Noun Generation for Nominalization in Academic Writing

Nominalization is a common technique in academic writing for producing abstract and formal text. Since it often involves paraphrasing a clause with a verb or adjectival phrase into a noun phrase, an important task is to generate the noun to replace the original verb or adjective. Given that a verb or adjective may have multiple nominalized forms with similar meaning, the system needs to be able to automatically select the most appropriate one. We propose an unsupervised algorithm that makes the selection with BERT, a state-of-the-art neural language model. Experimental results show that it significantly outperforms baselines based on word frequencies, word2vec and doc2vec.