The traditional machine learning approach usually needs human experts to label the data examples (e.g., document, images, signals, etc.) to train a model to perform classification or regression. The human labeling process is normally expensive in terms of both time and money. Especially for the case of deep models, where the size of the training data could be extremely large.

One alternative approach is called distant supervision, where the training data is generated by utilizing the existing database such as Freebase. For example, if our target is to extract the relation of friends, the item in Freebase that includes Buzz Lightyear and Woody Pride would be a positive example. By this mean, we can easily generate a large amount of labeled training data. However, for the model training, having only the positive examples are not enough. A more critical issue is how to generating the negative examples from the large-scale database. Please elaborate at least two ways to generating the negative examples in distant supervision.