Datasets:

mediabiasgroup
/

MBDef

# Which terms are kept

Part of [MBDef](https://huggingface.co/datasets/mediabiasgroup/MBDef). Step 5 decides for every term-in-paper whether it belongs to a media-bias taxonomy. This page says how; the [paper](https://arxiv.org/abs/XXXX.XXXXX) says how well it works.

## Three scores

Three methods score each term-in-paper; higher means more relevant.

- **Embedding classifier.** Each definition is embedded with Qwen3-Embedding-8B, and the term-in-paper's vectors are averaged. A logistic regression trained on the labelled records turns that vector into a probability.
- **LLM decision model** (basal-omni-1.0-ultra, 8-bit weights). For each term and its definitions it answers one question: is this a type of bias (*core*), a concept needed to describe, detect or explain bias (*satellite*), or neither (*irrelevant*)? The score is the probability of core plus satellite.
- **Retrieval judge.** Gemma 4 12B is shown the 32 most similar labelled records (found with Qwen3-Embedding-8B) and asked whether the term is relevant. Thinking is off and the temperature is 0.

## Combining them

1. **Put each score on a shared scale.** A member's score is replaced by the share of that member's own scores on the 1,700 labelled training and development records that lie below it (ties count half). For the example row above, the classifier's score 0.988 is higher than 97.5% of the classifier's scores on those records, so it becomes 0.975. This makes each term's score independent of which other terms are scored with it.
2. **Average them.** The weights are classifier 0.344, decision model 0.339 and judge 0.316. They were learned on the training and development records: a stacked logistic regression, chosen from 315 candidate combinations. For the example row: 0.344 × 0.975 + 0.339 × 0.936 + 0.316 × 0.991 = 0.967, its `ensemble_score`.
3. **Keep or drop.** A term-in-paper is kept when its combined score is at least **0.4150**. That cut keeps 95% of the relevant training and development records. It was set with each record scored by a model that had not seen its paper.

The kept rows are the `instances` table; their three scores, reference ranks and combined score are columns of it ([Columns](https://huggingface.co/datasets/mediabiasgroup/MBDef/blob/main/docs/columns.md)).

	# Which terms are kept

	Part of [MBDef](https://huggingface.co/datasets/mediabiasgroup/MBDef). Step 5 decides for every term-in-paper whether it belongs to a media-bias taxonomy. This page says how; the [paper](https://arxiv.org/abs/XXXX.XXXXX) says how well it works.

	## Three scores

	Three methods score each term-in-paper; higher means more relevant.

	- Embedding classifier. Each definition is embedded with Qwen3-Embedding-8B, and the term-in-paper's vectors are averaged. A logistic regression trained on the labelled records turns that vector into a probability.
	- LLM decision model (basal-omni-1.0-ultra, 8-bit weights). For each term and its definitions it answers one question: is this a type of bias (core), a concept needed to describe, detect or explain bias (satellite), or neither (irrelevant)? The score is the probability of core plus satellite.
	- Retrieval judge. Gemma 4 12B is shown the 32 most similar labelled records (found with Qwen3-Embedding-8B) and asked whether the term is relevant. Thinking is off and the temperature is 0.

	## Combining them

	1. Put each score on a shared scale. A member's score is replaced by the share of that member's own scores on the 1,700 labelled training and development records that lie below it (ties count half). For the example row above, the classifier's score 0.988 is higher than 97.5% of the classifier's scores on those records, so it becomes 0.975. This makes each term's score independent of which other terms are scored with it.
	2. Average them. The weights are classifier 0.344, decision model 0.339 and judge 0.316. They were learned on the training and development records: a stacked logistic regression, chosen from 315 candidate combinations. For the example row: 0.344 × 0.975 + 0.339 × 0.936 + 0.316 × 0.991 = 0.967, its `ensemble_score`.
	3. Keep or drop. A term-in-paper is kept when its combined score is at least 0.4150. That cut keeps 95% of the relevant training and development records. It was set with each record scored by a model that had not seen its paper.

	The kept rows are the `instances` table; their three scores, reference ranks and combined score are columns of it ([Columns](https://huggingface.co/datasets/mediabiasgroup/MBDef/blob/main/docs/columns.md)).