Learning a distinct representation for each sense of an ambiguous word could lead to more powerful and fine-grained models of vector-space representations. Yet while `multi-sense' methods have been proposed and tested on artificial word-similarity tasks, we don't know if they improve real natural language understanding tasks. In this paper we introduce a pipelined architecture for incorporating multi-sense embeddings into language understanding, and test the performance of a state-of-the-art multi-sense embedding model (based on Chinese Restaurant Processes). We apply the model to part-of-speech tagging, named entity recognition, sentiment analysis, semantic relation identification and semantic relatedness. We find that---if we carefully control for the number of dimensions--- sense-specific embeddings, whether alone or concatenated with standard (one vector for all senses) embeddings, introduce slight performance boost in semantic-related tasks, but is of little in others that depend on correctly identifying a few key words such as sentiment analysis.

本文介绍一种基于中餐馆过程的多义词嵌入模型，测试表明它在人类词语相似度判断方面取得了最先进的性能，并提出了一种管道式架构，用于将多义词嵌入结合到语言理解中，结果表明在某些任务（如词性标注、语义关系识别、语义相关性）中，多义词嵌入确实能够提高性能，但在其他任务（如命名实体识别、各种形式的情感分析）中却没有提高。

多义词嵌入是否提高自然语言理解？