Structural correspondence learning for parse disambiguation

2000

In this paper we argue in favour of an integration between statistically and syntactically based parsing, where syntax is intended in terms of shallow parsing with elementary trees. None of the statistically based analyses produce an accuracy level comparable to the one obtained by means of linguistic rules . Of course their data are strictly referred to English, with the exception of [2, 3, 4]. As to Italian, purely statistically based approaches are inefficient basically due to great sparsity of tag distribution -50% or less of unambiguous tags when punctuation is subtracted from the total count as reported by . We shall discuss our general statistical and syntactic framework and then we shall report on an experiment with four different setups: the first two approaches are bottom-up driven, i.e. from local tag combinations: A. Statistics only tag disambiguation; B. Stastistics plus syntactic biases; C. Syntactic-driven disambiguation with no statistics; D. Syntactic-driven disambiguation with conditional probabilities computed on syntactic constituents. The second two approaches are top-down driven, i.e. driven from syntactic structural cues in terms of elementary trees: In a preliminary experiment we made with automatic tagger, we obtained 99% accuracy in the training set and 98% in the test set using combined approaches: data derived from statistical tagging is well below 95% even when referred to the training set, and the same applies to syntactic tagging.

2005

My thesis work would not have been possible without the help of my advisor, other collaborators, and fellow students. I am especially fortunate to have been advised by Chris Manning. Firstly, I am grateful to him for teaching me almost everything I know about doing research and being part of the academic community. Secondly, I deeply appreciate his constant support and advice on many levels. The work in this thesis was profoundly shaped by numerous insightful discussions with him. I am also very happy to have been able to collaborate with Andrew Ng on random walk models for word dependency distributions. It has been a source of inspiration to interact with someone having such far-reaching research goals and being an endless source of ideas. He also gave me valuable advice on multiple occasions. Many thanks to Dan Jurafsky for initially bringing semantic role labeling to my attention as an interesting domain my research would fit in, contributing useful ideas, and helping with my dissertation on a short notice. Thanks also to the other members of my thesis defense committee-Trevor Hastie and Francine Chen. I am also grateful to Dan Flickinger and Stephan Oepen for numerous discussions on my work in hpsg parsing. Being part of the NLP and the greater AI group at Stanford has been extremely stimulating and fun. I am very happy to have shared an office with Dan Klein and Roger Levy for several years, and with Bill McCartney for one year. Dan taught me, among other things, to always aim to win when entering a competition, and to understand things from first principles. Roger was an example of how to be an excellent researcher while maintaining a balanced life with many outside interests. vi I will miss the heated discussions about research and the fun at NLP lunch. And thanks to Jenny for the Quasi-Newton numerical optimization code. Thanks to Aria Haghighi for collaborating on semantic role labeling models and for staying up very early in the morning to finish writing up papers. I would also like to take the chance to express my gratitude to my undergraduate advisor Galia Angelova and my high-school math teacher Georgi Nikov. Although they did not contribute directly to the current effort, I wouldn't be writing this without them. I am indebted in many ways to Penka Markova-most importantly, for being my friend for many years and for teaching me optimism and self-confidence, and additionally, for collaborating with me on the hpsg parsing work. I will miss the foosball games and green tea breaks with Rajat Raina and Mayur Naik. Thanks also to Jason Townsend and Haiyan Liu for being my good friends. Thanks to Galen Andrew for making my last year at Stanford the happiest. Finally, many thanks to my parents Diana and Nikola, my sister Maria, and my nieces Iana and Diana, for their support, and for bringing me great peace and joy. I gratefully acknowledge the financial support through the ROSIE project funded by Scottish Enterprise under the Stanford-Edinburgh link programme and through ARDA's Advanced Question Answering for Intelligence (AQUAINT) program.

Arxiv preprint cs/9906006, 1999

Late in an evening in November 1993, I received a bizarre phone-call concerning a research position in a project on parsing natural language. I was told that the project is about resolving ambiguity, that it is for two years only (stressing that a PhD is not the goal) and that it pays better than being a PhD-student (an "immoral" approach :-)). It sounded like adventure because I had already met Remko Scha a couple of times the year before, when I was writing my Master's thesis on ambiguity. During one of these times I asked Remko "how do you people in natural language processing get rid of ambiguity from a natural language grammar", Remko answered tersely "we are not interested in making natural language grammars unambiguous". As a computer scientist I was puzzled; I felt that Computer Science is a "safer" place to be than those "ambiguous linguistic environments". In an interview for the job I also met Rens Bod and Steven Krauwer, who was the intended project leader. The week before the interview I had read the papers on DOP. Because I was told that there were no polynomial-time parsing algorithms for DOP, I sat down and designed such an algorithm. During the interview I explained some of the details of the algorithm, Remko and Steven were interested in seeing this written down first, Rens was surprised and did not believe it was possible. Despite of that, I was hired to develop a parser for DOP in a two year project called CLASK. Meanwhile, Remko and his group were involved in a national project ("OVIS") of the Netherlands organization for Scientific Research (NWO). The results of CLASK constituted my "visa" for joining "OVIS" for one year. After that, Remko and I decided that it is time to concentrate on writing a thesis; NWO and the Foundation for Language and Speech (STT) decided to support this proposal. This thesis exists thanks to various project proposals submitted together with Remko Scha. Without the support of Remko Scha (ILLC), Steven Krauwer (STT), Jan Landsbergen (OTS) and Alice Dijkstra (NWO), this thesis would have remained virtual. Our proposals would not have become projects without additional support from Loe Boves, Martin Everaart, Gertjan van Noord, Eric Reuland, and the STT-board. I am grateful to my promoters for the involvement and the supervision. They listened, discussed, read, commented and corrected always with so much patience. I am especially indebted to Christer Samuelsson and Remko Bonnema who read and commented on earlier versions of all chapters; in particular, Christer detected and suggested corrections to a serious error in the original paper that led to chapter 3. I thank also Ameen Abu-Hanna, ix Yaser Yacoob and Yoad Winter for reading and commenting on earlier versions. Apart from the aforementioned people, this thesis benefited from discussions with Erik Aarts,

2001

Structural ambiguity is one of the most difficult problems in natural language processing. Two disambiguation mechanisms for unrestricted text analysis are commonly used: lexical knowledge and context considerations. Our parsing method includes three different mechanisms to reveal syntactic struc-tures and an additional voting module to obtain the most probable structures for a sentence. The developed tools do not require any tagging or syntactic marking of texts.

1993

Inthis paper we will showt hat Grammatical Inference is applicable to Natural Language Processing. Given the wide and complexrange of structures appearing in an unrestricted Natural Language likeE nglish, full Grammatical Inference, yielding a comprehensive syntactic and semantic definition of English, is too much to hope for at present. Instead, we focus on techniques for dealing with ambiguity resolution by probabilistic ranking; this does not require a full formal Chomskyan grammar.W eg iv e as hort overviewo ft he different levels and methods being investigated at CCALAS for probabilistic ranking of candidates in ambiguous English input. Grammatical Inference from English corpora. An earlier title for this paper was "Overviewo fg rammar acquisition research at CCALAS, Leeds University", but this was modified to avoid the impression of an incoherent set of research strands with no integrated, focussed common techniques or applications. The researchers in our group have nod etailed development plan imposed 'from above', but are working on independent PhD programmes; however, there are common theoretical tennets, ideas, and potential applications linking individual projects. In fact, preparing for the Colloquia on Grammatical Inference has helped us to appreciate these overarching, linking themes, as we realised that the definitions stated in the Programme clearly applied to our own work at CCALAS: 'Grammatical Inference ... has suffered from the lack of a focused research community ... Simply stated, the grammatical inference problem is to learn an efficient description that captures the essence of a set of data. This description may be used subsequently to classify data, or to generate further examples of similar data.' The data in our case is unrestricted English input, as exemplified by a Corpus or large collection of text samples. This renders a much harder challenge to Grammatical Inference than artificial languages, or selected examples of wellformed English sentences. The range of lexical items and grammatical constructs appearing in an unrestricted English Corpus is very large; and the problem is not just one of scale. The Corpus-based approach carries with it a blurring of the classical Chomskyan distinction between 'grammatical' and 'ungrammatical' English sentences. Indeed, [Sampson 87] went to the extreme of positing that there is NO boundary between grammatical and ungrammatical sentences in English; this might seem to imply that it is hopeless and eveni nv alid to attempt to infer a grammar for English. Furthermore, the Corpus-based approach eschews the use of 'intuitively constructed' examples in training: a learning algorithm should be trained with 'real' sentences from a Corpus. It would seem to followf rom this that we are also proscribed from artificially constructing negative counterexamples for our learning algorithms: we cannot guarantee that such counterexamples are truly illegal.