precinct 0.8.0

Approximate nearest-neighbor search over region embeddings
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195

//! In-domain validation: precinct's subsumption query over real *trained* box
//! embeddings.
//!
//! Loads a subsume box-embedding checkpoint (entities as `{mu, delta}` boxes,
//! `min = mu - exp(delta)/2`, `max = mu + exp(delta)/2`) trained on a WordNet
//! hypernym subset, then asks precinct to recover each concept's hypernym
//! ancestor two ways -- membership (`containing` the child's center) and strict
//! subsumption (`subsumers`) -- on real learned regions rather than synthetic or
//! clustered boxes. The gap between the two is itself the finding (see output).
//!
//! Data (gitignored): run `scripts/fetch_wordnet_boxes.sh` (it builds the
//! checkpoint with subsume's `save_checkpoint` example). Without it this example
//! prints instructions and exits 0.
//!
//! Run: cargo run --release --example wordnet_boxes

use precinct::{AxisBox, IndexParams, Region, RegionIndex, SearchParams};
use serde_json::Value;

const CHECKPOINT: &str = "data/wordnet_boxes.json";

/// WordNet `child parent` hypernym edges (the subset subsume trains on). Source:
/// WordNet noun hierarchy; `child ⊑ parent` means parent is the more general
/// concept, so its box should contain the child's.
const EDGES: &str = "\
dog.n.01 canine.n.02
canine.n.02 carnivore.n.01
carnivore.n.01 placental.n.01
placental.n.01 mammal.n.01
mammal.n.01 vertebrate.n.01
vertebrate.n.01 chordate.n.01
chordate.n.01 animal.n.01
animal.n.01 organism.n.01
organism.n.01 entity.n.01
cat.n.01 feline.n.01
feline.n.01 carnivore.n.01
wolf.n.01 canine.n.02
fox.n.01 canine.n.02
lion.n.01 feline.n.01
tiger.n.01 feline.n.01
horse.n.01 equine.n.01
equine.n.01 placental.n.01
eagle.n.01 bird_of_prey.n.01
bird_of_prey.n.01 bird.n.01
bird.n.01 vertebrate.n.01
sparrow.n.01 passerine.n.01
passerine.n.01 bird.n.01
salmon.n.01 fish.n.01
fish.n.01 vertebrate.n.01
trout.n.01 fish.n.01
oak.n.01 tree.n.01
tree.n.01 plant.n.02
plant.n.02 organism.n.01
pine.n.01 tree.n.01
rose.n.01 flower.n.01
flower.n.01 plant.n.02
tulip.n.01 flower.n.01
car.n.01 vehicle.n.01
vehicle.n.01 artifact.n.01
artifact.n.01 entity.n.01
truck.n.01 vehicle.n.01
bicycle.n.01 vehicle.n.01
whale.n.01 placental.n.01
dolphin.n.01 placental.n.01
snake.n.01 reptile.n.01
reptile.n.01 vertebrate.n.01
lizard.n.01 reptile.n.01
penguin.n.01 bird.n.01
bat.n.01 placental.n.01
spider.n.01 arthropod.n.01
arthropod.n.01 animal.n.01";

fn main() {
    let path = std::env::var("WORDNET_BOXES").unwrap_or_else(|_| CHECKPOINT.to_string());
    let Some(boxes_by_idx) = load_checkpoint(&path) else {
        eprintln!("Trained WordNet boxes not found at {path}.");
        eprintln!("Build with: scripts/fetch_wordnet_boxes.sh");
        return; // data-gated: a clean no-op when the dataset is absent.
    };

    // Reproduce subsume's entity interning: first appearance over the edges,
    // head then tail per line. Box index = interning order.
    let edges: Vec<(String, String)> = EDGES
        .lines()
        .filter_map(|l| {
            let mut it = l.split_whitespace();
            Some((it.next()?.to_string(), it.next()?.to_string()))
        })
        .collect();
    let mut id_of: std::collections::HashMap<String, u32> = std::collections::HashMap::new();
    let intern = |s: &str, m: &mut std::collections::HashMap<String, u32>| -> u32 {
        let n = m.len() as u32;
        *m.entry(s.to_string()).or_insert(n)
    };
    for (h, t) in &edges {
        intern(h, &mut id_of);
        intern(t, &mut id_of);
    }

    let dim = boxes_by_idx.values().next().map(|b| b.dim()).unwrap_or(0);
    let mut idx = RegionIndex::<AxisBox>::new(dim, IndexParams::default()).expect("index");
    for (&id, b) in &boxes_by_idx {
        if id_of.values().any(|&v| v == id) {
            idx.add(id, b.clone()).expect("add");
        }
    }
    idx.build().expect("build");
    println!(
        "Loaded {} trained WordNet concept boxes (dim {dim}).",
        idx.len()
    );

    // For each edge child ⊑ parent: does precinct surface `parent` as a subsumer
    // of `child`? (subsumers = regions whose box contains the child's box.)
    let params = || SearchParams {
        ef: 64,
        overretrieve: 16,
    };
    let min_prob = 0.3;
    let (mut strict, mut soft, mut member, mut total) = (0usize, 0usize, 0usize, 0usize);
    for (child, parent) in &edges {
        let (Some(&cid), Some(&pid)) = (id_of.get(child), id_of.get(parent)) else {
            continue;
        };
        let (Some(cbox), Some(_)) = (boxes_by_idx.get(&cid), boxes_by_idx.get(&pid)) else {
            continue;
        };
        total += 1;
        if idx
            .subsumers(cbox, params())
            .unwrap_or_default()
            .contains(&pid)
        {
            strict += 1;
        }
        if idx
            .subsumers_soft(cbox, min_prob, params())
            .unwrap_or_default()
            .iter()
            .any(|(id, _)| *id == pid)
        {
            soft += 1;
        }
        if idx
            .containing(cbox.center(), params())
            .unwrap_or_default()
            .contains(&pid)
        {
            member += 1;
        }
    }
    let pct = |h: usize| h as f64 / total as f64 * 100.0;
    println!("hypernym-ancestor recall over {total} edges:");
    println!(
        "  membership  (parent encloses child center):     {:.0}%",
        pct(member)
    );
    println!(
        "  soft subsumers (entailment_prob >= {min_prob}):       {:.0}%",
        pct(soft)
    );
    println!(
        "  strict subsumers (parent box ⊇ child box):      {:.0}%",
        pct(strict)
    );
    println!(
        "The trained Gumbel boxes nest *softly*: a child's center lands inside its\n\
         parent, but the child's full box pokes outside. So membership and the soft\n\
         subsumption query recover the is-a ancestor while strict box-containment\n\
         does not -- the index is sound; the trained embedding is soft."
    );
}

/// Parse a subsume box checkpoint (`{boxes: {idx: {mu, delta}}, dim}`) into
/// `AxisBox`es keyed by entity index.
fn load_checkpoint(path: &str) -> Option<std::collections::BTreeMap<u32, AxisBox>> {
    let text = std::fs::read_to_string(path).ok()?;
    let doc: Value = serde_json::from_str(&text).ok()?;
    let boxes = doc.get("boxes")?.as_object()?;
    let mut out = std::collections::BTreeMap::new();
    for (k, v) in boxes {
        let id: u32 = k.parse().ok()?;
        let mu = as_f32_vec(v.get("mu")?)?;
        let delta = as_f32_vec(v.get("delta")?)?;
        out.insert(id, AxisBox::from_mu_delta(mu, delta));
    }
    (!out.is_empty()).then_some(out)
}

fn as_f32_vec(v: &Value) -> Option<Vec<f32>> {
    v.as_array()?
        .iter()
        .map(|x| x.as_f64().map(|f| f as f32))
        .collect()
}