## The Design Principles of a Weighted Finite-State Transducer Library (2000)

Venue: | THEORETICAL COMPUTER SCIENCE |

Citations: | 105 - 22 self |

### BibTeX

@ARTICLE{Mohri00thedesign,

author = {Mehryar Mohri and Fernando Pereira and Michael Riley},

title = {The Design Principles of a Weighted Finite-State Transducer Library},

journal = {THEORETICAL COMPUTER SCIENCE},

year = {2000},

volume = {231},

pages = {17--32}

}

### Years of Citing Articles

### OpenURL

### Abstract

We describe the algorithmic and software design principles of an object-oriented library for weighted finite-state transducers. By taking advantage of the theory of rational power series, we were able to achieve high degrees of generality, modularity and irredundancy, while attaining competitive efficiency in demanding speech processing applications involving weighted automata of more than 10^7 states and transitions. Besides its mathematical foundation, the design also draws from important ideas in algorithm design and programming languages: dynamic programming and shortest-paths algorithms over general semirings, object-oriented programming, lazy evaluation and memoization.

### Citations

9061 | Introduction to Algorithms
- Cormen, Leiserson, et al.
- 2001
(Show Context)
Citation Context ...pairs shortest-paths problems --- closed semirings --- is compatible with the abstract notion of weights we use, we were able to include an efficient version of the generic algorithm of FloydWarshall =-=[1,3]-=- in our library. Using the same algorithm and code, we can provide the all-pairs shortest distances when weights are real numbers representing, for example, probabilities, but also when they are strin... |

2564 |
h~ Design and Analysis of Computer Algorithms
- Hopcroft, Ullman
- 1974
(Show Context)
Citation Context ...pairs shortest-paths problems --- closed semirings --- is compatible with the abstract notion of weights we use, we were able to include an efficient version of the generic algorithm of FloydWarshall =-=[1,3]-=- in our library. Using the same algorithm and code, we can provide the all-pairs shortest distances when weights are real numbers representing, for example, probabilities, but also when they are strin... |

347 | Regular models of phonological rule systems - Kaplan, Kay - 1994 |

320 | Finite-State Transducers in Language and Speech Processing
- Mohri
- 1997
(Show Context)
Citation Context ...ted automata; finite-state transducers; rational power series; speech recognition. 1 Introduction Finite-state techniques have proven valuable in a variety of natural-language processing applications =-=[5--11,14,16,18,19,29,33,34,37,39,40]-=-. However, speech processing imposes requirements that were not met by any existing finite-state library. In particular, speech recognition requires a general means for managingsuncertainty: all level... |

193 |
Automata-Theoretic Aspects of Formal Power Series
- Salomaa, Soittola
- 1978
(Show Context)
Citation Context ...\Delta; 0; 1). can be given for a weighted transducer T , except that S(T ) is now a rational power series over a semiring of rational power series, those mapping transducer output strings to weights =-=[35]-=-. Most of the algorithms of our library work with arbitrary semirings or with semirings from mathematically-defined subclasses (closed semirings, k-closed semirings [20]). To instantiate the library f... |

173 | Constraint Grammar: A Language-Independent System for Parsing Unrestricted Text. Mouton de Gruyter - Karlsson, Voutilainen, et al. - 1995 |

142 |
Morphology and Computation
- Sproat
- 1992
(Show Context)
Citation Context ...ted automata; finite-state transducers; rational power series; speech recognition. 1 Introduction Finite-state techniques have proven valuable in a variety of natural-language processing applications =-=[5--11,14,16,18,19,29,33,34,37,39,40]-=-. However, speech processing imposes requirements that were not met by any existing finite-state library. In particular, speech recognition requires a general means for managingsuncertainty: all level... |

127 | Speech recognition by composition of weighted finite automata
- Pereira, Riley
- 1997
(Show Context)
Citation Context ... [24,25,31]. This paper focuses on the overall design of the library starting from its mathematical foundation, rather than on specific algorithms or applications, which have been described elsewhere =-=[18,21,24--26,28,31]-=-. Although our initial motivation was to improve the tools available for speech recognition, we aimed always for the highest degree of generality compatible with the mathematical foundation and with t... |

122 |
On the definition of a family of automata
- Schützenberger
- 1961
(Show Context)
Citation Context ...e singleton power series given by (S; x) = k; (S; y) = 0 if x 6= y for x 2 \Sigma ; k 2 K. The rational power series are exactly those formal power series that can be represented by weighted automata =-=[36]-=-. Weighted automata are a generalization of the notion of automaton: each transition of a weighted automaton is assigned a weight in addition to the usual label(s). More formally, a weighted acceptor ... |

106 |
Context-dependent phonetic hidden markov models for speaker-independent continuous speech recognition
- Lee
- 1990
(Show Context)
Citation Context ...ch-recognition systems rely on a variety of probabilistic finitestate models, for instance n-gram language models [30], multiple-pronunciation dictionaries [13], and context-dependent acoustic models =-=[12]-=-. However, most speech recognizers do not take advantage of the shared properties of the information sources they use. Instead, they rely on special-purpose algorithms for specific representations. Th... |

89 |
STL Tutorial and Reference Guide
- Musser, Saini
(Show Context)
Citation Context ...++ version, we will use compile-time type-checking where possible. 4 This design philosophy has some similarities with that of other modern software toolkits such as the C++ Standard Template Library =-=[27]-=-. 5 The user can always copy this lazy automaton into an instance of a static automata class that supports the numstates operation. In other words, we favor explicit conversions to implicit ones. 11 T... |

86 |
Dictionnaires électroniques et analyse automatique de textes: le système INTEX
- Silberztein
- 1993
(Show Context)
Citation Context ...ted automata; finite-state transducers; rational power series; speech recognition. 1 Introduction Finite-state techniques have proven valuable in a variety of natural-language processing applications =-=[5--11,14,16,18,19,29,33,34,37,39,40]-=-. However, speech processing imposes requirements that were not met by any existing finite-state library. In particular, speech recognition requires a general means for managingsuncertainty: all level... |

84 | Weighted automata in text and speech processing
- Mohri, Pereira, et al.
- 1996
(Show Context)
Citation Context ...d composition are local, therefore they can be performed on demand, without needing to perform explicitly the replacement of T i by T 0 i . More details on filtered composition can be found elsewhere =-=[22,28]-=-. 9 4 Software Design Our library was designed to meet two important requirements: ffl Algorithms that operate on automata should do so only through abstract accessor and mutator operations, which in ... |

82 | Two-level morphology with composition - Karttunen, Kaplan, et al. - 1992 |

76 | An efficient compiler for weighted rewrite rules
- Mohri, Sproat
- 1996
(Show Context)
Citation Context ...states in the edit transducer encode appropriate context conditions. Furthermore, a set of weighted edit rules like those in Figure 11 can be directly compiled into an appropriate weighted transducer =-=[26]-=-. 6 Conclusion We presented a very general finite-state library based on the notions of semiring and of rational power series, which allowed us to use the same code for a variety of different applicat... |

58 | Minimization algorithms for sequential transducers
- Mohri
(Show Context)
Citation Context ... [24,25,31]. This paper focuses on the overall design of the library starting from its mathematical foundation, rather than on specific algorithms or applications, which have been described elsewhere =-=[18,21,24--26,28,31]-=-. Although our initial motivation was to improve the tools available for speech recognition, we aimed always for the highest degree of generality compatible with the mathematical foundation and with t... |

55 |
Minimisation of acyclic deterministic automata in linear time
- Revuz
- 1992
(Show Context)
Citation Context |

51 | On some applications of finite/state automata theory to natural language processing, Natural Language Engineering I
- Mohri
- 1996
(Show Context)
Citation Context ...nt semirings because of their foundation in the theory of rational power series [18]. For example, the same power series determinization algorithm and code [18] can be used to determinize transducers =-=[17]-=-, weighted transducers, weighted automata encountered in speech processing [24] and weighted automata using the probability operations. To do so, one just needs to use the algorithm with the string se... |

48 | Finite-state parsing and disambiguation - Koskenniemi - 1990 |

42 | Full expansion of context-dependent networks in large vocabulary speech recognition
- Mohri, Riley, et al.
- 1998
(Show Context)
Citation Context ...y management operations in speech recognition while creating the opportunity for hitherto unrecognized optimizations and achieving competitive or superior performance in many speech recognition tasks =-=[24,25,31]-=-. This paper focuses on the overall design of the library starting from its mathematical foundation, rather than on specific algorithms or applications, which have been described elsewhere [18,21,24--... |

37 |
Limited subsets of the free monoid
- Simon
- 1978
(Show Context)
Citation Context ...in line 8, is the semiring sum of such extended transitions with a given source, destination and label. In most speech-processing applications, the appropriate weight algebra is the tropical semiring =-=[38]-=-. Weights are positive real numbers representing negative logarithms of probabilities. Weights along a path are added; when several paths correspond to the same string, the weight of the string is the... |

36 |
Digital images and formal languages
- Culik, Kari
- 1997
(Show Context)
Citation Context ...re algorithms to the weighted case, and their implementation with the degree of efficiency required in speech recognition. 6 Weighted acceptors and transducers have also been used in image processing =-=[4]-=-. 14 Baseform Phone Weights Type a i b j w(a i ; b j ) ae eh 1 substitution d ffl 2 deletion ffl pr 1 insertion Fig. 9. Weighted Edit Distance 5.1 Simple Example: Alignment As a simple example of the ... |

28 |
General Algebraic Frameworks and Algorithms for Shortest-Distance Problems. Technical Memorandum 981210-10TM
- Mohri
- 1998
(Show Context)
Citation Context ...ucer output strings to weights [35]. Most of the algorithms of our library work with arbitrary semirings or with semirings from mathematically-defined subclasses (closed semirings, k-closed semirings =-=[20]-=-). To instantiate the library for a particular semiring K, we just need to give computational representations for the semiring elements and operations. Library algorithms, for instance composition, ff... |

28 |
ATT General-purpose finite-state machine software tools. http://www.research.att.com/sw/tools/fsm
- Mohri, Pereira, et al.
- 1997
(Show Context)
Citation Context ...tion is presented in terms of those commands for simplicity. These Unix shell-level commands are available for download for a variety of computer architectures from the AT&T Labs -- Research web site =-=[23]-=- along with documentation, tutorials, and exercises. 5 Language processing applications As noted in Section 1, finite-state methods have been used very successfully in a variety of language-processing... |

26 | Weighted determinization and minimization for large vocabulary speech recognition
- Mohri, Riley
- 1997
(Show Context)
Citation Context ...y management operations in speech recognition while creating the opportunity for hitherto unrecognized optimizations and achieving competitive or superior performance in many speech recognition tasks =-=[24,25,31]-=-. This paper focuses on the overall design of the library starting from its mathematical foundation, rather than on specific algorithms or applications, which have been described elsewhere [18,21,24--... |

23 | Minimization of Sequential Transducers
- Mohri
- 1994
(Show Context)
Citation Context ...e transformations: ffl-elimination, determinization [17,18] and minimization for unweighted (both the general case [1] and the more efficient acyclic case [29]) and weighted acceptors and transducers =-=[15,18]-=-, removal of inaccessible states and transitions; Search: best path [20], n-best paths, pruning (remove all states and transitions that occur only on paths of weight greater by a given threshold than ... |

23 |
Analyse syntaxique transformationnelle du français par transducteurs et lexique-grammaire
- Roche
- 1993
(Show Context)
Citation Context |

19 |
Non Deterministic Stochastic Language Models for Speech Recognition
- Riccardi, Bocchieri, et al.
- 1995
(Show Context)
Citation Context ...r library, those applications had not included speech speech recognition. Current speech-recognition systems rely on a variety of probabilistic finitestate models, for instance n-gram language models =-=[30]-=-, multiple-pronunciation dictionaries [13], and context-dependent acoustic models [12]. However, most speech recognizers do not take advantage of the shared properties of the information sources they ... |

17 | The use of finite automata in the lexical representation of natural language - Gross - 1989 |

16 | Syntactic analysis by local grammars automata: an efficient algorithm
- Mohri
- 1994
(Show Context)
Citation Context |

15 | Compact Representations by Finite-State Transducers
- Mohri
- 1994
(Show Context)
Citation Context |

13 |
Finite growth models
- Ristad, Yianilos
- 1996
(Show Context)
Citation Context ...tion, deletion and substitution weights between phonemes and phones. In a real application, those weights would be derived automatically from aligned examples using a suitable machine-learning method =-=[13,32]-=-. The minimum edit distance between two strings can be simply defined by the recurrences d(a 0 ; b 0 ) = 0 d s (a i ; b j ) = d(a i\Gamma1 ; b j \Gamma1 ) + w(a i ; b j ) (substitution) d d (a i ; b j... |

12 | Transducer composition for context-dependent network expansion
- Riley, Pereira, et al.
- 1997
(Show Context)
Citation Context ...y management operations in speech recognition while creating the opportunity for hitherto unrecognized optimizations and achieving competitive or superior performance in many speech recognition tasks =-=[24,25,31]-=-. This paper focuses on the overall design of the library starting from its mathematical foundation, rather than on specific algorithms or applications, which have been described elsewhere [18,21,24--... |

10 |
Optimal Speech Recognition Using Phone Recognition
- Ljolje, Riley
- 1992
(Show Context)
Citation Context ...uded speech speech recognition. Current speech-recognition systems rely on a variety of probabilistic finitestate models, for instance n-gram language models [30], multiple-pronunciation dictionaries =-=[13]-=-, and context-dependent acoustic models [12]. However, most speech recognizers do not take advantage of the shared properties of the information sources they use. Instead, they rely on special-purpose... |

10 |
On the use of sequential transducers in natural language processing
- Mohri
- 1997
(Show Context)
Citation Context |

10 |
A finite-state architecture for tokenization and grapheme-to-phoneme conver-sion for multilingual text analysis”. In F rom text to tags: Issues in multilingual language analysis
- Sproat
- 1995
(Show Context)
Citation Context |

4 | The Replace Operator. In 33rd Annual Meeting of the Association for Computational Linguistics - Karttunen - 1995 |

2 |
Two parsing methods by means of finite state transducers
- Roche
- 1994
(Show Context)
Citation Context |