"""Aspect keyword dictionary for clothing domain. Seed words for each aspect. Can be extended using metadata's `features`/`categories` text. """ import json import re from collections import Counter from typing import Dict, List, Iterable from .config import ASPECT_DICT_PATH # Seed dictionary - manually curated for clothing domain SEED_ASPECT_DICT: Dict[str, List[str]] = { "SIZE": [ "alterations", "baggy", "big", "boxy fit", "fit", "fit perfectly", "fits", "fitting", "great fit", "huge", "large", "length", "long", "loose", "loose fit", "narrow", "oversized", "perfect fit", "petite", "poor fit", "right size", "roomy", "runs big", "runs large", "runs small", "runs tight", "short", "shorter than expected", "size", "size down", "size up", "sizing", "small", "snug", "stretched out", "tight", "tiny", "too big", "too large", "too long", "too loose", "too short", "too small", "too tight", "true to size", "tts", "waist", "wide", "wrong size", ], "MATERIAL": [ "blend", "breathable", "bulky", "cashmere", "cheap fabric", "cheap material", "comfortable fabric", "cotton", "denim", "elastic", "fabric", "fabric feels", "feel", "feeling", "feels", "fleece", "heavy", "heavier", "heavier than expected", "heavy fabric", "itchy", "leather", "lightweight", "linen", "lycra", "material", "nylon", "polyester", "rayon", "rough", "scratchy", "see through", "see-through", "sheer", "silk", "soft", "spandex", "stiff", "stretch", "stretchy", "suede", "synthetic", "texture", "thick", "thin", "too heavy", "transparent", "uncomfortable fabric", "velvet", "warm", "weight", "wool", ], "QUALITY": [ "after wash", "after washing", "broke", "broken", "button", "buttons", "cheaply made", "construction", "craftsmanship", "defect", "defective", "durability", "durable", "faded", "faded after wash", "fading", "falling apart", "fell apart", "first wash", "flimsy", "fragile", "hem", "holes", "lasted", "lasts", "loose thread", "loose threads", "made well", "pilled", "pilling", "poor quality", "poorly made", "quality", "rip", "ripped", "seam", "seams", "shedding", "shrink", "shrunk", "shrunk after wash", "shrunk after washing", "stitch", "stitching", "sturdy", "tear", "threads", "tore", "torn", "well constructed", "well made", "well-made", "zipper", ], "APPEARANCE": [ "as advertised", "as described", "as pictured", "as shown", "beautiful", "bright", "color", "color is off", "color off", "colors", "colour", "darker", "design", "different color", "different from picture", "different from photo", "dull", "faded color", "graphic", "gorgeous", "lighter", "logo", "looked like", "looks like", "matches the picture", "misleading photo", "off color", "pattern", "photo", "picture", "print", "shade", "stunning", "tone", "true color", "true to color", "ugly", "vibrant", ], "STYLE": [ "boho", "casual", "chic", "classic", "collar", "compliment", "complimented", "compliments", "cut", "cute", "elegant", "fashion", "fashionable", "flattering", "formal", "modern", "neckline", "outfit", "preppy", "professional", "received compliments", "shape", "shapeless", "silhouette", "sleeves", "style", "stylish", "trendy", "unflattering", "vintage", ], "VALUE": [ "affordable", "bargain", "cheap price", "cost", "discount", "exchanged", "expensive", "for the price", "for this price", "good deal", "great deal", "money", "money back", "money well spent", "not worth", "not worth it", "overpriced", "price", "pricey", "refund", "return", "returning", "value", "value for money", "waste of money", "wasted money", "worth", "worth it", ], } def get_aspect_dict() -> Dict[str, List[str]]: """Return seed dict (callers can extend it).""" return {k: list(v) for k, v in SEED_ASPECT_DICT.items()} # Tokens / phrases we consider noise when mining metadata text _STOPWORDS = set(""" the and for with from this that have has had not your you our their these those there here when where what which who whom how why because all any its his her item product use used using also more most less other than just new cost free buy bought made make piece pieces set sets pack two one three inch inches cm mm gram fl oz lb pounds black white grey gray are can will would could should closure imported machine women men kids ladies amazon brand sleeve sleeves shirt shirts dress dresses pants jacket jackets """.split()) _GENERIC_MINED_TOKENS = { "are", "can", "will", "would", "could", "should", "closure", "imported", "machine", "women", "woman", "men", "kids", "ladies", "perfect", "nice", "great", "good", "comfortable", "casual", "design", "wear", "wash", "fit", "fabric", "material", "polyester", "cotton", "soft", } _AMBIGUOUS_SINGLE_TOKENS = { "cheap", "perfect", "comfortable", "design", "style", "fit", "wash", } def _tokenize(text: str) -> List[str]: return re.findall(r"[a-z][a-z\-]+", text.lower()) def extend_aspect_dict_from_metadata( meta_texts: Iterable[str], aspect_dict: Dict[str, List[str]] = None, top_k: int = 20, min_count: int = 5, ) -> Dict[str, List[str]]: """Mine frequent domain words from metadata text to extend aspect seed dict. For each aspect, if a metadata token co-occurs frequently with any seed word in the same metadata blob, add it to the aspect's vocabulary. """ if aspect_dict is None: aspect_dict = get_aspect_dict() # Build per-aspect co-occurrence counter aspect_cooc = {a: Counter() for a in aspect_dict} for text in meta_texts: if not isinstance(text, str) or not text: continue tokens = set(_tokenize(text)) if not tokens: continue # For each aspect, check if any seed appears in this blob for aspect, seeds in aspect_dict.items(): seed_set = set(_tokenize(" ".join(seeds))) if tokens & seed_set: # Co-occurring tokens -> aspect candidates for t in tokens: if t in _STOPWORDS or t in seed_set: continue if t in _GENERIC_MINED_TOKENS or t in _AMBIGUOUS_SINGLE_TOKENS: continue if len(t) < 4: continue aspect_cooc[aspect][t] += 1 # Add top-k high frequency co-occurring tokens to each aspect extended = {a: list(v) for a, v in aspect_dict.items()} for aspect, counter in aspect_cooc.items(): new_words = [w for w, c in counter.most_common(top_k * 3) if c >= min_count] # Filter words that look generic or ambiguous outside local context. new_words = [ w for w in new_words if w not in _STOPWORDS and w not in _GENERIC_MINED_TOKENS and w not in _AMBIGUOUS_SINGLE_TOKENS ][:top_k] extended[aspect].extend(new_words) # Dedup for a in extended: extended[a] = sorted(set(extended[a])) return extended def save_aspect_dict(aspect_dict: Dict[str, List[str]], path=ASPECT_DICT_PATH): with open(path, "w", encoding="utf-8") as f: json.dump(aspect_dict, f, indent=2) def load_aspect_dict(path=ASPECT_DICT_PATH) -> Dict[str, List[str]]: with open(path, encoding="utf-8") as f: return json.load(f) def compile_aspect_patterns(aspect_dict: Dict[str, List[str]]): """Compile regex patterns for each aspect (whole-phrase matching, case-insensitive).""" patterns = {} for aspect, keywords in aspect_dict.items(): # Sort longest first so multi-word phrases match before single tokens sorted_kws = sorted(set(keywords), key=lambda s: -len(s)) escaped = [re.escape(kw) for kw in sorted_kws] pattern = r"\b(?:" + "|".join(escaped) + r")\b" patterns[aspect] = re.compile(pattern, flags=re.IGNORECASE) return patterns