Neural Machine Translation (NMT) is an end-to-end learning approach for
automated translation, overcoming the weaknesses of conventional phrase-based translation
systems. Although NMT based systems have gained their popularity in commercial
translation applications, there is still plenty of room for improvement. Being the most
popular search algorithm in NMT, beam search is vital to the translation result. However,
traditional beam search can produce duplicate or missing translation due to its target
sequence selection strategy. Aiming to alleviate this problem, this paper proposed neural
machine translation improvements based on a novel beam search evaluation function. And
we use reinforcement learning to train a translation evaluation system to select better
candidate words for generating translations. In the experiments, we conducted extensive
experiments to evaluate our methods. CASIA corpus and the 1,000,000 pairs of bilingual
corpora of NiuTrans are used in our experiments. The experiment results prove that the
proposed methods can effectively improve the English to Chinese translation quality. 相似文献
The issue of a multicast search for a group of users is discussed in this study. Given the condition that the search is over
only after all the users in the group are found, this problem is called the Conference Call Search (CCS) problem. The goal
is to design efficient CCS strategies under delay and bandwidth constraints. While the problem of tracking a single user has
been addressed by many studies, to the best of our knowledge, this study is one of the first attempts to reduce the search
cost for multiple users. Moreover, as oppose to the single user tracking, for which one can always reduce the expected search
delay by increasing the expected search cost, for a multicast search the dependency between the delay and the search cost
is more complicated, as demonstrated in this study. We identify the key factors affecting the search efficiency, and the dependency
between them and the search delay. Our analysis shows that under tight bandwidth constraints, the CCS problem is NP-hard.
We therefore propose a search method that is not optimal, but has a low computational complexity. In addition, the proposed
strategy yields a low search delay as well as a low search cost. The performance of the proposed search strategy is superior
to the implementation of an optimal single user search on a group of users.
Amotz Bar-Noy received the B.Sc. degree in 1981 in Mathematics and Computer Science and the Ph.D. degree in 1987 in Computer Science, both
from the Hebrew University, Israel. From October 1987 to September 1989 he was a post-doc fellow in Stanford University, California.
From October 1989 to August 1996 he was a Research Staff Member with IBM T. J. Watson Research Center, New York. From February
1995 to September 2001 he was an associate Professor with the Electrical Engineering-Systems department of Tel Aviv University,
Israel. From September 1999 to December 2001 he was with AT research labs in New Jersey. Since February 2002 he is a Professor
with the Computer and Information Science Department of Brooklyn College - CUNY, Brooklyn New York.
Zohar Naor received the Ph.D. degree in Computer Science from Tel Aviv University, Tel Aviv, Israel, in 2000. Since 2003 he is with
the University of Haifa, Israel. His areas of interests include wireless networks, resource management of computer networks,
mobility, search strategies, and multiple access protocols. 相似文献
We consider the problem of rescheduling trains in the case where one track of a railway section consisting of two tracks
in opposing directions is closed due to construction activities. After presenting an appropriate model for this situation
we derive a polynomial algorithm for the subproblem of finding an optimal schedule with minimal latenesss if the subsequences
of trains for both directions outside the construction site are fixed. Based on this algorithm we propose a local search procedure
for the general problem of finding good schedules and report test results for some real world instances.
Received: December 8, 1999 / Accepted: May 2, 2001 相似文献
GENIUS-TF (Nucl. Instr. and Meth. A 511 (2003) 341; Nucl. Instr. and Meth. A 481 (2002) 149.) is a test-facility for the GENIUS project (GENIUS-Proposal, 20 November 1997; Z. Phys. A 359 (1997) 351; CERN Courier, November 1997, 16; J. Phys. G 24 (1998) 483; Z. Phys. A 359 (1997) 361; in: H.V. Klapdor-Kleingrothaus, H. Pas. (Eds.), First International Conference on Particle Physics Beyond the Standard Model, Castle Ringberg, Germany, 8–14 June 1997, IOP Bristol (1998) 485 and in Int. J. Mod. Phys. A 13 (1998) 3953; in: H.V. Klapdor-Kleingrothaus, I.V. Krivosheina (Eds.), Proceedings of the Second International Conference on Particle Physics Beyond the Standard Model BEYOND’ 99, Castle Ringberg, Germany 6–12 June 1999, IOP Bristol (2000) 915), a proposed large scale underground observatory for rare events which is based on operation of naked germanium detectors in liquid nitrogen for an extreme background reduction. Operation of naked Ge crystals in liquid nitrogen has been applied routinely already for more than 20 years by the CANBERRA Company for technical functions tests (CANBERRA Company, private communication, 5 March 2004.), but it never had found entrance into basic research. Only in 1997 first tests of application of this method for nuclear spectroscopy have been performed, successfully, in Heidelberg (Klapdor-Kleingrothaus et al., 1997, 1998; J. Hellmig and H.V. Klapdor-Kleingrothaus, 1997).
On May 5, 2003 the first four naked high-purity germanium detectors (total mass 10.52 kg) were installed in liquid nitrogen in the GENIUS Test Facility at the Gran Sasso underground laboratory. Since then the experiment has been running continuously, testing for the first time the novel technique in an underground laboratory and for a long-lasting period.
In this work, we present the first analysis of the GENIUS-TF background after the completion of the external shielding, which took place in December 2003. We focus especially on the background coming from 222Rn daughters. This is found to be at present by a factor of 200 higher than expected from simulation. It is still compatible with the scientific goal of GENIUS-TF, namely to search for cold dark matter by the modulation signal, but on the present level would cause serious problems for a full GENIUS—like experiment using liquid nitrogen. 相似文献
An efficient algorithm for the random packing of spheres can significantly save the cost of the preparation of an initial configuration often required in discrete element simulations. It is not trivial to generate such random packing at a large scale, particularly when spheres of various sizes and geometric domains of different shapes are present. Motivated by the idea of compression complemented by an efficient physical process to increase packing density, shaking, a new approach, termed compression algorithm, is proposed in this work to randomly fill any arbitrary polyhedral or cylindrical domains with spheres of various sizes. The algorithm features both simplicity and high efficiency. Tests show that it takes 181 s on a 1.4-GHz PC to complete the filling of a cylindrical domain with a total number of 26,787 spheres, achieving a packing density of 52.89%. 相似文献