US 20060195312 A1 Abstract A machine translation (MT) decoder may transform a translation problem into an integer programming problem, such as a Traveling Salesman Problem (TSP). The decoder may invoke an integer program (IP) solver to solve the integer programming problem and output a likely decoding based on the solution.
Claims(24) 1. A method comprising:
transforming a translation problem into an integer programming problem; and generating a translation in response to solving the integer programming problem. 2. The method of 3. The method of 4. The method of inputting a text segment including words in a source language, wherein each of said plurality of regions represents a corresponding one of the words in the source language, and wherein the one or more nodes in each of said plurality of regions represent possible translations of the word corresponding to said region. 5. The method of assigning a distance between each of said nodes, said distance comprising one or more linguistic constraints. 6. The method of 7. The method of 8. The method of 9. The method of 10. Apparatus comprising:
a transformation module operative to transform an input text segment in a source language into an integer programming problem; and an integer program solver operative to solve said integer programming problem. 11. The apparatus of a database including a plurality of linguistic constraints for a target language and a plurality of words in the target language corresponding to possible translations for a plurality of words in a source language. 12. The apparatus of 13. The apparatus of 14. The apparatus of 15. The apparatus of 16. An article comprising a machine-readable medium including machine-executable instructions, the instruction operative to cause the machine to:
transform a translation problem into an integer programming problem; and generate a translation in response to solving the integer programming problem. 17. The article of 18. The article of 19. The article of input a text segment including words in a source language, wherein each of said plurality of regions represents a corresponding one of the words in the source language, and wherein the one or more nodes in each of said plurality of regions represent possible translations of the word corresponding to said region. 20. The article of assign a distance between each of said nodes, said distance comprising one or more linguistic constraints. 21. The article of 22. The article of 23. The article of 24. The article of Description This application claims priority to U.S. Provisional Patent Application Ser. No. 60/295,182, filed on May 31, 2001. The research and development described in this application were supported by DARPA-ITO under grant number N66001-00-1-9814. The U.S. Government may have certain rights in the claimed inventions. Machine translation (MT) concerns the automatic translation of natural language sentences from a first language (e.g., French) into another language (e.g., English). Systems that perform MT techniques are said to “decode” the source language into the target language. One type of MT decoder is the statistical MT decoder. A statistical MT decoder that translates French sentences into English may include a language model (LM) that assigns a probability P(e) to any English string, a translation model (TM) that assigns a probability P(f|e) to any pair of English and French strings, and a decoder. The decoder may take a previously unseen sentence f and try to find the e that maximizes P(e|f), or equivalently maximizes P(e)·P(f|e). A “stack decoder” is a type of statistical MT decoder. In a stack decoder, possible translations are organized into a graph structure and then searched until an optimal solution (translation) is found. Although stack decoders tend to produce good results, they do so at a significant cost. Maintaining and searching a large potential solution space is expensive, both computationally and in terms of computer memory. A machine translation (MT) decoder may transform a translation problem into an integer programming problem, such as a Traveling Salesman Problem (TSP). The decoder may include a graph generator which generates a graph including a number of regions, or cities, corresponding to words in an input source language sentence or phrase. Each region may include a number of nodes, or hotels, corresponding to possible translations of the source language word corresponding to that region. The graph generator may use linguistic constraint information in a translation database to assign distances between hotels. The decoder may invoke an integer program (IP) solver to find a shortest tour on the graph, i.e., solve the TSP. The decoder may output a likely decoding which includes the words corresponding to the hotels visited aligned in the order the hotels were visited in the tour. A solution to an integer programming problem involves an assignment of variables. Solutions are constrained by inequalities involving linear combinations of variables. An optimal solution is one that respects the constraints and minimizes the value of the objective function, which is also a linear combination of variables. One type of linear integer programming problem is the Traveling Salesman Problem (TSP). The hypothetical situation posed by the TSP concerns a salesman who spends his time visiting a number of cities (or nodes) cyclically. In one tour the salesman visits each city just once, and finishes up where he started. The Traveling Salesman Problem is this: given a finite number of “cities” along with the cost of travel between each pair of them (e.g., distance), find the cheapest (e.g., shortest) way of visiting all the cities and returning to the starting point. Much effort has been dedicated to the TSP, and powerful tools have been developed to solve TSP and other integer programming problems. For example, the Center for Research on Parallel Computation (CRPC), a National Science Foundation Science and Technology Center established in 1989, has solved the TSP for 13,509 U.S. cities with populations of more than 500 people. Such an integer program (IP) solver may be utilized by the decoder The decoder may transform a decoding problem into a TSP format and then use an IP solver A graph generator A tour of cities may be defined as a sequence of hotels (starting at the sentence boundary hotel) that visits each city exactly once before returning to the starting point. If a hotel sits on the border between two cities, then staying at that hotel counts as visiting both cities. Each tour of cities corresponds to a potential decoding <e,a>, where “e” represents the words in the English string and “a” represents the alignment of the words. The owners of the hotels on the tour yield e, while the hotel locations yield a. The distances between hotels may represent various constraints of an integer program. Real-valued (asymmetric) distances may be established between pairs of hotels such that the length of any tour is exactly the negative of log (P(e)·P(a,f|e)) (block The distance assigned to each pair of hotels may be based on a translation model Assuming that the price has already been paid for visiting the “what” hotel, then the inter-hotel distance need only account for the partial alignment concerning “not”: -
- distance =
- -log (bigram(not | what)) //chance of “not” given the previous word was “what”
- -log (n(2 | not)) //chance that word “not” in English generates two French words
- -log (t(NE | not)) - log (t(PAS | not)) //chance that “not” would translate to “ne” and “pas”.
- -log (d
_{1}(+1 | class(what),class(NE))) - -log (d
_{>1}(+2 | class(PAS))) //given that “what” translates to the first French word, what is the chance that “not” will translate to the second French word.
These and constraints for different words may be stored in a translation database An infinite distance may be assigned in both directions between hotels that are located (even partially) in the same city because travel from one to the other can never be part of a tour. NULL-owned hotels may be treated specially. All non-NULL hotels must be visited before any NULL hotels and at most one NULL hotel may be visited on a tour. Since only one NULL hotel is allowed to be visited, the fertility of the NULL word is simply the number of cities that hotel straddles, and the length of f is the number of cities minus one. The tour selection may be cast as an integer programming problem (block This minimization may be subject to three classes of constraints. First, every city must be visited exactly once. That means exactly one tour segment must exit each city:
Second, the segments must be linked to one another, i.e., every hotel has either (a) one tour segment coming in and one going out, or (b) no segments in and none out. To put it another way, every hotel must have an equal number of tour segments going in and out:
Third, to prevent multiple independent sub-tours, require that every proper subset of cities have at least one tour segment leaving it:
There may be an exponential number of constraints in this third class. Once cast as an integer program, the IP solver may be invoked. Exemplary IP solvers include lp_solve, available free of charge at ftp://ftp.ics.ele.tue.nl/pub/lp_solve, and CPLEX, available from ILOG, Inc. of Mountain View, Calif. Mnemonic names may be assigned to the variables, and <e,a> extracted from the list of variables and their binary values. The sentence corresponding to the shortest tour may then be output (block A number of embodiments have been described. Nevertheless, it will be understood that various modifications may be made without departing from the spirit and scope of the invention. For example, blocks in the flowchart may be skipped or performed out of order and still produce desirable results. Accordingly, other embodiments are within the scope of the following claims. Referenced by
Classifications
Legal Events
Rotate |