ghigliottinAI MCQA

References:

• https://ghigliottin-ai.github.io/
• https://nlp4fun.github.io/

Starting from two different EVALITA tasks, nlp4fun (EVALITA 2018) and ghigliottin-AI (EVALITA 2020), we collected cc. 600 different games extracted from TV show and from BOARDGAME of "L'Eredità".

"La Ghigliottina" is a complex game, to be solved, it needs a very large comprehension of the italian cultural knowledge. It consists in: given five different, uncorrelated words, the solution is a word that is a shared concept between them. The original game itself is not well-posed, the solution is not unique, and list all the possible solution is not a affordable. We decided to reframe the problem as a Multi-choice QA, where four possible words are listed and between them all but one are incorrect answers to the game.

Distractor Generation

For each game the three distractor was chosen among all the possible italian words, the distractor was chosen to be aligned with 3 out of 5 hints and distant to the other ones (computing the cosine similarity in FastTest static embeddings). Moreover, the distractors was chosen to have lenght at most len(solution) + 1.

With this setting, we created three different words that are not the possible solution of the game, making a task relativelly simple to be solved by humans, but not that much for Language Models.

Example

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

{
  "w1": string, # text of the first hint
  "w2": string, # text of the second hint
  "w3": string, # text of the third hint
  "w4": string, # text of the fourth hint
  "w5": string, # text of the fifth hint
  "choices": list, # list of possible words, with the correct one plus 3 distractors
  "label": int, # index of the correct answer in the choices
}

Statistics

Training: 62

Test: 553

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:

Ti viene chiesto di risolvere il gioco della ghigliottina.\nIl gioco della ghigliottina consiste nel trovare un concetto che lega cinque parole date. Tale concetto è esprimibile tramite una singola parola.\n\n

MCQA style

Prompt:

Date le parole: {{w1}}, {{w2}}, {{w3}}, {{w4}}, {{w5}}\nDomanda: Quale tra i seguenti concetti è quello che lega le parole date?\nA. {{choices[0]}}\nB. {{choices[1]}}\nC. {{choices[2]}}\nD. {{choices[3]}}\nRisposta:

Cloze style

In this case the gold answer is not the corresponding letter but the word itself.

Date le parole: {{w1}}, {{w2}}, {{w3}}, {{w4}}, {{w5}}\nDomanda: Quale tra i seguenti concetti è quello che lega le parole date?\n{{choices[0]}}\n{{choices[1]}}\n{{choices[2]}}\n{{choices[3]}}\nRisposta:

Results

Here some results are reported from the two prompting strategies

GhigliottinAI-MCQA	ACCURACY (5-shots)
Gemma-2B	23.86
QWEN2-1.5B	39.24
Mistral-7B	42.49
ZEFIRO	40.86
Llama-3-8B	46.65
Llama-3-8B-IT	47.38
ANITA	41.95

GhigliottinAI-CLOZE	ACCURACY_norm (5-shots)
Gemma-2B	35.08
QWEN2-1.5B	33.81
Mistral-7B	39.60
ZEFIRO	41.22
Llama-3-8B	43.39
Llama-3-8B-IT	48.46
ANITA	48.64

Acknowledge

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

Further, We want to thanks the student of MNLP-2024 course, where with their first homework tried different interesting prompting strategies, reframing strategies, and distractor generation approaches.

The original dataset is freely available for download link_1, link_2.

License

No license found on original data.