Unverified50% confidenceCautionExact time
Collapse采用整词词表，无法泛化到未登录词（OOV），且每词单向量只能捕捉主导词义，多义词处理能力有限
1
Sources
50%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
无需MLP与注意力机制：点吸引子动力学如何学习词嵌入
redditr/learnmachinelearning7/10/2026
Related Claims
Unverified原始word2vec需要在整个词汇表上计算softmax归一化，计算复杂度为O(V)，负采样将问题转化为二分类以降低计算量68% similarUnverifiedUnigram 模型由 SentencePiece 库实现，从大词表出发逐步剔除对整体似然贡献最小的子词来缩减词表64% similarUnverifiedIndustry techniques to address character collapse include conversation summary compression, key information extraction, and vector database retrieval63% similarUnverified大模型处理文本以Token（词元）为最小处理单元，由分词器完成切分61% similarUnverified固定字符分块策略会截断句子、破坏语义完整性，应替换为智能语义分块或滑动窗口策略61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/491365API
curl https://kongchang.com/api/v1/knowledge/claims/491365MCP
get_claim(id=491365)