Initial commit

.gitignore

@@ -0,0 +1,164 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+.idea/
+
+.cache/
+
+.gradio/

LICENSE

@@ -0,0 +1,9 @@
+MIT License
+
+Copyright (c) 2024 Saeed Abbasi
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

README.md

@@ -11,4 +11,43 @@ license: apache-2.0
 short_description: TextTiling using LLM Embeddings for Text Segmentation
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# LLM TextTiling Demo
+
+This directory contains the **demo code** for an Extended TextTiling application, which segments text into coherent chunks by leveraging **LLM embeddings** (via Sentence Transformers) and a semantic shift probability threshold.
+
+## Live Demo on Hugging Face Spaces
+
+You can try out the [**LLM TextTiling Demo**](https://huggingface.co/spaces/saeedabc/llm-text-tiling-demo) in your browser—no setup required.
+
+## Overview
+
+- **Input Text**: Paste or type in any text you wish to segment.  
+- **Embedding Model**: Choose from supported Sentence Transformers models (e.g., `all-mpnet-base-v2`).
+- **Window Size (`k`)**: Controls how many sentences on the left and right are compared for similarity.
+- **Pooling Strategy** (`max`, `mean`, `min`): Determines how to combine similarity scores for sentences in the window.
+- **Threshold**: The semantic shift probability above which a sentence boundary is declared a **segment boundary**.
+
+## Functionality
+
+1. **Tokenization**: Splits the text into sentences.
+2. **Embedding**: Each sentence is converted into a vector representation using the chosen Sentence Transformers model.
+3. **Cosine Similarity**: Sliding window comparisons of sentence vectors to detect shifts in topic or meaning.
+4. **Segmentation**: If the **semantic shift probability** = `(1 - similarity)` exceeds the threshold, a new segment is started.
+5. **Output**:
+   - **Segmented text** sentences grouped by segment 
+   - **JSON data** containing segmentation metadata and chunk details  
+   - **Plot** depicting the segmentation boundaries over a probability curve
+
+## Demo Code Files
+
+- **`app.py`**: The Gradio app script that runs the demo.
+- **`util.py`**: Utility functions for sentence tokenization and related operations.
+- **`requirements.txt`**: Lists the dependencies used for this demo.
+
+## Contributing
+
+Feel free to open issues or PRs in the main repository if you have feedback or suggestions for improving this demo.
+
+## License
+
+The code in this demo is provided under the [MIT License](LICENSE).  

app.py

@@ -0,0 +1,236 @@
+import hashlib
+import pickle
+from pathlib import Path
+from itertools import zip_longest
+
+import gradio as gr
+import torch
+from sentence_transformers import SentenceTransformer, util
+import numpy as np
+import matplotlib.pyplot as plt
+import ruptures as rpt
+
+from util import sent_tokenize
+
+
+# _OPENAI_MODELS = ['text-embedding-ada-002', 'text-embedding-3-small', 'text-embedding-3-large']
+_ST_MODELS = ['all-mpnet-base-v2', 'multi-qa-mpnet-base-dot-v1', 'all-MiniLM-L12-v2']
+
+CACHE_DIR = '.cache'
+DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
+
+plt.rcParams.update({
+    'font.family': 'Times New Roman',  #'Arial',  # or 'Helvetica', 'Times New Roman'
+    'font.size': 12,  # General font size
+    'axes.titlesize': 13,  # Font size for titles
+    'axes.labelsize': 12,  # Font size for axis labels
+    'xtick.labelsize': 11,  # Font size for x-tick labels
+    'ytick.labelsize': 11,   # Font size for y-tick labels
+    'legend.fontsize': 11,   # Font size for legend
+    'legend.title_fontsize': 11   # Font size for legend title
+})
+
+
+def embed_sentences(sentences, embedder_fn, cache_path):
+    if Path(cache_path).exists():
+        print(f'Loading embeddings from cache: {cache_path}')
+        with open(cache_path, 'rb') as file:
+            embedded_sents = pickle.load(file)
+    else:
+        print(f'Embedding sentences and saving to cache: {cache_path}')
+        embedded_sents = embedder_fn(sentences)
+        assert len(embedded_sents) == len(sentences)
+        
+        with open(cache_path, 'wb') as file:
+            pickle.dump(embedded_sents, file)
+            
+    return embedded_sents
+
+
+def calculate_cosine_similarities(embedded_sents, k=1, pool='mean'):
+    def cosine_similarity(a, b):
+        sim = util.cos_sim(a, b)
+        if pool == 'mean':
+            return sim.mean().item()
+        elif pool == 'max':
+            return sim.max().item()
+        elif pool == 'min':
+            return sim.min().item()
+        else:
+            raise ValueError(f'Invalid pooling method: {pool}')
+
+    cosine_sims = []
+    for i in range(len(embedded_sents) - 1):
+        lctx = embedded_sents[max(0, i-k+1) : i+1]
+        rctx = embedded_sents[i+1 : i+k+1]
+        sim = cosine_similarity(lctx, rctx)
+        cosine_sims.append(sim)
+    # cosine_sims.append(0.0)
+    
+    assert len(cosine_sims) == len(embedded_sents) - 1, f'{len(cosine_sims)} != {len(embedded_sents)}'
+    return cosine_sims
+
+
+def predict_boundaries(cosine_sims, threshold):
+    probs = [1.0 - sim for sim in cosine_sims]
+    preds = [1 if prob >= threshold else 0 for prob in probs]
+    return preds, probs
+    
+
+def output_segments(sents, preds, probs):
+    assert len(sents) - 1 == len(preds) == len(probs), f'{len(sents)} - 1 != {len(preds)} != {len(probs)}'
+    
+    def iter_segments(sents, preds, probs):
+        segment = []
+        for i, (sent, pred, prob) in enumerate(zip_longest(sents, preds, probs)):
+            segment.append({
+                # 'id': i + 1, 
+                'text': sent, 
+                'prob': round(prob, 4) if prob is not None else None,
+            })
+            if pred == 1:
+                yield segment
+                segment = []
+        if len(segment) > 0:
+            yield segment
+            segment = []
+    
+    out = {
+        'metadata': {},
+        'chunks': [],
+    }
+    n_segs = 0
+    n_sents = 0
+    for _, segment in enumerate(iter_segments(sents, preds, probs)):
+        # out['chunks'].append({
+        #     'id': n_segs + 1, 
+        #     'chunk': segment,
+        # })
+        out['chunks'].append(segment)
+        n_segs += 1
+        n_sents += len(segment)
+
+    out['metadata'] = {
+        'n_chunks': n_segs,
+        'n_sents': n_sents,
+        'prob_mean': round(np.mean(probs), 4),
+        'prob_std': round(np.std(probs), 4),
+        'prob_min': round(min(probs), 4),
+        'prob_max': round(max(probs), 4),
+    }
+    
+    out_text = "\n-------------------------\n".join(["\n".join([sent['text'] for sent in segment]) for segment in out['chunks']])
+    
+    def plot_regimes(signal, preds):
+        def get_bkps_from_labels(labels):
+            return [i+1 for i, l in enumerate(labels) if l == 1]
+
+        # signal = signal[:-1]
+        preds = preds + [1]
+        bkps = get_bkps_from_labels(preds)
+        
+        # print(f'signal(#{len(signal)}): {signal}')
+        # print(f'bkps(#{len(bkps)}): {bkps}')
+        # if not bkps or bkps[-1] != len(signal):
+        #     print('Note: last segment is incomplete!')
+        
+        fig, [ax] = rpt.display(np.array(signal), bkps, figsize=(10, 5), dpi=250)
+        y_min = max(0.0, min(signal) - 0.1)
+        y_max = min(1.0, max(signal) + 0.1)
+        ax.set_ylim(y_min, y_max)
+        ax.set_title("Segment Regimes")
+        ax.set_xlabel("Sentence Index")
+        ax.set_ylabel("Semantic Shift Probability")
+        fig.tight_layout()
+        
+        return fig
+
+    fig = plot_regimes(probs, preds)
+    
+    return out_text, out, fig
+
+
+def text_segmentation(input_text, model_name, k, pool, threshold):
+    if model_name in _ST_MODELS:
+        model = SentenceTransformer(model_name, device=DEVICE)
+        embedder_fn = model.encode   
+    else:
+        raise ValueError(f'Invalid model name: {model_name}')
+
+    sents = sent_tokenize(input_text, method='nltk', initial_split_sep='\n')
+    
+    cache_id = hashlib.md5(input_text.encode()).hexdigest()
+    cache_path = Path(CACHE_DIR) / f'{cache_id}.pkl'
+    embedded_sents = embed_sentences(sents, embedder_fn, cache_path=cache_path)
+    
+    cosine_sims = calculate_cosine_similarities(embedded_sents, k=k, pool=pool)
+    
+    preds, probs = predict_boundaries(cosine_sims, threshold=threshold)
+
+    return output_segments(sents, preds, probs)
+    
+
+# with gr.Blocks(css=".custom-tab { padding: 20px; margin: 20px; }") as app:
+with gr.Blocks() as app:
+    gr.Markdown("""
+# LLM TextTiling Demo
+
+An **extended** approach to text segmentation that combines **TextTiling** with **LLM embeddings**.  
+Simply provide your text, choose an embedding model, and adjust segmentation parameters (window size, threshold, pooling).  
+The demo will split your text into coherent segments based on **semantic shifts**.
+
+[**View the code on GitHub**](https://github.com/saeedabc/llm-text-tiling/demo)
+""")
+
+    with gr.Row():
+        with gr.Column():
+            input_text = gr.Textbox(label="Input Text", placeholder="Enter your text here...", lines=15)
+            
+            with gr.Row():
+                with gr.Column():
+                    # model_name = gr.Radio(choices=_ST_MODELS, label="Embedding Model", value=_ST_MODELS[0])
+                    model_name = gr.Dropdown(choices=_ST_MODELS, label="Embedding Model", value=_ST_MODELS[0])
+                
+                with gr.Column():
+                    pool = gr.Dropdown(choices=['max', 'mean', 'min'], label="Pooling Strategy", value='max')
+                    
+            with gr.Row():
+                with gr.Column():
+                    threshold = gr.Slider(minimum=0, maximum=1, step=0.01, label="Threshold", value=0.5)
+                
+                with gr.Column():
+                    k = gr.Slider(minimum=1, maximum=10, step=1, label="Window Size", value=3)
+
+
+            submit_button = gr.Button("Chunk Text")
+
+        with gr.Column():
+            with gr.Tabs():
+                with gr.Tab("Output Text"):
+                    output_text = gr.Textbox(label="Output Text", placeholder="Chunks will appear here...", lines=22)
+                with gr.Tab("Output Json"):
+                    output_json = gr.Json(label="Output Json", open=False, max_height=500)
+                with gr.Tab("Output Visualization"):  #, elem_classes="custom-tab"):            
+                    output_fig = gr.Plot(label="Output Visualization")
+
+    submit_button.click(text_segmentation, inputs=[input_text, model_name, k, pool, threshold], outputs=[output_text, output_json, output_fig])
+
+    examples = gr.Examples(
+        examples=[
+            ["Rib Mountain is a census-designated place (CDP) in the town of Rib Mountain in Marathon County, Wisconsin, United States. "
+             "The population was 5,651 at the 2010 census. "
+             "The community is named for Rib Mountain. "
+             "According to the United States Census Bureau, the CDP has a total area of 33.8 km² (13.0 mi²). "
+             "31.4 km² (12.1 mi²) of it is land and 2.4 km² (0.9 mi²) of it (6.98%) is water. "
+             "As of the census of 2000, there were 6,059 people, 2,211 households, and 1,782 families residing in the CDP. "
+             "The population density was 193.0/km² (499.8/mi²). "
+             "There were 2,278 housing units at an average density of 72.6/km² (187.9/mi²).", "all-mpnet-base-v2", 3, 'max', 0.52],
+        ],
+        inputs=[input_text, model_name, k, pool, threshold],
+    )
+
+if __name__ == '__main__':
+    Path(CACHE_DIR).mkdir(exist_ok=True)
+    
+    # Launch the app
+    app.launch()  # share=True)

requirements.txt

@@ -0,0 +1,86 @@
+aiofiles==23.2.1
+annotated-types==0.7.0
+anyio==4.8.0
+certifi==2024.12.14
+charset-normalizer==3.4.1
+click==8.1.8
+contourpy==1.3.1
+cycler==0.12.1
+fastapi==0.115.6
+ffmpy==0.5.0
+filelock==3.16.1
+fonttools==4.55.3
+fsspec==2024.12.0
+gradio==5.12.0
+gradio_client==1.5.4
+h11==0.14.0
+httpcore==1.0.7
+httpx==0.28.1
+huggingface-hub==0.27.1
+idna==3.10
+Jinja2==3.1.5
+joblib==1.4.2
+kiwisolver==1.4.8
+markdown-it-py==3.0.0
+MarkupSafe==2.1.5
+matplotlib==3.10.0
+mdurl==0.1.2
+mpmath==1.3.0
+networkx==3.4.2
+nltk==3.9.1
+numpy==2.2.1
+nvidia-cublas-cu12==12.4.5.8
+nvidia-cuda-cupti-cu12==12.4.127
+nvidia-cuda-nvrtc-cu12==12.4.127
+nvidia-cuda-runtime-cu12==12.4.127
+nvidia-cudnn-cu12==9.1.0.70
+nvidia-cufft-cu12==11.2.1.3
+nvidia-curand-cu12==10.3.5.147
+nvidia-cusolver-cu12==11.6.1.9
+nvidia-cusparse-cu12==12.3.1.170
+nvidia-nccl-cu12==2.21.5
+nvidia-nvjitlink-cu12==12.4.127
+nvidia-nvtx-cu12==12.4.127
+orjson==3.10.14
+packaging==24.2
+pandas==2.2.3
+pillow==11.1.0
+pydantic==2.10.5
+pydantic_core==2.27.2
+pydub==0.25.1
+Pygments==2.19.1
+pyparsing==3.2.1
+python-dateutil==2.9.0.post0
+python-multipart==0.0.20
+pytz==2024.2
+PyYAML==6.0.2
+regex==2024.11.6
+requests==2.32.3
+rich==13.9.4
+ruff==0.9.1
+ruptures==1.1.9
+safehttpx==0.1.6
+safetensors==0.5.2
+scikit-learn==1.6.1
+scipy==1.15.1
+semantic-version==2.10.0
+sentence-transformers==3.3.1
+setuptools==75.8.0
+shellingham==1.5.4
+six==1.17.0
+sniffio==1.3.1
+starlette==0.41.3
+sympy==1.13.1
+threadpoolctl==3.5.0
+tokenizers==0.21.0
+tomlkit==0.13.2
+torch==2.5.1
+tqdm==4.67.1
+transformers==4.48.0
+triton==3.1.0
+typer==0.15.1
+typing_extensions==4.12.2
+tzdata==2024.2
+urllib3==2.3.0
+uvicorn==0.34.0
+websockets==14.1

util.py

@@ -0,0 +1,59 @@
+import os
+
+
+### NLTK ###
+import nltk
+if not os.path.exists(os.path.join(nltk.data.find('tokenizers'), 'punkt')):
+    nltk.download('punkt')
+
+def nltk_sent_tokenize(texts: list[str]):
+    return (sent for text in texts for sent in nltk.sent_tokenize(text))
+
+
+# ### Spacy ###
+# import spacy
+# try:
+#     spacy_nlp = spacy.load('en_core_web_sm')
+# except OSError:
+#     spacy.cli.download("en_core_web_sm")
+#     spacy_nlp = spacy.load('en_core_web_sm')
+
+# def spacy_sent_tokenize(texts: list[str]):
+#     # nlp = spacy.load('en_core_web_sm')
+#     return (sent.text for text in texts for sent in spacy_nlp(text).sents)
+
+
+# ### Segtok ###
+# from segtok.segmenter import split_single, split_multi
+
+# def segtok_sent_tokenize(texts: list[str]):
+#     return (sent for text in texts for sent in split_single(text))
+
+
+### Sentence Tokenization ###
+
+def sent_tokenize(text, method: str = 'nltk', initial_split_sep: str = None) -> list[str]:
+    def has_info(text: str):
+        return any(char.isalnum() for char in text)
+
+    texts = [text] if isinstance(text, str) else text
+    assert isinstance(texts, list)
+
+    if initial_split_sep:
+        texts = [sline 
+                 for text in texts 
+                 for line in text.split(initial_split_sep) 
+                 if (sline := line.strip())]
+
+    if method == 'nltk':
+        sents = nltk_sent_tokenize(texts)
+    # elif method == 'spacy':
+    #     sents = spacy_sent_tokenize(texts)
+    # elif method == 'segtok':
+    #     sents = segtok_sent_tokenize(texts)
+    elif method == 'none':
+        sents = texts
+    else:
+        raise ValueError(f"Invalid method: {method}")
+    
+    return [ssent for sent in sents if (ssent := sent.strip()) and has_info(ssent)]