Commit

When I removed parts from the trained model and evaluated it, it turned out that what was effective wasn't the distant context but the immediately preceding context. That's not what I expected.