## Capturing Practical Natural Language Transformations

Citations: | 27 - 0 self |

### BibTeX

@MISC{Knight_capturingpractical,

author = {Kevin Knight},

title = {Capturing Practical Natural Language Transformations},

year = {}

}

### OpenURL

### Abstract

We study automata for capturing transformations employed by practical natural language processing systems, such as those that translate between human languages. For several variations of finite-state string and tree transducers, we ask formal questions about expressiveness, modularity, teachability, and generalization.

### Citations

633 | Tree adjoining grammars
- Joshi
- 2003
(Show Context)
Citation Context ...generalize FST). Another future direction is to propose other desirable formal properties from a machine translation perspective, and to see whether more powerful, non-finite-state formalisms (e.g., (=-=Shieber and Schabes, 1990-=-)) have those properties. Translation models based on dependency grammars have also been proposed (e.g., (Shen et al., 2008)), and these may also be formalized. Finally, it would be useful to be able ... |

235 |
Tree Automata. Akadémiai Kiadó
- Gécseg, Steinby
- 1984
(Show Context)
Citation Context ...s rule re-orders the input children as it creates the output. In computer-readable format, the same rule looks like: q S(x0, x1, x2) → S(r x1, VP(s x0, q x2)) Tree transducer rules in the literature (=-=Gécseg and Steinby, 1984-=-) have a one-level LHS (left hand side) with a state, an input-tree symbol, and (optionally) a sequence of variables x0, x1 ... xn. The RHS (right hand side) shows what the rule emits. The RHS may be ... |

218 | An inequality with applications to statistical estimation for probabilistic functions of markov processes and to a model for ecology - Baum, Eagon - 1967 |

217 | What’s in a Translation Rule - Galley, Hopkins, et al. - 2004 |

169 |
Tree acceptors and some of their applications
- Doner
- 1970
(Show Context)
Citation Context ...glish output sentence. Figure 2 shows the reverse. The utility of hierarchical tree structure was noticed early by Chomsky, and as a result, automata theorists devised tree acceptors and transducers (=-=Doner, 1970-=-; Rounds, 1970; Thatcher, 1970), whose mathematical aim was to generalize the previously-developed string automata. Recently, NLP practitioners have been constructing weighted syntax models for machin... |

120 | Mappings and grammars on trees - Rounds - 1970 |

103 | Training Tree Transducers - Graehl, Knight, et al. |

98 | The Design Principles of a Weighted Finite-State Transducer Library - Mohri, Pereira, et al. - 2000 |

90 | A new string-to-dependency machine translation algorithm with a target dependency language model - Shen, Xu, et al. - 2008 |

64 | An overview of probabilistic tree transducers for natural language processing - Knight, Graehl - 2005 |

44 | Translation with finite-state devices - Knight, Al-Onaizan - 1998 |

32 | Learning non-isomorphic tree mappings for machine translation - Eisner - 2003 |

31 | 2006b. Tiburon: A weighted tree automata toolkit - May, Knight |

30 | A weighted finite state transducer implementation of the alignment template model for statistical machine translation - Kumar, Byrne - 2003 |

30 | The power of extended top-down tree transducers - Maletti, Graehl, et al. - 2009 |

22 | Synchronous grammars as tree transducers - Shieber - 2004 |

20 |
M.: Morphismes et bimorphismes d’arbres
- Arnold, Dauchet
- 1982
(Show Context)
Citation Context ...under composition. Because practitioners may be able to re-work their models into ǫ-free versions, it is worth asking whether ǫ-free xLNT is closed under composition. The answer is shown to be no in (=-=Arnold and Dauchet, 1982-=-), with the following example: c c f.tex; 20/08/2008; 10:49; p.1112 Kevin Knight h(t1, t2, h(t3, t4, h(...g(tn−1, tn)))) � τ1 g(t1, g(t2, g(...g(tn−1, tn)))) � τ2 g(t1, h(t2, t3, h(t4, t5, h(...h(tn−... |

18 |
A remark on finite transducers
- Schützenberger
- 1961
(Show Context)
Citation Context ...eights, we can compute an overall weight for any string pair <x, y>, allowing us to prefer one output over another. FSTs have nice computational properties, one of which is closure under composition (=-=Schutzenberger, 1961-=-; Mohri et al., 2000). This means that a pipeline of FSTs can always be re-built as a single FST, allowing a system designer to break a complex problem down into simple pieces, and to assemble those p... |

17 | Compositions of extended top-down tree transducers - Maletti - 2008 |

11 | Handbook of formal languages - Choffrut, Karhumäki - 1997 |

7 |
Carmel finite-state toolkit. http://www.isi.edu/licensed-sw/carmel
- Graehl
- 1997
(Show Context)
Citation Context ...t locally optimize the corpus probability in time O(n 2 ). Portable implementations of weighted FST composition, best path, and training can be found in software toolkits such as (Mohri et al., 2000; =-=Graehl, 1997-=-). Mohri et al (2000) provide an excellent overview of weighted FSTs for speech and NLP, and (Kumar and Byrne, 2003; Knight and Al-Onaizan, 1998) describe statistical machine translation systems based... |

1 | On extended tree transducers, their properties, and their connections with synchronous grammars - Graehl, Hopkins, et al. - 2007 |

1 | Eagon: 1967, ‘An Inequality with Application to Statistical Estimation for Probabilistic Functions of Markov Processes and to a Model for Ecology - Baum, A |

1 |
f.tex; 20/08/2008; 10:49; p.14 Practical Natural Language Transformations 15
- Knight, Al-Onaizan
- 1998
(Show Context)
Citation Context ...raining can be found in software toolkits such as (Mohri et al., 2000; Graehl, 1997). Mohri et al (2000) provide an excellent overview of weighted FSTs for speech and NLP, and (Kumar and Byrne, 2003; =-=Knight and Al-Onaizan, 1998-=-) describe statistical machine translation systems based on weighted FSTs. 1 The “o” operator indicates composition. A o B is the FST that captures the set of all string pairs <x, y> where there exist... |