In this paper, we compare the performance of a state-of-the-art statistical parser (Bikel, 2004) in parsing written and spoken language and in generating subcategorization cues from written and spoken language. Although Bikel’s parser achieves a higher accuracy for parsing written language, it achieves a higher accuracy when extracting subcategorization cues from spoken language. Our experiments also show that current technology for extracting subcategorization frames initially designed for written texts works equally well for spoken language. Additionally, we explore the utility of punctuation in helping parsing and extraction of subcategorization cues. Our experiments show that punctuation is of little help in parsing spoken language and extracting subcategorization cues from spoken language. This indicates that there is no need to add punctuation in transcribing spoken corpora simply in order to help parsers. 1

In this paper we show that an unsupervised method for ranking word senses automatically can be used to identify infrequently occurring senses. We demonstrate this using a ranking of noun senses derived from the BNC and evaluating on the sense-tagged text available in both SemCor and the SENSEVAL-2 English all-words task. We show that the method does well at identifying senses that do not occur in a corpus, and that those that are erroneously filtered but do occur typically have a lower frequency than the other senses. This method should be useful for word sense disambiguation systems, allowing effort to be concentrated on more frequent senses; it may also be useful for other tasks such as lexical acquisition. Whilst the results on balanced corpora are promising, our chief motivation for the method is for application to domain specific text. For text within a particular domain many senses from a generic inventory will be rare, and possibly redundant. Since a large domain specific corpus of sense annotated data is not available, we evaluate our method on domain-specific corpora and demonstrate that sense types identified for removal are predominantly senses from outside the domain.

Verb classes are sets of semantically-related verbs sharing a range of linguistic properties, such as: — possible realizations of arguments — interpretation associated with each possible argument realization The big question: What is behind verb classes that on the one hand makes them so appealing as a research tool and on the other hand explains their limitations? Overview: Part I: The appeal and limitations of verb classes

In this paper we present a verb sense disambiguation technique which is based on statistical clustering models which merge verbs with similar subcategorisation and selectional preferences into a cluster. The sense of a verb is disambiguated by (i) extracting the verb and its argument heads with a statistical parser from a given sentence, (ii) labeling the extracted verb-argument tuple with one or more clusters according to the clustering model, and (iii) assigning the verb to one of its possible senses based on this cluster information. Using only the cluster IDs as features, we obtained an accuracy of 57.06% which is close to the results of the best system in the Senseval-2 competition which used far more information. We also show that a generalization of the selectional preferences in terms of WordNet concepts leads to better performance due to a reduction of sparse data problems. Keywords: probabilistic verb clustering; verb sense disambiguation; selectional preferences.

The work presented in this paper explores a supervised method for learning a probabilistic model of a lexicon of VerbNet classes. We intend for the probabilistic model to provide a probability distribution of verb-class associations, over known and unknown verbs, including polysemous words. In our approach, training instances are obtained from an existing lexicon and/or from an annotated corpus, while the features, which represent syntactic frames, semantic similarity, and selectional preferences, are extracted from unannotated corpora. Our model is evaluated in type-level verb classification tasks: we measure the prediction accuracy of VerbNet classes for unknown verbs, and also measure the dissimilarity between the learned and observed probability distributions. We empirically compare several settings for model learning, while we vary the use of features, source corpora for feature extraction, and disambiguated corpora. In the task of verb classification into all VerbNet classes, our best model achieved a 10.69 % error reduction in the classification accuracy, over the previously proposed model. 1

A multiword predicate is the combination of a predicate (often a verb) with one or more of its arguments, that together form a single unit of predicative meaning. We focus on a broad class of multiword predicates, in which a verb combines with a noun in the direct object position (e.g., give a groan and shoot the breeze). The semantic interpretation of such multiword predicates involves a certain degree of idiosyncrasy; moreover, they are crosslinguistically frequent and appear in all text genres. Hence, they pose a great challenge to the current models of natural language processing. Most existing computational models treat multiword predicates as syntactically-dependent word sequences or collocations. Such a treatment ignores other important characteristics of these constructions, reflected in their distinct lexical and syntactic behaviour. Nonetheless, cues from the lexicosyntactic properties of multiword predicates have often been used in linguistic and psycholinguistic studies to explain their peculiar semantic behaviour. On the one hand, simple statistical approaches that only draw on the frequency of multiword predicates fail to account for much of the syntactic and semantic behaviour of these constructions. On the other hand, linguistic theories provide generalizations about the behaviour of multiword predicates that can be augmented with probabilistic knowledge about language in use. The main goal of the present study is to propose ways of combining the predictive power of linguistic theories with the coverage and robustness of statistical techniques to acquire linguistically-plausible and reliable corpus-drawn knowledge about multiword predicates. ii Dedication For Reza, who started me on this path many years before.

Improved automatic text understanding requires detailed linguistic information about the words that comprise the text. Particularly crucial is the knowledge about predicates, typically verbs, which communicate both the event being expressed and how participants are related to the event. Although the field of natural language processing (NLP) has yet to develop a clear consensus on guidelines for building a verb lexicon suitable for applications in NLP, class-based construction of verb lexicons (e.g. Levin verb classification) with explicitly stated syntactic and semantic information has proved beneficial to a wide range of NLP tasks in combating the pervasive problem of data sparsity and increasing coverage. Such broad coverage dictionaries and ontologies are difficult and costly to create and maintain by hand, it is therefore desirable to learn them from distributional information, such as can be obtained from unlabeled or sparsely labeled text corpora. To this end, this thesis will primarily address the following three questions: First, deriving Levin-style verb classifications from text corpora helps avoid the expensive hand-coding of such information, but appropriate features must be identified and

Lapata and Brew (2004) (hereafter LB04) obtain from untagged texts a statistical prior model that is able to generate class preferences for ambiguous Levin (1993) verbs (hereafter Levin). They also show that their informative priors, incorporated into a Naive Bayes classifier deduced from hand-tagged data, can aid in verb class disambiguation. We reanalyse LB04’s prior model and show that a single factor (the joint probability of class and frame) determines the predominant class for a particular verb in a particular frame. This means that the prior model cannot be sensitive to fine-grained lexical distinctions between different individual verbs falling in the same class. We replicate LB04’s supervised disambiguation experiments on large scale data, using deep parsers rather than the shallow parser of LB04. In addition, we introduce a method for training our classifier without using hand-tagged data. This relies on knowledge of Levin class memberships to move information from unambiguous to ambiguous instances of each class. We regard this system as unsupervised because it does not rely on human annotation of individual verb instances. Although our unsupervised verb class disambiguator does not match the performance of the ones that make use of hand-tagged data, it consistently outperforms the random baseline model. Our experiments also demonstrate that the informative priors derived from untagged texts help improve the performance of the classifier trained on untagged data. 1

We present an unsupervised learning method for disambiguating verbs that belong to more than one Levin verb class (1993) when occurring in a particular syntactic frame. We used examples that contain unambiguous verbs in each verb class as the training data for ambiguous verbs in that class. A Naive Bayesian classifier was employed for the disambiguation task using context words as features. Our experiments suggest that our unsupervised learning method does not match the supervised one in disambiguating Levin verbs, but it consistently outperforms the random baseline model. 1

Abstract. We studied a special case for the translation of English verbs in verb-object pairs. Researchers have studied the effects of the linguistic information about the verbs being translated, and many have reported how considering the objects of the verbs will facilitate the quality of translations. In this study, we took an extreme venue – assuming the availability of the Chinese translations of the English objects. We explored the issue with thousands of samples that we extracted from 2011 NTCIR PatentMT workshop. The results indicated that, when the English verbs and objects were known, the information about the object’s Chinese translation could still improve the quality of the verb’s translations but not quite significantly.