Towards discourse meaning: complexity of dependencies at the discourse level and at the sentence level

Aravind K. Joshi
Department of Computer and Information Science and Institute for Research in Cognitive Science
University of Pennsylvania Philadelphia PA, USA

DATE: May 13 2010, 12.30 p.m.
LOCATION: Room 20 – Faculty of Science – Povo

ABSTRACT:
The overall goal will be to discuss some issues concerning the dependencies at the discourse level and at the sentence level. However, first I will briefly describe the Penn Discourse Treebank (PDTB), a corpus with annotations for the discourse connectives (explicit, implicit, and alternate lexicalizations (AltLex)) and their arguments together with “attributions” of the arguments and the relations denoted by the connectives, and also the senses of the connectives. I will then focus on certain aspects of these dependencies such as (a) the complexity of the dependencies (b) attributions and their relationship to the dependencies, and (c) alternate lexicalizations of connectives (AltLex), among others. I will also discuss the implications of many of these issues for the representation of sentence structure itself, perhaps shedding some light on the transition from the sentence level to the level of immediate discourse.

BIOGRAPHY:
Joshi is the Henry K. Salvatore Professor of Computer and Cognitive Science at the University of Pennsylvania. He has previously received considerable recognition for his accomplishments. Three of his honors are particularly worthy of note. In 1997, he was the recipient of the highest honor in the field of artificial intelligence, the Research Excellence Award of the International Joint Conference of Artificial Intelligence (IJCAI), a distinction held by only eight other outstanding computer scientists. In 1999, he was appointed to the National Academy of Engineering, the only researcher in Natural Language Processing to have ever recieved this distinction. And just this year, Joshi was chosen to be the first recipient of the Lifetime Achievement Award given by the Association for Computational Linguistics. Joshi has contributed a number of key ideas to the formal science of language. Perhaps the best known of these is Tree Adjoining Grammar. His work on TAG has played an important role in both natural language processing and in theoretical linguistics. In both disciplines, it stands as a monument to the value of principled mathematical thinking. Two key ideas underlying TAG are, first, that the statement of local syntactic and semantic dependencies can be factored apart from recursion and, second, that a modest increase in power beyond context-free grammar is sufficient to characterize natural language syntax. The TAG adjoining operation, as defined by Joshi, achieves both of these results in a strikingly elegant way, providing a powerful tool for linguistic description that at the same time yields grammars guaranteed to be computationally tractable. A large body of mathematics, computational, empirical linguistic, and psycholinguisitc work by Joshi and numerous others has been developing the consequences of Joshi’s original insight for more than a quarter of century. Joshi’s work in mathematical linguistics over the years has had an extraordinary impact on linguistic theory, beyond the impact of TAGs themselves. To give just two examples here: (a) Joshi’s generalization of an earlier result of Stan Peters’ to show that arbitrary booleans of context sensitive filters on context free grammars still result in context free languages led directly to the development of Gerald Gazdar’s GPSG framework (actually first developed, we believe, while Gazdar was visiting Penn). (b) the generalization of TAGs to an entire class of languages (the so-called “Mildly Context Sensitive Languages”) provided a natural way to relate a number of superficially distinct linguistic theories from Combinatory Categorial Grammar to Head Grammar and HPSG to Government-Binding Theory and Minimalism. Another key contribution of Joshi’s to the science of language (along with Weinstein and Grosz) is Centering Theory, a computationally tractable model of attention during discourse. Centering Theory has attracted a wide following among linguists and computer scientists working on formal models of discourse. Its leading idea is that referring expressions can be ranked on the basis of various structural properties and that these rankings can predict the likely coreferents of anaphoric expressions in discourse. These predictions can be used in the automatic processing of discourse but they also have a fine-grained structure, so that the theory can be used to show how different choices of coreference produce different pragmatic effects. The theory is attractive in part because it provides a framework for capturing not only the relationship of a current utterance to previous utterances but also with expectations regarding utterances yet to come. Perhaps most strikingly, it has yielded the first successful objective definition of the notion of “topic” or “theme”, a concept long thought important by linguists but notoriously difficult to nail down. Centering Theory has been found relevant for modeling a number of properties related to discourse coherence, including anaphora resolution, the distribution of various types of pronouns, the felicity condidtions of marked syntactic forms, and aspects of prosody. Aside from his scientific work, Joshi has played a key organizational role in fostering the development of the new discipline of cognitive science. Over the past two decades and more, the University of Pennsylvania has developed a thriving program in cognitive science, largely due to the outstanding vision and tireless leadership that Joshi contributed to the effort. From his graduate student days onward, he was concerned with the interface between computation and cognition, working with early pioneers like Zellig Harris and Saul Gorn. By the late 1970′s Joshi had established an interdisciplinary faculty seminar that included psychologists and linguists, as well as computer scientists. This seminar was one of the early recipients of support from the Sloan Foundation’s cognitive science initiative. Later he led the effort at Penn to win an NSF Science and Technology Center for cognitive science and was founding co-director (with Lila Gleitman) of the Institute for Cognitive Science at Penn, a post which he held with great success until last year. His approach to these efforts was always to foster the broadest possible participation by researchers in different domains and with different orientations and always to emphasize the importance of educating young researchers and of supporting them morally and materially. IRCS is one of a small number of organizations at Penn that cross school lines and establishing it required great persistence and diplomatic skill. These were supplied by Joshi, whose commitment to a broad view of the field had convinced him that the effort was necessary.

CONTACT: Giuseppe Riccardi
giuseppe[DOT]riccardi[AT]unitn[DOT]it