anno

Information extraction for unstructured text: named entity recognition (NER), within-document coreference resolution, and structured pattern extraction.

Dual-licensed under MIT or Apache-2.0.

Tasks

Named entity recognition. Given input text, identify spans (start, end, type, confidence) where each span denotes a named entity [1, 2]. Entity types follow standard taxonomies (PER, ORG, LOC, MISC for CoNLL-style [2]) or caller-defined labels for zero-shot extraction. Offsets are character offsets (Unicode scalar values), not byte offsets; see CONTRACT.md.

Coreference resolution. Identify mention spans and group them into clusters, where each cluster tracks a consistent discourse referent within the document [3, 4]. Referents may be concrete or abstract, singular or plural, real-world or fictional — the constraint is within-document consistency, not ontological uniqueness. Coreferring mentions span proper names, definite descriptions, and pronouns: "Sophie Wilson", "the designer", and "She" form one cluster.

Relation extraction. Extract typed (head, relation, tail) triples from text. Available on RelationCapable backends; others fall back to co-occurrence edges for graph export.

Structured pattern extraction. Dates, monetary amounts, email addresses, URLs, phone numbers via deterministic regex grammars.

Backends

All backends produce the same output type: variable-length spans with character offsets.

ML backends are feature-gated (onnx or candle). Weights download from HuggingFace on first use. All backends expose model.capabilities() for runtime discovery. See BACKENDS.md for selection guidance and feature-flag details.

Install

cargo install --git https://github.com/arclabs561/anno --package anno-cli --bin anno --features "onnx eval"

From a local clone:

cargo install --path crates/anno-cli --bin anno --features "onnx eval"

ANNO_NO_DOWNLOADS=1 or HF_HUB_OFFLINE=1 forces cached-only behavior.

Examples

anno extract --text "Lynn Conway worked at IBM and Xerox PARC in California."

PER:1 "Lynn Conway"
ORG:2 "IBM" "Xerox PARC"
LOC:1 "California"

JSON output (schema-stable; uses pattern for offline reproducibility):

anno extract --model pattern --format json --text "Contact jobs@acme.com by March 15 for the \$50K role."

[
  {"text": "jobs@acme.com", "entity_type": "EMAIL", "start": 8, "end": 21, "confidence": 0.98},
  {"text": "March 15", "entity_type": "DATE", "start": 25, "end": 33, "confidence": 0.95},
  {"text": "$50K", "entity_type": "MONEY", "start": 42, "end": 46, "confidence": 0.95}
]

Zero-shot custom entity types (via GLiNER [5]):

anno extract --model gliner --extract-types "DRUG,SYMPTOM" \
  --text "Aspirin can treat headaches and reduce fever."

drug:1 "Aspirin" symptom:2 "headaches" "fever"

Coreference:

anno debug --coref -t "Sophie Wilson designed the ARM processor. She revolutionized mobile computing."

Coreference: "Sophie Wilson" → "She"

Downstream

Filter and pipe JSON output:

# All person entities, text only
anno extract --format json --file article.txt \
  | jq '[.[] | select(.entity_type == "PER") | .text]'

# Unique organizations, sorted
anno extract --format json --file article.txt \
  | jq '[.[] | select(.entity_type == "ORG") | .text] | unique | sort[]'

Batch a directory (parallel, cached):

anno batch --dir docs/ --parallel 4 --cache --output results/

# Stream stdin JSONL: {"id":"…","text":"…"} per line
cat corpus.jsonl | anno batch --stdin --parallel 4 --cache --output results/

Knowledge Graph (RDF)

anno export emits standard N-Triples or JSON-LD -- loadable into any RDF store (Oxigraph, Jena, Blazegraph, etc.). --base-uri sets the IRI namespace:

# Own namespace (recommended for private corpora)
anno export --input docs/ --output /tmp/kg/ --format ntriples \
  --base-uri https://myproject.example.com/kg/

# Align to DBpedia (widely used LOD namespace)
anno export --input docs/ --output /tmp/kg/ --format ntriples \
  --base-uri https://dbpedia.org/resource/

Default (urn:anno:) produces stable URNs suitable for local use without a registered domain. The output is standard W3C N-Triples, so any SPARQL-capable store can query it:

# Example: load into a SPARQL store and query
cat /tmp/kg/*.nt | your-rdf-store load --format ntriples

your-rdf-store query 'SELECT ?label WHERE {
  ?e a <https://myproject.example.com/kg/vocab#PERType> ;
     <http://www.w3.org/2000/01/rdf-schema#label> ?label .
}'

Semantic relation triples — RelationCapable backends (tplinker, gliner2) produce typed (head, relation, tail) triples instead of co-occurrence edges. Use --format graph-ntriples (--features graph) for routing through the internal graph substrate:

anno export --input docs/ --output /tmp/kg/ \
  --format graph-ntriples --model gliner2 \
  --base-uri https://myproject.example.com/kg/
# → <entity/person/0_Lynn_Conway_0> <rel/works_for> <entity/org/1_IBM_13> .

Property Graph CSV

Export node and edge tables as CSV for import into any property-graph database:

# Semantic edges (gliner2 or tplinker)
anno export --input docs/ --output /tmp/kg/ --format kuzu --model gliner2

# Co-occurrence edges (any other backend)
anno export --input docs/ --output /tmp/kg/ --format kuzu

Each file produces {stem}-nodes.csv + {stem}-edges.csv with columns:

• Nodes: id, entity_type, text, start, end, source
• Edges: from, to, rel_type, confidence

Library

[dependencies]
anno-lib = "0.3"

The crate is published as anno-lib on crates.io; the Rust import name is anno.

Feature flags

The onnx feature (default) pulls in ort (ONNX Runtime bindings), which requires a C++ runtime. For minimal builds, use default-features = false.

Basic extraction:

use anno::{Model, StackedNER};

let m = StackedNER::default();
let ents = m.extract_entities("Sophie Wilson designed the ARM processor.", None)?;
assert!(!ents.is_empty());
# Ok::<(), anno::Error>(())

Zero-shot custom types via DynamicLabels (GLiNER, GLiNER2, NuNER):

use anno::{DynamicLabels, GLiNEROnnx};

let m = GLiNEROnnx::new("onnx-community/gliner_small-v2.1")?;
let ents = m.extract_with_labels(
    "Aspirin treats headaches.",
    &["drug", "symptom"],
    None,
)?;
# Ok::<(), anno::Error>(())

Runtime capability discovery — all backends now report accurately:

use anno::Model;

fn print_caps(m: &dyn Model) {
    let c = m.capabilities();
    println!(
        "batch={} streaming={} gpu={} relations={} zero-shot={}",
        c.batch_capable, c.streaming_capable, c.gpu_capable,
        c.relation_capable, c.dynamic_labels,
    );
}

Architecture

Pipeline: Text → Extract → Coalesce → structured output. See ARCHITECTURE.md.

Evaluation

anno-eval provides dataset loading, backend-vs-dataset compatibility gating, and CoNLL-style [2] span-level evaluation (precision, recall, F1) with label mapping between backend and dataset taxonomies.

anno benchmark --model gliner --dataset conll2003

anno sampler (--features eval) provides a randomized matrix sampler with two modes: triage (worst-first) and measure (ML-only stable measurement).

Sampler examples

# Deterministic offline decision-loop smoke example (also run in CI).
cargo run -p anno-eval --example muxer_decision_loop --features eval

# Run the CI sampler harness locally (uses ~/.anno_cache for history/cache).
just ci-matrix-local 42 ner

Scope

Inference-time extraction only. Training is out of scope — use upstream frameworks (Hugging Face Transformers, Flair, etc.) and export ONNX weights for consumption.

Documentation

• QUICKSTART
• CONTRACT — offset semantics, scope, feature gating
• BACKENDS — backend selection, architecture, feature flags
• ARCHITECTURE — crate layout, dependency flow
• REFERENCES — full bibliography (NER, coref, relation extraction, software)
• API docs
• Changelog

References

Key citations; see docs/REFERENCES.md for the full list with links.

1. Grishman & Sundheim, COLING 1996 (MUC-6 NER). [PDF]
2. Tjong Kim Sang & De Meulder, CoNLL 2003 (CoNLL benchmark). [PDF]
3. Lee et al., EMNLP 2017 (end-to-end coreference). [arXiv]
4. Jurafsky & Martin, SLP3 2024 (coreference fundamentals). [Online]
5. Zaratiana et al., NAACL 2024 (GLiNER). [arXiv]
6. Bogdanov et al., 2024 (NuNER). [arXiv]
7. Li et al., AAAI 2022 (W2NER). [arXiv]
8. Devlin et al., NAACL 2019 (BERT). [arXiv]
9. Lafferty et al., ICML 2001 (CRF).
10. Wang et al., COLING 2020 (TPLinker). [arXiv]
11. Zaratiana et al., 2025 (GLiNER2). [arXiv]
12. Rabiner, Proceedings of the IEEE 1989 (HMM tutorial). [PDF]

Citeable via CITATION.cff.

License

Dual-licensed under MIT or Apache-2.0.