better text extraction error handling

[jrhizor]

I don't think we're running into this now, but if underlying schemas changed we wouldn't be able to extract the text we need and it wouldn't fail.