One of the prominent methods for explaining the decision of a machine-learning classifier is by a counterfactual example. Most current algorithms for generating such examples in the textual domain are based on generative language models. Generative models, however, are trained to minimize a specific loss function in order to fulfill certain requirements for the generated texts. Any change in the requirements may necessitate costly retraining, thus potentially limiting their applicability. In this paper, we present a general search-based framework for generating counterfactual explanations in the textual domain. Our framework is model-agnostic, domain-agnostic, anytime, and does not require retraining in order to adapt to changes in the user requirements. We model the task as a search problem in a space where the initial state is the classified text, and the goal state is a text in a given target class. Our framework includes domain-independent modification operators, but can also exploit domain-specific knowledge through specialized operators. The search algorithm attempts to find a text from the target class with minimal user-specified distance from the original classified object.

One of the main methods for computational interpretation of a text is mapping it into a vector in some embedding space.Such vectors can then be used for a variety of textual processing tasks.Recently, most embedding spaces are a product of training large language models (LLMs).One major drawback of this type of representation is their incomprehensibility to humans.Understanding the embedding space is crucial for several important needs, including the need to debug the embedding method and compare it to alternatives, and the need to detect biases hidden in the model.In this paper, we present a novel method of understanding embeddings by transforming a latent embedding space into a comprehensible conceptual space.We present an algorithm for deriving a conceptual space with dynamic ondemand granularity.We devise a new evaluation method, using either human rater or LLMbased raters, to show that the conceptualized vectors indeed represent the semantics of the original latent ones.We show the use of our method for various tasks, including comparing the semantics of alternative models and tracing the layers of the LLM.The code is available online 1 .

We tackle the task of Term Set Expansion (TSE): given a small seed set of example terms from a semantic class, finding more members of that class.The task is of great practical utility, and also of theoretical utility as it requires generalization from few examples.Previous approaches to the TSE task can be characterized as either distributional or pattern-based.We harness the power of neural masked language models (MLM) and propose a novel TSE algorithm, which combines the pattern-based and distributional approaches.Due to the small size of the seed set, fine-tuning methods are not effective, calling for more creative use of the MLM.The gist of the idea is to use the MLM to first mine for informative patterns with respect to the seed set, and then to obtain more members of the seed class by generalizing these patterns.Our method outperforms stateof-the-art TSE algorithms.

Human labeling of data can be very time-consuming and expensive, yet, in many cases it is critical for the success of the learning process. In order to minimize human labeling efforts, we propose a novel active learning solution that does not rely on existing sources of unlabeled data. It uses a small amount of labeled data as the core set for the synthesis of useful membership queries (MQs) — unlabeled instances generated by an algorithm for human labeling. Our solution uses modification operators, functions that modify instances to some extent. We apply the operators on a small set of instances (core set), creating a set of new membership queries. Using this framework, we look at the instance space as a search space and apply search algorithms in order to generate new examples highly relevant to the learner. We implement this framework in the textual domain and test it on several text classification tasks and show improved classifier performance as more MQs are labeled and incorporated into the training set. To the best of our knowledge, this is the first work on membership queries in the textual domain.

We address the task of Named Entity Disambiguation (NED) for noisy text.We present WikilinksNED, a large-scale NED dataset of text fragments from the web, which is significantly noisier and more challenging than existing newsbased datasets.To capture the limited and noisy local context surrounding each mention, we design a neural model and train it with a novel method for sampling informative negative examples.We also describe a new way of initializing word and entity embeddings that significantly improves performance.Our model significantly outperforms existing state-ofthe-art methods on WikilinksNED while achieving comparable performance on a smaller newswire dataset.

In this paper we introduce a novel method that can efficiently estimate a family of hierarchical dense sets in high-dimensional distributions. Our method can be regarded as a natural extension of the one-class SVM (OCSVM) algorithm that finds multiple parallel separating hyperplanes in a reproducing kernel Hilbert space. We call our method q-OCSVM, as it can be used to estimate $q$ quantiles of a high-dimensional distribution. For this purpose, we introduce a new global convex optimization program that finds all estimated sets at once and show that it can be solved efficiently. We prove the correctness of our method and present empirical results that demonstrate its superiority over existing methods.

Humans have an uncanny ability to learn new concepts with very few examples. Cognitive theories have suggested that this is done by utilizing prior experience of related tasks. We propose to emulate this process in machines, by transforming new problems into old ones. These transformations are called metaphors. Obviously, the learner is not given a metaphor, but must acquire one through a learning process. We show that learning metaphors yield better results than existing transfer learning methods. Moreover, we argue that metaphors give a qualitative assessment of task relatedness.

The task of automatically determining the correct sense of a polysemous word has remained a challenge to this day. In our research, we introduce Concept-Based Disambiguation (CBD), a novel framework that utilizes recent semantic analysis techniques to represent both the context of the word and its senses in a high-dimensional space of natural concepts. The concepts are retrieved from a vast encyclopedic resource, thus enriching the disambiguation process with large amounts of domain-specific knowledge. In such concept-based spaces, more comprehensive measures can be applied in order to pick the right sense. Additionally, we introduce a novel representation scheme, denoted anchored representation, that builds a more specific text representation associated with an anchoring word. We evaluate our framework and show that the anchored representation is more suitable to the task of word-sense disambiguation (WSD). Additionally, we show that our system is superior to state-of-the-art methods when evaluated on domain-specific corpora, and competitive with recent methods when evaluated on a general corpus.

We propose an efficient, generalized, nonparametric, statistical Kolmogorov-Smirnov test for detecting distributional change in high-dimensional data. To implement the test, we introduce a novel, hierarchical, minimum-volume sets estimator to represent the distributions to be tested. Our work is motivated by the need to detect changes in data streams, and the test is especially efficient in this context. We provide the theoretical foundations of our test and show its superiority over existing methods.

The problem we tackle in this work is, given a present news event, to generate a plausible future event that can be caused by the given event. We present a new methodology for modeling and predicting such future news events using machine learning and data mining techniques. Our Pundit algorithm generalizes examples of causality pairs to infer a causality predictor. To obtain precise labeled causality examples, we mine 150 years of news articles, and apply semantic natural language modeling techniques to titles containing certain predefined causality patterns. For generalization, the model uses a vast amount of world knowledge ontologies mined from LinkedData, containing ~200 datasets with approximately 20 billion relations. Empirical evaluation on real news articles shows that our Pundit algorithm reaches a human-level performance.

Domain-independent planning is one of the foundational areas in the field of Artificial Intelligence. A description of a planning task consists of an initial world state, a goal, and a set of actions for modifying the world state. The objective is to find a sequence of actions, that is, a plan, that transforms the initial world state into a goal state. In optimal planning, we are interested in finding not just a plan, but one of the cheapest plans. A prominent approach to optimal planning these days is heuristic state-space search, guided by admissible heuristic functions. Numerous admissible heuristics have been developed, each with its own strengths and weaknesses, and it is well known that there is no single "best'' heuristic for optimal planning in general. Thus, which heuristic to choose for a given planning task is a difficult question. This difficulty can be avoided by combining several heuristics, but that requires computing numerous heuristic estimates at each state, and the tradeoff between the time spent doing so and the time saved by the combined advantages of the different heuristics might be high. We present a novel method that reduces the cost of combining admissible heuristics for optimal planning, while maintaining its benefits. Using an idealized search space model, we formulate a decision rule for choosing the best heuristic to compute at each state. We then present an active online learning approach for learning a classifier with that decision rule as the target concept, and employ the learned classifier to decide which heuristic to compute at each state. We evaluate this technique empirically, and show that it substantially outperforms the standard method for combining several heuristics via their pointwise maximum.

Given a current news event, we tackle the problem of generating plausible predictions of future events it might cause. We present a new methodology for modeling and predicting such future news events using machine learning and data mining techniques. Our Pundit algorithm generalizes examples of causality pairs to infer a causality predictor. To obtain precisely labeled causality examples, we mine 150 years of news articles and apply semantic natural language modeling techniques to headlines containing certain predefined causality patterns. For generalization, the model uses a vast number of world knowledge ontologies. Empirical evaluation on real news articles shows that our Pundit algorithm performs as well as non-expert humans.

In binary-utility games, an agent can have only two possible utility values for final states, 1 (win) and 0 (lose). An adversarial binary-utility game is one where for each final state there must be at least one winning and one losing agent. We define an unbiased rational agent as one that seeks to maximize its utility value, but is equally likely to choose between states with the same utility value. This induces a probability distribution over the outcomes of the game, from which an agent can infer its probability to win. A single adversary binary game is one where there are only two possible outcomes, so that the winning probabilities remain binary values. In this case, the rational action for an agent is to play minimax. In this work we focus on the more complex, multiple-adversary environment. We propose a new algorithmic framework where agents try to maximize their winning probabilities. We begin by theoretically analyzing why an unbiased rational agent should take our approach in an unbounded environment and not that of the existing Paranoid or MaxN algorithms. We then expand our framework to a resource-bounded environment, where winning probabilities are estimated, and show empirical results supporting our claims.

Computing the degree of semantic relatedness of words is a key functionality of many language applications such as search, clustering, and disambiguation. Previous approaches to computing semantic relatedness mostly used static language resources, while essentially ignoring their temporal aspects. We believe that a considerable amount of relatedness information can also be found in studying patterns of word usage over time. Consider, for instance, a newspaper archive spanning many years. Two words such as "war" and "peace" might rarely co-occur in the same articles, yet their patterns of use over time might be similar. In this paper, we propose a new semantic relatedness model, Temporal Semantic Analysis (TSA), which captures this temporal information. The previous state of the art method, Explicit Semantic Analysis (ESA), represented word semantics as a vector of concepts. TSA uses a more refined representation, where each concept is no longer scalar, but is instead represented as time series over a corpus of temporally-ordered documents. To the best of our knowledge, this is the first attempt to incorporate temporal evidence into models of semantic relatedness. Empirical evaluation shows that TSA provides consistent improvements over the state of the art ESA results on multiple benchmarks.

Information retrieval systems traditionally rely on textual keywords to index and retrieve documents. Keyword-based retrieval may return inaccurate and incomplete results when different keywords are used to describe the same concept in the documents and in the queries. Furthermore, the relationship between these related keywords may be semantic rather than syntactic, and capturing it thus requires access to comprehensive human world knowledge. Concept-based retrieval methods have attempted to tackle these difficulties by using manually built thesauri, by relying on term cooccurrence data, or by extracting latent word relationships and concepts from a corpus. In this article we introduce a new concept-based retrieval approach based on Explicit Semantic Analysis (ESA), a recently proposed method that augments keyword-based text representation with concept-based features, automatically extracted from massive human knowledge repositories such as Wikipedia. Our approach generates new text features automatically, and we have found that high-quality feature selection becomes crucial in this setting to make the retrieval more focused. However, due to the lack of labeled data, traditional feature selection methods cannot be used, hence we propose new methods that use self-generated labeled training data. The resulting system is evaluated on several TREC datasets, showing superior performance over previous state-of-the-art results.

It is well known that there cannot be a single "best" heuristic for optimal planning in general. One way of overcoming this is by combining admissible heuristics (e.g. by using their maximum), which requires computing numerous heuristic estimates at each state. However, there is a tradeoff between the time spent on computing these heuristic estimates for each state, and the time saved by reducing the number of expanded states. We present a novel method that reduces the cost of combining admissible heuristics for optimal search, while maintaining its benefits. Based on an idealized search space model, we formulate a decision rule for choosing the best heuristic to compute at each state. We then present an active online learning approach for that decision rule, and employ the learned model to decide which heuristic to compute at each state. We evaluate this technique empirically, and show that it substantially outperforms each of the individual heuristics that were used, as well as their regular maximum.

Adequate representation of natural language semantics requires access to vast amounts of common sense and domain-specific world knowledge. Prior work in the field was based on purely statistical techniques that did not make use of background knowledge, on limited lexicographic knowledge bases such as WordNet, or on huge manual efforts such as the CYC project. Here we propose a novel method, called Explicit Semantic Analysis (ESA), for fine-grained semantic interpretation of unrestricted natural language texts. Our method represents meaning in a high-dimensional space of concepts derived from Wikipedia, the largest encyclopedia in existence. We explicitly represent the meaning of any text in terms of Wikipedia-based concepts. We evaluate the effectiveness of our method on text categorization and on computing the degree of semantic relatedness between fragments of natural language text. Using ESA results in significant improvements over the previous state of the art in both tasks. Importantly, due to the use of natural concepts, the ESA model is easy to explain to human users.

The novel task we aim at in this work is to predict top terms that will prominently appear in the future news. This is a difficult task that nobody attempted before, as far as we know. We present a novel methodology for using patterns of user queries to predict future events. Query history is obtained from Web resources such as Google trends. In order to predict whether a term will appear in tomorrow's news, we examine if the terms in today's queries indicated this term in the past. We provide empirical support for the effectiveness of our method by showing its prediction power on news archives.

Machine learning techniques are gaining prevalence in the production of a wide range of classifiers for complex real-world applications with nonuniform testing and misclassification costs. The increasing complexity of these applications poses a real challenge to resource management during learning and classification. In this work we introduce ACT (anytime cost-sensitive tree learner), a novel framework for operating in such complex environments. ACT is an anytime algorithm that allows learning time to be increased in return for lower classification costs. It builds a tree top-down and exploits additional time resources to obtain better estimations for the utility of the different candidate splits. Using sampling techniques, ACT approximates the cost of the subtree under each candidate split and favors the one with a minimal cost. As a stochastic algorithm, ACT is expected to be able to escape local minima, into which greedy methods may be trapped. Experiments with a variety of datasets were conducted to compare ACT to the state-of-the-art cost-sensitive tree learners. The results show that for the majority of domains ACT produces significantly less costly trees. ACT also exhibits good anytime behavior with diminishing returns.

Machine learning techniques are increasingly being used to produce a wide-range of classifiers for complex real-world applications that involve nonuniform testing costs and misclassification costs. As the complexity of these applications grows, the management of resources during the learning and classification processes be- comes a challenging task. In this work we introduce ACT (Anytime Cost-sensitive Trees), a novel framework for operating in such environments. ACT is an anytime algorithm that allows trading computation time for lower classification costs. It builds a tree top-down and exploits additional time resources to obtain better esti- mations for the utility of the different candidate splits. Using sampling techniques ACT approximates for each candidate split the cost of the subtree under it and fa- vors the one with a minimal cost. Due to its stochastic nature ACT is expected to be able to escape local minima, into which greedy methods may be trapped. Ex- periments with a variety of datasets were conducted to compare the performance of ACT to that of the state of the art cost-sensitive tree learners. The results show that for most domains ACT produces trees of significantly lower costs. ACT is also shown to exhibit good anytime behavior with diminishing returns.

This paper presents a new framework for anytime heuristic search where the task is to achieve as many goals as possible within the allocated resources. We show the inadequacy of traditional distance-estimation heuristics for tasks of this type and present alternative heuristics that are more appropriate for multiple-goal search. In particular, we introduce the marginal-utility heuristic, which estimates the cost and the benefit of exploring a subtree below a search node. We developed two methods for online learning of the marginal-utility heuristic. One is based on local similarity of the partial marginal utility of sibling nodes, and the other generalizes marginal-utility over the state feature space. We apply our adaptive and non-adaptive multiple-goal search algorithms to several problems, including focused crawling, and show their superiority over existing methods.

The majority of the existing algorithms for learning decision trees are greedy---a tree is induced top-down, making locally optimal decisions at each node. In most cases, however, the constructed tree is not globally optimal. Furthermore, the greedy algorithms require a fixed amount of time and are not able to generate a better tree if additional time is available. To overcome this problem, we present two lookahead-based algorithms for anytime induction of decision trees, thus allowing tradeoff between tree quality and learning time. The first one is depth-k lookahead, where a larger time allocation permits larger k. The second algorithm uses a novel strategy for evaluating candidate splits; a stochastic version of ID3 is repeatedly invoked to estimate the size of the tree in which each split results, and the one that minimizes the expected size is preferred. Experimental results indicate that for several hard concepts, our proposed approach exhibits good anytime behavior and yields significantly better decision trees when more time is available.

Text categorization algorithms usually represent documents as bags of words and consequently have to deal with huge numbers of features. Most previous studies found that the majority of these features are relevant for classification, and that the performance of text categorization with support vector machines peaks when no feature selection is performed. We describe a class of text categorization problems that are characterized with many redundant features. Even though most of these features are relevant, the underlying concepts can be concisely captured using only a few features, while keeping all of them has substantially detrimental effect on categorization accuracy. We develop a novel measure that captures feature redundancy, and use it to analyze a large collection of datasets. We show that for problems plagued with numerous redundant features the performance of C4.5 is significantly superior to that of SVM, while aggressive feature selection allows SVM to beat C4.5 by a narrow margin.

Although text categorization is a burgeoning area of IR research, readily available test collections in this field are surprisingly scarce. We describe a methodology and system (named ACCIO) for automatically acquiring labeled datasets for text categorization from the World Wide Web, by capitalizing on the body of knowledge encoded in the structure of existing hierarchical directories such as the Open Directory. We define parameters of categories that make it possible to acquire numerous datasets with desired properties, which in turn allow better control over categorization experiments. In particular, we develop metrics that estimate the difficulty of a dataset by examining the host directory structure. These metrics are shown to be good predictors of categorization accuracy that can be achieved on a dataset, and serve as efficient heuristics for generating datasets subject to user's requirements. A large collection of automatically generated datasets are made available for other researchers to use.

The performance of anytime algorithms can be improved by simultaneously solving several instances of algorithm-problem pairs. These pairs may include different instances of a problem (such as starting from a different initial state), different algorithms (if several alternatives exist), or several runs of the same algorithm (for non-deterministic algorithms). In this paper we present a methodology for designing an optimal scheduling policy based on the statistical characteristics of the algorithms involved. We formally analyze the case where the processes share resources (a single-processor model), and provide an algorithm for optimal scheduling. We analyze, theoretically and empirically, the behavior of our scheduling algorithm for various distribution types. Finally, we present empirical results of applying our scheduling algorithm to the Latin Square problem.

The size and complexity of software and hardware systems have significantly increased in the past years. As a result, it is harder to guarantee their correct behavior. One of the most successful methods for automated verification of finite-state systems is model checking. Most of the current model-checking systems use binary decision diagrams (BDDs) for the representation of the tested model and in the verification process of its properties. Generally, BDDs allow a canonical compact representation of a boolean function (given an order of its variables). The more compact the BDD is, the better performance one gets from the verifier. However, finding an optimal order for a BDD is an NP-complete problem. Therefore, several heuristic methods based on expert knowledge have been developed for variable ordering. We propose an alternative approach in which the variable ordering algorithm gains 'ordering experience' from training models and uses the learned knowledge for finding good orders. Our methodology is based on offline learning of pair precedence classifiers from training models, that is, learning which variable pair permutation is more likely to lead to a good order. For each training model, a number of training sequences are evaluated. Every training model variable pair permutation is then tagged based on its performance on the evaluated orders. The tagged permutations are then passed through a feature extractor and are given as examples to a classifier creation algorithm. Given a model for which an order is requested, the ordering algorithm consults each precedence classifier and constructs a pair precedence table which is used to create the order. Our algorithm was integrated with SMV, which is one of the most widely used verification systems. Preliminary empirical evaluation of our methodology, using real benchmark models, shows performance that is better than random ordering and is competitive with existing algorithms that use expert knowledge. We believe that in sub-domains of models (alu, caches, etc.) our system will prove even more valuable. This is because it features the ability to learn sub-domain knowledge, something that no other ordering algorithm does.

One of the most common mechanisms used for speeding up problem solvers is macro-learning. Macros are sequences of basic operators acquired during problem solving. Macros are used by the problem solver as if they were basic operators. The major problem that macro-learning presents is the vast number of macros that are available for acquisition. Macros increase the branching factor of the search space and can severely degrade problem-solving efficiency. To make macro learning useful, a program must be selective in acquiring and utilizing macros. This paper describes a general method for selective acquisition of macros. Solvable training problems are generated in increasing order of difficulty. The only macros acquired are those that take the problem solver out of a local minimum to a better state. The utility of the method is demonstrated in several domains, including the domain of NxN sliding-tile puzzles. After learning on small puzzles, the system is able to efficiently solve puzzles of any size.

It is common to view programs as a combination of logic and control: the logic part defines what the program must do, the control part -- how to do it. The Logic Programming paradigm was developed with the intention of separating the logic from the control. Recently, extensive research has been conducted on automatic generation of control for logic programs. Only a few of these works considered the issue of automatic generation of control for improving the efficiency of logic programs. In this paper we present a novel algorithm for automatic finding of lowest-cost subgoal orderings. The algorithm works using the divide-and-conquer strategy. The given set of subgoals is partitioned into smaller sets, based on co-occurrence of free variables. The subsets are ordered recursively and merged, yielding a provably optimal order. We experimentally demonstrate the utility of the algorithm by testing it in several domains, and discuss the possibilities of its cooperation with other existing methods.

. Agents that operate in a multi-agent system need an efficient strategy to handle their encounters with other agents involved. Searching for an optimal interaction strategy is a hard problem because it depends mostly on the behaviour of the others. One way to deal with this problem is to endow the agents with the ability to adapt their strategies based on their interaction experience. This work views interaction as a repeated game and presents a general architecture for a model-based agent that learns models of the rival agents for exploitationin future encounters. First, we describe a method for inferring an optimal strategy against a given model of another agent. Second, we present an unsupervised algorithm that infers a model of the opponent's strategy from its interaction behaviour in the past. We then present a method for incorporating exploration strategies into model-based learning. We report experimental results demonstrating the superiority of the model-based learning agent over non-adaptive agents and over reinforcement-learning agents.

Human chess players exhibit a large variation in the amount of time they allocate for each move. Yet, the problem of devising resource allocation strategies for game playing has not received enough attention. In this paper we present a framework for studying resource allocation strategies. We define allocation strategy and identify three major types of strategies: static, semi‐dynamic, and dynamic. We then describe a method for learning semi‐dynamic strategies from self‐generated examples. We present an algorithm for assigning classes to the examples based on the utility of investing extra resources. The method was implemented in the domain of checkers, and experimental results show that it is able to learn strategies that improve game‐playing performance.

In recent years there is much interest in word cooccurrence relations, such as n-grams, verb-object combinations, or cooccurrence within a limited context. This paper discusses how to estimate the probability of cooccurrences that do not occur in the training data. We present a method that makes local analogies between each specific unobserved cooccurrence and other cooccurrences that contain similar words, as determined by an appropriate word similarity metric. Our evaluation suggests that this method performs better than existing smoothing methods, and may provide an alternative to class based models.