📖 第20课：CRF序列标注

import numpy as np from collections import defaultdict class SimpleCRF: # Simplified CRF with gradient descent def __init__(self, num_tags, num_features, lr=0.01, reg=0.01): self.num_tags = num_tags self.num_features = num_features self.W = np.random.randn(num_features, num_tags, num_tags) * 0.01 self.lr = lr self.reg = reg def score(self, features, tags): # Compute sequence score total = 0 for t in range(len(tags)): f = features[t] if t == 0: total += np.sum(self.W[f, 0, tags[t]]) else: total += np.sum(self.W[f, tags[t-1], tags[t]]) return total def predict(self, features): # Viterbi decoding n = len(features) tags = list(range(self.num_tags)) V = [{} for _ in range(n)] back = [{} for _ in range(n)] for tag in tags: V[0][tag] = np.sum(self.W[features[0], 0, tag]) back[0][tag] = 0 for t in range(1, n): for tag in tags: best_score, best_prev = -np.inf, 0 for prev in tags: s = V[t-1][prev] + np.sum(self.W[features[t], prev, tag]) if s > best_score: best_score = s best_prev = prev V[t][tag] = best_score back[t][tag] = best_prev best_tag = max(V[-1], key=V[-1].get) result = [best_tag] for t in range(n-1, 0, -1): best_tag = back[t][best_tag] result.insert(0, best_tag) return result # 模拟序列标注任务：BIO实体标注 tag_names = ['O', 'B-PER', 'I-PER', 'B-LOC', 'I-LOC'] tag2idx = {t: i for i, t in enumerate(tag_names)} # 特征：简单的词索引（实际应用中需要更丰富的特征） vocab = {'张':0, '三':1, '在':2, '北':3, '京':4, '学':5, '习':6, '上':7, '海':8, '工':9, '作':10} num_features = len(vocab) num_tags = len(tag_names) crf = SimpleCRF(num_tags, num_features) # 手动设置权重模拟训练后的效果 # B-PER 通常在姓的位置，I-PER 在名的位置 crf.W[0, 0, tag2idx['B-PER']] = 2.0 # 张 -> B-PER crf.W[1, tag2idx['B-PER'], tag2idx['I-PER']] = 2.0 # 三 -> I-PER after B-PER crf.W[3, 0, tag2idx['B-LOC']] = 1.5 # 北 -> B-LOC crf.W[4, tag2idx['B-LOC'], tag2idx['I-LOC']] = 1.5 # 京 -> I-LOC after B-LOC crf.W[7, 0, tag2idx['B-LOC']] = 1.5 # 上 -> B-LOC crf.W[8, tag2idx['B-LOC'], tag2idx['I-LOC']] = 1.5 # 海 -> I-LOC # 预测 test_seqs = [ [vocab['张'], vocab['三'], vocab['在'], vocab['北'], vocab['京']], [vocab['上'], vocab['海'], vocab['工'], vocab['作']], ] print("=== CRF序列标注 ===") words_list = [['张','三','在','北','京'], ['上','海','工','作']] for words, feat in zip(words_list, test_seqs): tags = crf.predict(feat) tag_strs = [tag_names[t] for t in tags] print(f" {' '.join(words)} → {' '.join(tag_strs)}")