feat: try to get demo working after 2 years

2025-10-30 14:16:04 -07:00 · 2025-10-30 14:16:04 -07:00 · 8e2865c5ac
commit 8e2865c5ac
7 changed files with 1358 additions and 0 deletions
--- a/python3/.gitignore
+++ b/python3/.gitignore
@ -0,0 +1,12 @@
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+.venv/
+venv/
+ENV/
+
+# NLTK Data (uncomment if you want to download on each deployment)
+nltk_data/
--- a/python3/README.md
+++ b/python3/README.md
@ -0,0 +1,3 @@
+
+
+uv run flask --app salience run
--- a/python3/pyproject.toml
+++ b/python3/pyproject.toml
@ -0,0 +1,21 @@
+[project]
+name = "salience"
+version = "0.0.0"
+description = ""
+authors = [
+    { name = "Matt Neary", email = "neary.matt@gmail.com" }
+]
+license = { text = "MIT" }
+readme = "README.md"
+requires-python = ">=3.11"
+dependencies = [
+    "flask>=2.3.2,<3.0.0",
+    "transformers>=4.30.2,<5.0.0",
+    "nltk>=3.8.1,<4.0.0",
+    "sentence-transformers>=2.2.2,<3.0.0",
+    "numpy>=1.25.0,<2.0.0",
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
--- a/python3/salience/init.py
+++ b/python3/salience/init.py
@ -0,0 +1,18 @@
+from flask import Flask
+import numpy as np
+from .salience import extract
+import json
+
+app = Flask(__name__)
+
+with open('./transcript.txt', 'r') as file:
+    source_text = file.read().strip()
+sentence_ranges, adjacency = extract(source_text)
+
+@app.route("/salience")
+def salience_view():
+    return json.dumps({
+        'source': source_text,
+        'intervals': sentence_ranges,
+        'adjacency': np.nan_to_num(adjacency.numpy()).tolist(),
+    })
--- a/python3/salience/salience.py
+++ b/python3/salience/salience.py
@ -0,0 +1,63 @@
+import numpy as np
+import torch
+from sentence_transformers import SentenceTransformer
+import nltk.data
+import nltk
+import os
+
+# Set NLTK data path to project directory
+PROJECT_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+NLTK_DATA_DIR = os.path.join(PROJECT_DIR, 'nltk_data')
+
+# Add to NLTK's search path
+nltk.data.path.insert(0, NLTK_DATA_DIR)
+
+# Download to the custom location
+nltk.download('punkt', download_dir=NLTK_DATA_DIR)
+
+model = SentenceTransformer('all-mpnet-base-v2')
+sent_detector = nltk.data.load('tokenizers/punkt/english.pickle')
+
+def cos_sim(a, b):
+    sims = a @ b.T
+    a_norm = np.linalg.norm(a, axis=-1)
+    b_norm = np.linalg.norm(b, axis=-1)
+    a_normalized = (sims.T / a_norm.T).T
+    sims = a_normalized / b_norm
+    return sims
+
+def degree_power(A, k):
+    degrees = np.power(np.array(A.sum(1)), k).ravel()
+    D = np.diag(degrees)
+    return D
+
+def normalized_adjacency(A):
+    normalized_D = degree_power(A, -0.5)
+    return torch.from_numpy(normalized_D.dot(A).dot(normalized_D))
+
+def get_sentences(source_text):
+    sentence_ranges = list(sent_detector.span_tokenize(source_text))
+    sentences = [source_text[start:end] for start, end in sentence_ranges]
+    return sentences, sentence_ranges
+
+def text_rank(sentences):
+    vectors = model.encode(sentences)
+    adjacency = torch.tensor(cos_sim(vectors, vectors)).fill_diagonal_(0.)
+    adjacency[adjacency < 0] = 0
+    return normalized_adjacency(adjacency)
+
+def terminal_distr(adjacency, initial=None):
+    sample = initial if initial is not None else torch.full((adjacency.shape[0],), 1.)
+    scores = sample.matmul(torch.matrix_power(adjacency, 10)).numpy().tolist()
+    return scores
+
+def extract(source_text):
+    sentences, sentence_ranges = get_sentences(source_text)
+    adjacency = text_rank(sentences)
+    return sentence_ranges, adjacency
+
+def get_results(sentences, adjacency):
+    scores = terminal_distr(adjacency)
+    for score, sentence in sorted(zip(scores, sentences), key=lambda xs: xs[0]):
+        if score > 1.1:
+            print('{:0.2f}: {}'.format(score, sentence))
--- a/python3/salience/static/index.html
+++ b/python3/salience/static/index.html
@ -0,0 +1,120 @@
+CTYPE HTML>
+<html>
+  <head>
+    <meta charset="utf8" />
+    <title>Salience</title>
+    <script src="https://cdnjs.cloudflare.com/ajax/libs/mathjs/11.8.0/math.js" integrity="sha512-VW8/i4IZkHxdD8OlqNdF7fGn3ba0+lYqag+Uy4cG6BtJ/LIr8t23s/vls70pQ41UasHH0tL57GQfKDApqc9izA==" crossorigin="anonymous" referrerpolicy="no-referrer"></script>
+    <style>
+      body {
+        display: flex;
+        flex-direction: column;
+        height: 100vh;
+        margin: 0;
+      }
+      p {
+        width: 700px;
+        margin: 1em auto;
+        color: #4d4d4d;
+        font-family: sans-serif;
+        font-size: 15px;
+        line-height: 1.33em;
+        flex: 1;
+        overflow-y: scroll;
+      }
+      h1 {
+        width: 700px;
+        text-align: left;
+        margin: 15px auto;
+        margin-bottom: 0;
+        color: #000;
+        font-family: sans-serif;
+        font-size: 24px;
+      }
+      h1 span {
+        display: block;
+        font-size: 0.7em;
+        font-weight: normal;
+        color: #a0a0a0;
+      }
+      span.sentence {
+        --salience: 1;
+        background-color: rgba(249, 239, 104, var(--salience));
+      }
+      span.highlight {
+        background-color: rgb(185, 225, 244);
+      }
+      ::selection {
+        background: transparent;
+      }
+    </style>
+  </head>
+  <body>
+    <h1>
+      Salience
+      <span>automatic sentence highlights based on their significance to the document</span>
+    </h1>
+    <p id="content"></p>
+    <script type="text/javascript">
+      const content = document.querySelector('#content')
+      let adjacency = null
+      function scale(score) {
+        return Math.max(0, Math.min(1, score ** 3 - 0.95))
+      }
+      let exponent = 5
+      const redraw = () => {
+        if (!adjacency) return
+        const sentences = document.querySelectorAll('span.sentence')
+        if (!window.getSelection().isCollapsed) {
+          const sel = window.getSelection()
+          const fromNode = sel.anchorNode.parentNode
+          const toNode = sel.extentNode.parentNode
+          const fromIdx = Array.from(sentences).indexOf(fromNode)
+          const toIdx = Array.from(sentences).indexOf(toNode)
+          const range = [fromIdx, toIdx]
+          console.log('range', range)
+          range.sort((a, b) => a - b)
+          const vec = adjacency.map((x, i) => (i >= range[0] && i <= range[1]) ? 1 : 0)
+          const vec_sum = vec.reduce((a, x) => a + x, 0)
+          const scores = math.multiply(vec, adjacency).map(x => x * adjacency.length / vec_sum)
+          Array.from(sentences).forEach((node, i) => {
+            node.style.setProperty('--salience', scale(scores[i]))
+            if (i >= range[0] && i <= range[1]) node.classList.add('highlight')
+            else node.classList.remove('highlight')
+          })
+        } else {
+          const initial = adjacency.map(() => 1)
+          const scores = math.multiply(initial, math.pow(adjacency, exponent))
+          Array.from(sentences).forEach((node, i) => {
+            node.style.setProperty('--salience', scale(scores[i]))
+            node.classList.remove('highlight')
+          })
+        }
+      }
+      // Disabled functionality to center highlights on a selected fragment
+      // document.addEventListener('mousemove', redraw)
+      // document.addEventListener('mouseup', redraw)
+      fetch('/salience').then(async res => {
+        const data = await res.json()
+        console.log(data)
+        const source = data.source
+        const intervals = data.intervals
+        const tokens = intervals.map(([start, end]) => source.substr(start, end - start))
+        adjacency = data.adjacency
+        tokens.forEach((t, i) => {
+          const token = document.createElement('span')
+          token.innerText = t
+          token.classList.add('sentence')
+          content.appendChild(token)
+          if (tokens[i+1] && intervals[i+1][0] > intervals[i][1]) {
+            const intervening = document.createElement('span')
+            const start = intervals[i][1]
+            intervening.innerText = source.substr(start, intervals[i+1][0] - start)
+            content.appendChild(intervening)
+          }
+        })
+        redraw()
+      })
+    </script>
+  </body>
+</html>
+
--- a/python3/uv.lock
+++ b/python3/uv.lock