Common use of Cube Pruning Clause in Contracts

Cube Pruning. − 1. Since the corner hypothesis might fail the con- straint check, rule cube ordering is based on the score of the nearest hypothesis to the corner that satisifies its constraints (if any exists). This hypothesis is found by exploring neighbours in order of estimated score (that is, without calcu- lating the full language model score) starting at the corner. 2. When a hypothesis is popped from a cube and its neighbours created, constraint-failing neigh- bours are added to a ‘bad neighbours’ queue. 3. If a cube cannot produce a new hypothesis be- cause all of the neighbours fail constraints, it starts exploring neighbours of the bad neigh- bours. We place an arbitrary limit of 10 on the number of consecutive constraint-failing hypotheses to con- sider before discarding the cube. We anticipate that decoding for a highly in- flected target language will result in a less mono- tonic search space due to the increased formation of inflectionally-inconsistent combinations.