We present a novel method for clustering syntactic structures in raw text. The method involves first generating an ensemble of clas-sifiers, by training a statistical parser on samples from the training data. Then, the clas-sifier outputs for each instance are used as input for clustering. The resulting clusters group together instances whose internal representations in the statistical model are similar , and therefore depend on the feature set it uses. We apply our method to the simple task of clustering noun phrases. Experiments show that most encouraging results are obtained when the training samples are small. Possible applications of the method include speeding up error analysis of a supervised system, and adapting a supervised system to a new domain.

We conduct large-scale experiments to investigate optimal features for classification of verbs in biomedical texts. We introduce a range of feature sets and associated extraction techniques, and evaluate them thoroughly using a robust method new to the task: cost-based framework for pairwise clustering. Our best results compare favourably with earlier ones. Interestingly, they are obtained with sophisticated feature sets which include lexical and semantic information about selectional preferences of verbs. The latter are acquired automatically from corpus data using a fully unsupervised method.

Previous research has shown that syntactic features are the most informative features in automatic verb classification. We investigate their optimal characteristics by comparing a range of feature sets extracted from data where the proportion of verbal arguments and adjuncts is controlled. The data are obtained from different versions of VALEX [1] - a large SCF lexicon for English which was acquired automatically from several corpora and theWeb.We evaluate the feature sets thoroughly using four supervised classifiers and one unsupervised method. The best performing feature set includes rich syntactic information about both arguments and adjuncts of verbs. When combined with our best performing classifier (a novel Gaussian classifier), it yields the promising accuracy of 64.2% in classifying 204 verbs to 17 Levin (1993) classes. We discuss the impact of our results on the state-or-art and propose avenues for future work.

Previous research has shown that syntactic features are the most informative features in automatic verb classification. We experiment with a new, rich feature set, extracted from a large automatically acquired subcategorisation lexicon for English, which incorporates information about arguments as well as adjuncts. We evaluate this feature set using a set of supervised classifiers, most of which are new to the task. The best classifier (based on Maximum Entropy) yields the promising accuracy of 60.1% in classifying 204 verbs to 17 Levin (1993) classes. We discuss the impact of this result on the stateof-art, and propose avenues for future work.

We present a methodology for error analysis in entity annotation. To increase the accuracy in corpora, there is a need for an analysis method for detecting human annotation and schema errors. We use easiness statistics and information gain to gain insights into possible causes of error in the GENIA corpus of MEDLINE abstracts.

Lexical classes, when tailored to the application and domain in question, can provide an effective means to deal with a number of natural language processing (NLP) tasks. While manual construction of such classes is difficult, recent research shows that it is possible to automatically induce verb classes from cross-domain corpora with promising accuracy. We report a novel experiment where similar technology is applied to the important, challenging domain of biomedicine. We show that the resulting classification, acquired from a corpus of biomedical journal articles, is highly accurate and strongly domain-specific. It can be used to aid BIO-NLP directly or as useful material for investigating the syntax and semantics of verbs in biomedical texts.

We introduce a large computational subcategorization lexicon which includes subcategorization frame (SCF) and frequency information for 6,397 English verbs. This extensive lexicon was acquired automatically from five corpora and the Web using the current version of the comprehensive subcategorization acquisition system of Briscoe and Carroll (1997). The lexicon is provided freely for research use, along with a script which can be used to filter and build sub-lexicons suited for different natural language processing (NLP) purposes. Documentation is also provided which explains each sub-lexicon option and evaluates its accuracy.

We describe a system for BioCreative Task 2.1: finding evidence that suppports a GO term annotation for a given protein in a given biomedical paper. We approach the problem as a question answering task, where the query is constructed from a protein name, a GO term and its definition.

Previous research has demonstrated the utility of clustering in inducing semantic verb classes from undisambiguated corpus data. We describe a new approach which involves clustering subcategorization frame (SCF) distributions using the Information Bottleneck and nearest neighbour methods. In contrast to previous work, we particularly focus on clustering polysemic verbs. A novel evaluation scheme is proposed which accounts for the effect of polysemy on the clusters, offering us a good insight into the potential and limitations of semantically classifying undisambiguated SCF data.

Some statistical learning systems are evaluated using measures of distributional similarity. To deal with the problem of zero events in the distributions under comparison, smoothing is frequently performed before similarity measures are applied. Smoothing alters the information in the original distribution, and may add noise to the results. Here, we investigate the sensitivity of entropy-based similarity measures to noise from uninformative smoothing. Our experiments with two subcategorization acquisition systems show that similarity measures vary in their robustness. While some are led astray by noise from smoothing, others are more resilient.

Error analysis is a key step in developing statistical parsers. In doing this, we manually discover typical cases by examining parser output. In this paper we argue that the process can be speeded up by considering the output from an ensemble of parsers. We do this by resampling small proportions (10% and up) from the training data, and exploiting the high diversity of the resulting parsers - resulting from the sparseness of natural-language data. Varying the sample size, we can trace the gradual learning of each instance and classify instances into a few types. This division helps in distinguishing instances which are hard for the system, from instances which may be learned in principle. We suggest that such analysis can yield a qualitative approach to evaluation of statistical parsers.

Statistical NLP systems are frequently evaluated and compared on the basis of their performances on a single split of training and test data. Results obtained using a single split are, however, subject to sampling noise. In this paper we argue in favour of reporting a distribution of performance figures, obtained by resampling the training data, rather than a single number. The additional information from distributions can be used to make statistically quantified statements about differences across parameter settings, systems, and corpora.

We use seven machine learning algorithms for one task: identifying base noun phrases. The results have been processed by different system combination methods and all of these outperformed the best individual result. We have applied the seven learners with the best combinator, a majority vote of the top five systems, to a standard data set and managed to improve the best published result for this data set.

In this paper, a memory-based parsing method is extended for handling compositional structures. The method is oriented for learning to parse any selected subset of target syntactic structures. It is local, yet can handle also compositional structures. Parts of speech as well as embedded instances are being used simultaneously. The output is a partial parse in which instances of the target structures are marked.

Recognizing shallow linguistic patterns, such as basic syntactic relationships between words, is a common task in applied natural language and text processing. The common practice for approaching this task is by tedious manual definition of possible pattern structures, often in the form of regular expressions or finite automata. This paper presents a novel memory-based learning method that recognizes shallow patterns in new text based on a bracketed training corpus. The training data are stored as-is, in efficient suffix-tree data structures. Generalization is performed on-line at recognition time by comparing subsequences of the new text to positive and negative evidence in the corpus. This way, no information in the training is lost, as can happen in other learning systems that construct a single generalized model at the time of training. The paper presents experimental results for recognizing noun phrase, subject-verb and verb-object patterns in English. Since the learning approach enables easy porting to new domains, we plan to apply it to syntactic patterns in other languages and to sub-language patterns for information extraction.

Incorporating external information during a learning process is expected to improve its efficiency. We study a method for incorporating noun-class information, in the context of learning to resolve Prepositional Phrase Attachment (PPA) disambiguation. This is done within a recently introduced architecture, SNOW, a sparse network of threshold gates utilizing the Winnow learning algorithm. That architecture has already been demonstrated to perform remarkably well on a number of natural language learning tasks. The knowledge sources used were compiled from the WordNet database for general linguistic purposes, irrespective of the PPA problem, and are being incorporated into the learning algorithm by enriching its feature space. We study two strategies of using enriched features and the effects of using class information at different granularities, as well as randomly-generated knowledge which serves as a control set. Incorporating external knowledge sources within SNOW yields a statistically significant performance improvement. In addition, we find an interesting relation between the granularity of the knowledge sources used and the magnitude of the improvement. The encouraging results with noun-class data provide a motivation for carrying out more work on generating better linguistic knowledge sources. 1 I n t r o d u c t i o n A variety of inductive learning techniques have been used in recent years in natural language processing. Given a large training corpus as input and relying on statistical properties of language usage, statistics-based and machine learning algorithms are used to induce a classifier which can be used to resolve a disambiguation task. Applications of this line of research include ambiguity resolution at different levels of sentence analysis: part-of speech tagging, word-sense disambiguation, word selection in machine translation, context-sensitive spelling correction, word selection in speech recognition, and identification of discourse markers. Many natural language inferences, however, seem to rely heavily on semantic and pragmatic knowledge about the world and the language, that is not explicit in the training data. The ability to incorporate knowledge from other sources of information, be it knowledge that is acquired across modalities: prepared by a teacher or by an expert, is crucial for going beyond low level natural language inferences. Within Machine Learning, the use of knowledge is often limited to that of constraining the hypothesis space (either before learning or by probabilistically biasing the search for the hypothesis) or to techniques such as EBL (DeJong, 1981; Mitchell et al., 1986; DeJong and Mooney, 1986) which rely on explicit domain knowledge that can be used to explain (usually, prove deductively) the observed examples. The knowledge needed to perform languageunderstanding related tasks, however, does not exist in any explicit form that is amenable to techniques of this sort, and many believe that it will never be available in such explicit forms. An enormous amount of useful "knowledge" may be available, though. Pieces of information that may be found valuable in language-understanding related tasks may include: the root form of a verb; a list of nouns that are in some relation (e.g., are all countries) and can thus appear in similar contexts; a list of verbs that can be followed by a food item; a list of items you can see through, things that are furniture, a list of dangerous things, etc. This rich collection of information pieces does not form any domain theory to speak of and cannot be acquired from a single source of information. This knowledge is noisy, incomplete and ambiguous. While some of it may be acquired from text, a lot if it may only be acquired from other modalities, as those used by humans. We believe that integration of such knowledge is essential for NLP to attain high-level natural-language inference. Contrary to this intuition, experiments in text retrieval and natural language have not shown much improvement when incorporating information of the kind humans seem to use (Krovetz and Croft, 1992; Kosmynin and Davidson, 1996; Kar0v and Edelman,

We consider the high eccentricity, 0.63, of the newly discovered planet orbiting 16 Cyg B, using the fact that the parent star is part of a wide binary. We show that the high eccentricity of the planet could be the result of tidal forces exerted on 16 Cyg B and its planet by 16 Cyg A, the distant companion in the system. By following numerically stellar triple systems with parameters similar to those of 16 Cyg, we have established that the orbital eccentricity of the planet could have gone through strong modulation, with an amplitude of 0.8 or even larger, with typical timescale of tens of millions of years.The amplitude of the modulation of the orbital eccentricity strongly depends on the relative inclination between the plane of motion of the planet and that of the wide binary 16 Cyg AB. To account for the present eccentricity of the planet, we have to assume that the angle between the two planes of motion is large, at least 60 degrees. We argue that this assumption is not unreasonable for wide binaries like 16 Cyg AB.

We report the discovery of the spectroscopic triple system G38-13. Its outer orbit has one of the shortest known periods, 645 d, for a late-type triple system. The shortness of the outer period, and the small ratio of the periods of the outer and the inner orbits, 28: 1, make this system very interesting for the study of long-term modulations of the orbital elements of triple systems.

Several binary stars detected by the Center for Astrophysics (CfA) radial-velocity surveys were found to be members of triple systems. We present two examples, each requires a different analysis to discover its multiplicity.One example is G176-46, a double-lined halo star of the Carney & Latham (1987) high proper-motion survey. The secondary star (G176-46b) displays large radial velocity variations, in contrast with the primary (G176-46a), which is constant within the error limits. Figure 1 shows two cross correlations of the stellar spectra against the same calculated template taken at different times, which indicate that only the secondary’s peak changes its position. A similar variation was observed previously for ADS 8811 (Mazeh & Latham 1988).We have found the secondary radial velocity to vary with a period of 10.44 days, and therefore conclude that Gl76-46b is a member of a short-period binary system. The orbital solution has an amplitude of 38 km s−1 and eccentricity of 0.05.