TELEN

TELEN is an experimental Vietnamese legal retrieval architecture. It combines a frozen BKAI bi-encoder, a Legal Concept Graph, a three-layer GCN, attention pooling, and a HyperNetwork-generated rank-16 projection update.

The associated paper, "TELEN: Temporal Evolving Legal Embedding Network for Dynamic Vietnamese Law Retrieval," has been accepted for presentation at ICSEC 2026.

Intended use

The checkpoint supports research on Vietnamese legal document embeddings, graph-conditioned retrieval, and temporal corpus updates. It is not a legal advice system. Its benchmark labels indicate whether two articles share a parent document; they do not indicate that an article answers a legal question or remains legally effective.

Verified results

The cleaned evaluation corpus has 60,142 articles from 3,196 legal documents. The primary test uses 104 queries and a 1,406-article candidate pool from 2021.

Method	NDCG@10	MRR@10
Okapi BM25	0.5929	0.7154
Frozen PhoBERT-base-v2	0.4408	0.6059
DEk21	0.7012	0.7865
TELEN base-only inference ablation	0.8058	0.8742
TELEN	0.8068	0.8734

TELEN differs from the base-only inference ablation by 0.0010 NDCG@10. The law-clustered 95% bootstrap interval is [-0.0023, 0.0046], with (p=0.5529). The ablation bypasses the generated shift in the same checkpoint and is not a separately trained graph-free model.

Stress test	TELEN NDCG@10	Base-only NDCG@10
2021 pool, text only	0.7993	0.8003
Full 60,142-article pool	0.4606	0.4609

The title-free condition retains most of the restricted-pool score. Expanding the candidate pool to the full history produces a large drop.

Graph-update result

Adding all 59 documents from 2021 and recomputing the graph state leaves NDCG@10 at 0.8068 for both fresh and stale indexes. The generated-matrix change is (1.43 \times 10^{-8}) relative to the base matrix. The 59 new nodes receive (1.55 \times 10^{-5}) total attention mass. These measurements do not validate useful continuous adaptation for this checkpoint.

If a later update changes the projection, all query and document vectors must use the same state. Re-encode and rebuild the document index unless a separate compatibility method has been validated.

Files

• data/checkpoints/telen/telen_best.pt: audited checkpoint selected at epoch 9

• data/checkpoints/telen/graph_data_through_2020_v2.pt: graph cache containing documents available through 2020

• camera_ready_work/camera_ready_results.json: TELEN results, ablations, update tests, diagnostics, and stress tests

• camera_ready_work/camera_ready_baselines.json: BM25, PhoBERT, and DEk21 results

The audited checkpoint uses BKAI revision 84f9d9ada0d1a3c37557398b9ae9fcedcdf40be0. The evaluation uses VMTEB revision a3c6685dd0e97ff94ef896f894ec23ce7f859e76, DEk21 revision 517f1af7dd04a57194f1de2990f0c6ede0a3109b, and PhoBERT revision 86cd7fd4c148980922ac11a2cf5e257f2ba639e1.

Usage

Clone the GitHub repository, download this model repository, and keep the checkpoint paths used below.

from inference import TELENInference

model = TELENInference(
    checkpoint_path="data/checkpoints/telen/telen_best.pt",
    graph_cache_path="data/checkpoints/telen/graph_data_through_2020_v2.pt",
)

vectors = model.encode([
    "Điều 1: Thông tư này quy định về quản lý thuế...",
    "Điều 2: Đối tượng áp dụng là tổ chức và cá nhân...",
])

The tokenizer performs Vietnamese word segmentation with PyVi before calling the BKAI backbone. The model returns L2-normalized 768-dimensional vectors.

Training configuration

The BKAI backbone is frozen. Stage 1 trains the base projection for five epochs with triplet and InfoNCE losses. Stage 2 uses temporally ordered episodes and trains the projection, projection LayerNorm, GCN, state encoder, and HyperNetwork. The adaptation rank is 16, maximum sequence length is 480, and the reported inference path is deterministic.

Limitations

• Relevance is a same-parent-document proxy, not expert legal relevance.

• The BKAI model card reports pretraining on 80% of the Zalo 2021 challenge training set. Overlap with this corpus cannot be excluded.

• The full-history NDCG@10 result is 0.4606, substantially below the restricted 2021-pool result.

• The graph update does not produce a measurable ranking benefit for the audited checkpoint.

• Results come from one training seed.

Citation

@inproceedings{dinh2026telen,
  title     = {TELEN: Temporal Evolving Legal Embedding Network for Dynamic Vietnamese Law Retrieval},
  author    = {Dang Dinh},
  booktitle = {2026 30th International Computer Science and Engineering Conference (ICSEC)},
  year      = {2026},
  note      = {Accepted for presentation at ICSEC 2026}
}

Proceedings metadata should replace the provisional booktitle and note after publication.