Hai Hu

A (computational) linguist

[Hai Hu @thisisHuHai]

#NLProc #ACL2021NLP

[Hai Hu @thisisHuHai]

Excited to introduce our ACL Findings paper on investigating multilingual transfer learning of XLM-R with 4 new Chinese natural language inference datasets, including some uniquely Chinese linguistic phenomena. (1/n)

[Hai Hu @thisisHuHai]

https://t.co/8pY1nUxrnm

[Hai Hu @thisisHuHai]

For many phenomena, all models continue to struggle, highlighting the need for our new diagnostics to help benchmark Chinese and cross-lingual models. All new datasets/code are released at https://t.co/4bN99I8524 (6/n)

[Hai Hu @thisisHuHai]

These results, however, come with important caveats: cross-lingual models often perform best when trained on a mixture of English and high-quality monolingual NLI data (OCNLI), and are often hindered by automatically translated resources (XNLI-zh). (5/n)

[Hai Hu @thisisHuHai]

We find that cross-lingual models trained soley on English NLI data do transfer well across our Chinese tasks, i.e., in 3/4 of our challenge categories, they perform as well/better than the best monolingual models, even on 3/5 uniquely Chinese linguistic phenomena. (4/n)

[Hai Hu @thisisHuHai]

Most interestingly, for the diagnostics, we added NLI problems including several uniquely Chinese linguistic phenomena: Chinese idioms (Chengyu), pro-drop, non-core argument, etc. (3/n)

[Hai Hu @thisisHuHai]

Specifically, we created 1) Chinese HANS (based on ONCLI), 2) Chinese stress tests, 3) Expanded CLUE diagnostics, 4) Chinese semantic fragments. (2/n)

[Hai Hu @thisisHuHai]

"Hello, World!" from my Twitter bot... #C343Rocks