mayabot / mynlp

一个生产级、高性能、模块化、可扩展的中文NLP工具包。(中文分词、平均感知机、fastText、拼音、新词发现、分词纠错、BM25、人名识别、命名实体、自定义词典)
https://mynlp.mayabot.com/
Apache License 2.0
675 stars 90 forks source link

预测标签概率和不等于1? #29

Open Energy9502 opened 4 years ago

Energy9502 commented 4 years ago

训练数据集为正面、负面数据。但是预测数据的结果:正面的得分(概率)与负面,得分(概率)之和不等于1?请问是什么原因。难道每个标签的得分是(0,1)区间吗? 正面,[[labelpos,0.49219814],[labelneg,0.11921292]],RT @ABC: Fireworks greet Joe Biden and Kamala Harris following Biden's acceptance of the Democratic nomination for president. … Fireworks greet Joe Biden and Kamala Harris following Biden's acceptance of the Democratic nomination for president. #DemConvention 负面,[[labelneg,0.28777784],[labelpos,0.09808932]],RT @American_Bridge: The people who know Donald Trump best are sounding the alarm: our country cannot survive four more years of a Trump pr… The people who know Donald Trump best are sounding the alarm: our country cannot survive four more years of a Trump presidency.In our latest ad, a former member of his inner circle & fixer — @MichaelCohen212 — has a dire warning for us all. #RNC2020

jimichan commented 4 years ago

是的,按照我的理解,每个数据都计算和tag在向量空间的cos夹角,所以各自分布在[0,1]之间