Across all fields of academic study, well-respected experts cite their sources when sharing information. While large language models (LLMs) excel at synthesizing information, they do not provide reliable citation to sources, making it difficult to trace and verify the origins of the information they present. In contrast, search engines make sources readily accessible to users and place the burden of synthesizing information on the user. In our user survey results below, we find that users prefer search engines over LLMs for high-stakes queries, where concerns regarding information provenance outweigh the perceived utility of LLM responses.
To rigorously examine the interplay between the verifiability and utility of information sharing tools, we introduce the extractive-abstractive spectrum shown below, in which search engines and LLMs are extreme endpoints encapsulating multiple unexplored intermediate operating points. Search engines are extractive because they respond to queries with snippets of sources with links (citations) to the original webpages. LLMs are abstractive because they address queries with answers that synthesize and logically transform relevant information from training and in-context sources without reliable citation.
We define five operating points that span the extractive-abstractive spectrum and conduct human evaluations on seven systems across four diverse query distributions that reflect real-world QA settings: web search, language simplification, multi-step reasoning, and medical advice. As outputs become more abstractive, our results below demonstrate that perceived utility improves by as much as 200%, while the proportion of properly cited sentences decreases by as much as 50% and users take up to 3 times as long to verify cited information.
In this repository, we share our code to obtain generations from the seven systems, run human evaluation on a Streamlit and Supabase webapp, and visualize the results. We also provide our human evaluation data, processing pipeline, and visualization code.
Update the to-do's in ea_spectrum.yml and create the conda environment.
conda env create --name ea_spectrum --file=ea_spectrum.yml
This should take under ten minutes.
Download the Natural Questions (NQ), Wikipedia and Wikidata Multi-Hop (2WikiMH), and Multiple Answer Spans Healthcare (MASH) query datasets to extractive-abstractive-spectrum/data.
- NQ: https://ai.google.com/research/NaturalQuestions/download
- 2WikiMH: https://github.com/Alab-NII/2wikimultihop?tab=readme-ov-file
- MASH: https://github.com/mingzhu0527/MASHQA?tab=readme-ov-file
To obtain the snippet, quoted, paraphrased, entailed, and abstractive generations for the NQ (nq), Eta3G (eli5_nq), 2WikiMH (multihop), and MASH (mash) datasets:
python citation_systems/generateOperatingPoints.py --start_n 0 --n 20 --project_name example --data nq --best_of_k 10 --retriever google
To use gold-standard sources for MASH (--data mash) and 2WikiMH (--data multihop), replace --retriever google with --gold.
To obtain the cited gemini generations from files scraped using save_gemini_html.js and stored in example_directory:
python citation_systems/generateGeminiOutputs.py --start_n 0 --n 20 --project_name example --data nq --html_directory_path example_directory
To obtain the GPT-4 + Vertex generations, start by installing the packages below.
pip install vertexai
pip install google-cloud-discoveryengine
Then, download and install gcloud following https://cloud.google.com/sdk/docs/install. Following this, complete the authorization steps below:
./google-cloud-sdk/bin/gcloud init
./google-cloud-sdk/bin/gcloud auth application-default login
Create a project on https://console.cloud.google.com/ and add its gcloud project_id to generatePostHocOutputs.py.
Finally, generate the GPT-4 + Vertex outputs:
python citation_systems/generatePostHocOutputs.py --start_n 0 --n 20 --project_name example --data nq
Each of the scripts above produce a .json file of results where each row corresponds to one query with five cited generations. To obtain a file with rows corresponding to individual responses for a query, use the citation_systems/annotation_processing_for_sl.py script as described in the next section. None of the example commands above should take more than an hour to run.
To use the annotation interface, create a new Supabase project and create the tables specified in annotation_interface/table_schemas.txt.
Generations from generateOperatingPoint.py, generatePostHocOutputs.py, and generateGeminiOutputs.py must be processed with the following script to obtain example_generations_byQueryOP.csv:
python citation_systems/annotation_processing_for_sl.py --filename example_generations.jsonl --start_n 0 --n 20
Instances to annotate from the csv obtained in the previous step must be loaded to Supabase:
python annotation_interface/load_instances_to_annotate.py --filename example_generations_byQueryOP.csv --db instances_to_annotate
The same example_generations_byQueryOP.csv needs to be uploaded to Google Drive as a Google Sheets file. Change the Google sheets sharing link to "Anyone with the link" and add it as a gsheets connection in annotation_interface/.streamlit/secrets.toml, like so:
[connections.gsheets_example]
spreadsheet = "<Google Sheets sharing link>"
Credentials and API keys for Supabase and MTurk should also be stored in the secrets.toml file.
Run the app locally by calling streamlit run annotation_interface/sl_app.py. The webapp can be deployed through Streamlit.
Processed human evaluation results for the five reference operating point generations (_ops_) as well as the GPT-4 + Vertex and Gemini generations (_baselines_) are available in mturk_results/processed_results. Files ending in _needs_citation only contain results for sentences that require citation, as judged by the Vertex API. Files ending in _all only contain results for all sentences. The human evaluation results before processing are stored in mturk_results/unprocessed_results.
We provide our processing pipeline visualize_results/clean_mturk_results.ipynb that obtains the processed result files from the unprocessed results. The processed result files are visualized, as presented in the paper, in visualize_results/plotting_by_metric.ipynb. We also provide the processing pipeline as a script: visualize_results/clean_results.py.
The user preference survey results are available in prolific_results and visualized in visualize_results/prolific_platform_preference.ipynb.
@misc{worledge2024extractiveabstractivespectrumuncoveringverifiability,
title={The Extractive-Abstractive Spectrum: Uncovering Verifiability Trade-offs in LLM Generations},
author={Theodora Worledge and Tatsunori Hashimoto and Carlos Guestrin},
year={2024},
eprint={2411.17375},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2411.17375},
}
This project is released under the Apache 2.0 License.