated human judgment statistics.

Column	Type	Description
questionario	string	Questionnaire identifier in which the item appeared (e.g., Q1–Q10)
m_id	string	Metaphor identifier (e.g., M1, M42)
gruppo_bilanciamento	string	Balancing group (G1–G7) corresponding to the syntactic pattern distribution
item	string	Full sentence containing the metaphor
v_term	string	Lexical item annotated as metaphorical
pattern	string	Syntactic pattern of the metaphor (e.g., N₁ di N₂, V ∼ N)
valenza	string	Valency of the metaphorical element (nessuno, intransitiva, transitiva)
classe_lessicale	string	Part of speech of the metaphorical element (e.g., N₁, Verbo, Aggettivo)
prompt	string	Sentence frame shown to participants for interpretation elicitation
avg_conv	numeric	Mean conventionality rating (1–5)
avg_cntxt	numeric	Mean context adequacy rating (1–5)
entropia di shannon	numeric	Shannon entropy of interpretation distribution (measure of variability)
consenso_max	numeric	Maximum consensus score (relative frequency of most common interpretation)
n_int	integer	Number of human interpretations collected for the item

---

Balanced syntactic groups (dataset)

Pattern	Valency	Metaphorical element (PoS)	Group size (n = 140)
N₁ di N₂	None	Noun₁	20
N ∼ Adj	None	Noun	20
N ∼ Adj	None	Adjective	20
N₁ = N₂	None	Noun₂	20
V ∼ N (intransitive)	Intransitive	Verb	20
V ∼ N (transitive)	Transitive	Verb	20
V ∼ N (transitive)	Transitive	Verb and Noun	20

---

❝ Cite

@inproceedings{mazzoli-etal-2025-language,
    title = "Language Models and the Magic of Metaphor: A Comparative Evaluation with Human Judgments",
    author = "Mazzoli, Simone  and
      Suozzi, Alice  and
      Lebani, Gianluca",
    editor = "Bosco, Cristina  and
      Jezek, Elisabetta  and
      Polignano, Marco  and
      Sanguinetti, Manuela",
    booktitle = "Proceedings of the Eleventh Italian Conference on Computational Linguistics (CLiC-it 2025)",
    month = sep,
    year = "2025",
    address = "Cagliari, Italy",
    publisher = "CEUR Workshop Proceedings",
    url = "https://aclanthology.org/2025.clicit-1.68/",
    pages = "710--721",
    ISBN = "979-12-243-0587-3"
}

---