This model projects crossword clues and candidate answers into a shared 768-dimensional latent space, enabling efficient retrieval via FAISS inner-product search.

Model Description

Architecture: Asymmetric Dual Encoder with two separate XLM-RoBERTa encoders (one for clues, one for answers), a shared LayerNorm, and a shared linear projection head.
Base encoder: sentence-transformers/paraphrase-multilingual-mpnet-base-v2 (278M parameters per tower)
Training objective: Symmetric contrastive loss (InfoNCE) with in-batch hard negative mining and learnable temperature
Embedding dimension: 768
Training data: Italian crossword clues and dictionary definitions paired with their answers/defined words

For more details, see the paper: Crossword Space: Latent Manifold Learning for Italian Crosswords and Beyond (CLiC-it 2025). This model corresponds to our best configuration, which includes dictionary items during training.

Usage

Loading the model

python

from transformers import AutoModel

model = AutoModel.from_pretrained(
    "cruciverb-it/crossword-space-mpnet-base-ade",
    trust_remote_code=True,
)

Or, if you have the crossword-space repository cloned locally:

python

from model import DualEncoderADE

model = DualEncoderADE.from_pretrained("cruciverb-it/crossword-space-mpnet-base-ade")

Full example

python

import torch
import torch.nn.functional as F
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained(
    "cruciverb-it/crossword-space-mpnet-base-ade",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("cruciverb-it/crossword-space-mpnet-base-ade")
model.eval()

clues = ["Capitale d'Italia", "Fiume che attraversa Roma"]
answers = ["ROMA", "TEVERE"]

clue_enc = tokenizer(clues, padding=True, truncation=True, max_length=64, return_tensors="pt")
ans_enc = tokenizer(answers, padding=True, truncation=True, max_length=16, return_tensors="pt")

with torch.no_grad():
    clue_emb, ans_emb = model(
        def_input_ids=clue_enc["input_ids"],
        def_attention_mask=clue_enc["attention_mask"],
        ans_input_ids=ans_enc["input_ids"],
        ans_attention_mask=ans_enc["attention_mask"],
    )

# L2-normalize for cosine similarity / inner-product search
clue_emb = F.normalize(clue_emb, dim=-1)
ans_emb = F.normalize(ans_emb, dim=-1)

# Similarity matrix
similarity = clue_emb @ ans_emb.T
print(similarity)

Evaluation Results

Retrieval performance on four Italian test sets. Length-filtered metrics restrict the FAISS index to candidates matching the expected answer length, which is the standard setting for crossword solving.

Full Index

Test Set	Acc@1	Acc@10	Acc@100	Acc@1000	MRR
Crossword	31.8	63.4	81.3	91.0	42.7
Dictionary	17.5	40.5	63.3	82.2	25.3
ONLI	11.8	34.9	61.0	81.5	19.7
Neologisms	9.0	25.0	59.0	82.0	14.6

Length-filtered Index

Test Set	Acc@1	Acc@10	Acc@100	Acc@1000	MRR
Crossword	54.0	80.2	91.3	97.2	63.6
Dictionary	35.6	62.6	82.5	95.7	45.0
ONLI	36.0	65.0	84.3	95.8	45.8
Neologisms	28.0	60.0	84.0	94.0	38.1

Citation

bibtex

@inproceedings{ciaccio-etal-2025-crossword-space,
    title = "Crossword Space: Latent Manifold Learning for Italian Crosswords and Beyond",
    author = "Ciaccio, Cristiano and Sarti, Gabriele and Miaschi, Alessio and Dell'Orletta, Felice",
    booktitle = "Proceedings of the Eleventh Italian Conference on Computational Linguistics (CLiC-it 2025)",
    year = "2025",
    url = "https://aclanthology.org/2025.clicit-1.26/"
}

License

CC BY 4.0