An Adversarial Joint Learning Model for Low-Resource Language Semantic Textual Similarity.

Semantic Textual Similarity (STS) of low-resource language is a challenging research problem with practical applications. Traditional solutions employ machine translation techniques to translate the low-resource languages to some resource-rich languages such as English. Hence, the final performance is highly dependent on the quality of machine translation. To decouple the machine translation dependency while still take advantage of the data in resource-rich languages, this work proposes to jointly learn the low-resource language STS task and that of a resource-rich one, which only relies on multilingual word embeddings. In particular, we project the low-resource language word embeddings into the semantic space of the resource-rich language via a translation matrix. To make the projected word embeddings resemble that of the resource-rich language, a language discriminator is introduced as an adversarial teacher. Thus the parameters of sentence similarity neural networks of two tasks can be effectively shared. The plausibility of our model is demonstrated by extensive experimental results.