BERT Activations for Verb-Particle Constructions from the British National Corpus

A research study by Hassane Kissane examining neural representations of verb-particle constructions in BERT and CxG-BERT models. The analysis uses data from the British National Corpus, applying multidimensional scaling, generalized discrimination value, and permutation tests to model activations. The study was last updated in April 2026.

Use Cases

• Analyzing layer-specific syntactic structure encoding in BERT based on verb-particle constructions.
• Comparing linguistic representation between standard BERT and a construction-aware variant (CxG-BERT).
• Investigating variation in neural processing across different verb types mentioned in the study.
• Testing assumptions of uniform linguistic processing in neural networks using the described methodology.

Strengths

• Study is licensed under CC-BY-4.0, allowing for open sharing and reuse.
• Analysis is based on data from the British National Corpus, a known linguistic resource.
• Includes results from multiple analytical methods: MDS, GDV, and permutation tests.

Limitations

• The dataset is 5.9 MB, which is a small scale for raw model activations.
• Row count and column-level documentation are absent; field semantics must be inferred after download.
• The primary files are PDF and DOCX formats, which may require extraction of underlying data.

Provenance

Source

Hassane Kissane via figshare.

Collection Method

Model activations extracted from BERT and CxG-BERT using data from the British National Corpus.

Time Range

The temporal coverage of the source British National Corpus data is not specified.

Freshness

Last updated 2026-04-09 18:36:45; freshness should be verified.

Geography

The spatial coverage of the source British National Corpus data is not specified.