Parallel Active Chart Parsing

Introduction In the field of linguistics, Active Chart Parsing (ACP) is an algorithm that generates all possible parsings of a sentence, given an ambiguous grammar. Therefore, ACP would be one part of a natural language interface to a database server, if real-time processing could be achieved. Parallel processing is one way to achieve real-time performance. A previously reported parallelizations of ACP [1] on a Network of Workstations (NOW) showed no speedup, rather the reverse, an unsurprising conclusion, as ACP is inherently finegrained and synchronous. Partial results, i.e. tentative or definite identification of phrases, are stored for later use in the sequential algorithm, and continually exchanged in any parallel algorithm. A further difficulty for any parallelization is that there is an ordering constraint, as, at least in the English language, aggregated phrases are formed from adjacent words or phrases. The reported implementation was through the Orca language, which adopts a shared-object model with implicit message passing. Replication of shared object state is reported to have resulted in a slow-down, as local changes of state required a broadcast. Therefore, the ACP algorithm is appropriate for a physically shared-memory address space, where no replication is required. In this paper, we report a parallelization on a commodity four-processor Dell multiprocessor (700 MHz with 1 GB memory, 1 MB L2 cache). Such processors, which are cost effective due to their commercial use as web servers, are programmed in parallel mode using multithreading.