# The terms view

Part of [MBDef](https://huggingface.co/datasets/mediabiasgroup/MBDef). `terms` gathers, for each concept, what all its papers say about it, field by field, and merges items that say the same thing. It builds on the [other extracted fields](#other-extracted-fields) of each row.

## How items merge

- **Names** (`synonyms`, `related_terms`, `subtypes`, `supertypes`) merge by the rule that merges the terms themselves ([Term merging](https://huggingface.co/datasets/mediabiasgroup/MBDef/blob/main/docs/term-merging.md), with the vocabulary of all corpus terms): “misinformation” and “Mis-information” (among the related terms of “disinformation”) become one item, stated in 188 papers. A name whose spelling is one of the dataset's concepts links to it (`links_to`), so the supertypes and subtypes form links between concepts.
- **Sentences** (`definitions`, `description`, `examples`) merge only when they read the same after case, punctuation and spacing are set aside, as when several papers quote one source.
- **`papers`** counts the distinct papers that state an item: “disinformation” as a supertype of “fake news” is stated in 38 papers. `variants` records every merge, and `sources` gives each paper's own wording and citations.
- **Fields are kept as each paper puts them.** The same name can sit under several fields: for “fake news”, “disinformation” is a synonym in 60 papers, a related term in 147 papers, a subtype in 24 papers, a supertype in 38 papers. The view does not choose between them.

**Worked example.** The `terms` row of “fake news” (659 papers) says, most stated first:

- `supertypes`: “disinformation” (38 papers; the concept “disinformation”), “misinformation” (29 papers; the concept “misinformation”), “information disorder” (7 papers; the concept “information disorder”), and 29 more
- `subtypes`: “propaganda” (28 papers; the concept “propaganda”), “disinformation” (24 papers; the concept “disinformation”), “misinformation” (23 papers; the concept “misinformation”), and 253 more
- `synonyms`: “disinformation” (60 papers; the concept “disinformation”), “misinformation” (49 papers; the concept “misinformation”), “false news” (37 papers; the concept “false news”), and 137 more
- `related_terms`: “misinformation” (153 papers; the concept “misinformation”), “disinformation” (147 papers; the concept “disinformation”), “rumors” (48 papers; the concept “rumor”), and 240 more
- `definitions`: 1,707 after merging; the one stated by most papers (25): “fabricated information that mimics news media content in form but not in organizational…”

| Field | Items | After merging | Merged from several forms | Stated by 2+ papers | Name a concept |
|---|---:|---:|---:|---:|---:|
| `definitions` | 103,476 | 99,259 | 3,172 | 562 | — |
| `description` | 41,762 | 41,377 | 252 | 57 | — |
| `synonyms` | 6,434 | 5,498 | 146 | 252 | 2,878 |
| `related_terms` | 21,269 | 18,000 | 373 | 978 | 12,299 |
| `subtypes` | 12,682 | 11,160 | 467 | 620 | 6,496 |
| `supertypes` | 9,105 | 7,849 | 335 | 538 | 6,167 |
| `examples` | 10,190 | 9,904 | 132 | 109 | — |

144 name items name their own concept (for example a term listed as its own synonym); they are kept as extracted.

## Other extracted fields

Besides its definitions, the extraction lists what a paper says about a term in six more fields, the same fields the ground truth annotates. They come from the same extraction runs as the definitions and stay with their term: Step 5 judged the term, not each item.

- Every item is exactly what the extraction model wrote into that field for that paper.
- Only the definitions were checked by the second model.
- Nothing moves an item to another field or votes across papers: a name a paper lists under several fields (for example as both a synonym and a related term) stays under each, and when papers disagree, every paper's choice is kept.

| Field | Items | Rows with any |
|---|---:|---:|
| `description` | 41,762 | 21,314 |
| `synonyms` | 6,434 | 4,495 |
| `related_terms` | 21,269 | 12,277 |
| `subtypes` | 12,682 | 3,941 |
| `supertypes` | 9,105 | 8,011 |
| `examples` | 10,190 | 5,095 |

In all, 101,442 items on 34,336 of the 62,066 rows. The extraction's quality figures ([How it was made](https://huggingface.co/datasets/mediabiasgroup/MBDef/blob/main/docs/how-it-was-made.md), step 3) are measured on definitions; the other fields have no such measurement.