Large language models accurately identify decision reasons in verbal reports