Low-bias extraction of domain-specific concepts