Robert E. Mercer

The Task Rehearsal Method of Life-Long Learning: Overcoming Impoverished Data

The task rehearsal method (TRM) is introduced as an approach to life-long learning that uses the representation of previously learned tasks as a source of inductive bias. This inductive bias enables TRM to generate more accurate hypotheses for new tasks that have small sets of training examples. TRM has a knowledge retention phase during which the neural network representation of a successfully learned task is stored in a domain knowledge database, and a knowledge recall and learning phase during which virtual examples of stored tasks are generated from the domain knowledge. The virtual examples are rehearsed as secondary tasks in parallel with the learning of a new (primary) task using the ?MTL neural network algorithm, a variant of multiple task learning (MTL). The results of experiments on three domains show that TRM is effective in retaining task knowledge in a representational form and transferring that knowledge in the form of virtual examples. TRM with ?MTL is shown to develop more accurate hypotheses for tasks that suffer from impoverished training sets.

Towards an Automated Citation Classifier

Described here is a first attempt to classify citations according to function in a fully automatic manner, that is, complete journal articles in electronic form are input to the citation classifier and a set of citations with their suggested function (chosen from a previously proposed scheme of functions) is output. The description consists of a brief introduction to the classification scheme, a description of the classifier, and a summary of the results of a test of the classifier on real data.

PLATYPUS: a platform for distributed answer set solving

We propose a model to manage the distributed computation of answer sets within a general framework. This design incorporates a variety of software and hardware architectures and allows its easy use with a diverse cadre of computational elements. Starting from a generic algorithmic scheme, we develop a platform for distributed answer set computation, describe its current state of implementation, and give some experimental results.

Minimal forward checking

Forward Checking (FC) is a highly regarded complete search algorithm used to solve constraint satisfaction problems. In this paper a lazy variant of FC called minimal forward checking (MFC) is introduced. MFC is a natural marriage of incremental FC and backchecking. Given a variable selection heuristic which does not depend on domain size MPCs worst case performance on any CSP instance is the number of constraint checks performed by FC. Experiments using hard random problems are presented which show that MFC outperforms FC especially for problems with large domain sizes and/or a large number of variables.<<ETX>>

Planning animation cinematography and shot structure to communicate theme and mood

Standard techniques, such as soundtrack recording, storyboarding and key-framing, are used to create animation adaptations of narratives. Many aspects of the narrative, such as moods, themes, character motivations and plot, must he captured in the audio-visual medium. Our work focusses on achieving the communication of moods and themes solely through the application of well-known cinematography techniques. We present a planning system that transforms a description of animator intentions and character actions into a series of camera shots which portray these intentions. The planner accomplishes this portrayal by utilizing lighting, framing, camera motion, colour choice and shot pacing. The final output is an animation that is intended to produce a viewer impression to support the animators description of the mood and theme of the narrative.

Classifying Spam Emails Using Text and Readability Features

Supervised machine learning methods for classifying spam emails are long-established. Most of these methods use either header-based or content-based features. Spammers, however, can bypass these methods easily-especially the ones that deal with header features. In this paper, we report a novel spam classification method that uses features based on email content-language and readability combined with the previously used content-based task features. The features are extracted from four benchmark datasets viz. CSDMC2010, Spam Assassin, Ling Spam, and Enron-Spam. We use five well-known algorithms to induce our spam classifiers: Random Forest (RF), BAGGING, ADABOOSTM1, Support Vector Machine (SVM), and Naïve Bayes (NB). We evaluate the classifier performances and find that BAGGING performs the best. Moreover, its performance surpasses that of a number of state-of-the-art methods proposed in previous studies. Although applied only to English language emails, the results indicate that our method may be an excellent means to classify spam emails in other languages, as well.

The Frequency of Hedging Cues in Citation Contexts in Scientific Writing

Citations in scientific writing fulfill an important role in creating relationships among mutually relevant articles within a research field. These inter-article relationships reinforce the argumentation structure that is intrinsic to all scientific writing. Therefore, determining the nature of the exact relationship between a citing and cited paper requires an understanding of the rhetorical relations within the argumentative context in which a citation is placed. To determine these relations automatically in scientific writing, we have suggested that stylistic and rhetorical cues will be significant. One type of cue that we have studied is the discourse cue, which provides cohesion among textual components. Another form of rhetorical cue involves hedging to modify the affect of a scientific claim. Hedging in scientific writing has been extensively studied by Hyland, including cataloging the pragmatic functions of the various types of cues. In this paper we show that the hedging cues proposed by Hyland occur more frequently in citation contexts than in the text as a whole. With this information we conjecture that hedging cues are an important aspect of the rhetorical relations found in citation contexts and that the pragmatics of hedges may help in determining the purpose of citations.

Identifying Explicit Discourse Connectives in Text

Explicit discourse relations in text are signalled by discourse connectives like since, because, however, etc. Identifying discourse connectives is a part of the bigger task called discourse parsing in which discourse coherence relations are extracted from text. In this paper we report improvements to the state-of-the-art for identifying explicit discourse connectives in the Penn Discourse Treebank and the Biomedical Discourse Relation Bank. These improvements have been achieved with maximum entropy (logistic regression) classifiers by combining machine learning features from previous approaches with new surface level features that capture information about a connective’s surrounding phrases and new syntactic features that add more information from the path in the syntax tree connecting the root to the connective and from the clause following the connective by means of its syntactic head.

Monotonic Answer Set Programming

Answer set programming (ASP) does not allow for incrementally constructing answer sets or locally validating constructions like proofs by only looking at a part of the given program. In this article, we elaborate upon an alternative approach to ASP that allows for incremental constructions. Our approach draws its basic intuitions from the area of default logics. We investigate the feasibility of the concept of semi-monotonicity known from default logics as a basis of incrementality. On the one hand, every logic program has at least one answer set in our alternative setting, which moreover can be constructed incrementally based on generating rules. On the other hand, the approach may produce answer sets lacking characteristic properties of standard answer sets, such as being a model of the given program. We show how integrity constraints can be used to re-establish such properties, even up to correspondence with standard answer sets. Furthermore, we develop an SLD-like proof procedure for our incremental approach to ASP, which allows for query-oriented computations. Also, we provide a characterization of our definition of answer sets via a modification of Clarks completion. Based on this notion of program completion, we present an algorithm for computing the answer sets of a logic program in our approach.

Using Hedges to Classify Citations in Scientific Articles

Citations in scientific writing fulfil an important role in creating relationships among mutually relevant articles within a research field. These inter-article relationships reinforce the argumentation structure intrinsic to all scientific writing. Therefore, determining the nature of the exact relationship between a citing and cited paper requires an understanding of the rhetorical relations within the argumentative context in which a citation is placed. To determine these relations automatically, we have suggested that various stylistic and rhetorical cues will be significant. One such cue that we are studying is the use of hedging to modify the affect of a scientific claim. We provide evidence that hedging occurs more frequently in citation contexts than in the text as a whole. With this information we conjecture that hedging is a significant aspect of the rhetorical structure of citation contexts and that the pragmatics of hedges may help in determining the rhetorical purpose of citations. A citation indexing tool for biomedical literature analysis is introduced.