Reading: Gaussian Mixture Embeddings for Multiple Word Prototypes

[a1da4]

0. Paper

authors: Xinchi Chen, Xipeng Qiu, Jingxiang Jiang, Xuanjing Huang paper: arxiv

1. What is it?

They proposed gaussian mixture embeddings for words.

2. What is amazing compared to previous works?

Their model represents polysemous words efficiently and outperforms previous baselines.

3. Where is the key to technologies and techniques?

3.1 Gaussian Mixture Skip-Gram (GMSG)

• set K meanings each word (fixed)
• calculate the similarity between two distributions of two words

3.2 Dynamic Gaussian Mixture Skip-Gram (D-GMSG)

• compute the similarity between two distributions of two words
• if the similarity is lower than a threshold, add a new random Gaussian component (sense)

4. How did evaluate it?

4.1 Word Similarity

4.2 Word Similarity in Context (#202)

5. Is there a discussion?

6. Which paper should read next?

[a1da4]

206

Topical information

[a1da4]

208

Gaussian Mixture × Skip-Gram

[a1da4]

209

Gaussian Mixture × fasttext