3-Way Composition of Weighted Finite-State Transducers

Composition of weighted transducers is a fundamental algorithm used in many applications, including for computing complex edit-distances between automata, or string kernels in machine learning, or to combine different components of a speech recognition, speech synthesis, or information extraction system. We present a generalization of the composition of weighted transducers, 3-way composition, which is dramatically faster in practice than the standard composition algorithm when combining more than two transducers. The worst-case complexity of our algorithm for composing three transducers $T_1$, $T_2$, and $T_3$ resulting in $T$, \ignore{depending on the strategy used, is $O(|T|_Q d(T_1) d(T_3) + |T|_E)$ or $(|T|_Q d(T_2) + |T|_E)$,} is $O(|T|_Q \min(d(T_1) d(T_3), d(T_2)) + |T|_E)$, where $|\cdot|_Q$ denotes the number of states, $|\cdot|_E$ the number of transitions, and $d(\cdot)$ the maximum out-degree. As in regular composition, the use of perfect hashing requires a pre-processing step with linear-time expected complexity in the size of the input transducers. In many cases, this approach significantly improves on the complexity of standard composition. Our algorithm also leads to a dramatically faster composition in practice. Furthermore, standard composition can be obtained as a special case of our algorithm. We report the results of several experiments demonstrating this improvement. These theoretical and empirical improvements significantly enhance performance in the applications already mentioned.

💡 Research Summary

The paper addresses a fundamental bottleneck in the manipulation of weighted finite‑state transducers (WFSTs): the composition of more than two transducers. While binary composition is a well‑studied operation, real‑world pipelines—such as speech recognition, speech synthesis, information extraction, and string‑kernel based machine learning—often require the sequential composition of three or more WFSTs. The naïve approach composes the first two transducers, producing an intermediate transducer that can be extremely large, and then composes this intermediate result with the third transducer. This two‑step method has a worst‑case time complexity of O(|T₁|·|T₂|·|T₃|) and can be prohibitive both in time and memory.

The authors propose a direct “3‑way composition” algorithm that builds the final transducer T = T₁ ∘ T₂ ∘ T₃ without ever materialising the intermediate product. The core idea is to explore the three‑dimensional state space (q₁, q₂, q₃) using a queue. For each pair of transitions e₁ ∈ E