🌐 第20课：跨语言知识图谱

import numpy as np from collections import defaultdict class CrossLingualAligner: """跨语言实体对齐器""" def __init__(self): self.lang_entities = defaultdict(dict) # {lang: {id: {name, type, ...}}} self.translations = {} # {(lang1, id1): (lang2, id2)} self.embeddings = {} # {(lang, id): vector} def add_entity(self, lang, entity_id, name, etype, aliases=None): self.lang_entities[lang][entity_id] = { "name": name, "type": etype, "aliases": aliases or [] } def add_translation(self, lang1, id1, lang2, id2): ">>>添加已知翻译对（种子数据）""" self.translations[(lang1, id1)] = (lang2, id2) self.translations[(lang2, id2)] = (lang1, id1) def set_embedding(self, lang, entity_id, vector): self.embeddings[(lang, entity_id)] = vector def align_by_name(self, lang1, lang2): """基于名称翻译的对齐""" alignments = [] for id1, e1 in self.lang_entities[lang1].items(): for id2, e2 in self.lang_entities[lang2].items(): if e1["type"] != e2["type"]: continue # 检查名称是否在别名列表中 names1 = {e1["name"]} | set(e1.get("aliases", [])) names2 = {e2["name"]} | set(e2.get("aliases", [])) if names1 & names2: alignments.append(((lang1, id1), (lang2, id2), 1.0)) return alignments def align_by_embedding(self, lang1, lang2, top_k=5): """基于嵌入相似度的对齐""" emb1 = {k: v for k, v in self.embeddings.items() if k[0] == lang1} emb2 = {k: v for k, v in self.embeddings.items() if k[0] == lang2} alignments = [] for k1, v1 in emb1.items(): scores = [] for k2, v2 in emb2.items(): sim = np.dot(v1, v2) / (np.linalg.norm(v1) * np.linalg.norm(v2) + 1e-8) scores.append((k2, sim)) scores.sort(key=lambda x: -x[1]) for k2, sim in scores[:top_k]: if sim > 0.5: alignments.append((k1, k2, sim)) return alignments def transfer_knowledge(self, source_lang, target_lang, source_triples): ">>>知识迁移：将源语言的KB三元组迁移到目标语言""" transferred = [] for sh, r, st in source_triples: # 找到源语言实体的目标语言对应 target_h = self.translations.get((source_lang, sh)) target_t = self.translations.get((source_lang, st)) if target_h and target_t: if target_h[0] == target_lang and target_t[0] == target_lang: transferred.append((target_h[1], r, target_t[1])) return transferred # ========== 测试 ========== cla = CrossLingualAligner() # 中文实体 cla.add_entity("zh", "鲁迅", "鲁迅", "PER", ["Lu Xun", "周树人"]) cla.add_entity("zh", "北京", "北京", "LOC", ["Beijing", "Peking"]) cla.add_entity("zh", "呐喊", "呐喊", "WORK", ["Call to Arms"]) # 英文实体 cla.add_entity("en", "lu_xun", "Lu Xun", "PER", ["鲁迅"]) cla.add_entity("en", "beijing", "Beijing", "LOC", ["北京", "Peking"]) cla.add_entity("en", "call_to_arms", "Call to Arms", "WORK") # 种子翻译对 cla.add_translation("zh", "鲁迅", "en", "lu_xun") cla.add_translation("zh", "北京", "en", "beijing") print("=== 基于名称的对齐 ===") for k1, k2, score in cla.align_by_name("zh", "en"): print(f" {k1} ↔ {k2}: {score}") # 知识迁移 zh_triples = [("鲁迅", "创作", "呐喊"), ("鲁迅", "出生地", "北京")] transferred = cla.transfer_knowledge("zh", "en", zh_triples) print(" === 知识迁移 zh→en ===") for h, r, t in transferred: print(f" ({h}, {r}, {t})") # 嵌入对齐 np.random.seed(42) for lang, eid in [("zh","鲁迅"),("en","lu_xun"),("zh","北京"),("en","beijing")]: cla.set_embedding(lang, eid, np.random.randn(10)) print(" === 嵌入对齐 ===") for k1, k2, score in cla.align_by_embedding("zh", "en"): print(f" {k1} ↔ {k2}: {score:.4f}")