Topical Word Embeddings

论文 《 Topical Word Embeddings 》 记录

paper
code

Word Embedding面临的问题

homonymy and polysemy

解决homonymy and polysemy方法

multi-prototype： 对每个word赋予多个embedding

当前multi-prototype方法的缺点

1). These models generate multi-prototype vectors for each word in isolation, ignoring complicated correlations among words as well as their contexts. 说的很抽象
2). In multi-prototype setting, contexts of a word are divided into clusters with no overlaps. In reality, a word’s several senses may correlate with each other, and there is not clear semantic boundary between them.

解决上述缺点的方法(提出三个模型)

TWE三个模型的缺点

• TWE-1: TWE-1 does not consider the immediate interaction between a word and its assigned topic for learning（单词和主题向量没有直接的交互）
• TWE-2： TWE-2 considers the inner interaction of a word-topic pair by simply regarding the pair as a pseudo word, but it suffers from the sparsity issue because the occurrences of each word are rigidly discriminated into different topics.（假设单词在语料中出现N次， 每个主题下的单词平均只能学习到$N/T$次）
• TWE-3: TWE-3 provides trade-off between discrimination and sparsity. But during the learning process of TWE-3, topic embeddings will influence the corresponding word embeddings, which may make those words in the same topic less discriminative.($T<<W$)

训练细节

Initialization is important for learning TWE models. In TWE-1, we first learn word embeddings using Skip-Gram. Afterwards, we initialize each topic vector with the average over all words assigned to this topics, and learn topic embeddings while keeping word embeddings unchanged. In TWE-2, we initialize the vector of each topic-word pair with the corresponding word vector from Skip-Gram, and learn TWE models. In TWE-3, we initialize word vectors using those from Skip-Gram, and topic vectors using those from TWE-1, and learn TWE models.

Experiments

Contextual Word Similarity

考虑到每个单词只有在上下文的条件下才可以区分， 所以在评价multi-prototype模型的时候，采用Contextual Word Similarity任务，试验结果如下：

个人总结： AvgSimC优于MaxSimC, 反映出单词之间的语义还是有交集的， 正如作者所说In reality, a word’s several senses may correlate with each other, and there is not clear semantic boundary between them;

Text Classification

macro-average and micro-average(precision, recall, F1-measure)

个人感觉只适用multi-class classification

二分类

precision=P, recall=R