The research in Artificial Intelligence is rapidly productized and there are plenty of online solutions available at your disposal. Some of the Computer Vision related tasks (like segmentation, or image classification) give you an acceptable accuracy without training or customizing the models. However, if your data is a bit specific, close-ups, medical images, or product images, you will find fine-tuned models to work much better.
If you work with text then you probably noticed that such models do not generalize well across different styles of documents. To give you a concrete example our recent work on sentiment analysis for German proved how significant the custom approach is to achieve good performance. The table below summarises results on sentiment analysis tasks of GermEval (2017), the most popular German NLP competition.
|Solution||Task 1a||Task 1b|
|Sayyed et al. (2017)||0.733||0.750|
|Naderalvojoud et al. (2017)||0.749||0.736|
While we’re proud of getting the best accuracy in both tasks, it is much more important to notice how poorly the generic solution performed compared with the custom models.