基于对抗样本与分布分离度的质检模型训练方法及系统
本发明涉及质检模型训练技术领域,尤其涉及一种基于对抗样本与分布分离度的质检模型训练方法及系统。获取待训练文本质检模型与训练集,通过对正常文本施加语义扰动生成对抗样本并通过迭代更新形成候选池,将候选池与训练集混合后提取向量表示,基于类别标签构建正常类凸包与异常类凸包并筛选位于边界区域的重点样本集,构建包含分类损失与凸包间隔损失的训练目标对混合样本及重点样本进行训练得到训练完成的模型。该方法能够提升文本质检模型在不确定区域的分类性能和鲁棒性。
APPLICANTS
Browse source applicant names and unchanged name fields from imported patent records.
These are name-based results, not verified legal entities. Identical names may refer to different people or organizations; aliases and historical names are not merged.
This name was extracted from a distinct applicant node in the source XML.
1 indexed patent records
Patent counts refer to distinct records held by this site, not an applicant’s complete portfolio. A joint name field is not a single company.
Verified corpus date coverage and last import time are not currently available. Publication dates on patent records are not import dates. Cached directory results may lag behind new imports.
Search the local original applicant field containing this full text. Results may include joint lists or longer names; this is not an exact identity search.
Search patents with this applicant condition