Identification and Tracing of Ambiguous Names: Discriminative and Generative Approaches

Save this PDF as:
 WORD  PNG  TXT  JPG

Size: px
Start display at page:

Download "Identification and Tracing of Ambiguous Names: Discriminative and Generative Approaches"

Transcription

1 Ientification an Tracing of Ambiguous Names: Discriminative an Generative Approaches Xin Li Paul Morie Dan Roth Department of Computer Science University of Illinois, Urbana, IL Abstract A given entity representing a person, a location or an organization may be mentione in text in multiple, ambiguous ways. Unerstaning natural language requires ientifying whether ifferent mentions of a name, within an across ocuments, represent the same entity. We present two machine learning approaches to this problem, which we call the Robust Reaing problem. Our first approach is a iscriminative approach, traine in a supervise way. Our secon approach is a generative moel, at the heart of which is a view on how ocuments are generate an how names (of ifferent entity types) are sprinkle into them. In its most general form, our moel assumes: (1) a joint istribution over entities (e.g., a ocument that mentions Presient Kenney is more likely to mention Oswal or White House than Roger Clemens ), (2) an author moel, that assumes that at least one mention of an entity in a ocument is easily ientifiable, an then generates other mentions via (3) an appearance moel, governing how mentions are transforme from the representative mention. We show that both approaches perform very accurately, in the range of 90% 95% F 1 measure for ifferent entity types, much better than previous approaches to (some aspects of) this problem. Our extensive experiments exhibit the contribution of relational an structural features an, somewhat surprisingly, that the assumptions mae within our generative moel are strong enough to yiel a very powerful approach, that performs better than a supervise approach with limite supervise information. Introuction Reaing an unerstaning text requires the ability to isambiguate at several levels, abstracting away etails an using backgroun knowlege in a variety of ways. One of the ifficulties that humans resolve instantaneously an unconsciously is that of reaing names. Most names of people, locations, organizations an others, have multiple writings that are being use freely within an across ocuments. The variability in writing a given concept, along with the fact that ifferent concepts may have very similar writings, poses a significant challenge to progress in natural language processing. Consier, for example, an open omain question answering system (Voorhees 2002) that attempts, given Copyright c 2004, American Association for Artificial Intelligence ( All rights reserve. a question like: When was Presient Kenney born? to search a large collection of articles in orer to pinpoint the concise answer: on May 29, The sentence, an even the ocument that contains the answer, may not contain the name Presient Kenney ; it may refer to this entity as Kenney, JFK or John Fitzgeral Kenney. Other ocuments may state that John F. Kenney, Jr. was born on November 25, 1960, but this fact refers to our target entity s son. Other mentions, such as Senator Kenney or Mrs. Kenney are even closer to the writing of the target entity, but clearly refer to ifferent entities. Even the statement John Kenney, born turns out to refer to a ifferent entity, as one can tell observing that the ocument iscusses Kenney s batting statistics. A similar problem exists for other entity types, such as locations an organizations. A hoc solutions to this problem, as we show, fail to provie a reliable an accurate solution. This paper presents two learning approaches to the problem of Robust Reaing, compares them to existing approaches an shows that an unsupervise learning approach can be applie very successfully to this problem, provie that it is use along with strong but realistic assumptions on the nature of the use of names in ocuments. Our first moel is a iscriminative approach that moels the problem as that of eciing whether any two names mentione in a collection of ocuments represent the same entity. This straightforwar moelling of the problem results in a classification problem as has been one by several other authors (Cohen, Ravikumar, & Fienberg 2003; Bilenko & Mooney 2003) allowing us to compare our results with these. This is a stanar pairwise classification task, uner a supervise learning protocol; our main contribution in this part is to show how relational (string an token-level) features an structural features, representing transformations between names, can improve the performance of this classifier. Several attempts have been mae in the literature to improve the results by performing some global optimization, with the above mentione pairwise classifier as the similarity metric. The results of these attempts were not conclusive an we provie some explanation for why that is. We prove that, assuming optimal clustering, this approach reuces the error of a pairwise classifier in the case of two classes (corresponing to two entities); however, in the more general NATURAL LANGUAGE PROCESSING & INFORMATION EXTRACTION 419

2 case, when the number of entities (classes) is greater than 2, global optimization mechanisms coul be worse than pairwise classification. Our experiments concur with this. However, as we show, if one first splits the ata in some coherent way e.g., to groups of ocuments originate at about the same time perio this can ai clustering significantly. This observation motivates our secon moel. We evelope a global probabilistic moel for Robust Reaing, etaile in (Li, Morie, & Roth 2004). Here we briefly illustrate one of its instantiations an concentrate on its experimental stuy an a comparison to the iscriminative moels. At the heart of this approach is a view on how ocuments are generate an how names (of ifferent types) are sprinkle into them. In its most general form, our moel assumes: (1) a joint istribution over entities, so that a ocument that mentions Presient Kenney is more likely to mention Oswal or White House than Roger Clemens ; (2) an author moel, that makes sure that at least one mention of a name in a ocument is easily ientifiable (after all, that s the author s goal), an then generates other mentions via (3) an appearance moel, governing how mentions are transforme from the representative mention. Our goal is to learn the moel from a large corpus an use it to support Robust Reaing. Given a collection of ocuments we learn the moel in an unsupervise way; that is, the system is not tol uring training whether two mentions represent the same entity. We only assume, as in the iscriminative moel above, the ability to recognize names, using a name entity recognizer run as a preprocessor. Our experimental results are somewhat surprising; we show that the unsupervise approach can solve the problem accurately, giving accuracies (F 1 ) above 90%, an better than our iscriminative classifier (obviously, with a lot more ata). In the rest of the paper, we further stuy the iscriminative an generative moel in an attempt to provie explanations to the success of the unsupervise approach, an she light on the problem of Robust Reaing an hopefully, many relate problems. Our stuy aresses two issues: (1) Learning Protocol: A supervise learning approach is traine on a training corpus, an teste on a ifferent one, necessarily resulting in some egraation in performance. An unsupervise metho learns irectly on the target corpus. The ifference, as we show, is significant. (2) Structural assumptions: we show that our generative moel benefits from the assumptions mae in esigning it. This is shown by evaluating a fairly weak initialization of moel parameters, which still results in a moel for Robust Reaing with respectable results. We leave open the question of how much ata is require for a supervise learning methos before it can match the results of our supervise learner. There is little previous work we know of that irectly aresses the problem of Robust Reaing in a principle way, but some relate problems have been stuie. From the natural language perspective, there has been a lot of work on the relate problem of coreference resolution (Soon, Ng, & Lim 2001; Ng & Carie 2003; Kehler 2002) which aims at linking occurrences of noun phrases an pronouns, typically occurring in a close proximity within a short ocument or a paragraph, base on their appearance an local context. In the context of atabases, several works have looke at the problem of recor linkage - recognizing uplicate recors in a atabase (Cohen & Richman 2002; Hernanez & Stolfo 1995; Bilenko & Mooney 2003; Doan et al. 2003). Specifically, (Pasula et al. 2002) consiers the problem of ientity uncertainty in the context of citation matching an suggests a probabilistic moel for that. One of the few works we are aware of that works irectly with text ata an across ocuments, is (Mann & Yarowsky 2003), which consiers one aspect of the problem that of istinguishing occurrences of ientical names in ifferent ocuments, an only of people (e.g., o occurrences of Jim Clark in ifferent ocuments refer to the same person?). The rest of this paper first presents the experimental methoology an ata use in our evaluation; it then presents the esign of our pairwise name classifier, a comparison to other classifiers in the literature, an a iscussion of clustering on top of a pairwise classifier. Finally, we present the generative moel an compare the iscriminative an generative approaches along several imensions. Experimental Methoology In our experimental stuy we evaluate ifferent moels on the problem of robust reaing for three entity types: People (Peop), Locations (Loc) an Organizations (Org). We collecte 8, 000 names from ranomly sample New York Times articles in the TREC corpus (Voorhees 2002) (about 4,000 personal names 1, 2,000 locations an 2, 000 organizations). The ocuments were annotate by a name entity tagger for these entity types. The annotation was then manually correcte an each name mention was labelle with its corresponing entity by two annotators. 5 pairs of training an test sets with 600 names in each ata set are constructe for each of the three entity types by ranomly picking from the 8,000 names. For the iscriminative classifier, given a training set, we generate positive training examples using all co-referring pairs of names, an negative examples by ranomly selecting (10 times that number of) pairs of names that o not refer to the same entity. The probabilistic moel is traine using a larger corpus. The results in all the experiments in this paper are evaluate using the same test sets, except when comparing the clustering schemes. For a comparative evaluation, the outcomes of each approach on a test set of names are converte to a classification over all possible pairs of names (incluing non-matching pairs). Since most pairs are trivial negative examples, an the classification accuracy can always reach nearly 100%, only the set M p of examples that are preicate to belong to the same entity (positive preictions) are use in evaluation, an are compare with the set M a of examples annotate as positive. The performance of an approach is then evaluate by Precision P = Mp Ma M p, Recall R = Mp Ma M a, an F 1 = 2PR P+R. Only F 1 values are shown an compare in this paper. 1 Honorifics an postfixes like Jr. are parts of a personal name. 420 NATURAL LANGUAGE PROCESSING & INFORMATION EXTRACTION

3 A Discriminative Moel The Robust Reaing problem can be formalize as pairwise classification: Fin a function f : N N {0,1} which classifies two strings (representing entity writings) in the name space N, as to whether they represent the same entity (0) or not (1). Most prior work aopte fixe string similarity metrics an create the esire function by simply thresholing the similarity between two names. (Cohen, Ravikumar, & Fienberg 2003) compare experimentally a variety of string similarity metrics on the task of matching entity names an foun that, overall, the best-performing metho is a hybri scheme (SoftTFIDF) combining a TFIDF weighting scheme of tokens with the Jaro-Winkler stringistance scheme. (Bilenko & Mooney 2003) propose a learning framework (Marlin) for improving entity matching using trainable measures of textual similarity. They compare a learne eit istance measure an a learne vector space base measure that employs a classifier (SVM) with fixe measures (but not the ones mentione above) an showe some improvements in performance. We propose a learning approach, LMR, that focuses on representing a given pair of names using a collection of relational (string an token-level) an structural features. Over these we learn, for each entity type, a linear classifier, that we train using the SNoW (Sparse Network of Winnows (Carlson et al. 1999)) learning architecture. A feature extractor automatically extracts active features in a atariven way for each pair of names. Our ecision is thus of the form: f(n 1, n 2) = arg max f l (n 1, n 2) = arg max l {0,1} l {0,1} k w i,l f i (1) where w i,l is the weight of feature f i in the function f l. Feature Extraction An example is generate by extracting features from a pair of names. Two types of features are use: relational features representing mappings between tokens in the two names, an structural features, representing the structural transformations of tokens in one name into tokens of the other. Each name is moelle as a partition in a bipartite graph, with each token in that name as a vertex (see Fig. 1). At most one token-base relational feature is extracte for each ege in the graph, by traversing a prioritize list of feature types until a feature is activate; if no active features are foun, it goes to the next pair of tokens. This scheme guarantees that only the most important (expressive) feature is activate for each pair of tokens. An aitional constraint is that each token in the smaller partition can only activate one feature. We use 13 types of token-base features. Some of the interesting types are: Honorific Equal: active if both tokens are honorifics an equal. An honorific is a title such as Mr., Mrs., Presient, Senator or Professor. Honorific Equivalence: active if both tokens are honorifics, not equal, but equivalent ( Professor, Prof. ). i=1 Honorific Mismatch : active for ifferent honorifics. NickName: active if tokens have a nickname relation ( Thomas an Tom ). John John Kenney J. Kenney Figure 1: There are two possible features for token 1 in the smaller partition but only the higher priority Equality feature is activate. F 1 (%) Marlin SoftTFIDF LMR Peop Loc Org Table 1: Performance of Different Pairwise Classifiers in an Ientical Setting. Results are evaluate in the average F 1 value over the 5 test sets for each entity type with 600 names in each of them. The columns are ifferent classifiers an LMR is our classifier. The learne classifiers are traine using corresponing training sets with 600 names. Equality: both names are equal. Eit Distance: the tokens have an eit-istance of 1. Symbol Map: one token is a symbolic representative of the other ( an an & ). Relational features are not sufficient, since a nonmatching pair of names coul activate exactly the same set of token-base features as a matching pair. Consier, for example, two names that are all the same except that one has an aitional token. Our structural features were esigne to avoi this phenomenon. These features encoe information on the relative orer of tokens between the two names, by recoring the location of the participating tokens in the partition. E.g., for the pairs ( John Kenney, John Kenney ) an ( John Kenney, John Kenney Davis ), the active relational features are ientical; but, the first pair activates the structural features (1, 2) an (1, 2), while the secon pair activates (1,3) an (1,,2). Experimental Comparison Table 1 presents the average F 1 using three ifferent pairwise classifiers on the 5 test sets we escribe before. The best results of each classifier using a uniform threshol for each entity type are shown here. The LMR classifier outperforms the non-learning SoftTFIDF classifier an the learne classifier (Marlin), uner ientical conitions. Table 2 shows the contribution of our feature types to the performance of LMR. The Baseline classifier only uses string-eit-istance features an Equality features. The Token-Base classifier uses all relational token-base features while the Structural classifier uses, in aition, the F 1 (%) Baseline Token-Base Structural Peop Loc Org Table 2: Contribution of Different Feature Sets The LMR classifier is traine with ifferent feature sets using the 5 training sets. Results are evaluate in the average F 1 value over the 5 test sets for each entity type with 600 names in each of them. NATURAL LANGUAGE PROCESSING & INFORMATION EXTRACTION 421

4 structural features. Aing relational an structural features types is very significant, an more so to People ue to the larger amount of overlapping strings in ifferent entities. Does Clustering Help? There is a long-hel intuition that using clustering on top of a pairwise classifier might reuce its error; several works (Cohen, Ravikumar, & Fienberg 2003; McCallum & Wellner 2003) have thus applie clustering in similar tasks. However, we prove here that while optimal clustering always helps to reuce the error of a pairwise classifier when there are two clusters (corresponing to two entities), in general, for k classes an k > 2, it is not the case. A typical clustering algorithm views ata points n N to be clustere as feature vectors n = (f 1,f 2,...,f ) in a -imensional feature space. A ata point n is in one of k classes C = {C 1,C 2,...,C k }. From a generative view, the observe ata points D = {n 1,n 2,...,n t } are sample i.i.. from a joint probability istribution P efine over tuples (n,c) N C (P is a mixture of k moels). A istance metric ist(n 1,n 2 ) is use to quantify the ifference between two points. Clustering algorithms iffer in how they approximate the hien generative moels given D. Definition 1 Given two i.i. sample ata points (n 1,c 1 ) an (n 2,c 2 ) N C with only n 1,n 2 observe an c 1,c 2 hien, an P N C, the problem of Robust Reaing is to fin a function f : N N {0,1} which satisfies: f(n 1,n 2 ) = 0 iff c 1 = c 2 (2) Definition 2 The pairwise classification in this setting is: f p (n 1,n 2 ) = 0 ist(n 1,n 2 ) T (3) where T 0 is a threshol. The clustering base ecision: f c (n 1,n 2 ) = 0 argmax i p(n 1 C i ) = argmax i p(n 2 C i ) (4) Definition 3 The error rate of the pair function f is: err(f) = E P (f(n 1,n 2 ) I(n 1,n 2 )) (5) where I(n 1,n 2 ) = 0 iff c 1 = c 2 ;1, otherwise. Lemma 1 Assume ata is generate accoring to a uniform mixture of k Gaussians G = {g 1,g 2,...,g k } with the same covariance (i.e., a ata point is generate by first choosing one of k moels with probability p k (p k = 1/k), an then sampling accoring to the i-th Gaussian chosen.), an suppose we fin the correct k Gaussian istributions after clustering, then for k = 2 an T, err(f c ) < err(f p ) (6) However, this oesn t hol in general for k > 2. Proof (sketch): 2 It s easy to see that the probability ensity over tuples in N C is f(n,c i ) = 1 k g i(n). The error rates err(f c ) an err(f p ) can be compute using the ensity function. Thus, for k = 2, we get err(f c ) = A2, 2 The assumption of Gaussian istributions can be relaxe. where A = R [g 1(X) g 2 (X)]X, an R is the area in the feature space that satisfies g 2 (X) > g 1 (X) (X N). R is a half space here. We also have: err(f p) = M(X 1,T) R [g2(x1) g1(x1)]x1 [g2(x2) g1(x2)]x2 > err(fc) where M(X 1,T) is the sphere area in the feature space whose point X satisfies ist(x 1,X) T. This is so since M(X [g 1,T) 2(X 2 ) g 1 (X 2 )]X 2 < R [g 2(X 2 ) g 1 (X 2 )]X 2. For k > 2, we can compute err(f c ) an err(f p ) in a similar way an compare them. We foun that in this case, each can be smaller than the other in ifferent cases. Specifically, when k i=1 g i(n) > 2 g j (n) for all j an n, an T 0, we have err(f c ) > err(f p ). To stuy this issue experimentally, we esigne an compare several clustering schemes for the Robust Reaing task 3. In a Direct clustering approach, we cluster name entities from a collection of ocuments with regar to the target entities, all at once. That is, the entities are viewe as the hien classes that generate the observe name entities in text. The clustering algorithm use is a stanar agglomerative clustering algorithm base on complete-link (which performe better than other clustering methos we attempte). The evaluation, column 2 of Table 3, shows a egraation in the results relative to pairwise classification. Although, in general, clustering oes not help, we attempte to see if clustering can be helpe by exploiting some structural properties of the omain. We split the set of ocuments into three groups, each originate at about the same time perio, an clustere again. In this case (columns Hier (Date) Hierarchically clustering accoring to Dates), the results are significantly better than Direct clustering. We then performe a control experiment (Hier (Ranom)), in which we split the ocument set ranomly into three sets of the same sizes as before; the eterioration in the results in this case inicate that exploiting the structure is significant. The ata set use here was ifferent from the one use in other experiments. They were create by ranomly selecting 600 names for each entity type from the ocuments of years The 1,800 names for each entity type were ranomly split into equal training an test set. We traine the LMR pairwise classifier for each entity type using the corresponing annotate training set an cluster the test set with LMR evaluate as a similarity metric. A Generative Moel for Robust Reaing Motivate by the above observation, we escribe next a generative moel for Robust Reaing, esigne to exploit ocuments structure an assumptions on how they are generate. The etails of this moel are escribe in (Li, Morie, & Roth 2004). Here we briefly escribe one instantiation 3 Given the activation values of the classifier, we efine the istance metric as follows. Let p, n be the positive an negative activation values, respectively, for two names n 1 an n 2; then, ist(n 1, n 2) = en. e p +e n 422 NATURAL LANGUAGE PROCESSING & INFORMATION EXTRACTION

5 F 1 (%) LMR Direct Hier. (Date) Hier. (Ranom) Peop Loc Org Table 3: Best Performance of Different Clustering Approaches (Various parameter settings, incluing ifferent numbers of clusters were experimente within Direct Clustering.) LMR represents our pairwise classifier. The other three columns are ifferent clustering schemes. Results are evaluate by the F 1 values using the best threshol in experiments. The test set has 900 names for each entity type. John Fitzgeral Kenney M i e John Fitzgeral Kenney e i Presient John F. Kenney r i {Presient Kenney, Kenney, JFK} E House of Representatives E House of Representatives R House of Representatives M {House of Representatives, The House} Figure 2: Generating a ocument of the moel (Moel II there) an concentrate on its experimental stuy an a comparison to the iscriminative moels. We efine a probability istribution over ocuments = {E,R,M }, by escribing how ocuments are being generate. In its most general form the moel has the following three components: (1) A joint probability istribution P(E ) that governs how entities (of ifferent types) are istribute into a ocument an reflects their co-occurrence epenencies. (When initializing the moel escribe here, we assume that entities are chosen inepenently of each other.) (2) One mention of a name in a ocument, a representative r is easily ientifiable an other mentions are generate from it. The number of entities in a ocument, size(e ), an the number of mentions of each entity in E, size(mi ), are assume in the current evaluation to be istribute uniformly over a small plausible range so that they can be ignore in later inference. (3) The appearance probability of a name generate (transforme) from its representative is moelle as a prouct istribution over relational transformations of attribute values. This moel captures the similarity between appearances of two names. In the current evaluation the same appearance moel is use to calculate both the probability P(r e) that generates a representative r given an entity e an the probability P(m r) that generates a mention m given a representative r. Attribute transformations are relational, in the sense that the istribution is over transformation types an inepenent of the specific names. Fig. 2 epicts the generation process of a ocument. Thus, in orer to generate a ocument, after picking size(e ) an {size(m1 ),size(m2 ),...}, each entity e i is selecte into inepenently of others, accoring to P(e i ). Next, F 1 (%) Marlin SoftTFIDF LMR Generative Peop Loc Org Table 4: Discriminative an Generative Moels Results are evaluate by the average F 1 values over the 5 test sets for each entity type. Marlin, SoftTFIDF an LMR are the three pairwise classifiers; Generative is the generative moel. F 1 (%) LMR Generative (all) Generative (unseen ata) Peop Loc Org Table 5: Results of Different Learning Protocols for the Generative Moel. The table shows the results of our supervise classifier (LMR) traine with 600 names, the generative moel traine with all the 8, 000 names an the generative moel traine with the part of 8, 000 names not use in the corresponing test set. Results are evaluate an average over 5 test sets for each entity type. the representative ri for each entity e i is selecte accoring to P(ri e i ) an for each representative the actual mentions are selecte inepenently accoring to P(m j r j ). Assuming conitional inepenency between M an E given R, the probability istribution over ocuments is therefore P() = P(E, R, M ) = P(E )P(R E )P(M R ) E [P(e i )P(ri e i )] i=1 (r j,m j ) P(m j r j) after we ignore the size components. Given a mention m in a ocument (M is the set of observe mentions in ), the key inference problem is to etermine the most likely entity e m that correspons to it. This is one by computing: E = argmax E EP(E, R, M θ), (7) where θ is the learne moel s parameters. This gives the assignment of the most likely entity e m for m. An the probability of the ocument collection D is: P(D) = D P(). The moel parameters are learne in an unsupervise way given a collection of ocuments: that is, the information whether two mentions represent the same entity is unavailable uring training. A greey search algorithm moifie after the stanar EM algorithm is aopte here to simplify the computation. Given a set of ocuments D to be stuie an the observe mentions M in each ocument, this F 1 (%) Generative Initial Peop Loc Org Table 6: Performance of Simple Initialization. Generative the generative moel learne in a normal way. Initial the parameters of the generative moel initialize using some simple heuristics an use to cluster names. Results are evaluate by the average F 1 values over the 5 test sets for each entity type. NATURAL LANGUAGE PROCESSING & INFORMATION EXTRACTION 423

6 algorithm iteratively upates the moel parameter θ (several unerlying probabilistic istributions escribe before) an the structure (that is, E an R ) of each ocument by improving P(D θ). Unlike the stanar EM algorithm, in the E-step, it seeks the most likely E an R for each ocument rather than the expecte assignment. There is an initial step to acquire a coarse guess of ocument structures an to seek the set of entities E in a close collection of ocuments D. A local clustering is performe to group all mentions insie each ocument. A set of simple heuristics of matching attributes 4 is applie to computing the similarity between mentions, an pairs of mentions with similarity above a threshol are clustere together. A Comparison between the Moels We traine the generative moel in an unsupervise way with all 8, 000 names. The somewhat surprising results are shown in Table 4. The generative moel outperforme the supervise classifier for People an Organizations. That is, by incorporating a lot of unlabelle ata, the unsupervise learning coul o better. To unerstan this better, an to compare our iscriminative an generative approaches further, we aresse the following two issues: Learning protocol: A supervise learning approach is traine on a training corpus, an teste on a ifferent one resulting, necessarily, in some egraation in performance. An unsupervise metho learns irectly on the target corpus. This, as we show, can be significant. In a secon experiment, in which we o not train the generative moel on names it will see in the test set, results clearly egrae (Table 5). Structural assumptions: Our generative moel benefits from the structural assumptions mae in esigning the moel. We exhibit this by evaluating a fairly weak initialization of the moel, an showing that, nevertheless, this results in a Robust Reaing moel with respectable results. Table 6 shows that after initializing the moel parameters with the heuristics use in the EM-like algorithm, an without further training (but with the inference of the probabilistic moel), the generative moel can perform reasonably well. Conclusion The paper presents two learning approaches to the robust reaing problem cross-ocument ientification of names espite ambiguous writings. In aition to a stanar moelling of the problem as a classification task, we evelope a moel that escribes the natural generation process of a ocument an the process of how names are sprinkle into them, taking into account epenencies between entities across types an an author moel. We have shown that both moels gain a lot from incorporating structural an relational information features in the classification moel; coherent ata splits for clustering an the natural generation process in the case of the probabilistic moel. The robust reaing problem is a very significant barrier on our way towar supporting robust natural language processing, an the reliable an accurate results we show are 4 E.g., one example of a heuristics is: If mention n 1 is a substring of mention n 2, then their similarity is 0.8. thus very encouraging. Also important is the fact that our unsupervise moel performs so well, since the availability of annotate ata is, in many cases, a significant obstacle to goo performance in NLP tasks. In aition to further stuies of the iscriminative moel, incluing going beyon the current noisy supervision (given at a global annotation level, although learning is one locally). exploring how much ata is neee for a supervise moel to perform as well as the unsupervise moel, an whether the initialization of the unsupervise moel can gain from supervision, there are several other critical issues we woul like to aress from the robust reaing perspective. These inclue (1) incorporating more contextual information (like time an place) relate to the target entities, both to support a better moel an to allow temporal tracing of entities; (2) stuying an incremental approach to learning the moel an (3) integrating this work with other aspect of coreference resolution (e.g., pronouns). Acknowlegment This research is supporte by NSF grants IIS , ITR IIS an EIA , an ONR MURI Awar an an equipment onation from AMD. References Bilenko, M., an Mooney, R Aaptive uplicate etection using learnable string similarity measures. In KDD. Carlson, A.; Cumby, C.; Rosen, J.; an Roth, D The SNoW learning architecture. Technical Report UIUCDCS-R , UIUC Computer Science Department. Cohen, W., an Richman, J Learning to match an cluster large high-imensional ata sets for ata integration. In KDD. Cohen, W.; Ravikumar, P.; an Fienberg, S A comparison of string metrics for name-matching tasks. In IIWeb Workshop Doan, A.; Lu, Y.; Lee, Y.; an Han, J Profile-base object matching for information integration. IEEE Intelligent Systems 18(5): Hernanez, M., an Stolfo, S The merge/purge problem for large atabases. In SIGMOD. Kehler, A Coherence, Reference, an the Theory of Grammar. CSLI Publications. Li, X.; Morie, P.; an Roth, D Robust reaing: Ientification an tracing of ambiguous names. In HLT-NAACL. Mann, G., an Yarowsky, D Unsupervise personal name isambiguation. In CoNLL. McCallum, A., an Wellner, B Towar conitional moels of ientity uncertainty with application to proper noun coreference. In IJCAI Workshop on Information Integration on the Web. Ng, V., an Carie, C Improving machine learning approaches to coreference resolution. In ACL. Pasula, H.; Marthi, B.; Milch, B.; Russell, S.; an Shpitser, I Ientity uncertainty an citation matching. In NIPS. Soon, W.; Ng, H.; an Lim, D A machine learning approach to coreference resolution of noun phrases. Computational Linguistics (Special Issue on Computational Anaphora Resolution) 27: Voorhees, E Overview of the TREC-2002 question answering track. In TREC, NATURAL LANGUAGE PROCESSING & INFORMATION EXTRACTION

Robust Reading: Identification and Tracing of Ambiguous Names

Robust Reading: Identification and Tracing of Ambiguous Names Robust Reaing: Ientification an Tracing of Ambiguous Names Xin Li Paul Morie Dan Roth Department of Computer Science University of Illinois, Urbana, IL 61801 {xli1,morie,anr}@uiuc.eu Abstract A given entity,

More information

Robust Reading of Ambiguous Writing

Robust Reading of Ambiguous Writing In submission; please o not istribute. Abstract A given entity, representing a person, a location or an organization, may be mentione in text in multiple, ambiguous ways. Unerstaning natural language requires

More information

UCLA STAT 13 Introduction to Statistical Methods for the Life and Health Sciences. Chapter 9 Paired Data. Paired data. Paired data

UCLA STAT 13 Introduction to Statistical Methods for the Life and Health Sciences. Chapter 9 Paired Data. Paired data. Paired data UCLA STAT 3 Introuction to Statistical Methos for the Life an Health Sciences Instructor: Ivo Dinov, Asst. Prof. of Statistics an Neurology Chapter 9 Paire Data Teaching Assistants: Jacquelina Dacosta

More information

Data Center Power System Reliability Beyond the 9 s: A Practical Approach

Data Center Power System Reliability Beyond the 9 s: A Practical Approach Data Center Power System Reliability Beyon the 9 s: A Practical Approach Bill Brown, P.E., Square D Critical Power Competency Center. Abstract Reliability has always been the focus of mission-critical

More information

Applications of Global Positioning System in Traffic Studies. Yi Jiang 1

Applications of Global Positioning System in Traffic Studies. Yi Jiang 1 Applications of Global Positioning System in Traffic Stuies Yi Jiang 1 Introuction A Global Positioning System (GPS) evice was use in this stuy to measure traffic characteristics at highway intersections

More information

Open World Face Recognition with Credibility and Confidence Measures

Open World Face Recognition with Credibility and Confidence Measures Open Worl Face Recognition with Creibility an Confience Measures Fayin Li an Harry Wechsler Department of Computer Science George Mason University Fairfax, VA 22030 {fli, wechsler}@cs.gmu.eu Abstract.

More information

State of Louisiana Office of Information Technology. Change Management Plan

State of Louisiana Office of Information Technology. Change Management Plan State of Louisiana Office of Information Technology Change Management Plan Table of Contents Change Management Overview Change Management Plan Key Consierations Organizational Transition Stages Change

More information

Modelling and Resolving Software Dependencies

Modelling and Resolving Software Dependencies June 15, 2005 Abstract Many Linux istributions an other moern operating systems feature the explicit eclaration of (often complex) epenency relationships between the pieces of software

More information

A New Evaluation Measure for Information Retrieval Systems

A New Evaluation Measure for Information Retrieval Systems A New Evaluation Measure for Information Retrieval Systems Martin Mehlitz martin.mehlitz@ai-labor.e Christian Bauckhage Deutsche Telekom Laboratories christian.bauckhage@telekom.e Jérôme Kunegis jerome.kunegis@ai-labor.e

More information

View Synthesis by Image Mapping and Interpolation

View Synthesis by Image Mapping and Interpolation View Synthesis by Image Mapping an Interpolation Farris J. Halim Jesse S. Jin, School of Computer Science & Engineering, University of New South Wales Syney, NSW 05, Australia Basser epartment of Computer

More information

Unsteady Flow Visualization by Animating Evenly-Spaced Streamlines

Unsteady Flow Visualization by Animating Evenly-Spaced Streamlines EUROGRAPHICS 2000 / M. Gross an F.R.A. Hopgoo Volume 19, (2000), Number 3 (Guest Eitors) Unsteay Flow Visualization by Animating Evenly-Space Bruno Jobar an Wilfri Lefer Université u Littoral Côte Opale,

More information

Springs, Shocks and your Suspension

Springs, Shocks and your Suspension rings, Shocks an your Suspension y Doc Hathaway, H&S Prototype an Design, C. Unerstaning how your springs an shocks move as your race car moves through its range of motions is one of the basics you must

More information

Purpose of the Experiments. Principles and Error Analysis. ε 0 is the dielectric constant,ε 0. ε r. = 8.854 10 12 F/m is the permittivity of

Purpose of the Experiments. Principles and Error Analysis. ε 0 is the dielectric constant,ε 0. ε r. = 8.854 10 12 F/m is the permittivity of Experiments with Parallel Plate Capacitors to Evaluate the Capacitance Calculation an Gauss Law in Electricity, an to Measure the Dielectric Constants of a Few Soli an Liqui Samples Table of Contents Purpose

More information

Sensor Network Localization from Local Connectivity : Performance Analysis for the MDS-MAP Algorithm

Sensor Network Localization from Local Connectivity : Performance Analysis for the MDS-MAP Algorithm Sensor Network Localization from Local Connectivity : Performance Analysis for the MDS-MAP Algorithm Sewoong Oh an Anrea Montanari Electrical Engineering an Statistics Department Stanfor University, Stanfor,

More information

Average-Case Analysis of a Nearest Neighbor Algorithm. Moett Field, CA USA. in accuracy. neighbor.

Average-Case Analysis of a Nearest Neighbor Algorithm. Moett Field, CA USA. in accuracy. neighbor. From Proceeings of the Thirteenth International Joint Conference on Articial Intelligence (1993). Chambery, France: Morgan Kaufmann Average-Case Analysis of a Nearest Neighbor Algorithm Pat Langley Learning

More information

JON HOLTAN. if P&C Insurance Ltd., Oslo, Norway ABSTRACT

JON HOLTAN. if P&C Insurance Ltd., Oslo, Norway ABSTRACT OPTIMAL INSURANCE COVERAGE UNDER BONUS-MALUS CONTRACTS BY JON HOLTAN if P&C Insurance Lt., Oslo, Norway ABSTRACT The paper analyses the questions: Shoul or shoul not an iniviual buy insurance? An if so,

More information

Lecture 8: Expanders and Applications

Lecture 8: Expanders and Applications Lecture 8: Expaners an Applications Topics in Complexity Theory an Pseuoranomness (Spring 013) Rutgers University Swastik Kopparty Scribes: Amey Bhangale, Mrinal Kumar 1 Overview In this lecture, we will

More information

Ch 10. Arithmetic Average Options and Asian Opitons

Ch 10. Arithmetic Average Options and Asian Opitons Ch 10. Arithmetic Average Options an Asian Opitons I. Asian Option an the Analytic Pricing Formula II. Binomial Tree Moel to Price Average Options III. Combination of Arithmetic Average an Reset Options

More information

Enterprise Resource Planning

Enterprise Resource Planning Enterprise Resource Planning MPC 6 th Eition Chapter 1a McGraw-Hill/Irwin Copyright 2011 by The McGraw-Hill Companies, Inc. All rights reserve. Enterprise Resource Planning A comprehensive software approach

More information

A Case Study of Applying SOM in Market Segmentation of Automobile Insurance Customers

A Case Study of Applying SOM in Market Segmentation of Automobile Insurance Customers International Journal of Database Theory an Application, pp.25-36 http://x.oi.org/10.14257/ijta.2014.7.1.03 A Case Stuy of Applying SOM in Market Segmentation of Automobile Insurance Customers Vahi Golmah

More information

A Comparison of Performance Measures for Online Algorithms

A Comparison of Performance Measures for Online Algorithms A Comparison of Performance Measures for Online Algorithms Joan Boyar 1, Sany Irani 2, an Kim S. Larsen 1 1 Department of Mathematics an Computer Science, University of Southern Denmark, Campusvej 55,

More information

The one-year non-life insurance risk

The one-year non-life insurance risk The one-year non-life insurance risk Ohlsson, Esbjörn & Lauzeningks, Jan Abstract With few exceptions, the literature on non-life insurance reserve risk has been evote to the ultimo risk, the risk in the

More information

Detecting Possibly Fraudulent or Error-Prone Survey Data Using Benford s Law

Detecting Possibly Fraudulent or Error-Prone Survey Data Using Benford s Law Detecting Possibly Frauulent or Error-Prone Survey Data Using Benfor s Law Davi Swanson, Moon Jung Cho, John Eltinge U.S. Bureau of Labor Statistics 2 Massachusetts Ave., NE, Room 3650, Washington, DC

More information

On Adaboost and Optimal Betting Strategies

On Adaboost and Optimal Betting Strategies On Aaboost an Optimal Betting Strategies Pasquale Malacaria 1 an Fabrizio Smerali 1 1 School of Electronic Engineering an Computer Science, Queen Mary University of Lonon, Lonon, UK Abstract We explore

More information

Modeling and Predicting Popularity Dynamics via Reinforced Poisson Processes

Modeling and Predicting Popularity Dynamics via Reinforced Poisson Processes Proceeings of the Twenty-Eighth AAAI Conference on Artificial Intelligence Moeling an Preicting Popularity Dynamics via Reinforce Poisson Processes Huawei Shen 1, Dashun Wang 2, Chaoming Song 3, Albert-László

More information

INFLUENCE OF GPS TECHNOLOGY ON COST CONTROL AND MAINTENANCE OF VEHICLES

INFLUENCE OF GPS TECHNOLOGY ON COST CONTROL AND MAINTENANCE OF VEHICLES 1 st Logistics International Conference Belgrae, Serbia 28-30 November 2013 INFLUENCE OF GPS TECHNOLOGY ON COST CONTROL AND MAINTENANCE OF VEHICLES Goran N. Raoičić * University of Niš, Faculty of Mechanical

More information

Towards a Framework for Enterprise Architecture Frameworks Comparison and Selection

Towards a Framework for Enterprise Architecture Frameworks Comparison and Selection Towars a Framework for Enterprise Frameworks Comparison an Selection Saber Aballah Faculty of Computers an Information, Cairo University Saber_aballah@hotmail.com Abstract A number of Enterprise Frameworks

More information

Cross-Over Analysis Using T-Tests

Cross-Over Analysis Using T-Tests Chapter 35 Cross-Over Analysis Using -ests Introuction his proceure analyzes ata from a two-treatment, two-perio (x) cross-over esign. he response is assume to be a continuous ranom variable that follows

More information

Firewall Design: Consistency, Completeness, and Compactness

Firewall Design: Consistency, Completeness, and Compactness C IS COS YS TE MS Firewall Design: Consistency, Completeness, an Compactness Mohame G. Goua an Xiang-Yang Alex Liu Department of Computer Sciences The University of Texas at Austin Austin, Texas 78712-1188,

More information

! # % & ( ) +,,),. / 0 1 2 % ( 345 6, & 7 8 4 8 & & &&3 6

! # % & ( ) +,,),. / 0 1 2 % ( 345 6, & 7 8 4 8 & & &&3 6 ! # % & ( ) +,,),. / 0 1 2 % ( 345 6, & 7 8 4 8 & & &&3 6 9 Quality signposting : the role of online information prescription in proviing patient information Liz Brewster & Barbara Sen Information School,

More information

9.3. Diffraction and Interference of Water Waves

9.3. Diffraction and Interference of Water Waves Diffraction an Interference of Water Waves 9.3 Have you ever notice how people relaxing at the seashore spen so much of their time watching the ocean waves moving over the water, as they break repeately

More information

RUNESTONE, an International Student Collaboration Project

RUNESTONE, an International Student Collaboration Project RUNESTONE, an International Stuent Collaboration Project Mats Daniels 1, Marian Petre 2, Vicki Almstrum 3, Lars Asplun 1, Christina Björkman 1, Carl Erickson 4, Bruce Klein 4, an Mary Last 4 1 Department

More information

Capacitive mtouch Sensing Solutions

Capacitive mtouch Sensing Solutions Capacitive mtouch Sensing Solutions Design Guielines 2008 Microchip Technology Incorporate. ll Rights Reserve. Slie 1 Hello, my name is Marc McComb, Technical Training Engineer in the Security Microcontroller

More information

A New Pricing Model for Competitive Telecommunications Services Using Congestion Discounts

A New Pricing Model for Competitive Telecommunications Services Using Congestion Discounts A New Pricing Moel for Competitive Telecommunications Services Using Congestion Discounts N. Keon an G. Ananalingam Department of Systems Engineering University of Pennsylvania Philaelphia, PA 19104-6315

More information

An intertemporal model of the real exchange rate, stock market, and international debt dynamics: policy simulations

An intertemporal model of the real exchange rate, stock market, and international debt dynamics: policy simulations This page may be remove to conceal the ientities of the authors An intertemporal moel of the real exchange rate, stock market, an international ebt ynamics: policy simulations Saziye Gazioglu an W. Davi

More information

1 High-Dimensional Space

1 High-Dimensional Space Contents High-Dimensional Space. Properties of High-Dimensional Space..................... 4. The High-Dimensional Sphere......................... 5.. The Sphere an the Cube in Higher Dimensions...........

More information

FAST JOINING AND REPAIRING OF SANDWICH MATERIALS WITH DETACHABLE MECHANICAL CONNECTION TECHNOLOGY

FAST JOINING AND REPAIRING OF SANDWICH MATERIALS WITH DETACHABLE MECHANICAL CONNECTION TECHNOLOGY FAST JOINING AND REPAIRING OF SANDWICH MATERIALS WITH DETACHABLE MECHANICAL CONNECTION TECHNOLOGY Jörg Felhusen an Sivakumara K. Krishnamoorthy RWTH Aachen University, Chair an Insitute for Engineering

More information

6.3 Microbial growth in a chemostat

6.3 Microbial growth in a chemostat 6.3 Microbial growth in a chemostat The chemostat is a wiely-use apparatus use in the stuy of microbial physiology an ecology. In such a chemostat also known as continuous-flow culture), microbes such

More information

Stock Market Value Prediction Using Neural Networks

Stock Market Value Prediction Using Neural Networks Stock Market Value Preiction Using Neural Networks Mahi Pakaman Naeini IT & Computer Engineering Department Islamic Aza University Paran Branch e-mail: m.pakaman@ece.ut.ac.ir Hamireza Taremian Engineering

More information

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.436J/15.085J Fall 2008 Lecture 14 10/27/2008 MOMENT GENERATING FUNCTIONS

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.436J/15.085J Fall 2008 Lecture 14 10/27/2008 MOMENT GENERATING FUNCTIONS MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.436J/15.085J Fall 2008 Lecture 14 10/27/2008 MOMENT GENERATING FUNCTIONS Contents 1. Moment generating functions 2. Sum of a ranom number of ranom variables 3. Transforms

More information

The Elastic Capacitor and its Unusual Properties

The Elastic Capacitor and its Unusual Properties 1 The Elastic Capacitor an its Unusual Properties Michael B. Partensky, Department of Chemistry, Braneis University, Waltham, MA 453 partensky@attbi.com The elastic capacitor (EC) moel was first introuce

More information

n-parameter families of curves

n-parameter families of curves 1 n-parameter families of curves For purposes of this iscussion, a curve will mean any equation involving x, y, an no other variables. Some examples of curves are x 2 + (y 3) 2 = 9 circle with raius 3,

More information

A Monte Carlo Simulation of Multivariate General

A Monte Carlo Simulation of Multivariate General A Monte Carlo Simulation of Multivariate General Pareto Distribution an its Application 1 1 1 1 1 1 1 1 0 1 Luo Yao 1, Sui Danan, Wang Dongxiao,*, Zhou Zhenwei, He Weihong 1, Shi Hui 1 South China Sea

More information

ThroughputScheduler: Learning to Schedule on Heterogeneous Hadoop Clusters

ThroughputScheduler: Learning to Schedule on Heterogeneous Hadoop Clusters ThroughputScheuler: Learning to Scheule on Heterogeneous Haoop Clusters Shehar Gupta, Christian Fritz, Bob Price, Roger Hoover, an Johan e Kleer Palo Alto Research Center, Palo Alto, CA, USA {sgupta, cfritz,

More information

CHAPTER 5 : CALCULUS

CHAPTER 5 : CALCULUS Dr Roger Ni (Queen Mary, University of Lonon) - 5. CHAPTER 5 : CALCULUS Differentiation Introuction to Differentiation Calculus is a branch of mathematics which concerns itself with change. Irrespective

More information

DEVELOPMENT OF A BRAKING MODEL FOR SPEED SUPERVISION SYSTEMS

DEVELOPMENT OF A BRAKING MODEL FOR SPEED SUPERVISION SYSTEMS DEVELOPMENT OF A BRAKING MODEL FOR SPEED SUPERVISION SYSTEMS Paolo Presciani*, Monica Malvezzi #, Giuseppe Luigi Bonacci +, Monica Balli + * FS Trenitalia Unità Tecnologie Materiale Rotabile Direzione

More information

Solving Problems Involving Line Plots

Solving Problems Involving Line Plots Lesson -A Common Core Lesson Overview Domain Measurement an Data Cluster Represent an interpret ata. Stanar.MD. Make a line plot to isplay a ata set of measurements in fractions of a unit ( _, _, _ ).

More information

Guidelines for the use of UHPLC Instruments

Guidelines for the use of UHPLC Instruments Guielines for the use of UHPLC Instruments Requirements for UHPLC instruments, metho evelopment in UHPLC an metho transfer from regular HPLC to UHPLC Authors: Dr. Davy Guillarme, Prof. Jean-Luc Veuthey

More information

A Data Placement Strategy in Scientific Cloud Workflows

A Data Placement Strategy in Scientific Cloud Workflows A Data Placement Strategy in Scientific Clou Workflows Dong Yuan, Yun Yang, Xiao Liu, Jinjun Chen Faculty of Information an Communication Technologies, Swinburne University of Technology Hawthorn, Melbourne,

More information

Learning-Based Summarisation of XML Documents

Learning-Based Summarisation of XML Documents Learning-Base Summarisation of XML Documents Massih R. Amini Anastasios Tombros Nicolas Usunier Mounia Lalmas {name}@poleia.lip6.fr {first name}@cs.qmul.ac.uk University Pierre an Marie Curie Queen Mary,

More information

Minimum-Energy Broadcast in All-Wireless Networks: NP-Completeness and Distribution Issues

Minimum-Energy Broadcast in All-Wireless Networks: NP-Completeness and Distribution Issues Minimum-Energy Broacast in All-Wireless Networks: NP-Completeness an Distribution Issues Mario Čagal LCA-EPFL CH-05 Lausanne Switzerlan mario.cagal@epfl.ch Jean-Pierre Hubaux LCA-EPFL CH-05 Lausanne Switzerlan

More information

Using research evidence in mental health: user-rating and focus group study of clinicians preferences for a new clinical question-answering service

Using research evidence in mental health: user-rating and focus group study of clinicians preferences for a new clinical question-answering service DOI: 10.1111/j.1471-1842.2008.00833.x Using research evience in mental health: user-rating an focus group stuy of clinicians preferences for a new clinical question-answering service Elizabeth A. Barley*,

More information

Performance And Analysis Of Risk Assessment Methodologies In Information Security

Performance And Analysis Of Risk Assessment Methodologies In Information Security International Journal of Computer Trens an Technology (IJCTT) volume 4 Issue 10 October 2013 Performance An Analysis Of Risk Assessment ologies In Information Security K.V.D.Kiran #1, Saikrishna Mukkamala

More information

Introduction to Integration Part 1: Anti-Differentiation

Introduction to Integration Part 1: Anti-Differentiation Mathematics Learning Centre Introuction to Integration Part : Anti-Differentiation Mary Barnes c 999 University of Syney Contents For Reference. Table of erivatives......2 New notation.... 2 Introuction

More information

Consumer Referrals. Maria Arbatskaya and Hideo Konishi. October 28, 2014

Consumer Referrals. Maria Arbatskaya and Hideo Konishi. October 28, 2014 Consumer Referrals Maria Arbatskaya an Hieo Konishi October 28, 2014 Abstract In many inustries, rms rewar their customers for making referrals. We analyze the optimal policy mix of price, avertising intensity,

More information

Product Differentiation for Software-as-a-Service Providers

Product Differentiation for Software-as-a-Service Providers University of Augsburg Prof. Dr. Hans Ulrich Buhl Research Center Finance & Information Management Department of Information Systems Engineering & Financial Management Discussion Paper WI-99 Prouct Differentiation

More information

A Universal Sensor Control Architecture Considering Robot Dynamics

A Universal Sensor Control Architecture Considering Robot Dynamics International Conference on Multisensor Fusion an Integration for Intelligent Systems (MFI2001) Baen-Baen, Germany, August 2001 A Universal Sensor Control Architecture Consiering Robot Dynamics Frierich

More information

Image compression predicated on recurrent iterated function systems **

Image compression predicated on recurrent iterated function systems ** 1 Image compression preicate on recurrent iterate function systems ** W. Metzler a, *, C.H. Yun b, M. Barski a a Faculty of Mathematics University of Kassel, Kassel, F. R. Germany b Faculty of Mathematics

More information

DETERMINING OPTIMAL STOCK LEVEL IN MULTI-ECHELON SUPPLY CHAINS

DETERMINING OPTIMAL STOCK LEVEL IN MULTI-ECHELON SUPPLY CHAINS HUNGARIAN JOURNA OF INDUSTRIA CHEMISTRY VESZPRÉM Vol. 39(1) pp. 107-112 (2011) DETERMINING OPTIMA STOCK EVE IN MUTI-ECHEON SUPPY CHAINS A. KIRÁY 1, G. BEVÁRDI 2, J. ABONYI 1 1 University of Pannonia, Department

More information

Optimizing Multiple Stock Trading Rules using Genetic Algorithms

Optimizing Multiple Stock Trading Rules using Genetic Algorithms Optimizing Multiple Stock Traing Rules using Genetic Algorithms Ariano Simões, Rui Neves, Nuno Horta Instituto as Telecomunicações, Instituto Superior Técnico Av. Rovisco Pais, 040-00 Lisboa, Portugal.

More information

Universal Gravity Based on the Electric Universe Model

Universal Gravity Based on the Electric Universe Model Universal Gravity Base on the Electric Universe Moel By Frerik Nygaar frerik_nygaar@hotmail.com Nov 015, Porto Portugal. Introuction While most people are aware of both Newton's Universal Law of Gravity

More information

Heat-And-Mass Transfer Relationship to Determine Shear Stress in Tubular Membrane Systems Ratkovich, Nicolas Rios; Nopens, Ingmar

Heat-And-Mass Transfer Relationship to Determine Shear Stress in Tubular Membrane Systems Ratkovich, Nicolas Rios; Nopens, Ingmar Aalborg Universitet Heat-An-Mass Transfer Relationship to Determine Shear Stress in Tubular Membrane Systems Ratkovich, Nicolas Rios; Nopens, Ingmar Publishe in: International Journal of Heat an Mass Transfer

More information

Traffic Delay Studies at Signalized Intersections with Global Positioning System Devices

Traffic Delay Studies at Signalized Intersections with Global Positioning System Devices Traffic Delay Stuies at Signalize Intersections with Global Positioning System Devices THIS FEATURE PRESENTS METHODS FOR APPLYING GPS DEVICES TO STUDY TRAFFIC DELAYS AT SIGNALIZED INTERSECTIONS. MOST OF

More information

Digital barrier option contract with exponential random time

Digital barrier option contract with exponential random time IMA Journal of Applie Mathematics Avance Access publishe June 9, IMA Journal of Applie Mathematics ) Page of 9 oi:.93/imamat/hxs3 Digital barrier option contract with exponential ranom time Doobae Jun

More information

Lecture 17: Implicit differentiation

Lecture 17: Implicit differentiation Lecture 7: Implicit ifferentiation Nathan Pflueger 8 October 203 Introuction Toay we iscuss a technique calle implicit ifferentiation, which provies a quicker an easier way to compute many erivatives we

More information

Jitter effects on Analog to Digital and Digital to Analog Converters

Jitter effects on Analog to Digital and Digital to Analog Converters Jitter effects on Analog to Digital an Digital to Analog Converters Jitter effects copyright 1999, 2000 Troisi Design Limite Jitter One of the significant problems in igital auio is clock jitter an its

More information

If you have ever spoken with your grandparents about what their lives were like

If you have ever spoken with your grandparents about what their lives were like CHAPTER 7 Economic Growth I: Capital Accumulation an Population Growth The question of growth is nothing new but a new isguise for an age-ol issue, one which has always intrigue an preoccupie economics:

More information

Di usion on Social Networks. Current Version: June 6, 2006 Appeared in: Économie Publique, Numéro 16, pp 3-16, 2005/1.

Di usion on Social Networks. Current Version: June 6, 2006 Appeared in: Économie Publique, Numéro 16, pp 3-16, 2005/1. Di usion on Social Networks Matthew O. Jackson y Caltech Leeat Yariv z Caltech Current Version: June 6, 2006 Appeare in: Économie Publique, Numéro 16, pp 3-16, 2005/1. Abstract. We analyze a moel of i

More information

Improving Direct Marketing Profitability with Neural Networks

Improving Direct Marketing Profitability with Neural Networks Volume 9 o.5, September 011 Improving Direct Marketing Profitability with eural etworks Zaiyong Tang Salem State University Salem, MA 01970 ABSTRACT Data mining in irect marketing aims at ientifying the

More information

Legal Claim Identification: Information Extraction with Hierarchically Labeled Data

Legal Claim Identification: Information Extraction with Hierarchically Labeled Data Legal Claim Ientification: Information Extraction with Hierarchically Labele Data Mihai Sureanu, Ramesh Nallapati an Christopher Manning Stanfor University {mihais,nmramesh,manning}@cs.stanfor.eu Abstract

More information

Stochastic Modeling of MEMS Inertial Sensors

Stochastic Modeling of MEMS Inertial Sensors BULGARIAN ACADEMY OF SCIENCES CYBERNETICS AND INFORMATION TECHNOLOGIES Volume 10, No Sofia 010 Stochastic Moeling of MEMS Inertial Sensors Petko Petkov, Tsonyo Slavov Department of Automatics, Technical

More information

Exploratory Optimal Latin Hypercube Designs for Computer Simulated Experiments

Exploratory Optimal Latin Hypercube Designs for Computer Simulated Experiments Thailan Statistician July 0; 9() : 7-93 http://statassoc.or.th Contribute paper Exploratory Optimal Latin Hypercube Designs for Computer Simulate Experiments Rachaaporn Timun [a,b] Anamai Na-uom* [a,b]

More information

MSc. Econ: MATHEMATICAL STATISTICS, 1995 MAXIMUM-LIKELIHOOD ESTIMATION

MSc. Econ: MATHEMATICAL STATISTICS, 1995 MAXIMUM-LIKELIHOOD ESTIMATION MAXIMUM-LIKELIHOOD ESTIMATION The General Theory of M-L Estimation In orer to erive an M-L estimator, we are boun to make an assumption about the functional form of the istribution which generates the

More information

_Mankiw7e_CH07.qxp 3/2/09 9:40 PM Page 189 PART III. Growth Theory: The Economy in the Very Long Run

_Mankiw7e_CH07.qxp 3/2/09 9:40 PM Page 189 PART III. Growth Theory: The Economy in the Very Long Run 189-220_Mankiw7e_CH07.qxp 3/2/09 9:40 PM Page 189 PART III Growth Theory: The Economy in the Very Long Run 189-220_Mankiw7e_CH07.qxp 3/2/09 9:40 PM Page 190 189-220_Mankiw7e_CH07.qxp 3/2/09 9:40 PM Page

More information

Increasing Packet Delivery Ratio in DSR by Link Prediction

Increasing Packet Delivery Ratio in DSR by Link Prediction Increasing Packet Delivery Ratio in DSR by Link Preiction Liang Qin Thomas Kunz Department of Systems an Computer Engineering Carleton University, Ottawa, Ontario, Canaa K1S 5B6 {lqin,tkunz}@sce.carleton.ca

More information

5 Isotope effects on vibrational relaxation and hydrogen-bond dynamics in water

5 Isotope effects on vibrational relaxation and hydrogen-bond dynamics in water 5 Isotope effects on vibrational relaxation an hyrogen-bon ynamics in water Pump probe experiments HDO issolve in liqui H O show the spectral ynamics an the vibrational relaxation of the OD stretch vibration.

More information

Minimizing Makespan in Flow Shop Scheduling Using a Network Approach

Minimizing Makespan in Flow Shop Scheduling Using a Network Approach Minimizing Makespan in Flow Shop Scheuling Using a Network Approach Amin Sahraeian Department of Inustrial Engineering, Payame Noor University, Asaluyeh, Iran 1 Introuction Prouction systems can be ivie

More information

GPRS performance estimation in GSM circuit switched services and GPRS shared resource systems *

GPRS performance estimation in GSM circuit switched services and GPRS shared resource systems * GPRS performance estimation in GSM circuit switche serices an GPRS share resource systems * Shaoji i an Sen-Gusta Häggman Helsinki Uniersity of Technology, Institute of Raio ommunications, ommunications

More information

. A UML/MARTE DETECTION OF STARVATION AND DEADLOCKS AT THE DESIGN LEVEL IN CONCURRENT SYSTEM

. A UML/MARTE DETECTION OF STARVATION AND DEADLOCKS AT THE DESIGN LEVEL IN CONCURRENT SYSTEM . A UML/MARTE DETECTION OF STARVATION AND DEADLOCKS AT THE DESIGN LEVEL IN CONCURRENT SYSTEM C.Revath Department of computer science an Engineering, Karunya University, comibatore,inia.rcswathiathi@gmail.com

More information

Using WordNet for Text Categorization

Using WordNet for Text Categorization 16 The International Arab Journal of Information Technology, Vol. 5, No. 1, January 2008 Using WorNet for Text Categorization Zakaria Elberrichi 1, Abelattif Rahmoun 2, an Mohame Amine Bentaalah 1 1 EEDIS

More information

Week 4 - Linear Demand and Supply Curves

Week 4 - Linear Demand and Supply Curves Week 4 - Linear Deman an Supply Curves November 26, 2007 1 Suppose that we have a linear eman curve efine by the expression X D = A bp X an a linear supply curve given by X S = C + P X, where b > 0 an

More information

A Blame-Based Approach to Generating Proposals for Handling Inconsistency in Software Requirements

A Blame-Based Approach to Generating Proposals for Handling Inconsistency in Software Requirements International Journal of nowlege an Systems Science, 3(), -7, January-March 0 A lame-ase Approach to Generating Proposals for Hanling Inconsistency in Software Requirements eian Mu, Peking University,

More information

Optimal Control Policy of a Production and Inventory System for multi-product in Segmented Market

Optimal Control Policy of a Production and Inventory System for multi-product in Segmented Market RATIO MATHEMATICA 25 (2013), 29 46 ISSN:1592-7415 Optimal Control Policy of a Prouction an Inventory System for multi-prouct in Segmente Market Kuleep Chauhary, Yogener Singh, P. C. Jha Department of Operational

More information

Supporting Adaptive Workflows in Advanced Application Environments

Supporting Adaptive Workflows in Advanced Application Environments Supporting aptive Workflows in vance pplication Environments Manfre Reichert, lemens Hensinger, Peter Daam Department Databases an Information Systems University of Ulm, D-89069 Ulm, Germany Email: {reichert,

More information

Performance of 2D versus 3D Topographic Representations for Different Task Types

Performance of 2D versus 3D Topographic Representations for Different Task Types Performance of versus Topographic Representations for Different s Debra MacIvor Savage, Eric N. Wiebe an Hugh A. Devine North Carolina State University Raleigh, North Carolina In this stuy, a performance

More information

The Quick Calculus Tutorial

The Quick Calculus Tutorial The Quick Calculus Tutorial This text is a quick introuction into Calculus ieas an techniques. It is esigne to help you if you take the Calculus base course Physics 211 at the same time with Calculus I,

More information

Implementing IP Traceback in the Internet An ISP Perspective

Implementing IP Traceback in the Internet An ISP Perspective Implementing IP Traceback in the Internet An ISP Perspective Dong Wei, Stuent Member, IEEE, an Nirwan Ansari, Senior Member, IEEE Abstract--Denial-of-Service (DoS) attacks consume the resources of remote

More information

ISSN: 2277-3754 ISO 9001:2008 Certified International Journal of Engineering and Innovative Technology (IJEIT) Volume 3, Issue 12, June 2014

ISSN: 2277-3754 ISO 9001:2008 Certified International Journal of Engineering and Innovative Technology (IJEIT) Volume 3, Issue 12, June 2014 ISSN: 77-754 ISO 900:008 Certifie International Journal of Engineering an Innovative echnology (IJEI) Volume, Issue, June 04 Manufacturing process with isruption uner Quaratic Deman for Deteriorating Inventory

More information

Patch Complexity, Finite Pixel Correlations and Optimal Denoising

Patch Complexity, Finite Pixel Correlations and Optimal Denoising Patch Complexity, Finite Pixel Correlations an Optimal Denoising Anat Levin Boaz Naler Freo Duran William T. Freeman Weizmann Institute MIT CSAIL Abstract. Image restoration tasks are ill-pose problems,

More information

10.2 Systems of Linear Equations: Matrices

10.2 Systems of Linear Equations: Matrices SECTION 0.2 Systems of Linear Equations: Matrices 7 0.2 Systems of Linear Equations: Matrices OBJECTIVES Write the Augmente Matrix of a System of Linear Equations 2 Write the System from the Augmente Matrix

More information

Seeing the Unseen: Revealing Mobile Malware Hidden Communications via Energy Consumption and Artificial Intelligence

Seeing the Unseen: Revealing Mobile Malware Hidden Communications via Energy Consumption and Artificial Intelligence Seeing the Unseen: Revealing Mobile Malware Hien Communications via Energy Consumption an Artificial Intelligence Luca Caviglione, Mauro Gaggero, Jean-François Lalane, Wojciech Mazurczyk, Marcin Urbanski

More information

Option Pricing for Inventory Management and Control

Option Pricing for Inventory Management and Control Option Pricing for Inventory Management an Control Bryant Angelos, McKay Heasley, an Jeffrey Humpherys Abstract We explore the use of option contracts as a means of managing an controlling inventories

More information

New Hypothesis Testing-Based Methods for Fault Detection for Smart Grid Systems

New Hypothesis Testing-Based Methods for Fault Detection for Smart Grid Systems New Hypothesis esting-base Methos for Fault Detection for Smart Gri Systems Qian He an Rick S. Blum ECE Department, Lehigh University 9 Memorial Drive West, Bethlehem, PA 85 {qih27; rblum}@lehigh.eu Abstract

More information

Math 230.01, Fall 2012: HW 1 Solutions

Math 230.01, Fall 2012: HW 1 Solutions Math 3., Fall : HW Solutions Problem (p.9 #). Suppose a wor is picke at ranom from this sentence. Fin: a) the chance the wor has at least letters; SOLUTION: All wors are equally likely to be chosen. The

More information

Optimized Capacity Bounds for the Half-Duplex Gaussian MIMO Relay Channel

Optimized Capacity Bounds for the Half-Duplex Gaussian MIMO Relay Channel Optimize Capacity Bouns for the Half-Duplex Gaussian MIMO elay Channel Lennart Geres an Wolfgang Utschick International ITG Workshop on mart Antennas (WA) February 211 c 211 IEEE. Personal use of this

More information

Professional Level Options Module, Paper P4(SGP)

Professional Level Options Module, Paper P4(SGP) Answers Professional Level Options Moule, Paper P4(SGP) Avance Financial Management (Singapore) December 2007 Answers Tutorial note: These moel answers are consierably longer an more etaile than woul be

More information

An Introduction to Event-triggered and Self-triggered Control

An Introduction to Event-triggered and Self-triggered Control An Introuction to Event-triggere an Self-triggere Control W.P.M.H. Heemels K.H. Johansson P. Tabuaa Abstract Recent evelopments in computer an communication technologies have le to a new type of large-scale

More information

Procedure to Measure the Data Transmission Speed in Mobile Networks in Accordance with the LTE Standard (Methodical procedure)

Procedure to Measure the Data Transmission Speed in Mobile Networks in Accordance with the LTE Standard (Methodical procedure) Prague, 15 August 2013 Proceure to Measure the Data Transmission Spee in Mobile Networks in Accorance with the LTE Stanar (Methoical proceure Publishe in connection with the Tener for the awar of the rights

More information

Stress Concentration Factors of Various Adjacent Holes Configurations in a Spherical Pressure Vessel

Stress Concentration Factors of Various Adjacent Holes Configurations in a Spherical Pressure Vessel 5 th Australasian Congress on Applie Mechanics, ACAM 2007 10-12 December 2007, Brisbane, Australia Stress Concentration Factors of Various Ajacent Holes Configurations in a Spherical Pressure Vessel Kh.

More information

(We assume that x 2 IR n with n > m f g are twice continuously ierentiable functions with Lipschitz secon erivatives. The Lagrangian function `(x y) i

(We assume that x 2 IR n with n > m f g are twice continuously ierentiable functions with Lipschitz secon erivatives. The Lagrangian function `(x y) i An Analysis of Newton's Metho for Equivalent Karush{Kuhn{Tucker Systems Lus N. Vicente January 25, 999 Abstract In this paper we analyze the application of Newton's metho to the solution of systems of

More information