hcm777/query-embedding-mix-word-mix

🌐 [ACL 2026] Query Embedding Mix Word-Mix Query Bundles

Tongyao Zhu, Chao-Ming Huang, and Min-Yen Kan

  

<p align="center"> <img src="https://raw.githubusercontent.com/tongyao-zhu/query-embedding-mix/main/assets/figures/teaser_diagram.png" alt="Study protocol: monolingual, word-mixed, and embedding-mixed query representations" width="72%"> <br> _{This dataset packages the word-mix validation artifacts released with the ACL 2026 study.} </p>

This dataset repository accompanies our ACL 2026 paper on mixed-language queries in multilingual dense retrieval. It packages the word-level code-mixed query bundles used in the appendix validation workflow, where word-mix is used as a probe to check whether embedding-level interpolation follows similar ratio trends.

The release covers four validation pairs:

EN-ZH, EN-VI, ZH-VI, HI-ID

Each pair is available as:

• —a Hugging Face configuration backed by a normalized Parquet file under data/
• —the original TSV bundle under raw/
• —per-pair metadata and checksums under metadata/

What You Will Find Here 📦

data/                  Parquet files for Hub loading and dataset viewer support
raw/                   Original TSV bundles copied from the released artifact directories
metadata/              Pair manifests, checksums, and staging summary
README.md              Dataset card tying this release back to the paper and code repo

Pair Summary 🔎

The Parquet files are the recommended interface for load_dataset(). The raw TSV bundles are preserved for provenance and direct inspection.

Load From Hugging Face 🤗

Load one pair as a Hugging Face dataset configuration:

from datasets import load_dataset

ds = load_dataset("hcm777/query-embedding-mix-word-mix", "en-zh", split="train")
common_band = ds.filter(lambda row: row["band_label"] == "40-60" and row["is_common_qid"])

Download the raw TSV bundle for a pair:

from huggingface_hub import snapshot_download

local_dir = snapshot_download(repo_id="hcm777/query-embedding-mix-word-mix", repo_type="dataset")

Row Schema

Each Parquet row corresponds to one (source_file, qid, text) entry from the raw bundle.

Repository Structure 🗂️

README.md
data/
  en-zh.parquet
  en-vi.parquet
  zh-vi.parquet
  hi-id.parquet
raw/
  en-zh/
  en-vi/
  zh-vi/
  hi-id/
metadata/
  en-zh.json
  en-vi.json
  zh-vi.json
  hi-id.json
  summary.json

How This Maps To The Repo 🧭

Provenance 📚

• —Source code repository: https://github.com/tongyao-zhu/query-embedding-mix
• —Paper: Tongyao Zhu, Chao-Ming Huang, and Min-Yen Kan. "When Does Mixing Help? Analyzing Query Embedding Interpolation in Multilingual Dense Retrieval." Proceedings of the Annual Meeting of the Association for Computational Linguistics, 2026.
• —Generator used for the raw bundles: query_embedding_mix/generate_cm_bands.py

Notes

• —This dataset is a paper-facing artifact release, not a train/test benchmark package.
• —The raw TSV files are preserved under raw/ for provenance and direct inspection.
• —The Parquet files are the recommended interface for loading from the Hub.
• —Before publishing publicly, confirm redistribution of the source query text and generated variants is compatible with the upstream data terms you intend to rely on.

Citation ✍️

If this dataset is useful for your research, please cite our ACL 2026 paper:

@inproceedings{zhu2026queryembeddingmix,
  title     = {When Does Mixing Help? Analyzing Query Embedding Interpolation in Multilingual Dense Retrieval},
  author    = {Zhu, Tongyao and Huang, Chao-Ming and Kan, Min-Yen},
  booktitle = {Proceedings of the Annual Meeting of the Association for Computational Linguistics},
  year      = {2026}
}