semantic-compressor 2.3__tar.gz → 2.5__tar.gz

{semantic_compressor-2.3/semantic_compressor.egg-info → semantic_compressor-2.5}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: semantic_compressor
-Version: 2.3
+Version: 2.5
 Author: Carlo Moro
 Author-email: Carlo Moro <cnmoro@gmail.com>
 Classifier: Programming Language :: Python :: 3
@@ -9,13 +9,14 @@ Classifier: Operating System :: OS Independent
 Requires-Python: >=3.7
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: numpy<2
+Requires-Dist: numpy
 Requires-Dist: nltk
 Requires-Dist: scikit-learn
-Requires-Dist: fasttext
+Requires-Dist: lingua-language-detector
 Requires-Dist: model2vec
 Requires-Dist: pyspellchecker
 Dynamic: author
+Dynamic: license-file
 Dynamic: requires-python
 
 ```python

semantic_compressor-2.5/compressor/semantic.py

@@ -0,0 +1,414 @@
+import os, sys, importlib.resources, functools, re
+
+_NLTK_DATA_PATH = None
+try:
+    _NLTK_DATA_PATH = str(importlib.resources.files('compressor').joinpath('resources/nltk_data'))
+except Exception:
+    _NLTK_DATA_PATH = os.path.join(os.path.dirname(__file__), 'resources', 'nltk_data')
+
+os.environ['NLTK_DATA'] = _NLTK_DATA_PATH
+
+from collections import Counter
+
+_PUNCT_REATTACH = re.compile(r'\s+([.!,\?;:])')
+_PUNCT_BOUNDARY = re.compile(r'([.!,\?;:])(?=\S)')
+_HYPHENATION = re.compile(r'(\w)-\s*\n\s*(\w)')
+_NOISE_CHARS = re.compile(r'[\|\•\[\]\(\)\"“”]')
+_LEADING_HYPHEN = re.compile(r'(?m)^\s*-\s*')
+_STRAY_HYPHEN = re.compile(r'(?<!\w)-(?!\w)')
+_REPEATED_PUNCT = re.compile(r'([!?.,;:]){2,}')
+_MULTI_SPACE = re.compile(r'[ \t]+')
+_MULTI_NEWLINE = re.compile(r'\n{2,}')
+_AGGRESSIVE_CLEAN = re.compile(r'[^A-Za-zÀ-ÿ\s\.\,\;\:\?\!]')
+_MULTI_SPACE2 = re.compile(r'\s{2,}')
+
+_EN_STOPWORDS_PATH = str(importlib.resources.files('compressor').joinpath('resources/en_stopwords.pkl'))
+_PT_STOPWORDS_PATH = str(importlib.resources.files('compressor').joinpath('resources/pt_stopwords.pkl'))
+
+
+@functools.lru_cache(maxsize=1)
+def _ensure_nltk_ready():
+    import nltk.data
+    nltk.data.path.insert(0, _NLTK_DATA_PATH)
+
+
+@functools.lru_cache(maxsize=1)
+def _get_tokenizer():
+    from compressor.minbpe.regex import RegexTokenizer
+    return RegexTokenizer()
+
+
+@functools.lru_cache(maxsize=1)
+def _get_english_stemmer():
+    _ensure_nltk_ready()
+    from nltk.stem import PorterStemmer
+    return PorterStemmer()
+
+
+@functools.lru_cache(maxsize=1)
+def _get_portuguese_stemmer():
+    _ensure_nltk_ready()
+    from nltk.stem import RSLPStemmer
+    return RSLPStemmer()
+
+
+@functools.lru_cache(maxsize=1)
+def _get_language_detector():
+    from lingua import Language, LanguageDetectorBuilder
+    return LanguageDetectorBuilder.from_languages(
+        Language.ENGLISH, Language.PORTUGUESE
+    ).build()
+
+
+@functools.lru_cache(maxsize=1)
+def _get_language_enums():
+    from lingua import Language
+    return Language
+
+
+@functools.lru_cache(maxsize=1)
+def _get_english_stopwords():
+    import pickle
+    return pickle.load(open(_EN_STOPWORDS_PATH, "rb"))
+
+
+@functools.lru_cache(maxsize=1)
+def _get_portuguese_stopwords():
+    import pickle
+    return pickle.load(open(_PT_STOPWORDS_PATH, "rb"))
+
+
+@functools.lru_cache(maxsize=1)
+def _get_embedding_model():
+    from model2vec import StaticModel
+    return StaticModel.from_pretrained("cnmoro/static-nomic-eng-ptbr-tiny")
+
+
+@functools.lru_cache(maxsize=1)
+def _get_hashing_vectorizer():
+    from sklearn.feature_extraction.text import HashingVectorizer
+    return HashingVectorizer(ngram_range=(1, 6), analyzer='char', n_features=512)
+
+
+@functools.lru_cache(maxsize=1)
+def _get_sent_tokenize():
+    _ensure_nltk_ready()
+    from nltk.tokenize import sent_tokenize
+    return sent_tokenize
+
+
+def clean_text(text: str) -> str:
+    text = _HYPHENATION.sub(r'\1\2', text)
+    text = _NOISE_CHARS.sub(' ', text)
+    text = _LEADING_HYPHEN.sub('', text)
+    text = _STRAY_HYPHEN.sub(' ', text)
+    text = _REPEATED_PUNCT.sub(r'\1', text)
+    text = _MULTI_SPACE.sub(' ', text)
+    text = _MULTI_NEWLINE.sub('\n', text).strip()
+
+    alpha_ratio = sum(c.isalpha() for c in text) / max(len(text), 1)
+    if alpha_ratio < 0.8:
+        text = _AGGRESSIVE_CLEAN.sub(' ', text)
+        text = _MULTI_SPACE2.sub(' ', text).strip()
+
+    text = _PUNCT_REATTACH.sub(r'\1', text)
+    text = _PUNCT_BOUNDARY.sub(r'\1 ', text)
+    return text
+
+
+def extract_textual_embeddings(text):
+    v = _get_hashing_vectorizer()
+    import numpy as np
+    return np.asarray(v.transform([text]).sum(axis=0)).ravel().tolist()
+
+
+def extract_semantic_embeddings(text):
+    return _get_embedding_model().encode([text])[0]
+
+
+def structurize_text(full_text, tokens_per_chunk=300, chunk_overlap=0):
+    tok = _get_tokenizer()
+    chunks = []
+    current_chunk = []
+    current_chunk_length = 0
+    tokens = tok.encode(full_text)
+    for i, token in enumerate(tokens):
+        if current_chunk_length + 1 > tokens_per_chunk:
+            chunks.append(current_chunk)
+            current_chunk = tokens[i - chunk_overlap:i] if i > chunk_overlap else []
+            current_chunk_length = len(current_chunk)
+        current_chunk.append(token)
+        current_chunk_length += 1
+    chunks.append(current_chunk)
+    return [tok.decode(chunk) for chunk in chunks]
+
+
+def count_tokens(text):
+    return len(_get_tokenizer().encode(text))
+
+
+def detect_language(text):
+    Language = _get_language_enums()
+    lang = _get_language_detector().detect_language_of(text)
+    return 'pt' if lang == Language.PORTUGUESE else 'en'
+
+
+def compute_and_remove_repeated_ngrams(text, ngram_size=3, threshold=3):
+    words = text.split()
+    n = len(words)
+    if n < ngram_size:
+        return text
+
+    ngram_tuples = [tuple(words[i:i + ngram_size]) for i in range(n - ngram_size + 1)]
+    counter = Counter(ngram_tuples)
+    repeated = [ng for ng, count in counter.items() if count > threshold]
+    if not repeated:
+        return text
+
+    for ng in repeated:
+        first = True
+        i = 0
+        while i <= len(words) - ngram_size:
+            if tuple(words[i:i + ngram_size]) == ng:
+                if first:
+                    first = False
+                    i += ngram_size
+                else:
+                    del words[i:i + ngram_size]
+            else:
+                i += 1
+    return ' '.join(words)
+
+
+def calculate_similarity(embed1, embed2):
+    from sklearn.metrics.pairwise import cosine_similarity
+    return cosine_similarity([embed1], [embed2])[0][0]
+
+
+def _get_stopwords(lang):
+    if lang == 'pt':
+        return _get_portuguese_stopwords()
+    return _get_english_stopwords()
+
+
+def semantic_compress_text(full_text, compression_rate=0.7, num_topics=5, reference_text: str = None, perform_cleaning: bool = True):
+    import warnings
+    from sklearn.feature_extraction.text import TfidfVectorizer
+    from sklearn.decomposition import TruncatedSVD
+    import numpy as np
+    import traceback
+
+    try:
+        if perform_cleaning:
+            full_text = clean_text(full_text)
+
+        sent_tokenize = _get_sent_tokenize()
+        sentences = sent_tokenize(full_text)
+
+        final_sentences = []
+        for s in sentences:
+            final_sentences.extend(s.split('\n'))
+        sentences = final_sentences
+        n_sentences = len(sentences)
+
+        text_lang = detect_language(full_text)
+        stopwords = _get_stopwords(text_lang)
+
+        if n_sentences >= 3:
+            n_topics = min(num_topics, max(2, n_sentences // 5))
+            max_features = min(3000, max(500, n_sentences * 10))
+
+            vectorizer = TfidfVectorizer(stop_words=stopwords, max_features=max_features)
+            doc_term_matrix = vectorizer.fit_transform(sentences)
+            svd = TruncatedSVD(n_components=n_topics, random_state=42)
+            with warnings.catch_warnings():
+                warnings.filterwarnings('ignore', category=RuntimeWarning, message='.*divide by zero.*')
+                svd.fit(doc_term_matrix)
+            topic_scores = np.abs(svd.transform(vectorizer.transform(sentences)))
+        else:
+            topic_scores = np.ones((n_sentences, 1)) * 0.5
+
+        doc_embedding = extract_semantic_embeddings(full_text)
+
+        if reference_text is not None:
+            reference_text_embedding = extract_semantic_embeddings(reference_text)
+            doc_embedding = 0.6 * doc_embedding + 0.4 * reference_text_embedding
+
+        sentence_embeddings = _get_embedding_model().encode(sentences)
+
+        sentence_scores = []
+        for i, sentence in enumerate(sentences):
+            sentence_embedding = sentence_embeddings[i]
+            semantic_similarity = calculate_similarity(doc_embedding, sentence_embedding)
+
+            topic_importance = float(np.max(topic_scores[i]))
+
+            words = sentence.split()
+            unique_words = set(w.lower() for w in words if w.lower() not in stopwords)
+            lexical_diversity = len(unique_words) / len(words) if words else 0
+
+            importance = 0.6 * semantic_similarity + 0.3 * topic_importance + 0.2 * lexical_diversity
+            sentence_scores.append((sentence, importance))
+
+        sorted_sentences = sorted(sentence_scores, key=lambda x: x[1], reverse=True)
+
+        total_words = sum(len(s.split()) for s in sentences)
+        target_words = int(total_words * compression_rate)
+
+        compressed_text = []
+        current_words = 0
+        for sentence, _ in sorted_sentences:
+            sentence_words = len(sentence.split())
+            if current_words + sentence_words <= target_words:
+                compressed_text.append(sentence)
+                current_words += sentence_words
+            else:
+                break
+
+        if not compressed_text:
+            compressed_text = [sentences[0]]
+
+        compressed_text.sort(key=lambda x: sentences.index(x))
+        compressed_text = [s.capitalize() for s in compressed_text]
+
+        cleaned_compressed_text = ' '.join(compressed_text).replace('  ', ' ').strip()
+        cleaned_compressed_text = compute_and_remove_repeated_ngrams(cleaned_compressed_text)
+        return cleaned_compressed_text
+    except Exception:
+        traceback.print_exc()
+    return full_text
+
+
+def compress_text(text, *, target_token_count=None, compression_rate=0.7, reference_text_steering=None, perform_cleaning=True):
+    import traceback
+    try:
+        if target_token_count is None:
+            compression_rate = 1 - compression_rate
+        else:
+            original_token_count = count_tokens(text)
+            if original_token_count <= target_token_count:
+                return text
+            compression_rate = target_token_count / original_token_count
+
+        return semantic_compress_text(
+            full_text=text,
+            compression_rate=compression_rate,
+            reference_text=reference_text_steering,
+            perform_cleaning=perform_cleaning
+        )
+    except Exception:
+        traceback.print_exc()
+    return text
+
+
+def stem_text(text, lang='en'):
+    if lang == 'en':
+        stemmer = _get_english_stemmer()
+    else:
+        stemmer = _get_portuguese_stemmer()
+    return ' '.join(stemmer.stem(word) for word in text.split())
+
+
+def correct_spelling(sentence, detected_lang="pt"):
+    from spellchecker import SpellChecker
+    spell = SpellChecker(language=detected_lang)
+    words = sentence.split()
+    fixed = [spell.correction(word) for word in words]
+    result = []
+    for original, fixed_word in zip(words, fixed):
+        result.append(fixed_word if fixed_word is not None else original)
+    return ' '.join(result)
+
+
+def preprocess_and_extract_textual_embedding(block, use_stemming, lang):
+    processed_block = block.lower() if not use_stemming else stem_text(block.lower(), lang)
+    return extract_textual_embeddings(processed_block)
+
+
+def find_needle_in_haystack(
+        *, haystack: str, needle: str, block_size=300,
+        embedding_mode: str = 'both',
+        semantic_embeddings_weight: float = 0.3,
+        textual_embeddings_weight: float = 0.7,
+        use_stemming: bool = False,
+        correct_spelling_needle: bool = False
+    ):
+    import traceback
+    try:
+        if embedding_mode not in {'semantic', 'textual', 'both'}:
+            raise ValueError("Invalid embedding_mode. Choose 'semantic', 'textual', or 'both'.")
+
+        blocks = structurize_text(haystack, tokens_per_chunk=block_size)
+
+        lang = detect_language(f"{needle}\n\n{haystack}")
+
+        if correct_spelling_needle:
+            needle = correct_spelling(needle, lang)
+
+        needle_semantic_embedding = None
+        needle_textual_embedding = None
+
+        if embedding_mode in {'semantic', 'both'}:
+            needle_semantic_embedding = extract_semantic_embeddings(needle)
+
+        if embedding_mode in {'textual', 'both'}:
+            needle_textual_embedding = extract_textual_embeddings(
+                needle.lower() if not use_stemming else stem_text(needle, lang)
+            )
+
+        haystack_semantic_embeddings = []
+        haystack_textual_embeddings = []
+
+        if embedding_mode in {'semantic', 'both'}:
+            if len(blocks) == 1:
+                haystack_semantic_embeddings = [extract_semantic_embeddings(blocks[0])]
+            else:
+                from concurrent.futures import ProcessPoolExecutor
+                with ProcessPoolExecutor() as executor:
+                    haystack_semantic_embeddings = list(executor.map(extract_semantic_embeddings, blocks))
+
+        if embedding_mode in {'textual', 'both'}:
+            if len(blocks) == 1:
+                haystack_textual_embeddings = [preprocess_and_extract_textual_embedding(blocks[0], use_stemming, lang)]
+            else:
+                from concurrent.futures import ProcessPoolExecutor
+                from multiprocessing import cpu_count
+                with ProcessPoolExecutor(max_workers=int(cpu_count() // 1.5)) as executor:
+                    haystack_textual_embeddings = list(
+                        executor.map(preprocess_and_extract_textual_embedding, blocks, [use_stemming] * len(blocks), [lang] * len(blocks))
+                    )
+
+        semantic_similarities = []
+        textual_similarities = []
+
+        if embedding_mode in {'semantic', 'both'}:
+            semantic_similarities = [
+                calculate_similarity(needle_semantic_embedding, be)
+                for be in haystack_semantic_embeddings
+            ]
+
+        if embedding_mode in {'textual', 'both'}:
+            textual_similarities = [
+                calculate_similarity(needle_textual_embedding, be)
+                for be in haystack_textual_embeddings
+            ]
+
+        if embedding_mode == 'semantic':
+            sorted_blocks = sorted(zip(blocks, semantic_similarities), key=lambda x: x[1], reverse=True)
+        elif embedding_mode == 'textual':
+            sorted_blocks = sorted(zip(blocks, textual_similarities), key=lambda x: x[1], reverse=True)
+        else:
+            sorted_blocks = sorted(
+                zip(blocks, semantic_similarities, textual_similarities),
+                key=lambda x: x[1] * semantic_embeddings_weight + x[2] * textual_embeddings_weight,
+                reverse=True
+            )
+
+        most_similar_block = sorted_blocks[0][0]
+        most_similar_block_index = blocks.index(most_similar_block)
+        start_index = most_similar_block_index - 1 if most_similar_block_index > 0 else 0
+        needle_region = blocks[start_index:most_similar_block_index + 2]
+        return ''.join(needle_region).strip()
+    except Exception:
+        traceback.print_exc()
+    return haystack

{semantic_compressor-2.3 → semantic_compressor-2.5}/pyproject.toml

@@ -1,10 +1,10 @@
 [build-system]
-requires = ["setuptools>=61.0", "numpy<2", "nltk", "scikit-learn", "fasttext", "onnxruntime", "onnxruntime-extensions", "pyspellchecker"]
+requires = ["setuptools>=61.0"]
 build-backend = "setuptools.build_meta"
 
 [project]
 name = "semantic_compressor"
-version = "2.3"
+version = "2.5"
 authors = [
   { name="Carlo Moro", email="cnmoro@gmail.com" },
 ]
@@ -17,10 +17,10 @@ classifiers = [
     "Operating System :: OS Independent",
 ]
 dependencies = [
-    "numpy<2",
+    "numpy",
     "nltk",
     "scikit-learn",
-    "fasttext",
+    "lingua-language-detector",
     "model2vec",
     "pyspellchecker"
 ]

{semantic_compressor-2.3 → semantic_compressor-2.5/semantic_compressor.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: semantic_compressor
-Version: 2.3
+Version: 2.5
 Author: Carlo Moro
 Author-email: Carlo Moro <cnmoro@gmail.com>
 Classifier: Programming Language :: Python :: 3
@@ -9,13 +9,14 @@ Classifier: Operating System :: OS Independent
 Requires-Python: >=3.7
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: numpy<2
+Requires-Dist: numpy
 Requires-Dist: nltk
 Requires-Dist: scikit-learn
-Requires-Dist: fasttext
+Requires-Dist: lingua-language-detector
 Requires-Dist: model2vec
 Requires-Dist: pyspellchecker
 Dynamic: author
+Dynamic: license-file
 Dynamic: requires-python
 
 ```python

{semantic_compressor-2.3 → semantic_compressor-2.5}/semantic_compressor.egg-info/SOURCES.txt

@@ -11,7 +11,6 @@ compressor/minbpe/basic.py
 compressor/minbpe/regex.py
 compressor/resources/__init__.py
 compressor/resources/en_stopwords.pkl
-compressor/resources/lid.176.ftz
 compressor/resources/pt_stopwords.pkl
 compressor/resources/nltk_data/corpora/stopwords.zip
 compressor/resources/nltk_data/corpora/wordnet.zip
@@ -184,4 +183,16 @@ semantic_compressor.egg-info/PKG-INFO
 semantic_compressor.egg-info/SOURCES.txt
 semantic_compressor.egg-info/dependency_links.txt
 semantic_compressor.egg-info/requires.txt
-semantic_compressor.egg-info/top_level.txt
+semantic_compressor.egg-info/top_level.txt
+tests/__init__.py
+tests/conftest.py
+tests/test_benchmark.py
+tests/test_clean_text.py
+tests/test_compression.py
+tests/test_correct_spelling.py
+tests/test_embeddings.py
+tests/test_find_needle_in_haystack.py
+tests/test_language_and_stemming.py
+tests/test_ngrams.py
+tests/test_semantic_embeddings.py
+tests/test_tokenizer.py

{semantic_compressor-2.3 → semantic_compressor-2.5}/semantic_compressor.egg-info/requires.txt

@@ -1,6 +1,6 @@
-numpy<2
+numpy
 nltk
 scikit-learn
-fasttext
+lingua-language-detector
 model2vec
 pyspellchecker

{semantic_compressor-2.3 → semantic_compressor-2.5}/semantic_compressor.egg-info/top_level.txt

@@ -1 +1,2 @@
 compressor
+tests

{semantic_compressor-2.3 → semantic_compressor-2.5}/setup.py

@@ -2,7 +2,7 @@ from setuptools import setup, find_packages
 
 setup(
     name='semantic_compressor',
-    version='2.3',
+    version='2.5',
     author='Carlo Moro',
     author_email='cnmoro@gmail.com',
     description="Semantic text compression",
@@ -12,10 +12,10 @@ setup(
     },
     include_package_data=True,
     install_requires=[
-        "numpy<2",
+        "numpy",
         "nltk",
         "scikit-learn",
-        "fasttext",
+        "lingua-language-detector",
         "model2vec",
         "pyspellchecker"
     ],

semantic_compressor-2.5/tests/conftest.py

@@ -0,0 +1,92 @@
+import os
+import sys
+import pytest
+import importlib.resources
+
+_NLTK_DATA_PATH = None
+try:
+    _NLTK_DATA_PATH = str(importlib.resources.files('compressor').joinpath('resources/nltk_data'))
+except Exception:
+    _NLTK_DATA_PATH = os.path.join(
+        os.path.dirname(__file__), '..', 'compressor', 'resources', 'nltk_data'
+    )
+
+os.environ['NLTK_DATA'] = _NLTK_DATA_PATH
+
+from compressor.semantic import (
+    clean_text,
+    detect_language,
+    stem_text,
+    count_tokens,
+    structurize_text,
+    extract_textual_embeddings,
+    calculate_similarity,
+    compute_and_remove_repeated_ngrams,
+    correct_spelling,
+)
+
+
+def pytest_configure(config):
+    config.addinivalue_line(
+        "markers",
+        "need_model: mark test as requiring the model2vec model (skipped if not available)",
+    )
+
+
+def _check_model_available():
+    try:
+        from compressor.semantic import _get_embedding_model
+        _get_embedding_model()
+        return True
+    except Exception:
+        return False
+
+
+def pytest_collection_modifyitems(config, items):
+    model_available = _check_model_available()
+    if not model_available:
+        skip_need_model = pytest.mark.skip(reason="model2vec model not available")
+        for item in items:
+            if "need_model" in item.keywords:
+                item.add_marker(skip_need_model)
+
+
+@pytest.fixture
+def sample_text_en():
+    return (
+        "The quick brown fox jumps over the lazy dog. "
+        "This is a test sentence for the semantic compressor. "
+        "Natural language processing is a fascinating field. "
+        "Machine learning algorithms can analyze text data efficiently."
+    )
+
+
+@pytest.fixture
+def sample_text_pt():
+    return (
+        "O rato roeu a roupa do rei de Roma. "
+        "Esta é uma frase de teste para o compressor semântico. "
+        "Processamento de linguagem natural é uma área fascinante. "
+        "Algoritmos de aprendizado de máquina podem analisar texto eficientemente."
+    )
+
+
+@pytest.fixture
+def sample_text_noisy():
+    return "Hello,   World!!!   This is... a   very  noisy??? text---with | weird • characters [and] (parens)."
+
+
+@pytest.fixture
+def sample_text_hyphenated():
+    return "This is a hyphen- ated word that should be re- paired.\n\nSecond paragraph here."
+
+
+@pytest.fixture
+def sample_text_dense():
+    return (
+        "Artificial intelligence has transformed the modern world. "
+        "Deep learning models can recognize patterns in complex data. "
+        "Neural networks are inspired by the human brain. "
+        "Natural language understanding enables machines to read text. "
+        "Computer vision allows machines to interpret images and video."
+    )