Dataset Card for "DiMB-RE"

DiMB-RE (Diet-Microbiome Relation Extraction) corpus is a resource for mining diet-microbiome associations from scientific literature.

Dataset Details

Dataset Description

Curated by: Gibong Hong, Veronica Hindle, Nadine M. Veasley, Hannah D. Holscher, Halil Kilicoglu
Funded by: University of Illinois Personalized Nutrition Initiative Seed Grant, National Center for Complementary and Integrative Health (NCCIH), Office of Data Science Strategy (ODSS)
Shared by: ScienceNLP Lab, University of Illinois Urbana-Champaign
Language(s) (NLP): English
License: [More Information Needed]

Dataset Sources

Repository: https://github.com/ScienceNLP-Lab/DiMB-RE
Paper: DiMB-RE: Mining the Scientific Literature for Diet-Microbiome Associations

Uses

Direct Use

[More Information Needed]

Out-of-Scope Use

[More Information Needed]

Dataset Structure

The dataset is provided in a JSON format and has two configurations: a sentence-level view and a document-level view. Each data point in the sentence-level configuration contains the following fields:

id: A unique string identifier for the sentence (e.g., PMC4994979_sent_0).
doc_key: A string identifying the source document (e.g., PMC4994979).
tokens: A list of strings representing the words in the sentence.
ner: A list of dictionaries for each named entity, containing its start and end token indices and its entity type.
ner_tags: A sequence of strings representing the BIO (Beginning, Inside, Outside) tag for each token.
triggers: A list of dictionaries for each relation trigger, containing its start and end token indices and its type (which corresponds to a relation type).
relations: A list of dictionaries, where each dictionary defines a relationship between two entities (a head and a tail), including their token spans, types, the relation type, and the factuality level.
triplets: A list of dictionaries linking a head entity, a tail entity, and a trigger by their token spans.

Example Data Point (sentence_level)

{
    "doc_key": "PMC4994979",
    "id": "PMC4994979_sent_0",
    "ner": [
        {"end": 7, "start": 2, "type": "Nutrient"},
        {"end": 11, "start": 9, "type": "Physiology"}
    ],
    "ner_tags": ["O", "O", "B-Nutrient", "I-Nutrient", "I-Nutrient", "I-Nutrient", "I-Nutrient", "O", "O", "B-Physiology", "I-Physiology", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O", "O"],
    "relations": [
        {
            "factuality": "Unknown",
            "head": 0,
            "head_end": 7,
            "head_start": 2,
            "head_type": "Nutrient",
            "tail": 1,
            "tail_end": 11,
            "tail_start": 9,
            "tail_type": "Physiology",
            "type": "AFFECTS"
        }
    ],
    "tokens": ["Effect", "of", "vitamin", "E", "with", "therapeutic", "iron", "supplementation", "on", "iron", "repletion", "and", "gut", "microbiome", "in", "U", ".", "S", ".", "iron", "deficient", "infants", "and", "toddlers", ":", "a", "randomized", "control", "trial"],
    "triggers": [
        {"end": 392, "start": 391, "type": "Nutrient"}
    ],
    "triplets": [
        {
            "head_end": 7,
            "head_start": 2,
            "relation": "0",
            "tail_end": 11,
            "tail_start": 9,
            "trigger_end": 1,
            "trigger_start": 0
        }
    ]
}

The document-level view additional contains the following field:

sentences: A list of dictionaries for each sentence, containing its start and end token indices.

Entity Types

The dataset is annotated with 15 entity types:

Food, Nutrient, DietPattern, Microorganism, DiversityMetric, Metabolite, Physiology, Disease, Measurement, Enzyme, Gene, Chemical, Methodology, Population, Biospecimen

Relation Types

Relations capture the interactions between entities and are categorized into 13 types:

AFFECTS, IMPROVES, WORSENS, ASSOCIATED_WITH, POS_ASSOCIATED_WITH, NEG_ASSOCIATED_WITH, INTERACTS_WITH, INCREASES, DECREASES, CAUSES, PREVENTS, PREDISPOSES, HAS_COMPONENT

Annotations also include relation triggers (the specific word or phrase indicating the relation, e.g., "increased") and factuality levels (Factual, Probable, Possible, Doubtful, Negated, and Unknown).

Dataset Creation

Curation Rationale

The motivation for creating DiMB-RE was the recognition that while the scientific literature contains vast amounts of evidence on diet-microbiome interactions, this knowledge is locked in unstructured text. Manually curated databases are often limited and not scalable. This dataset was created to enable the use of NLP to automatically machine-read the literature, structure this information, and ultimately facilitate knowledge-guided analysis to advance personalized nutrition.

Source Data

Data Collection and Processing

The source data consists of titles and abstracts from 165 publications and the full-text Results sections from 30 of those publications. The articles were retrieved from PubMed using a manually crafted search string developed by domain experts in food science and human nutrition. The search terms focused on key concepts in diet-microbiome research.

Who are the source data producers?

The source data was produced by the authors of the scientific articles included in the corpus. These are researchers from various institutions globally who have published on the topic of diet and the microbiome.

Annotations

Annotation process

The annotation was performed by two graduate students in food science and nutrition and a senior investigator with expertise in NLP and biomedical informatics, using the Brat annotation tool. The process involved multiple stages of annotation, calculation of inter-annotator agreement (IAA), and adjudication of disagreements to refine the annotation guidelines. The final annotations were verified for consistency and accuracy. IAA for entities was reasonable (mean F1-score of 0.69 exact, 0.80 partial), while relation agreement was more modest (mean F1-score of 0.41 exact, 0.54 partial), highlighting the task's difficulty.

Who are the annotators?

The annotators were Veronica Hindle and Nadine M. Veasley, and Halil Kilicoglu. Gibong Hong also participated in verifying the final annotations.

Personal and Sensitive Information

[More Information Needed]

Bias, Risks, and Limitations

[More Information Needed]

Recommendations

[More Information Needed].

Citation

BibTeX:

@misc{hong2024dimbreminingscientificliterature,
      title={DiMB-RE: Mining the Scientific Literature for Diet-Microbiome Associations}, 
      author={Gibong Hong and Veronica Hindle and Nadine M. Veasley and Hannah D. Holscher and Halil Kilicoglu},
      year={2024},
      eprint={2409.19581},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2409.19581}, 
}

APA:

Hong, G., Hindle, V., Veasley, N. M., Holscher, H. D., & Kilicoglu, H. (2024). DiMB-RE: Mining the Scientific Literature for Diet-Microbiome Associations. ArXiv. /abs/2409.19581

Glossary [optional]

[More Information Needed]

More Information [optional]

[More Information Needed]

Dataset Card Contributions

Thanks to @phucdev for adding this dataset.