Upload README.md with huggingface_hub

README.md

@@ -11,151 +11,50 @@ tags:
 - statute-retrieval
 - ukrainian
 - benchmark
+- temporal-degradation
 pretty_name: UA-StatuteRetrieval
 size_categories:
 - 1K<n<10K
 configs:
-- config_name: ablation_comparison
-  data_files:
-  - split: train
-    path: ablation_comparison/train-*
 - config_name: article_performance
   data_files:
   - split: train
     path: article_performance/train-*
-- config_name: embedding_drift
+- config_name: ablation_comparison
   data_files:
   - split: train
-    path: embedding_drift/train-*
+    path: ablation_comparison/train-*
 - config_name: sliding_window
   data_files:
   - split: train
     path: sliding_window/train-*
-dataset_info:
-- config_name: ablation_comparison
-  features:
-  - name: year
-    dtype: int64
-  - name: mrr_original
-    dtype: float64
-  - name: mrr_fixed_article
-    dtype: float64
-  - name: mrr_train_test
-    dtype: float64
-  splits:
-  - name: train
-    num_bytes: 288
-    num_examples: 9
-  download_size: 2525
-  dataset_size: 288
-- config_name: article_performance
-  features:
-  - name: target_article
-    dtype: string
-  - name: n_predictions
-    dtype: int32
-  - name: degree
-    dtype: int32
-  - name: mean_rank_cn
-    dtype: float64
-  - name: mean_rank_aa
-    dtype: float64
-  - name: hit5_cn
-    dtype: float64
-  - name: hit5_aa
-    dtype: float64
-  - name: mrr_cn
-    dtype: float64
-  - name: mrr_aa
-    dtype: float64
-  - name: law_number
-    dtype: string
-  - name: law_article
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 750943
-    num_examples: 3667
-  download_size: 225585
-  dataset_size: 750943
 - config_name: embedding_drift
-  features:
-  - name: article_key
-    dtype: string
-  - name: law_number
-    dtype: string
-  - name: law_article
-    dtype: string
-  - name: n_snippets_2012
-    dtype: int64
-  - name: n_snippets_2024
-    dtype: int64
-  - name: cosine_similarity
-    dtype: float64
-  - name: drift
-    dtype: float64
-  splits:
-  - name: train
-    num_bytes: 21522
-    num_examples: 116
-  download_size: 7923
-  dataset_size: 21522
-- config_name: sliding_window
-  features:
-  - name: eval_year
-    dtype: int64
-  - name: window_years
-    dtype: int64
-  - name: window_label
-    dtype: string
-  - name: train_years
-    dtype: string
-  - name: mrr_aa
-    dtype: float64
-  - name: mrr_cn
-    dtype: float64
-  - name: hit10_aa
-    dtype: float64
-  - name: hit10_cn
-    dtype: float64
-  - name: n_predictions
-    dtype: int64
-  - name: n_cases
-    dtype: int64
-  - name: n_articles
-    dtype: int64
-  - name: n_train_cases
-    dtype: int64
-  splits:
-  - name: train
-    num_bytes: 2986
-    num_examples: 30
-  download_size: 7095
-  dataset_size: 2986
+  data_files:
+  - split: train
+    path: embedding_drift/train-*
 ---
 
-# UA-StatuteRetrieval: Legal Article Prediction Benchmark
+# UA-StatuteRetrieval: A 20-Year Statute Retrieval Benchmark from 396M Ukrainian Court Citations
 
-A large-scale benchmark for evaluating legal statute retrieval methods on Ukrainian court decisions. Ground truth is derived exhaustively from 37M+ codex citations extracted from the Ukrainian Unified State Register of Court Decisions (EDRSR).
+A large-scale benchmark for evaluating temporal stability of legal statute retrieval methods. Ground truth is derived exhaustively from 396 million codex-article citations extracted from 101 million Ukrainian court decisions (2007--2026).
 
-## Dataset Description
+**Paper:** *Temporal Decay of Co-Citation Predictability: A 20-Year Statute Retrieval Benchmark from 396M Ukrainian Court Citations* (Ovcharov, 2025)
 
-Instead of manual annotations, this benchmark exploits **implicit relevance judgments**: when a judge cites Article 625 of the Civil Code, that constitutes a relevance judgment for that article given the case context.
+## Dataset Configs
 
-### Key Statistics (2024)
-- **37.5M** codex citations extracted
-- **2.25M** court decisions
-- **3,671** statutory articles evaluated
-- **1.8M** retrieval predictions (leave-one-out protocol)
+| Config | Description | Rows |
+|--------|-------------|-----:|
+| `article_performance` | Per-article MRR, Hit@5, degree for 3,667 articles (2024 snapshot) | 3,667 |
+| `ablation_comparison` | Original / fixed-article / train-test MRR per year (2008--2024) | 9 |
+| `sliding_window` | Mitigation experiment: MRR by eval year x window size (1, 3, 5, 10, all) | 30 |
+| `embedding_drift` | Per-article semantic drift (cosine distance 2012 -> 2024) via E5-large | 116 |
 
-## Files
+## Key Findings
 
-| File | Description |
-|------|-------------|
-| `article_retrieval_performance.parquet` | Per-article metrics: degree, mean rank, Hit@5, MRR for 3,667 articles (year 2024) |
-| `temporal_metrics.parquet` | Retrieval metrics by year (2007-2026) showing temporal degradation |
-| `temporal_metrics.csv` | Same as above in CSV format |
-| `difficulty_stratification.csv` | Performance by article frequency bin (hub/high/mid/low/rare) |
+1. **Co-citation predictability decays 33--47% over 12 years** (Mann-Kendall p < 0.005)
+2. **Decay is non-uniform**: criminal procedure remains stable (MRR ~0.40); civil law degrades from 0.35 to 0.15
+3. **Neither BM25 nor dense retrieval (E5-large, BGE-M3) escapes temporal degradation**
+4. **Sliding-window mitigation improves MRR by 3--28%** over cumulative indexing
 
 ## Evaluation Protocol
 
@@ -166,56 +65,64 @@ For each court decision with cited articles {a1, ..., an}:
 2. Score all candidate articles using remaining citations as seed
 3. Compute rank of masked article among non-seed candidates
 
-### Baselines
+### Baselines (2024 snapshot)
 
 | Method | Hit@10 | MRR |
 |--------|:------:|:---:|
 | Adamic-Adar | 0.545 | 0.272 |
 | Common Neighbors | 0.534 | 0.266 |
+| E5-large (dense) | 0.192 | 0.090 |
+| BGE-M3 (dense) | 0.240 | 0.096 |
+| BM25 (lexical) | 0.082 | 0.047 |
 | Degree Baseline | 0.111 | 0.059 |
 
 ### Difficulty Stratification
 
-| Bin | Articles | Hit@10 (AA) | MRR (AA) |
-|-----|:--------:|:-----------:|:--------:|
-| Hub (>100K cites) | 21 | 0.890 | 0.536 |
-| High (10K-100K) | 354 | 0.601 | 0.274 |
-| Mid (1K-10K) | 864 | 0.138 | 0.074 |
-| Low (100-1K) | 1,689 | 0.041 | 0.020 |
-| Rare (<100) | 739 | 0.012 | 0.010 |
-
-### Temporal Degradation
-
-MRR degrades from 0.42-0.56 (2008-2012) to 0.27-0.28 (2024-2026), demonstrating that co-citation patterns become less predictive over time as legal practice evolves.
+| Bin | Articles | MRR (AA) |
+|-----|:--------:|:--------:|
+| Hub (>100K cites) | 21 | 0.536 |
+| High (10K--100K) | 354 | 0.274 |
+| Mid (1K--10K) | 864 | 0.074 |
+| Low (100--1K) | 1,689 | 0.020 |
+| Rare (<100) | 739 | 0.010 |
 
 ## Usage
 
 ```python
 from datasets import load_dataset
 
-ds = load_dataset("overthelex/ua-statute-retrieval")
+# Per-article performance (2024)
+articles = load_dataset("overthelex/ua-statute-retrieval", "article_performance", split="train")
+print(articles.to_pandas()[["target_article", "degree", "mrr_aa"]].head(10))
+
+# Temporal ablation (Fig 2 in paper)
+ablation = load_dataset("overthelex/ua-statute-retrieval", "ablation_comparison", split="train")
+print(ablation.to_pandas())
+
+# Sliding-window mitigation (Table 6 in paper)
+sw = load_dataset("overthelex/ua-statute-retrieval", "sliding_window", split="train")
+print(sw.to_pandas().pivot(index="eval_year", columns="window_label", values="mrr_aa"))
 
-# Per-article performance
-articles = ds["train"].to_pandas()
-print(articles[["target_article", "degree", "mrr_aa", "hit5_aa"]].head(20))
+# Embedding drift (Fig 9 in paper)
+drift = load_dataset("overthelex/ua-statute-retrieval", "embedding_drift", split="train")
+print(drift.to_pandas().groupby("law_number")["drift"].mean().sort_values(ascending=False))
 ```
 
 ## Citation
 
 ```bibtex
-@article{keshchenko2025statute,
-  title={UA-StatuteRetrieval: A Full-Graph Benchmark for Legal Article Prediction from 37M Citation Records},
-  author={Keshchenko, Volodymyr},
+@article{ovcharov2025statute,
+  title={Temporal Decay of Co-Citation Predictability: A 20-Year Statute Retrieval Benchmark from 396M Ukrainian Court Citations},
+  author={Ovcharov, Volodymyr},
   year={2025}
 }
 ```
 
 ## Source Data
 
-Derived from the Ukrainian Unified State Register of Court Decisions (EDRSR). Raw citation data (502M records) is not published.
+Derived from the Ukrainian Unified State Register of Court Decisions (EDRSR, https://reyestr.court.gov.ua). Raw citation data (502M records) remains proprietary; the co-citation graph is available separately at [overthelex/ua-court-citation-graph](https://huggingface.co/datasets/overthelex/ua-court-citation-graph).
 
 ## Related Datasets
 
-- [overthelex/ua-court-citation-graph](https://huggingface.co/datasets/overthelex/ua-court-citation-graph) - Co-citation graph
-- [overthelex/ukrainian-court-decisions](https://huggingface.co/datasets/overthelex/ukrainian-court-decisions) - Court decision metadata
-- [overthelex/ua-case-outcome](https://huggingface.co/datasets/overthelex/ua-case-outcome) - Case outcome prediction
+- [overthelex/ua-court-citation-graph](https://huggingface.co/datasets/overthelex/ua-court-citation-graph) -- Co-citation graph (99.5M decisions)
+- [overthelex/ukrainian-court-decisions](https://huggingface.co/datasets/overthelex/ukrainian-court-decisions) -- Court decision metadata