Nadav Eiron

SemTag and seeker: bootstrapping the semantic web via automated semantic annotation

This paper describes Seeker, a platform for large-scale text analytics, and SemTag, an application written on the platform to perform automated semantic tagging of large corpora. We apply SemTag to a collection of approximately 264 million web pages, and generate approximately 434 million automatically disambiguated semantic tags, published to the web as a label bureau providing metadata regarding the 434 million annotations. To our knowledge, this is the largest scale semantic tagging effort to date.We describe the Seeker platform, discuss the architecture of the SemTag application, describe a new disambiguation algorithm specialized to support ontological disambiguation of large-scale data, evaluate the algorithm, and present our final results with information about acquiring and making use of the semantic tags. We argue that automated large scale semantic tagging of ambiguous content can bootstrap and accelerate the creation of the semantic web.

Analysis of anchor text for web search

It has been observed that anchor text in web documents is very useful in improving the quality of web text search for some classes of queries. By examining properties of anchor text in a large intranet, we hope to shed light on why this is the case. Our main premise is that anchor text behaves very much like real user queries and consensus titles. Thus an understanding of how anchor text is related to a document will likely lead to better understanding of how to translate a user’s query into high quality search results. Our approach is experimental, based on a study of a large corporate intranet, including the content as well as a large stream of queries against that content. We conduct experiments to investigate several aspects of anchor text, including their relationship to titles, the frequency of queries that can be satisfied by anchortext alone, and the homogeneity of results fetched by anchor text.

A case for automated large-scale semantic annotation

Abstract This paper describes Seeker, a platform for large-scale text analytics, and SemTag, an application written on the platform to perform automated semantic tagging of large corpora. We apply SemTag to a collection of approximately 264 million web pages, and generate approximately 434 million automatically disambiguated semantic tags, published to the web as a label bureau providing metadata regarding the 434 million annotations. To our knowledge, this is the largest scale semantic tagging effort to date. We describe the Seeker platform, discuss the architecture of the SemTag application, describe a new disambiguation algorithm specialized to support ontological disambiguation of large-scale data, evaluate the algorithm, and present our final results with information about acquiring and making use of the semantic tags. We argue that automated large-scale semantic tagging of ambiguous content can bootstrap and accelerate the creation of the semantic web.

On the difficulty of approximately maximizing agreements

We address the computational complexity of learning in the agnostic framework. For a variety of common concept classes we prove that, unless P=NP, there is no polynomial time approximation scheme for finding a member in the class that approximately maximizes the agreement with a given training sample. In particular our results apply to the classes of monomials, axis-aligned hyper-rectangles, closed balls and monotone monomials. For each of these classes we prove the NP-hardness of approximating maximal agreement to within some fixed constant (independent of the sample size and of the dimensionality of the sample space). For the class of half-spaces, we prove that, for any † > 0, it is NP-hard to approximately maximize agreements to within a factor of (418=415 i †), improving on the best previously known constant for this problem, and using a simpler proof. An interesting feature of our proofs is that, for each of the classes we discuss, we find patterns of training examples that, while being hard for approximating agreement within that concept class, allow ecient agreement maximization within other concept classes. These results bring up a new aspect of the model selection problem ‐ they imply that the choice of hypothesis class for agnostic learning from among those considered in this paper can drastically eect the computational complexity of the learning process.

PAC learning with nasty noise

We introduce a new model for learning in the presence of noise, which we call the Nasty Noise model. This model generalizes previously considered models of learning with noise. The learning process in this model, which is a variant of the PAC model, proceeds as follows: Suppose that the learning algorithm during its execution asks for m examples. The examples that the algorithm gets are generated by a nasty adversary that works according to the following steps. First, the adversary chooses m examples (independently) according to a fixed (but unknown to the learning algorithm) distribution D as in the PAC-model. Then the powerful adversary, upon seeing the specific m examples that were chosen (and using his knowledge of the target function, the distribution D and the learning algorithm), is allowed to remove a fraction of the examples at its choice, and replace these examples by the same number of arbitrary examples of its choice; the m modified examples are then given to the learning algorithm. The only restriction on the adversary is that the number of examples that the adversary is allowed to modify should be distributed according to a binomial distribution with parameters η (the noise rate) and m.On the negative side, we prove that no algorithm can achieve accuracy of e > 2η in learning any non-trivial class of functions. We also give some lower bounds on the sample complexity required to achieve accuracy e = 2η + Δ. On the positive side, we show that a polynomial (in the usual parameters, and in 1/(e-2η)) number of examples suffice for learning any class of finite VC-dimension with accuracy e > 2η. This algorithm may not be efficient; however, we also show that a fairly wide family of concept classes can be efficiently learned in the presence of nasty noise.

Self-Directed Learning and Its Relation to the VC-Dimension and to Teacher-Directed Learning

We study the self-directed (SD) learning model. In this model a learner chooses examples, guesses their classification and receives immediate feedback indicating the correctness of its guesses. We consider several fundamental questions concerning this model: the parameters of a task that determine the cost of learning, the computational complexity of a student, and the relationship between this model and the teacher-directed (TD) learning model. We answer the open problem of relating the cost of self-directed learning to the VC-dimension by showing that no such relation exists. Furthermore, we refute the conjecture that for the intersection-closed case, the cost of self-directed learning is bounded by the VC-dimension. We also show that the cost of SD learning may be arbitrarily higher that that of TD learning.Finally, we discuss the number of queries needed for learning in this model and its relationship to the number of mistakes the student incurs. We prove a trade-off formula showing that an algorithm that makes fewer queries throughout its learning process, necessarily suffers a higher number of mistakes.

Learning monotone dnf from a teacher that almost does not answer membership queries

We present results concerning the learning of Monotone DNF (MDNF) from Incomplete Membership Queries and Equivalence Queries. Our main result is a new algorithm that allows efficient learning of MDNF using Equivalence Queries and Incomplete Membership Queries with probability of p=1-1/poly(n,t) of failing. Our algorithm is expected to make O((tn/(1-p))2) queries, when learning a MDNF formula with t terms over n variables. Note that this is polynomial for any failure probability p=1-1/poly(n,t). The algorithms running time is also polynomial in t,n, and 1/(1-p). In a sense this is the best possible, as learning with p=1-1/ω(poly(n,t)) would imply learning MDNF, and thus also DNF, from equivalence queries alone.1 1. An early version of this paper appeared as Bshouty and Eiron (2001).

The Computational Complexity of Densest Region Detection

We investigate the computational complexity of the task of detecting dense regions of an unknown distribution from unlabeled samples of this distribution. We introduce a formal learning model for this task that uses a hypothesis class as it “anti-overfitting” mechanism. The learning task in our model can be reduced to a combinatorial optimization problem. We can show that for some constants, depending on the hypothesis class, these problems are NP-hard to approximate to within these constant factors. We go on and introduce a new criterion for the success of approximate optimization geometric problems. The new criterion requires that the algorithm competes with hypotheses only on the points that are separated by some margin ? from their boundaries. Quite surprisingly, we discover that for each of the two hypothesis classes that we investigate, there is a “critical value” of the margin parameter ?. For any value below the critical value the problems are NP-hard to approximate, while, once this value is exceeded, the problems become poly-time solvable.

Matrix multiplication: a case study of enhanced data cache utilization

Modern machines present two challenges to algorithm engineers and compiler writers: They have superscalar, super-pipelined structure, and they have elaborate memory subsystems specifically designed to reduce latency and increase bandwidth. Matrix multiplication is a classical benchmark for experimenting with techniques used to exploit machine architecture and to overcome the limitations of contemporary memory subsystems.This research aims at advancing the state of the art of algorithm engineering by balancing instruction level parallelism, two levels of data tiling, copying to provably avoid any cache conflicts, and prefetching in parallel to computational operations, in order to fully exploit the memory bandwidth. Measurements on IBMs RS/6000 43P workstation show that the resultant matrix multiplication algorithm outperforms IBMs ESSL by 6.8-31.8%, is less sensitive to the size of the input data, and scales better.In this paper we introduce a cache aware algorithm for matrix multiplication. We also suggest generic guidelines that may be applied to compute intensive algorithm to efficiently utilize the data cache. We believe that some of our concepts may be embodied in compilers.

Links in Hierarchical Information Networks

We provide evidence that the inherent hierarchical structure of the web is closely related to the link structure. Moreover, we show that this relationship explains several important features of the web, including the locality and bidirectionality of hyperlinks, and the compressibility of the web graph. We describe how to construct data models of the web that capture both the hierarchical nature of the web as well as some crucial features of the link graph.