The computer can read but understand does it? When we read, we notice errors, both in the text itself, as in the previous sentence, and in our own interpretation. Modern natural language processing can extract information from vast amounts of text but are oblivious to errors. This leads to a problem called error propagation: errors from one part of the system are passed on to the next, accumulate and encourage even more errors. This book documents the pervasiveness of error propagation in natural language processing, systematizes existing solutions, and proposes new ones. I argue that error propagation can only be solved by teaching the computer to detect and fix its own errors and, for that, it needs to know what all humans know: commonsense knowledge about how the world works.
|Award date||28 May 2021|
|Place of Publication||s.l.|
|Publication status||Published - 28 May 2021|
- end-to-end model