Talk2Ref: A Dataset for Reference Prediction from Scientific Talks

Scientific talks are a growing medium for disseminating research, and automatically identifying relevant literature that grounds or enriches a talk would be highly valuable for researchers and students alike. We introduce Reference Prediction from Talks (RPT), a new task that maps long and unstructured scientific presentations to relevant papers.
To support research on RPT, we present Talk2Ref, the first large-scale dataset of its kind, containing 6,279 talks and 43,429 cited papers (26 per talk on average), where relevance is approximated by the papers cited in the talk’s corresponding source publication.

We establish strong baselines by evaluating state-of-the-art text embedding models in zero-shot retrieval scenarios and propose a dual-encoder architecture trained on Talk2Ref. We further explore strategies for handling long transcripts and training for domain adaptation.
Our results show that fine-tuning on Talk2Ref significantly improves citation prediction performance, demonstrating both the challenges of the task and the effectiveness of our dataset for learning semantic representations from spoken scientific content.

The dataset and trained models are released under an open license to foster future research on integrating spoken scientific communication into citation recommendation systems.

---

Dataset Summary

To the best of our knowledge, no existing dataset supports research on Reference Prediction from Talks (RPT). Talk2Ref is the first large-scale resource pairing scientific presentations with their corresponding relevant papers. Relevance is modeled using the citations in each talk’s source publication.

Talk2Ref includes:

• 6,279 scientific talks
• 43,429 cited papers
• ≈26 references per talk
• Spanning 2017–2022
• Covering ACL, NAACL, and EMNLP conferences

This dataset provides a foundation for systematically studying reference prediction from spoken scientific content at scale.

---

Dataset Structure

Split	Conferences	Years	Talks	Avg. Length (min)	Avg. Words	Avg. References	Total References
Train	ACL, NAACL, EMNLP	2017–2021	3,971	12.1	1615	26.75	31,064
Dev	ACL	2022	882	9.9	1327	26.05	11,805
Test	EMNLP, NAACL	2022	1,426	9.1	1186	25.66	16,935
Total	ACL, NAACL, EMNLP	2017–2022	6,279	11.1	1,478	26.4	43,429

Talks are partitioned chronologically by conference year.
Earlier years form the training split (2017–2021), and later years (2022) are used for development and testing, ensuring temporal consistency between splits.

---

Dataset Fields

Field	Type	Description
video_path	string	URL or path to the original conference talk video.
audio	audio	Audio waveform of the talk segment with sampling rate information.
sr	int	Sampling rate (Hz) of the audio recording.
abstract	string	Abstract of the corresponding scientific paper.
language	string	Language of the talk (English).
split	string	Split name (“train”, “dev”, or “test”).
duration	float	Duration of the audio in seconds.
conference	string	Conference name (ACL, NAACL, or EMNLP).
year	string	Year of the conference.
transcription	string	Automatic speech recognition (ASR) transcript of the talk.
title	string	Paper title associated with the talk.
references	list	List of structured metadata for cited papers, including title, authors, abstract, year.

---

Data Collection and Processing

1. Source Acquisition:
   Conference talks and associated papers were obtained from the ACL Anthology.

2. Audio Extraction:
   Audio tracks were extracted from videos and converted to .wav format using FFmpeg.

3. Transcription:
   Speech was transcribed using Whisper-Large-v3.

4. Reference Extraction:
   The corresponding paper PDFs were parsed with GROBID, extracting all cited references and metadata.

5. Abstract Retrieval:
   Missing abstracts were filled by querying CrossRef, arXiv, OpenAlex, and Semantic Scholar.

6. Filtering:
   Invalid or placeholder abstracts were removed.

This process results in a rich dataset linking each talk to its cited papers, including audio, transcript, and metadata.

---

Use Cases

Talk2Ref supports research on:

• Reference Prediction from Spoken Content
• Speech-to-Text and Speech-to-Abstract Generation
• Retrieval and Representation Learning

---

Licensing

The dataset is distributed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).
Users are free to share and adapt the dataset with appropriate attribution.

---

Citation

If you use this dataset, please cite the following paper:

@misc{broy2025talk2refdatasetreferenceprediction,
  title        = {Talk2Ref: A Dataset for Reference Prediction from Scientific Talks},
  author       = {Frederik Broy and Maike Züfle and Jan Niehues},
  year         = {2025},
  eprint       = {2510.24478},
  archivePrefix= {arXiv},
  primaryClass = {cs.CL},
  url          = {https://arxiv.org/abs/2510.24478}
}