博文

Domain portability myth in natural language processing (NLP)

已有 4942 次阅读 2014-8-1 09:50 |个人分类:立委科普|系统分类:科研笔记| domain, portability, 领域移植性

One widely held myth is about domain portability of an NLP system: a hand-crafted rule system has poor portability because once the domain changes, the system has to be re-crafted from scratch while a Machine Learning system can be re-trained on domain data keeping the algorithm unchanged. It all sounds too obvious and convincing. Our article to publish in CCCF uncovers the real picture behind the myth. The truth might well be just the opposite: a well architected rule system enjoys much more domain portability than a machine learning system does. Stay tuned.

Lei Zhang Interesting. But how can you handle thousands of domains by a rule system efficiently.8h ago
Wei Li That's why we need to write an article to answer such questions. To make it short, it depends on how the system is structured. If a domain independent deep parser is used as logical basis for domain apps, then fast domain porting is possible because only the relatively small domain specific component needs to be developed for each new domain. But with machine learning based on bag of words or on shallow processing, each new domain involves new data and training/tuning, which can be daunting. less7h ago
Lei Zhang So you assume much knowledge is shared across different domains. Looking forward to the article.7h ago
Wei Li Yes, for NLP, linguistic knowledge is largely common across domains. The core English grammar does not change whether it is used in this domain or that domain. On another note, in real life apps, it is rare that we need to handle thousands of domains at one time. The typical situation is that the product team decides to enter a new domain for a new offering, and the NLP development team needs to develop and port to that chosen domain for delivery. Then we can measure which approach makes the development faster in delivering the requirements from the product manager. less6h ago