main.py

import argparse, re, itertools
import numpy as np
from sentence_transformers import SentenceTransformer, util

def candidates(text):
    # crude noun-phrase-ish candidates: sequences of words (2-4 words) without punctuation
    words = re.findall(r"[A-Za-z][A-Za-z\-']+", text.lower())
    phrases = set()
    for n in range(1, 4):
        for i in range(len(words)-n+1):
            phrase = " ".join(words[i:i+n])
            if len(phrase) > 3:
                phrases.add(phrase)
    return list(phrases)

def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--text", type=str, default="Deep learning enables representation learning using neural networks, widely used in computer vision and NLP.")
    parser.add_argument("--top_k", type=int, default=8)
    args = parser.parse_args()

    model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
    cand = candidates(args.text)[:200]
    if not cand:
        print([]); return
    emb_text = model.encode([args.text], normalize_embeddings=True)
    emb_cand = model.encode(cand, normalize_embeddings=True)
    scores = util.cos_sim(emb_text, emb_cand).cpu().numpy()[0]
    top_idx = np.argsort(scores)[::-1][:args.top_k]
    result = [(cand[i], float(scores[i])) for i in top_idx]
    for k, s in result:
        print(f"{k}\t{s:.3f}")

if __name__ == "__main__":
    main()