philoberta_alignment lives in NuBerea/concept-primitives — embedding compute artifacts that require model inference belong in the primitive layer, not in the analytical layer.

Usage

from datasets import load_dataset

# Concept-level drift (Schlattmann & Vogl 2024)
drift = load_dataset("NuBerea/concept-analysis", "concept_drift", split="train")

# Lemma-level drift (Hamilton et al. 2016 / CADE) — filter to freq-stable rows
hamilton = load_dataset("NuBerea/concept-analysis", "concept_drift_hamilton", split="train")
stable = hamilton.filter(lambda x: x["freq_a"] >= 10 and x["freq_b"] >= 5)

# Concept trajectories — SPhilBerta 768-dim space
traj = load_dataset("NuBerea/concept-analysis", "concept_trajectory_sp", split="train")

# Concept aliases (English term → BERTopic concept_id)
aliases = load_dataset("NuBerea/concept-analysis", "concept_aliases", split="train")

# PhiloBERTA-style alignment scores live in concept-primitives (model-inference primitive)
philo = load_dataset("NuBerea/concept-primitives", "philoberta_alignment", split="train")
grc_lat = philo.filter(lambda x: x["lang_pair"] == "grc-lat" and not x["low_freq_flag"])

# Flan-T5 sense labels — natural-language gloss per concept per stratum
senses = load_dataset("NuBerea/concept-analysis", "concept_sense_labels", split="train")

License

CC BY 4.0. Derived from NuBerea/concept-graph and NuBerea/features (both CC BY 4.0). Methods implemented from Schlattmann & Vogl 2024 (CC BY / arXiv open access) and Hamilton et al. 2016.

Upstream Data Sources

Source	Role	License
NuBerea/concept-graph	Concept nodes, span edges, BERTopic centroids	CC BY 4.0
NuBerea/features	SPhilBerta + shlm-grc-en verse embeddings	CC BY 4.0
NuBerea/pericope-genealogy	`composition_dag` bt_layer + plate_segment for T8 configs	CC BY 4.0
NuBerea/macula-greek, NuBerea/macula-hebrew	Lemma corpora for nrsv_lemma_bridge	CC BY 4.0
NuBerea/nrsv	English verse text for nrsv_lemma_bridge + Flan-T5 prompts	CC BY 4.0
NuBerea/vulgate-primitives	Latin token corpora for Hamilton drift (Latin pair)	CC BY 4.0
NuBerea/AI-models	bi_encoder_v1 + question_bin_assignment.json for concept_aliases	CC BY 4.0
NuBerea/secondary-sources	Philo + Josephus Greek text segments	CC BY 4.0
NuBerea/second-temple	Aramaic Targum text (English translation filtered)	CC BY 4.0
Schlattmann & Vogl 2024 (arXiv:2501.00391)	Drift methodology	CC BY / arXiv
Hamilton et al. 2016 (arXiv:1605.09096)	Diachronic word2vec method	Open access
Giulianelli et al. 2023 (arXiv:2305.11993)	Flan-T5 sense label method	Open access

Boundary:

NuBerea/concept-graph holds topology: span nodes, edges, concept nodes, C2P/C2C edges
NuBerea/concept-analysis holds derived outputs: centroids, communities, trajectories, drift metrics

Configs

Trajectory configs — T8 (BT-layer indexed, pericope-genealogy join)

These two configs are T8 tier — they read NuBerea/pericope-genealogy (T7) and add Bradley-Terry compositional stratum to each verse embedding. Cardinal rule: all 59,478 verse-level rows from concept_verse_edge_sp are preserved; bt_layer is nullable (pd.NA) for NT verses not in the composition_dag anchor node set.

Config	Rows	Description
`concept_verse_edge_bt`	59,478	`concept_verse_edge_sp` + `bt_layer` (nullable Int64, 16 levels), `plate_segment`, `osisRef`, `match_grain='verse'`
`concept_trajectory_bt`	487	Concept × bt_layer centroids; 77 concepts × up to 14 bt_layers; `match_grain='pericope'` (derived view, does not replace `concept_trajectory_sp`)

Coverage: 59.5% overall bt_layer assignment (lxx 70.2%, nt 24.5%, vg 61.2%). NT partial coverage is expected — composition_dag tracks OT source composition, not NT pericopes. NT single-verse anchor nodes (intertextuality OSIS refs) account for the 24.5% NT coverage.

Schema additions in concept_verse_edge_bt vs. concept_verse_edge_sp:

Column	Type	Description
`osisRef`	str	Normalized OSIS dot-format ref (LXX `'1Chr 1:1'` → `'1Chr.1.1'`)
`bt_layer`	Int64 (nullable)	Bradley-Terry compositional stratum (0–15); null for unmatched verses
`plate_segment`	str	BIC-DP plate segment label from pericope-genealogy; null if unmatched
`match_grain`	str	Always `'verse'` — cardinal rule compliance marker

Schema for concept_trajectory_bt:

Column	Type	Description
`concept_id`, `concept_label`	str	Concept identifier and label
`bt_layer`	int	Compositional stratum (0–15)
`verse_count`	int	Verses contributing to centroid (min 2)
`centroid`	list[float]	Mean SPhilBERTa 768-dim embedding vector
`mean_cosine_to_centroid`	float	Intra-layer cohesion measure
`verse_refs`	list[str]	OSIS refs of contributing verses (deduplicated)
`model_id`	str	Embedding model
`match_grain`	str	Always `'pericope'` — derived view marker

Trajectory configs — T6 (SPhilBerta space, 768-dim, LXX + NT + Vulgate)

Config	Rows	Description
`concept_verse_edge_sp`	59,478	Concept × verse, all three strata; `bowphs/SPhilBerta` embeddings
`concept_trajectory_sp`	190	Concept × stratum centroids + `mean_cosine_to_centroid`; 3 strata

PSEU trajectory configs — T6 (SPhilBerta space, 768-dim, Deuterocanon + Pseudepigrapha)

These configs extend the diachronic chain into stratum 6 (LXX Deuterocanon) and stratum 7 (Second Temple Pseudepigrapha) using nearest-centroid assignment rather than pericope-containment topology. PSEU pericopes are not yet represented in NuBerea/concept-graph, so each verse is assigned to its single nearest concept centroid (top-1 cosine similarity) pooled from all LXX+NT+VG verse embeddings in concept_verse_edge_sp. The assignment_method column records 'nearest_centroid' to distinguish from the topology-driven 'pericope_containment' method used in concept_verse_edge_sp.

Source embeddings: NuBerea/features → deut_verse_embeddings_sp (5,857 rows) and pseu_verse_embeddings_sp (15,043 rows).

Config	Rows	Description
`concept_verse_edge_deut`	5,857	Concept × verse for LXX Deuterocanon (stratum 6); top-1 nearest centroid per verse
`concept_trajectory_deut`	78	Per-concept centroids for Deuterocanon stratum; all 78 concepts covered
`concept_verse_edge_pseu`	15,043	Concept × verse for Second Temple Pseudepigrapha (stratum 7); top-1 nearest centroid per verse
`concept_trajectory_pseu`	78	Per-concept centroids for Pseudepigrapha stratum; all 78 concepts covered

Schema additions vs. concept_verse_edge_sp:

Column	Type	Description
`assignment_method`	str	Always `'nearest_centroid'` for PSEU configs

edge_weight carries the cosine similarity of the verse to the assigned concept centroid (range 0–1; all verses assign cleanly at mean cosine ≈ 0.78 deut / 0.75 pseu).

Drift analysis across all five SPhilBerta strata (lxx → deut → pseu → nt → vg) is now possible by joining concept_trajectory_sp (lxx/nt/vg) with concept_trajectory_deut and concept_trajectory_pseu on concept_id.

Patristic concept edges — T6 (SPhilBerta space, 768-dim, patristic GRC)

Extends the diachronic concept chain from the canonical biblical strata (LXX → NT → Vg) into patristic Greek literature (~1st–7th c. CE). Uses the same nearest-centroid assignment method as the PSEU configs: each segment is assigned to its single nearest concept centroid (top-1 cosine similarity) pooled from the LXX+NT+VG reference space.

Source embeddings: NuBerea/concept-primitives → patr_grc_embeddings_sp (~40K rows).

Config	Rows	Description
`concept_verse_edge_patr`	~40K	Concept × patristic segment; top-1 nearest centroid per segment

Schema additions vs. concept_verse_edge_sp:

Column	Type	Description
`segment_id`	str	Patristic segment SHA-256 hash (replaces `osis_ref`)
`cts_urn`	str	CTS work URN (e.g. `urn:cts:greekLit:tlg2042.tlg001`)
`author_label`	str	Author name
`family_id`	str	NuBerea family identifier (e.g. `fam:origen`)
`assignment_method`	str	Always `'nearest_centroid'`

Secondary literature concept edges — T6 (SPhilBerta space, 768-dim)

Extends the concept chain into Philo, Josephus (Greek), Aramaic Targum, and Rabbinic literature using the same nearest-centroid assignment method as the PSEU configs. All segments are assigned top-1 cosine similarity against the 78 concept centroids derived from concept_verse_edge_sp.

Sources: NuBerea/secondary-sources (philo_grc, josephus_grc), NuBerea/second-temple (targum, filtered to rows with text_en), NuBerea/talmud (mishnah, tosefta, filtered to rows with text_en).

Config	Rows	Description
`concept_verse_edge_philo`	7,906	Concept × segment for Philo of Alexandria (Greek); top-1 nearest centroid
`concept_verse_edge_josephus`	12,735	Concept × segment for Josephus (Greek); top-1 nearest centroid
`concept_verse_edge_targum`	14,452	Concept × segment for Aramaic Targums (English translation); top-1 nearest centroid
`concept_verse_edge_talmud`	~37K	Concept × segment for Mishnah + Tosefta (English translation); top-1 nearest centroid
`concept_verse_edge_talmud_rabbinica`	16,435	Clean-license rebuild (rabbinica Kulp/Guggenheimer CC-BY): Mishnah + first-ever Yerushalmi arm; Bavli excluded (no English); top-1 nearest centroid
`concept_verse_edge_targum_psj`	5,816	Pseudo-Jonathan arm on Etheridge 1862 Public Domain English (osisRef-addressed); top-1 nearest centroid
`concept_verse_edge_nhc`	2,622	Concept × segment for Nag Hammadi Corpus (English); reuses SPhilBerta embeddings from `NuBerea/features → gnostic_ln_bridge`; top-1 nearest centroid
`concept_verse_edge_af`	2,236	Concept × segment for Apostolic Fathers (Greek); reuses SPhilBerta embeddings from `NuBerea/features → patr_grc_embeddings_sp` filtered by `APOSTOLIC_FAMILY_IDS + APOSTOLIC_CANONICAL_WORKS + lang=grc`; top-1 nearest centroid. Built 2026-05-23 to close BT coverage gap for E9 ontology cross-check.
`concept_verse_edge_ecw`	490	Concept × segment for Early Christian Apocrypha (English text → on-the-fly SPhilBerta embedding); source `NuBerea/heretical-texts → ecw_apocryphal_eng`; top-1 nearest centroid. Built 2026-05-23 to close BT coverage gap for E9 ontology cross-check.

Philo/Josephus schema (vs. concept_verse_edge_sp):

Column	Type	Description
`segment_id`	str	SHA-256 hash of segment text
`cts_urn`	str	CTS work URN (e.g. `urn:cts:greekLit:tlg0018.tlg001`)
`author_label`	str	"Philo of Alexandria" or "Josephus"
`work_label`	str	Work title
`stratum`	str	`'philo'` or `'josephus'`
`assignment_method`	str	Always `'nearest_centroid'`

Targum schema additions:

Column	Type	Description
`osis_ref`	str	OSIS ref derived from Targum book + section (nullable for non-biblical sections)
`targum_name`	str	Targum name (e.g. `'Targum Onkelos'`, `'Targum Jonathan'`)
`section`	str	Section header within the Targum
`book`	str	OT book name
`stratum`	str	Always `'targum'`

Talmud schema:

Column	Type	Description
`segment_id`	str	SHA-256 hash of segment text
`ref`	str	Sefaria ref (e.g. `'Mishnah Berakhot 1:1'`)
`title`	str	Tractate title
`tractate`	str	Tractate name
`order`	str	Seder (order) name
`stratum`	str	`'mishnah'` (~~200 CE) or `'tosefta'` (~~220 CE)
`assignment_method`	str	Always `'nearest_centroid'`

NHC schema:

Column	Type	Description
`segment_id`	str	SHA-256 segment identifier (from gnostic_ln_bridge)
`codex`	str	NHC codex identifier (e.g. `NHC_I`)
`tractate_num`	int	Tractate number within codex
`work_slug`	str	Work slug (e.g. `gospel_of_thomas`)
`work_title`	str	Full work title
`stratum`	str	Always `'nhc'`
`assignment_method`	str	Always `'nearest_centroid'`

Mean edge weights: Philo 0.707, Josephus 0.701, Targum 0.734, NHC 0.713.

Trajectory configs — T6 (shlm-grc-en space, 768-dim, LXX + NT only)

Config	Rows	Description
`concept_verse_edge`	28,909	Concept × verse, LXX and NT strata
`concept_trajectory`	112	Concept × stratum centroids; LXX and NT

Corpus-timeline series

concept_trajectory_series and concept_drift_sequential extend the diachronic chain to all 11 major strata in temporal order, from the Septuagint through Mishnah and Tosefta. This is the full cross-corpus drift picture for each concept.

Strata in order (stratum_order 1–11):

#	Stratum	Language	Source config	Notes
1	lxx	grc	concept_verse_edge_sp	Septuagint
2	deut	grc	concept_verse_edge_deut	LXX Deuterocanon
3	pseu	grc	concept_verse_edge_pseu	Second Temple Pseudepigrapha
4	philo	grc	concept_verse_edge_philo	Philo of Alexandria
5	nt	grc	concept_verse_edge_sp	New Testament
6	josephus	grc	concept_verse_edge_josephus	Josephus
7	patr	grc	concept_verse_edge_patr	Greek Patristics
8	targum	en	concept_verse_edge_targum	Aramaic Targums (EN)
9	mishnah	en	concept_verse_edge_talmud	Mishnah (~200 CE)
10	tosefta	en	concept_verse_edge_talmud	Tosefta (~220 CE)
11	vg	lat	concept_verse_edge_sp	Vulgate

Cross-language note: pairs 7 (patr→targum, grc→en) and 10 (tosefta→vg, en→lat) are cross-language transitions. same_language=False flags them so they can be excluded from within-language drift analysis. Within-language same-language pairs are clean: grc→grc (pairs 1–6) and en→en (pairs 8–9).

Config	Rows	Description
`concept_trajectory_series`	802	Per-concept centroids for all 11 strata (long format; 78 concepts × up to 11 strata)
`concept_drift_sequential`	693	Adjacent-pair drift across 10 stratum transitions; `same_language` flag; `centroid_drift = 1 − cosine(centroid_a, centroid_b)`

concept_trajectory_series schema:

Column	Type	Description
`concept_id`, `concept_label`	str	Concept identifier and label
`stratum`	str	Stratum name (lxx/deut/pseu/philo/nt/josephus/patr/targum/mishnah/tosefta/vg)
`stratum_order`	int	Temporal order (1–11)
`language`	str	Embedding language: `grc`, `en`, or `lat`
`verse_count`	int	Segments contributing to this concept's centroid in this stratum
`mean_edge_weight`	float	Mean cosine to the LXX+NT+VG reference centroid
`crystallization`	float	Mean cosine of member embeddings to their own per-stratum centroid
`centroid`	list[float]	768-dim SPhilBERTa mean centroid vector
`model_id`	str	Embedding model identifier

concept_drift_sequential schema:

Column	Type	Description
`concept_id`, `concept_label`	str	Concept identifier and label
`stratum_a`, `stratum_b`	str	Source and target stratum for this pair
`pair_order`	int	Pair index (1–10)
`language_a`, `language_b`	str	Language of each stratum
`same_language`	bool	False for cross-language pairs (patr→targum, tosefta→vg)
`centroid_cosine`	float	Cosine similarity between the two per-stratum centroids
`centroid_drift`	float	`1 − centroid_cosine`; range [0, 2]
`delta_crystallization`	float	`crystallization_b − crystallization_a`
`delta_mean_edge_weight`	float	`mean_edge_weight_b − mean_edge_weight_a`

Drift configs

Config	Rows	Description
`concept_drift`	78	Schlattmann & Vogl 2024 drift metrics per concept — lxx/nt/vg strata
`concept_drift_stj`	78	PSEU-extended drift — 5 pairs: lxx→deut, lxx→pseu, deut→nt, pseu→nt, deut→pseu
`concept_drift_hamilton`	5,023	Hamilton/CADE word2vec lemma drift; 3 language pairs

concept_drift_hamilton covers three independent analyses (cross-language alignment is not valid):

Language	Pair	Lemmas	Method
Hebrew	EBH → LBH	1,325	Hurvitz/Young classification; WLC lemmas
Greek	LXX → NT	1,657	lxx_verse_units + SBLGNT; NFKC-normalized
Latin	VG_OT → VG_NT	2,041	Vulgate token parquets

Columns: lemma, language, stratum_a, stratum_b, drift_score (1 − cosine after Procrustes), cosine_sim, freq_a, freq_b.

Community configs (migrated from concept-graph)

Config	Rows	Description
`community_candidate`	48	Community summaries: dominant anchor, corpus, book spread
`community_membership`	2,423	Pericope → community assignments
`concept_centroids`	86	L1 (78) + L2 (8) BERTopic centroids; `bi_encoder_v1` 384-dim

Alias config

Config	Rows	Description
`concept_aliases`	567	English theological term → `bt1_xxx` concept_id list; TF-IDF scored

Alignment config (PhiloBERTA-style)

philoberta_alignment lives in NuBerea/concept-primitives — it requires mBERT inference and therefore belongs in the primitive layer, not in concept-analysis. See that repo for schema and methodology.

Sense label config (Flan-T5)

Config	Rows	Description
`concept_sense_labels`	190	Flan-T5-large generated sense gloss per concept per stratum

One row per concept_trajectory_sp row (190 = ~63 concepts × 3 strata). Columns: concept_id, concept_label, stratum, sense_label, model_id, verse_sample (list of osis_refs used), prompt_token_count.

NRSV lemma bridge

Config	Rows	Description
`nrsv_lemma_bridge`	185,338	NRSV English token ↔ MACULA Greek/Hebrew lemma co-occurrence frequencies

Pure data join: NRSV verse text × MACULA lemma corpora joined on osis_ref. Columns: nrsv_word, sstrong, dict_form, testament, verse_count, lemma_total_verses, nrsv_word_total_verses, pct_of_lemma, pct_of_nrsv_word, low_freq_flag. Filter to verse_count >= 3 AND (pct_of_lemma >= 0.10 OR pct_of_nrsv_word >= 0.20) for stable mappings.

concept_aliases maps English terms (e.g. "baptism", "atonement") to one or more BERTopic concept IDs ranked by TF-IDF score. Generated by huggingface/candidates/build/concept-analysis/scripts/build_aliases_from_hf.py — fully HF-native: loads bi_encoder_v1 + training/question_bin_assignment.json from NuBerea/AI-models and L1 centroids from NuBerea/concept-graph (concept_node config), then assigns each question (filtered to topical_align >= 0.2) to its nearest L1 centroid and runs the per-cluster TF-IDF over question-title tokens. No local file dependencies. Regenerate via python scripts/build.py --config concept_aliases.

Drift methodology (`concept_drift`)

Implements two complementary methods from Schlattmann & Vogl 2024 ("Trajectories of Change", arXiv:2501.00391), adapted from time-sliced scientific corpora to biblical transmission strata.

Method 1 — Crystallization index (intra-stratum EDE): crystallization_X = mean_cosine_to_centroid from concept_trajectory_sp. High = tight cluster (crystallized, focused); low = diffuse (polysemous).

Method 2 — Cross-stratum density at reference centroid: cross_density(A→B) = mean cosine similarity of stratum B's verse embeddings to stratum A's centroid. Answers: do B's verses cluster near where A's concept was centred? Higher = concept preserved; lower = drifted.

drift_score = (centroid_drift + cross_drift) / 2

Why KDE/JSD was not used: The paper's Gaussian KDE approach requires same-corpus time slices with natural embedding overlap. LXX and NT verse embeddings for the same concept occupy non-overlapping regions in embedding space (different verse texts entirely), causing JSD to saturate near 1.0 for all concepts with no discrimination. cross_density is the correct adaptation.

Key findings (LXX → NT → Vulgate)

Transition	Mean centroid cosine	Interpretation
LXX → NT	0.910	Larger drift — theologically creative step
NT → VG	0.971	Smaller drift — primarily translation
LXX → VG	0.971	Vulgate broadly tracks Greek semantics

Most-drifted (LXX→NT): Spiritual/Gift/Spirit (0.373), Christian/Marriage/Sexual (0.322), Satan/Spirit/Demons (0.269)

Most stable (LXX→NT): Sons/Israelites/Children (0.134), Jonah/Fear/Sign (0.137), Jerusalem/Deuteronomy/Valley (0.148)

PSEU drift methodology (`concept_drift_stj`)

Extends concept_drift to cover LXX Deuterocanon (stratum 6) and Second Temple Pseudepigrapha (stratum 7). Same Schlattmann & Vogl 2024 methods; same 768-dim SPhilBerta embedding space.

Five transition pairs:

Pair	Description
lxx → deut	LXX canonical OT → Deuterocanon (Second Temple Jewish composition)
lxx → pseu	LXX canonical OT → Pseudepigrapha (apocalyptic / wisdom literature)
deut → nt	Deuterocanon → NT (deuterocanonical influence on NT authors)
pseu → nt	Pseudepigrapha → NT (Second Temple apocalyptic→NT continuity)
deut → pseu	Within-PSEU: Deuterocanon vs. Pseudepigrapha semantic overlap

Schema (78 rows, one per concept):

concept_id, concept_label
Per stratum (lxx, nt, vg, deut, pseu): verse_count_X, crystallization_X, dispersion_X
Per pair ({s_a}_{s_b}): centroid_cosine, centroid_drift, cross_density, cross_drift, delta_dispersion, drift_score (= mean of centroid_drift + cross_drift), plus cross_density_{s_b}_{s_a} (reverse direction)

Source embeddings: NuBerea/concept-analysis → concept_trajectory_sp, concept_verse_edge_sp, concept_trajectory_deut, concept_trajectory_pseu, concept_verse_edge_deut, concept_verse_edge_pseu.

Build: python scripts/build.py --config concept_drift_stj

Hamilton/CADE methodology (`concept_drift_hamilton`)

Implements Hamilton et al. 2016 (arXiv:1605.09096):

Train word2vec (100-dim, window=5) per corpus slice.
Align spaces via orthogonal Procrustes on shared vocabulary (Hamilton alignment step).
drift_score = 1 − cosine_sim(aligned_a, b) per shared lemma.

Three analyses are run independently — cross-language alignment is not valid.

Hebrew EBH/LBH classification follows Hurvitz/Young/Rezetko: LBH books are 1–2 Chronicles, Ezra, Nehemiah, Esther, Daniel, Ecclesiastes, Haggai, Zechariah, Malachi, Joel, Jonah, Obadiah; all others are EBH.

Greek NFKC normalization is applied to unify LXX polytonic OXIA accents (U+1F77 etc.) with SBLGNT monotonic TONOS accents (U+03AF etc.) before training. Without this, shared vocabulary collapses from ~1,600 to ~277 lemmas.

Note on drift_score > 1.0: Possible for low-frequency lemmas whose aligned vectors point in opposite directions (cosine_sim < 0). This is mathematically valid but indicates extreme drift or sampling noise; filter by freq_a >= 10 and freq_b >= 5 for stable estimates.

Reproducibility

The _sp trajectory configs are fully reproducible from HF:

python scripts/verify.py

Inputs:

NuBerea/concept-graph: concept_span_edge, span_node, concept_node
NuBerea/greek-embeddings-primitives: lxx_verse_embeddings_sp, nt_verse_embeddings_sp, vulgate_verse_embeddings

The concept_drift config is reproducible by running:

python huggingface/candidates/build/concept-trajectories/scripts/analyze_concept_drift.py

which reads the locally-built concept_trajectory_sp and concept_verse_edge_sp parquets.

The concept_drift_hamilton config is reproducible by running:

/tmp/gensim_venv/bin/python huggingface/candidates/build/concept-analysis/scripts/build_hamilton.py

Inputs: macula-hebrew-WLC.tsv, lxx_verse_units.jsonl, macula-greek-SBLGNT.tsv, vulgate-primitives token parquets.

The concept_sense_labels config is reproducible by running:

python huggingface/candidates/build/concept-analysis/scripts/build_sense_labels.py --model large

Inputs: concept_trajectory_sp (local), NuBerea/nrsv (HF download). Model: ltg/flan-t5-definition-en-large.

The philoberta_alignment config has moved to NuBerea/concept-primitives — see that repo for the build script and reproduction instructions.

The concept_verse_edge_bt and concept_trajectory_bt T8 configs are reproducible by running:

python huggingface/candidates/build/concept-analysis/scripts/build.py --config concept_verse_edge_bt
python huggingface/candidates/build/concept-analysis/scripts/build.py --config concept_trajectory_bt

Inputs: NuBerea/concept-analysis (concept_verse_edge_sp), NuBerea/pericope-genealogy (composition_dag), NuBerea/nrsv. Run concept_verse_edge_bt before concept_trajectory_bt.

The concept_aliases config is reproducible HF-natively by running:

python huggingface/candidates/build/concept-analysis/scripts/build.py --config concept_aliases

Inputs (all from HF): NuBerea/AI-models (models/bi_encoder_v1/ + training/question_bin_assignment.json), NuBerea/concept-graph (concept_node config). Model: bi_encoder_v1 (inference only; encodes ~3,521 question titles in seconds on MPS).

Citation

concept_drift implements Schlattmann & Vogl 2024:

@misc{schlattmann-vogl-2024-trajectories,
  title     = {Trajectories of Change: Measuring Semantic Drift in Scientific Discourse},
  author    = {Schlattmann, Philipp and Vogl, Tobias},
  year      = {2024},
  eprint    = {2501.00391},
  archivePrefix = {arXiv},
  url       = {https://arxiv.org/abs/2501.00391},
}

concept_drift_hamilton implements Hamilton et al. 2016 and the CADE variant (Bianchi et al. 2020):

@inproceedings{hamilton-etal-2016-diachronic,
  title     = {Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change},
  author    = {Hamilton, William L. and Leskovec, Jure and Jurafsky, Dan},
  booktitle = {Proceedings of ACL 2016},
  year      = {2016},
  eprint    = {1605.09096},
  archivePrefix = {arXiv},
  url       = {https://arxiv.org/abs/1605.09096},
}

@inproceedings{bianchi-etal-2020-cade,
  title     = {{CADE}: Context-Aware Diachronic Embeddings for Temporal Analysis},
  author    = {Bianchi, Federico and Terragni, Silvia and Hovy, Dirk},
  booktitle = {Proceedings of EMNLP 2020},
  year      = {2020},
  eprint    = {2004.06519},
  archivePrefix = {arXiv},
  url       = {https://arxiv.org/abs/2004.06519},
}

concept_sense_labels implements Giulianelli et al. 2023:

@inproceedings{giulianelli-etal-2023-interpretable,
  title     = {Interpretable Word Sense Representations via Definition Generation: The Case of Semantic Change Analysis},
  author    = {Giulianelli, Mario and Koplenig, Alexander and Piotrowski, Michael and Schlechtweg, Dominik},
  booktitle = {Proceedings of ACL 2023},
  year      = {2023},
  eprint    = {2305.11993},
  archivePrefix = {arXiv},
  url       = {https://arxiv.org/abs/2305.11993},
}

Migration note

This repo absorbs NuBerea/concept-trajectories (all 4 configs migrated intact) and three analytical configs previously in NuBerea/concept-graph (community_candidate, community_membership, concept_centroids).

NuBerea/concept-trajectories is deprecated; use NuBerea/concept-analysis for all new work.