SciFCheX: Developing a Scientific Fact-Checker with Hybrid Evidence Retrieval Approaches 📟

Note

This is the official public release of the SciFCheX, scientific fact-checking pipeline, authored by Filip J. Cierkosz. This research project has been pursued in completion of the BSc Computer Science degree at the University of Sheffield as a final-year project, which achieved a high-marked distinction. The development version of this project (which contains approx. 500 commits) will remain private. The introduced scientific verification pipeline consists of: evidence retrieval, rationale selection, and label prediction stages. Please note that this project was particularly focused on experimenting with novel approaches associated with the first stage of the pipeline, i.e., evidence retrieval. On top of that, the presented solution addresses the SCIVER shared task, and has been submitted to its official leaderboard hosted by Allen AI that can be found here. The project ranked in TOP 50 worldwide submissions at the time of its publication. More descriptions about the performance can be found in this section.

Table of Contents 📖

• Project Structure
• Approach
• Performance
• Dependencies & Experimental Environment
• Abstract Retrieval
• Pipeline
• Training
• Contribution

Project Structure 🔎

The section contains brief clarifications about the project resources that can be found within this repository. It is strongly advised to visit individual files for further context about implementations, since each of them has been annotated with quite a comprehensive amount of documentation:

• .github - Standard git setup that contains PR template, workflows, etc.
• data - Full copy of the SciFact dataset from the paper titled "Fact or Fiction: Verifying Scientific Claims" by Wadden et al. (2020). The dataset contains 1,409 samples provided with a corpus of 5,183 abstract documents. The data was used for training and evaluations of the presented pipeline.
• pipeline - The full-implementation of the SciFCheX system, which consists of the following packages:
  • eval - Implements the AbstractRetrievalEvaluator and PipelineEvaluator classes and their respective tools. This toolkit is essential to evaluate the model performance on different levels of complexity.
  • inference - Handles the pipeline flow when running predictions at inference time. It consists of: abstract_retieval -> rationale_selection -> label_prediction. Notably, the abstract_retrieval package contains all retrieval implementations tested during this research (i.e., TF-IDF, BM25, BM25 + BGE-M3, BM25 + BERT, BM25 + SciBERT, BM25 + BioBERT).
  • training - Includes the PyTorch implementation for the training of abstract retrieval (only for BERT-based retrieval settings), rationale selection, and label prediction, and their associated tools.
• results - The directory stores all the experimental results obtained from different experiments. These are automatically created by eval package, whenever evaluation is completed. results/abstract_retrieval contains AR-only experimental results, whereas results/pipeline has the ones for full-pipeline runs.
• resources - Contains project related resources, e.g., images.
• script - Implements all essential BASH scripts that can be used to run, e.g., abstract retrieval, pipeline, training, etc. All these can be run with different settings through specifying CLI args. Please follow the next sections of this README to find out more, or refer to script/ files to explore their respective documentation.

Approach 🧠

As briefly mentioned earlier, the SciFCheX pipeline follows the popular multi-step pipeline approach: evidence retrieval (referred as abstract retrieval), rationale selection, and label prediction. A high-level purpose of such a system can be analyzed in the image above. To clarify again - this fact verification system is specifically tailored to scientific claims.

---

The task addressed by SciFCheX can be defined as provided in the SCIVER shared task description (Wadden et al., 2021):

Given a valid scientific claim, the task is to (1) identify all abstracts in the corpus that are relevant to the claim, (2) label the relation of each relevant abstract to the claim as either "SUPPORTS" or "REFUTES", and (3) identify a set of evidentiary sentences (i.e. rationales) to justify each label.

---

While developing the final pipeline for SciFCheX, the experiments included experimenting with the following approaches:

• For abstract retrieval:
  • 2 different sparse methods: TF-IDF, BM25.
  • 4 different hybrid retrieval approaches (i.e., sparse + dense retrieval): BM25 + BGE-M3, BM25 + BERT, BM25 + SciBERT, BM25 + BioBERT-base; where: all BERT-based approaches involved fine-tuning a classifier on the train set of SciFact dataset.
• For rationale selection: the approaches involved utilizing RoBERTa-large, SciBERT, BioBERT-base, BioBERT-large; where: each transformer was fine-tuned on SciFact only.
• For label prediction: similarly as the line above, the approaches involved testing RoBERTa-large, SciBERT, BioBERT-base, BioBERT-large transformers; where: each transformer was fine-tuned on SciFact; in addition - there was tested a version of RoBERTa-large from VERISCI (Wadden et al., 2020), that was pre-trained on FEVER and further fine-tuned on SciFact by the authors of that model.

---

Thus, the most optimal settings concluded for the SciFCheX pipeline consists of:

• AR: BM25 + SciBERT classifier fine-tuned on SciFact. More about AR training can be found in this file.
• RS: SciBERT transformer fine-tuned on SciFact to select rationales. More about RS training can be found in this file.
• LP: RoBERTa-large transformer fine-tuned on FEVER+SciFact; this model was accessed from VERISCI resources, since this project did not include fine-tuning on FEVER (due to computational limitations associated with the scope of this project).

Performance 🏎️

Abstract Retrieval: The following table reports the results for all tested approaches, evaluated on the dev set of SciFact. The best performing AR approach in isolation, i.e., BM25 + BioBERT-base is annotated respectively. Still, the full-pipeline uses SciBERT, since it was found to perform better when integrated into the full-pipeline.

Approach	F1	Precision	Recall	Hit-one	Hit-all
TFIDF (K=3) [VERISCI]	0.26	0.16	0.69	0.85	0.83
BM25 (K=3)	0.29	0.18	0.78	0.90	0.88
BM25 (K=20) + BGE-M3	0.52	0.40	0.76	0.90	0.88
BM25 (K=20) + BERT	0.66	0.80	0.56	0.74	0.72
BM25 (K=20) + SciBERT [SCIFCHEX]	0.72	0.82	0.64	0.79	0.77
BM25(K=20) + BioBERT-base	0.73	0.86	0.63	0.78	0.76

---

Pipeline: The following tables report full-pipeline performance of SciFChex compared with the official VERISCI model, run on dev set of SciFact. Full results for different setups can be found in results/PIPELINE_DEV.txt (merged file), and in results/pipeline/ directory.

1. SciFCheX (AR: BM25 + SciBERT, RS: SciBERT, LP: RoBERTa-large) results on dev:

	Sentence Selection	Sentence Label	Abstract Label-only	Abstract Rationalized
Precision	0.79	0.72	0.86	0.82
Recall	0.46	0.42	0.49	0.46
F1	0.59	0.53	0.62	0.59

2. VERISCI (AR: TF-IDF, RS: RoBERTa-large, LP: RoBERTa-large) results on dev:

	Sentence Selection	Sentence Label	Abstract Label-only	Abstract Rationalized
Precision	0.53	0.47	0.56	0.53
Recall	0.44	0.39	0.47	0.45
F1	0.48	0.43	0.51	0.49

---

Overall, it can be observed that the improvements applied to evidence retrieval lead to major improvements in the full-pipeline performance, which was further confirmed in the official SCIVER Shared Task Leaderboard, where the individual SCIFCHEX report can be found here. At the time of the model submission (i.e., in early May 2024), SCIFCHEX placed in TOP 50 submissions worldwide.

Dependencies & Experimental Environment 🧪

Warning

The presented SciFCheX model has been tested, trained and evaluated using NVIDIA A100 GPU / NVIDIA L4 GPU using Google Colab Pro+ account. The initial development attempts involved working on HPC unit provided by the University of Sheffield (running on CentOS Linux), but there were issues associated with CUDA compatibility issues in relation to torch version required for this project (1.7.0). Thereby, it is recommended to use CUDA 11.1.X to run this project. Moreover, it is worth noting that the model is unlikely to successfuly execute on other OS, such as: Windows, or macOS (NB: the ARM architecture on macOS might not support dependencies for requirements_scifchex.txt).

1. Clone Repository: Clone the repository and access it through the following commands (NB: Skip the step if you were provided with the code base. Just make sure you navigate inside the root of the projet):

   git pull https://github.com/chizo4/SciFCheX.git && cd SciFCheX

2. GPU Activation: Before you move further make sure your GPU device is available. It's up to you to make sure it's activated, since different GPU require different setup. At the same time, it might be different if you run the model on HPC unit, which might require you to submit a job, rather than run it interactively. As noted earlier, this project was developed on NVIDIA A100. Once GPU resource is accessed/activated, make sure you use CUDA 11.1.X. It can be verified through:

   nvcc --version

3. Check conda Version: Next, ensure you have conda installed on your machine. You can verify its installation by checking its version:

   conda --version

   [!WARNING] If your shell fails to recognize the command (which is very unlikely if you are a computer scientist), please install conda following the official guidelines: conda docs.

4. Environment Setup: The project has been provided with an automation file, that will set up all the necessary environments without further supervision:

   bash ./SciFCheX/script/env-setup.sh

Abstract Retrieval 📑

Note

Each approach implemented for this stage can be found in pipeline/inference/abstract_retrieval directory. In addition, the implemented evaluation can be found in pipeline/eval/abstract_retrieval_evaluator.py. The evaluation script both prints the results at the end of the cell execution and saves it into path SciFCheX/results/abstract_retrieval.

---

Having completed all the configuration steps, it is finally the time to explore the capabilities of this project. The command below runs the custom-made script, that can be found in SciFCheX/script/abstract-retrieval.sh. It allows you to execute different retrieval approaches tested in this research project.

---

The AR script works in the following way:

bash ./script/abstract-retrieval.sh [retrieval] [dataset] [run_eval]

Where the options can be defined as:

• [retrieval] ➡️ biobert (best-performing approach), scibert, bert, bge_m3, bm25, tfidf.

• [dataset] ➡️ dev (recommended), test (skips evaluation, since no labels), train (please avoid using for BERT retrievals, as these models were fine-tuned on it and so their scores will be close to perfection).

• [run_eval] ➡️ true (shows evaluation results), false (skips evaluation).

---

❗The configuration provided below is set to run the best-performing retrieval developed in this project. Namely, it uses the HYBRID approach of first finding K=20 best matching docs with BM25, that are further classified by the SciBERT transformer model, fine-tuned on train set of the SciFact dataset.

bash ./script/abstract-retrieval.sh scibert dev true

Pipeline ✅

Note

Each approach implemented for full-pipeline can be found in pipeline/inference/ directory, which contains abstract_retrieval, rationale_selection, and label_prediction packages for predictions. In addition, the implemented evaluation can be found in pipeline/eval/pipeline_evaluator.py and its respective tools package. The evaluation script both prints the results at the end of the cell execution and saves it into path results/pipeline.

---

Now, it is the time to test the performance of the whole pipeline! The command in the next cell executes the custom-made script, that can be found in script/pipeline.sh. It makes it possible to run the pipeline for different abstract retrieval settings combined with different rationale selection and label prediction modules.

---

The script command works in the following way:

bash ./script/pipeline.sh [retrieval] [model] [dataset]

Where the options can be defined as:

• [retrieval] ➡️ As specified in Abstract Retrieval. In addition, can run oracle, which directly passes the gold selection into RS / LP stages (i.e., it imitates a "perfect" retrieval scenario).

• [model] ➡️ scifchex (for SciFCheX), baseline (for a VERISCI replica developed by FJC), verisci (for the actual VERISCI model from Allen AI).

• [dataset] ➡️ dev (recommended), test (skips evaluation, since no labels), train (please avoid using for BERT retrievals, as these models were fine-tuned on it and so their scores will be close to perfection).

---

The configuration provided below is set to run the best-performing SciFCheX pipeline, which consists of:

• Abstract Retrieval: Uses the aforementioned HYBRID approach that first finds K=20 best matching docs with BM25, that are further classified by the SciBERT transformer model, fine-tuned on train set of the SciFact dataset.
• Rationale Selection: Uses a SciFact fine-tuned SciBERT model.
• Label Prediction: Uses a FEVER+SciFact fine-tuned RoBERTa-large model (provided by VERISCI authors).

bash ./script/pipeline.sh scibert scifchex dev

Training 🔥

Note

This is just a brief insight about training approaches. In reality, it required much deeper exploration, including different epochs settings etc. Similarly as in other stages, FJC automated everything to the highest level of details to make the solution as efficient as possible. All the training implementation and its tools can be found in pipeline/training directory.

---

Warning

Running any of the scripts below on a free version of a notebook might take several hours, or might even directly fail, due to lack of memory on the T4 GPU. FJC used NVIDIA A100 and NVIDIA L4 GPUs for training all of the modules.

---

The AR stage required training for the following settings: biobert, scibert, bert. Notably, bge_m3 setting was used without any fine-tuning. Unsuprisingly, the sparse approaches, such as: tfidf, bm25, did not required training.

AR training can be executed via:

bash script/train-abstract-retrieval.sh [model]

Where the [model] option can be defined as:

• bert: For BERT.
• scibert: For SciBERT.
• biobert_base: For BioBERT-base (best performing in isolation).

---

The RS stage required training for all transformers.

RS training can be executed via:

cd SciFCheX/ && bash script/train-rationale-selection.sh [model]

Where the [model] option can be defined as:

• scibert: For scibert (best performing).
• biobert_large: For BioBERT-large.
• biobert_base: For BioBERT-base.
• roberta_large: For RoBERTa-large.

---

The LP stage required training for all transformers.

LP training can be executed via:

cd SciFCheX/ && bash script/train-label-prediction.sh [model]

Where the [model] option can be defined as:

• scibert: For scibert.
• biobert_large: For BioBERT-large.
• biobert_base: For BioBERT-base.
• roberta_large: For RoBERTa-large.

Contribution 🤝

Note

In case you had questions associated with this project, or spotted any issue (including experimental setup), please feel free to contact Filip J. Cierkosz via any of the links included on this GitHub page. You might also want to open a Pull Request with any suggested fixes, or questions.