Exploiting Non-Parallel Corpora for Statistical Machine Translation

Hoang Cuong, Le Anh Cuong, Nguyen Phuong Thai

Ho Tu Bao
School of Knowledge Science Japan Advanced Institute of Science and Technology (JAIST) Japan

AbstractConstructing a corpus of parallel sentence pairs is an important work in building a Statistical Machine Translation system. It impacts deeply how the quality of a Statistical Machine Translation could achieve. The more parallel sentence pairs we use to train the system, the better translations quality it is. Nowadays, comparable non-parallel corpora become important resources to alleviate scarcity of parallel corpora. The problem here is how to extract parallel sentence pairs automatically but accurately from comparable non-parallel corpora, which are usually very noisy. This paper presents how we can apply the reinforcement-learning scheme with our new proposed algorithm for detecting parallel sentence pairs. We specify that from an initial set of parallel sentences in a domain, the proposed model can extract a large number of new parallel sentence pairs from non-parallel corpora resources in different domains, concurrently increasing the systems translation ability gradually.

I. I NTRODUCTION Statistical Machine Translation (SMT) is a machine translation approach which depends on creating a parameter probabilistic model by analyzing parallel sentence pairs in a bilingual corpus. With the de-facto Moses SMT Engine [5], the effort in building an acceptable translation system quality is reduced by releasing a lot of works. Nowadays, extracting a large number of parallel sentence pairs is one of the most consuming-time and important works for building a good SMT system. Unfortunately, parallel corpora have been limited in size, language coverage, and language register [9]. Comparable non-parallel corpora are much more available from various resources in different domains, such as from Wikipedia, News websites, etc. However, these resources are very noisy environments. This paper focuses on the problem: From an initial training corpus which usually contains a small number of parallel sentence pairs, how do we can expand this training corpus, specially to a new knowledge domain. By this way, we can improve the SMT system repeatedly. In addition, the SMT system can also enhance its translations ability in new knowledge domains. In a general framework of extracting parallel corpora we rstly derive parallel sentence pair candidates, and then determine whether a pair is parallel or not based on the similarity measurement between the two sentences in the pair. Gale-Church [4] measured the rate of lengths between two bilingual sentences. This method is suitable in very clean

environments (purely parallel corpora); Michel Simard [12] submitted mixing length scheme from Gale-Church [4] and cognate criterion to a unify criterion to obtain better results in the environments which are little noise. But these two criteria are not powerful, robust, noise-protecting or accurate enough for extracting parallel sentence pairs from extremely noisy environments [7]. Some other studies such as Utsuro [13] and Zhao [17] used a statistical translation lexicon model. However, these works suffer much from the ambiguity of word translations, which is a reason of the error rate problem (i.e. the high error rate of extracted parallel sentence pairs). Building a statistical translation lexicon model from a set of available parallel sentence pairs for constructing a dictionary is an error-prone challenge, because its error rate measurement is extremely hard to control [10], [15], [8], [17]. Some methods tried to reduce the error rate by limiting top translations for each source language word, for example, Munteanu [9] used top ve translations of each of its words to reduce as much as possible the error rate of the parallel corpus extraction. However, it was not a steady solving, and thus there was still a risky error-prone challenge. Using this scheme almost rejects correct word-by-word translations in dictionary. The risk of the error rate is also bounded the power of other methods which are based on the statistical translation lexicon model, such as in ([8], [14], [9], [1]). They purely used the word translation model as the main criterion. Under our observation1 this model will not be sufcient enough to satisfy the requirement of extracting parallel sentences in a noisy environment. In this paper we improve the similarity measurement by proposing a new algorithm in which we combine length-based ltering, cognate condition, and content-based similarity measurement with some appropriate modications. Specially we improve the algorithm of measuring content-based similarity by using phrasal overlapping calculation denoted in [11] with the help of a complete phrase-based SMT system. We also isolate the translation phrases (also called segments) and use a constraint rule to prevent the unexpected overlapping phrases. The proposed algorithm will extend a large number of parallel
1 This

observation will be conrmed in the section Experiment.

sentence pairs and concurrently reduce considerably error rate, in comparison with previous studies. In addition, by using the scheme of reinforcement learning, we will show that the SMT systems translation quality is deeply increased gradually. The rest of this paper is organized as follows: Section II presents the reinforcement learning idea and its application to our task. Section III shows our method for detecting parallel sentence pairs. Section IV describes in detail the algorithm for measuring the content-based similarity between the two sentences. Section V presents our experimental evaluations to clarify our contributions. Finally, conclusion is derived in section VI. II. R EINFORCEMENT MODEL FOR EXPANDING PARALLEL
CORPORA

Fig. 1.

Architecture of the proposed model

The basic reinforcement-learning model typical consists of: a set of environment states S; a set of actions A; a set R of scalar immediate rewards. A reinforcement-learning agent interacts with its environment in discrete time steps. At each specic time t with the state st S, together with the set of available actions A(st ). It elects an action in the action set A(st ) and the environment walks to a new state st+1 with a reward rt+1 . Concerning to the environments which have a terminated state, the goal of the agent in reinforcement learning is to try to establish a plan : S A which collects as much reward as possible: R= r0 + r1 + ... + rn . The prior target of parallel sentence extracting at each specic time t (corresponding to the systems translation ability at that time C t ) could not extract all the parallel sentence pairs from a comparable non-parallel corpora resource. Our scheme is that the most highest priority of nding job at each specic time is extracting all of possible candidates based on the systems translation ability at that time and reextracting it latter to get the lack of parallel sentence pairs which could not achieve previously or extracting new nonparallel corpora in different domains due to the increase of the systems translation ability. Applying the simplest case of reinforcement learning model, where each state is represented as the systems translation ability at each specic time t, we have at each state only one available action (corresponding to extracting all of parallel sentence pairs as much as possible and we get the number of parallel sentence pairs reward achieving). After nding all possible parallel sentence pairs at each time t, we then retrain the SMT system and go to the new state st+1 . Fig. 1 shows the architecture of our reinforcement scheme that deals with both tasks: expanding training corpus of parallel sentence pairs and improving the corresponding SMT system. In more details, the detecting methods consist of two parts. The rst one is the ltering of the length and the cognate. The second one is the ltering based on the similarity measuring. The candidate passes the rst one condition will be translated into a new target language sentence by a complete phrase-based SMT system. The two target sentences (one

sentence is a normal human sentence and another sentence is a SMT system generated sentence) will be calculated to a similarity value based on the similarity measurement. If the similarity value passes a threshold called , the candidate will be added to the parallel sentence set. At a suitable time, we will append all of the new parallel sentence pairs to the SMT systems training set and retrain the SMT system again. We will have a new better SMT system to re-extract the resource again or extract other new domains. III. M ETHOD FOR D ETECTING PARALLEL S ENTENCE PAIRS The proposed method aims to detect parallel sentence pairs. It includes three steps as follows: Step 1 - Filtering candidates based on the ratio of lengths of the two bilingual sentences in each candidate. Step 2 - Filtering candidates based on the similarity of cognates in the two bilingual sentences. Step 3 - Measuring content-based similarity described in Figure 2, and then determine whether a candidate is parallel or not. In step 1, Gale-Church [4] indicated that the length measurement could be used in a dynamic programming framework in order to nd the maximum likelihood alignment of sentences. We use this measurement as a length ltering criterion by checking the ratio of candidate pairs length. From all parallel sentence pairs of the training corpus we calculate the mean and variance of all the length ratios. If a candidate has its ratio staying in the circle of this mean and variance, it passes the step 1 and goes to step 2. In step 2, we adjust Simard [12] idea by comparing the orders of cognate sequences between two sentences in a bilingual sentence candidate (the cognates are just non-translation symbols, such as question mark, bracket, parentheses, etc.). For example, if the structure of source language sentence is: A, B, C D. and target language sentences structure is: E, F, G H . In this step we compare the orders of cognate sequences between the two sentences in a pair and also check the length ltering condition in its sub-parts. This condition is satised if the cognate sequences are the same and its sub-parts pass

Fig. 2.

The content-based similarity measuring algorithm

the length-ltering condition (as in step 1), we then go to step 3. In step 3, we estimate a content based similarity between the two sentences in the candidate. This similarity is assigned a score, and if this score is not less than a threshold called we obtain a new parallel sentence pair. As in the Fig. 1, the detecting method consists of two parts. The rst part will remove candidates based on the conditions of length and cognate. The second one uses a content-based similarity measurement for ltering candidates. Firstly, If the candidates satisfy the predened conditions then they are moved the second part in the algorithm. Secondly, if these candidates also pass the conditions of content-based similarity measurement we will obtain new parallel sentence pairs. Each time when we complete expanding the training parallel corpus, we can re-train the SMT system and repeat the processing of expanding parallel corpus as well as improving the SMT system. IV. A LGORITHM FOR M EASURING THE C ONTENT-BASED S IMILARITY The content-based similarity measurement is an exciting problem. In our specic problem, we have to nd out a similarity value between two sentences. The special thing is, one of them is a normal human sentence and the other is a machinegenerated sentence. The machine-generated sentence is yielded by a complete phrase-based SMT system, which its generation is based on the language model and the translation model [6]. So we cannot treat it like a normal human sentence. We also cannot apply semantic metric or grammar metric to measure the similarity between two normal human sentences. The word overlapping metric seems to be appropriate for this problem. Banerjee and Pedersencite [3] introduced the lexical overlap measurement based on Zipfs law between the length of phrases and their frequencies in a text collection, which is called the multi-word phrases overlap measuring method. Ponzetto and Strube [11] used the sum of sentence lengths and apply the hyperbolic tangent function to minimize the effect of the outliers. In fact, the traditional lexical overlap measure-

ments treat sentences as a bag of words and does not regard highly the differences between single words overlap and multiword phrases overlap. The comparison of the content-based similarity measurements is credited by [2] and they pointed out that the multi-word phrases overlap measurement is the best measurement in the lexical overlap metric. For the work of determining parallel sentences in the two languages, we rst use a complete phrase-based SMT system to translate a sentence (called translation sentence) to the target language, then apply the phrasal overlap measurement for the two sentences in the same language. Different from other studies in which they measured the overlap of words/phrases between the two sentences, we here can utilize more information from outputs of the complete phrase-based SMT system (based on the MOSES Framework [5]). That will help to reduce noisy phrases, avoid inaccurate results and x aw as mentioned in [5], [16], [6]. The MOSES Framework [5] uses the training parallel corpus to extract all possible parallel phrases from the word-based alignment results. The following criterion proposed by Zens J [16] denes the set of bilingual phrases BP BP (f1 , eI , A) 1 J I of sentence pair (f1 , e1 ) consisting word alignment matrix A generated from IBM model:
j+m i+n , ei ) : (i , j ) A} : {(fj j j j+mii i+n

The basic formula for nding ebest (decoding step) in statistical phrase-based model (mixing several components which contribute to the overall score: the phrase translation probability , reordering model d and the language model p LM ), which gives all i = 1, ..., I input phrases fi and output phrase ei and their positions starti and endi : argmaxe
I i=1

(fi |ei )d(starti endi1 1)pLM (e)

For almost parallel sentence candidates, we have a very long number of words of each sentence in a pair. The number of words is usually more than 10-15 words per a sentence. By the way, surprisingly, Koehn [6] pointed out that limiting the length to a maximum of only three words per a phrase already gains top performance. Using longer phrases does not yield much improvement, and occasionally leads to worse results. In general, we also usually use trigram as the default language model parameter of a phrase-based SMT system [6]. We see that PLM is actually is not clue enough to be an relationship between all phrases ei . This means that we can assume the output of the complete phrase-based SMT system decoding step, all of the phrase element ei is independent with other elements, or there is no (or rarely) relationship between the elements. From the results of the MOSESs decoding process we can split the translation sentence into separate segments. For example, a translation sentence with trace2 has format sequence of segments like the following:
2 Running the MOSES decoder with the segmentation trace switch using -t option

t = |w1 w2 ...wk ||wk+1 wk+2 ...wn |... So that if we treat these segments independently we can avoid measuring the overlap on the phrases such as: wk wk+1 , wk1 wk wk+1 , . . . It means that we will not take the phrases in which their words appear in different translation segments. Note that in a noisy environment this phenomenon may cause many wrong results. Note that for computing phrase overlapping measurement between the sentences t and e (i.e. the contentbased similarity) we use the proposed formulate in [3] and [11], as follows: simoverlap,phrase (t, e) = tanh( where overlap(t, e) =
N n m overlapphrase (t,e) ) |t|+|e|

A. Data preparation For the initial parallel corpus, we extract from a Subtitle resource and obtain 50,000 parallel sentence pairs. To do that, we rst use the length-based ltering method [4] and then we manually check them again. The Wikipedia resource is chosen for expanding the parallel corpus. It is worth to emphasize that the knowledge domain of Wikipedia is far different from Subtitle domain. We found Wikipedia resource a clue for bilingual connection, like that: a Wikipedia page (in source language) will connect to (if exists) another wiki page (in target language) via Wikipedias hyperlink structure. By this evidence, we can collect a set of bilingual pages (in English and Vietnames). Then from a pair of bilingual pages, denoted as page A (containing m sentences) and page B (containing n sentences) we have nm candidates of parallel sentence pairs. B. Evaluation 1 This experiment comes from a development set with about 34,000 parallel sentence candidates getting from Wikipedia resource (they are both satised the conditions on length and cognation) and to be going through the content-based similarity measuring algorithm. With the obtained parallel sentence pairs, we will manually check them which is a true bitext pair or wrong bitext pair (i.e. the sentences in a pair are parallel or not). Table 1 shows the results when using 1-gram overlapping measurement; Table 2 shows the results when implementing the improved algorithm with phrasal overlapping, independent fragments, and the constraint rule. 0.45 0.5 0.55 0.6 Total 778 474 266 156 True 582 389 231 140 Wrong 196 85 35 16 Error(%) 25.19 17.93 13.15 10.25

n2

here m is a number of n-word phrases that appear in both sentences. From our observation, long overlapping phrases take a large proportion in the score of overlapping measurement between the two sentences. Therefore, the appearance of overlapping phrases in non-parallel sentences may cause much mis-detection of parallel sentence pairs. In a very noisy environment there easily exist overlapping phrases randomly. It is worth to notice that this phenomenon has not been mentioned in previous studies. To overcome this drawback, we added a constraint rule to the algorithm of measuring content-based similarity, that is: an overlapping phrase with N words (called N-word overlapping phrase) will be counted if there are at least N overlapping phrases/words which have their lengths shorter than N and does not appear in the fragment of the N-word overlapping phrase. With this constraint improving, our detecting model is extremely strong noise-ltering and better than previous studies. The quality of our detecting model based on constraintimproving will show more detail in the section Experiment. V. EXPERIMENT This experiment is deployed on an English - Vietnamese phrase-based SMT Project, using Moses framework [5]. We implement three evaluations to clarify major contributions of the proposed model.

Table 1. Error rate of detecting parallel sentence pairs using 1-gram overlapping 0.35 0.4 0.45 0.5 0.55 0.6 Total 595 404 272 172 108 83 True 586 401 272 172 108 83 Wrong 9 3 0 0 0 0 Error(%) 1.51 0.74 0.00 0.00 0.00 0.00

The rst evaluation estimates the threshold , concurrently shows the improvement of the proposed algorithm of measuring content-based similarity. The second evaluation compares our parallel sentence pair detecting method with a previous dictionary-based considered as the baseline. The third evaluation shows the ability of expanding the parallel corpus, consequently improving the SMT system via measuring its BLEU score.

Table 2. Error rate of detecting parallel sentence pairs using phrasal overlapping, independent fragment, and the constraint rule. The obtained results in Table 1 and Table 2 have shown that the proposed algorithm for detecting parallel sentences is much better than the normal algorithm. They shows that the error rate of using the normal overlapping method is much higher, meanwhile with the improved method we can achieve much lower error rate even there are no errors when is greater than 0.4. It is interesting that, the improved algorithm even bring

Fig. 3.

Comparing the proposed algorithm and the baseline algorithm

Fig. 4.

The system performance improvements gradually

a larger number of true parallel sentence pairs (586 pairs in comparison with 582 pairs) while we still obtain absolutely much higher accurate result. This evaluation also determines that = 0.35 is a suitable threshold. C. Evaluation 2 This evaluation was done in the same condition condition with the Evaluation 1 to compare how better our proposed method is in comparison with previous studies [13], [9] which used length, cognate, and dictionary for measuring content-based similarity. For this evaluation, we use top ve translations of each of its words as described in [9]. 0.35 0.4 0.45 0.5 0.55 0.6 Total 149 105 64 45 31 29 True 134 93 60 43 31 29 Wrong 15 12 4 2 0 0 Error(%) 10.06 11.43 6.25 4.65 0.00 0.00

D. Evaluation 3 This evaluation tests the re-training with the re-extracting capacity of our proposed model. We extract 9,998 parallel links from Wikipedia (via Wikipedias hyperlink structure) and use this resource for evaluating the scalability of our method. Note that by using step 1 and step 2 of the parallel sentence detecting algorithm, we remove a large number of parallel sentence pair candidates from the whole of candidates (tens million of parallel sentence pair candidates). Consequently, there are only about 958,800 candidates to be used for step 3. Starting up with a set of 50,000 available parallel sentence pairs collected from Subtitle resources we will test the capacity of extending parallel sentence pairs. Applying the reinforcement scheme and using the method for detecting parallel sentences do this. Each time when the training is extended we will retrain the SMT system, and then apply it to the candidates again to nd out new results. Table 4 shows the experimental result: Iterator Iter. 1 Iter. 2 Iter. 3 Iter. 4 Iter. 5 Training 50,000 72,835 89,340 94,082 95,212 BLEU(%) 8.92 21.80 23.01 23.96 24.07 Extract 22,835 16,505 4,742 1,130 0

Table 3. Error rate of detecting parallel sentence pairs using Dictionary. Table 3 shows the obtained results of error rate when detecting parallel sentence pairs using Dictionary. We can see that a dictionary created from lexical translation probability output is extremely noisy. In addition, Figure 3 points out that when the word-based translation models limit the number of possible word-translations to balance the error rate controlling, it will reduce intensely the number of extracted parallel sentences. In contrast, although our proposed method satises deeply the error rate controlling, meanwhile it could also extract more parallel sentence pairs.

Table 4. The Results of Integrating Reinforcement Learning with Our Detecting method We use a test set of parallel sentence pairs of 10,000 parallel sentence pairs handling 3 from Wikipedia resource to test the improvement of the BLEU score of our phrase-based SMT system.
3 We rst extract automatically a set of parallel sentence pair candidates based on the proposed algorithm, and then check them again by hand for obtaining exactly parallel sentence pairs.

At rst the BLEU score is 8.92% obtained by using the initial training set. Then at the rst iteration of the process of extending parallel corpus, we achieve 22,835 new parallel sentence pairs. To get more parallel sentences, we retrain our SMT system with the new training set containing 72,835 parallel sentence pairs the BLEU score is improved up to 21.80%. Then, the SMT system continues extracting more 16,505 new parallel sentence pairs which were not extracted at the previous iterations according to the lack of the capacity of our SMT system. This result denotes that the SMT system is now upgrading its translation ability. At the end, we can extract in the total of 45,212 new parallel sentence pairs and the BLEU score reaches to 24.07% which is far from the beginning system. Fig. 4 shows out more details the increasing of BLEU score value gradually. Another interesting thing is when we try to extract other resources, at a suitable time, we could turn back and could still extract more parallel sentence pairs from this resource. That is one of the most valuable merits of our proposed model - applying reinforcement learning scheme integrating with our detecting method. VI. C ONCLUSION This paper has proposed a model that integrates our new method of detecting parallel sentences into a reinforcement learning scheme for the purpose of extending a parallel corpus, and consequently improving statistical machine translation. Various experiments have been conducted and the obtained results have shown that the proposed algorithm of detecting parallel sentences could efciently extract a larger number of parallel sentence pairs and the error rate is much reduced in comparison with the previous results. This new algorithm is then applied into a reinforcement scheme, which allows the SMT system can be upgraded gradually with its improved ability of translation, especially for covering new knowledge domains. VII. ACKNOWLEDGMENT This work is partially supported by the CN.10.01 project at The University of Engineering and Technology, Vietnam National University, Hanoi. This work is also partially supported by the Vietnams National Foundation for Science and Technology Development (NAFOSTED), project code 102.99.35.09. R EFERENCES
[1] Sadaf AbduI-Rauf and Holger Schwenk. On the use of comparable corpora to improve smt performance. In Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics, EACL 09, pages 1623, Stroudsburg, PA, USA, 2009. Association for Computational Linguistics. [2] Palakorn Achananuparp, Xiaohua Hu, and Xiajiong Shen. The evaluation of sentence similarity measures. In Proceedings of the 10th international conference on Data Warehousing and Knowledge Discovery, DaWaK 08, pages 305316, Berlin, Heidelberg, 2008. Springer-Verlag. [3] Satanjeev Banerjee and Ted Pedersen. Extended gloss overlaps as a measure of semantic relatedness. In Proceedings of the 18th international joint conference on Articial intelligence, pages 805810, San Francisco, CA, USA, 2003. Morgan Kaufmann Publishers Inc.

[4] William A. Gale and Kenneth W. Church. A program for aligning sentences in bilingual corpora. Comput. Linguist., 19:75102, March 1993. [5] Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondej Bojar, Alexandra Conr stantin, and Evan Herbst. Moses: open source toolkit for statistical machine translation. In Proceedings of the 45th Annual Meeting of the ACL on Interactive Poster and Demonstration Sessions, ACL 07, pages 177180, Stroudsburg, PA, USA, 2007. Association for Computational Linguistics. [6] Philipp Koehn, Franz Josef Och, and Daniel Marcu. Statistical phrasebased translation. In Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology - Volume 1, NAACL 03, pages 4854, Stroudsburg, PA, USA, 2003. Association for Computational Linguistics. [7] I. Dan Melamed. Bitext maps and alignment via pattern recognition. Comput. Linguist., 25:107130, March 1999. [8] Robert C. Moore. Fast and accurate sentence alignment of bilingual corpora. In Proceedings of the 5th Conference of the Association for Machine Translation in the Americas on Machine Translation: From Research to Real Users, AMTA 02, pages 135144, London, UK, UK, 2002. Springer-Verlag. [9] Dragos Stefan Munteanu and Daniel Marcu. Improving machine translation performance by exploiting non-parallel corpora. Comput. Linguist., 31:477504, December 2005. [10] Franz Josef Och. Minimum error rate training in statistical machine translation. In Proceedings of the 41st Annual Meeting on Association for Computational Linguistics - Volume 1, ACL 03, pages 160167, Stroudsburg, PA, USA, 2003. Association for Computational Linguistics. [11] Simone Paolo Ponzetto and Michael Strube. Knowledge derived from wikipedia for computing semantic relatedness. J. Artif. Int. Res., 30:181 212, October 2007. [12] Michel Simard, George F. Foster, and Pierre Isabelle. Using cognates to align sentences in bilingual corpora. In Proceedings of the 1993 conference of the Centre for Advanced Studies on Collaborative research: distributed computing - Volume 2, CASCON 93, pages 10711082. IBM Press, 1993. [13] Takehito Utsuro, Hiroshi Ikeda, Masaya Yamane, Yuji Matsumoto, and Makoto Nagao. Bilingual text, matching using bilingual dictionary and statistics. In Proceedings of the 15th conference on Computational linguistics - Volume 2, COLING 94, pages 10761082, Stroudsburg, PA, USA, 1994. Association for Computational Linguistics. [14] Dekai Wu. Aligning a parallel english-chinese corpus statistically with lexical criteria. In Proceedings of the 32nd annual meeting on Association for Computational Linguistics, ACL 94, pages 8087, Stroudsburg, PA, USA, 1994. Association for Computational Linguistics. [15] Richard Zens and Hermann Ney. A comparative study on reordering constraints in statistical machine translation. In Proceedings of the 41st Annual Meeting on Association for Computational Linguistics - Volume 1, ACL 03, pages 144151, Stroudsburg, PA, USA, 2003. Association for Computational Linguistics. [16] Richard Zens, Franz Josef Och, and Hermann Ney. Phrase-based statistical machine translation. In Proceedings of the 25th Annual German Conference on AI: Advances in Articial Intelligence, KI 02, pages 1832, London, UK, 2002. Springer-Verlag. [17] Bing Zhao and Stephan Vogel. Adaptive parallel sentences mining from web bilingual news collection. In Proceedings of the 2002 IEEE International Conference on Data Mining, ICDM 02, pages 745, Washington, DC, USA, 2002. IEEE Computer Society.