归一化gensim方法most_similar_cosmul提供的相似性评分

如何解决归一化gensim方法most_similar_cosmul提供的相似性评分

我正在尝试决定是否对一个项目使用gensim方法most_similar（）和most_similar_cosmul（），在该项目中我试图找到一组与输入列表相似的单词。

虽然这两种方法都提供了一个用于限制结果集中单词数量的参数（默认值：10），但我正在尝试根据相似度阈值（例如0.5）来选择单词。如果有助于相似度的肯定词数量较少，则在这两种方法之间效果很好。

但是，尽管常规的most_similar（）方法（线性）针对每个输出对返回相对稳定的相似性评分，而与输入单词的数量无关，但most_similar_cosmul（）方法（乘法）返回的相似性却越来越小。 / p>

我想知道是否有任何方法可以标准化most_similar_cosmul方法提供的相似性，以便它们独立于输入单词的数量。 most_similar_cosmul方法基于Levy和Yoav Goldberg的论文“稀疏和显式单词表示中的语言规则”。

>>> print(model.wv.most_similar(['digital_transformation','digital']))
[('social_mobile',0.6574317216873169),('social_networking',0.653410792350769),('mobile',0.6508731245994568),('mobile_social',0.6483871340751648),('digitization',0.6388466358184814),('digital_platform',0.6366733908653259),('digitalization',0.6243988871574402),('omni-channel',0.6230252385139465),('multi-channel',0.6205648183822632),('digital_marketing',0.6161972284317017)]

>>> print(model.wv.most_similar_cosmul(['digital_transformation',0.6048797965049744),0.6038507223129272),0.5969080328941345),0.594704270362854),0.5940128564834595),0.5880148410797119),0.585797905921936),0.5844268798828125),0.5779333710670471),0.575140118598938)]



>>> print(model.wv.most_similar(['digital_transformation','digital','virtual']))
[('mobile',0.6831567883491516),0.6692748665809631),('social_mobile',0.6688156127929688),('cloud-based',0.6612646579742432),('cloud',0.6573182344436646),0.656197190284729),('connected',0.6240546107292175),('physical_digital',0.6219688653945923),0.6217926740646362),('digital_content',0.6217813491821289)]

>>> print(model.wv.most_similar_cosmul(['digital_transformation',0.4431973695755005),0.4394254982471466),0.43778157234191895),0.4324479103088379),0.43241411447525024),0.4277876913547516),0.4095992147922516),0.40511685609817505),0.40396806597709656),0.40359607338905334)]



>>> print(model.wv.most_similar(['digital_transformation','virtual','online']))
[('mobile',0.7285350561141968),('online_mobile',0.6857519745826721),0.6757094264030457),0.6697002053260803),0.6669844388961792),0.6654024720191956),0.6573283076286316),0.6509564518928528),0.6482810974121094),0.64772629737854)]

>>> print(model.wv.most_similar_cosmul(['digital_transformation',0.35193151235580444),0.3212329149246216),0.31998541951179504),0.31541353464126587),0.3134937584400177),0.31218039989471436),0.3066185712814331),0.3035270869731903),0.301998496055603),0.30137890577316284)]