CFG filtering for HPSG parsing was introduced by Torisawa torisawa-compiling,torisawa-filtering, and it is based on the observation that many failed unifications (between edges in the chart and daughters in the grammar rules in course of completion) occur in partial parse trees that will later be discarded. By eliminating in advance partial parse trees that do not contribute to the final parse tree, a significant amount of failed unification will be avoided. This is accomplished by analyzing parse trees for an approximate CFG extracted from the HPSG.
Several methods exist for generating CFGs from UBGs. The method proposed in [Torisawa et al.2000] is an improvement over previously existing methods. The extracted CFG takes into account dependencies between rules (by analyzing feature propagations across rules), which play an important role in restricting the possible parse trees in an HPSG. The authors report improvements of up to 80% for this method. However, this result was achieved by removing features from the original grammar, making these results less realistic. More than that, the cost of filtering the CFGs is reported to be a major drawback of this method.