Special Issue for the Natural Language Engineering Journal

Statistical Learning of Natural Language Structured Input and Output

Machine learning and statistical approaches have become indispensable for large part of Computational Linguistics and Natural Language Processing research. On one hand, they have enhanced systems' accuracy and have significantly sped-up some design phases, e.g. the inference phase. On the other hand, their use requires careful parameter tuning and, above all, engineering of machine-based representations of natural language phenomena, e.g. by means of features, which sometimes detach from the common sense interpretation of such phenomena.

These difficulties become more marked when the input/output data have a structured and relational form: the designer has both to engineer features for representing the system input, e.g. the syntactic parse tree of a sentence, and devise methods for generating the output, e.g. by building a set of classifiers, which provide boundaries and type (argument, function or concept type) of some of the parse-tree constituents.

Research in empirical Natural Language Processing has been tackling these complexities since the early work in the field, e.g. part-of-speech tagging is a problem in which the input --word sequences-- and output --POS-tag sequences-- are structured. However, the models initially designed were mainly based on local information. The use of such ad hoc solutions was mainly due to the lack of statistical and machine learning theory suggesting how models should be designed and trained for capturing dependencies among the items in the input/output structured data. In contrast, recent work in machine learning has provided several paradigms to globally represent and process such data: structural kernel methods, linear models for structure learning, graphical models, constrained conditional models, and re-ranking, among others.

However, none of the above approaches has been shown to be superior in general to the rest. A general expressivity-efficiency trade off is observed, making the best option usually task-dependant. Overall, the special issue is devoted to study engineering techniques for effectively using natural language structures in the input and in the output of typical computational linguistics applications. Therefore, the study on generalization of new or traditional methods, which allow for fast design in different or novel NLP tasks is one important aim of this special issue.

Finally, the special issue is also seeking for (partial) answers to the following questions:

• Is there any evidence (empirical or theoretical) that can establish the superiority of one class of learning algorithms/paradigms over the others when applied to some concrete natural language structures?
• When we use different classes of methods, e.g. SVMs vs CRFs, or different paradigms, what do we loose and what do we gain from a practical viewpoint (implementation, efficiency and accuracy)? This question is particularly interesting, when considering  different structure types: syntactic or semantic both shallow or deep.
• Can we empirically demonstrate that theoretically motivated algorithms, e.g. SVM-struct, improve simpler models, e.g. re-ranking, in the NLP case?
• Are there any other novel engineering approaches to NLP input and output structures?

Topics

For this special issue we invite submissions of papers describing novel and challenging work/results in theories, models, applications or empirical studies on statistical learning for natural language processing involving structured input and/or structured output. Therefore, the invited submission must concern with
(a) any kind of natural language problems; and (b) natural language structured data.

Assuming the target above, the range of topics to be covered will include, but will not be limited to the following:

• Practical and theoretical new learning approaches and architectures
  
• Experimental evaluation/comparison of different approaches
  
• Kernel Methods
  
• Algorithms for structure output (batch and on–line):
  – structured SVMs, Perceptron, etc.
  – on sequences, trees, graphs, etc.
  
• Bayesian Learning, Generative Models, Graphical Models 
  
• Relational Learning 
  
• Constraint Conditional models 
  
• Integer Linear Programming approaches 
  
• Graph-based algorithms 
  
• Ranking and Reranking 
  
• Scalability and effciency of ML methods 
  
• Robust approaches
  – noisy data, domain adaptation, small training sets, etc.
  
• Unsupervised and semi-supervised models
  
• Encoding of syntactic/semantic structures
  
• Structured data encoding deep semantic information and relations
  
• Relation between the syntactic and semantic layers in structured data
  

Important Dates

Call for papers:		30 November 2010
Submission of articles:		20 April 2011
Preliminary decisions to authors:		26 July 2011
Submission of revised articles:		28 Septmber 2011
Final decisions to authors:		8 December 2011
Final versions due from authors:		27 December 2011

Guest Editors

Lluís Màrquez
TALP Research Center, Technical University of Catalonia
lluism a t lsi d o t upc d o t edu

Alessandro Moschitti
Information Engineering and Computer Science Department, University of Trento
moschitti a t dit d o t unitn d o t it

Guest Editorial Board

Roberto Basili, University of Rome, Italy       
Ulf Brefeld, Yahoo-Reseatch, Spain
Razvan Bunescu, Ohio University, US           
Nicola Cancedda, Xerox, France               
Stephen Clark, University of Cambridge, UK       
Trevor Cohn, University of Sheffield, UK
Xavier Carreras, Universitat Politècnica de Catalunya, Spain       
Walter Daelemans, University of Antwerp, Belgium
Jason Eisner, John Hopkins University, US
Hal Daumé, University of Maryland, US   
James Henderson, University of Geneva, Switzerland   
Liang Huang, ISI, University of Southern California, US
Richard Johansson, University of Gothenburg, Sweden
Terry Koo, MIT CSAIL, US
Mirella Lapata, University of Edinburgh, UK       
Yuji Matsumoto, Nara Institute of Science and Technology, Japan   
Ryan McDonald, Microsoft Research, US
Raymon Money, University of Texas at Austin, US
Hwee Tou Ng, National University of Singapore, Singapore   
Sebastian Riedel, University of Massachusetts, US       
Dan Roth, University of Illinois at Urbana Champaign, US   
Mihai Surdeanu, Stanford University, US               
Ivan Titov, Saarland University, Germany           
Kristina Toutanova, Microsoft Research, US           
Jun'ichi Tsujii, University of Tokyo, Japan           
Antal van den Bosch, Tilburg University, The Netherlands   
Scott Yih, Microsoft Research, US               
Fabio Massimo Zanzotto, University of Rome, Italy       
Min Zhang, A-STAR, Singapore                   

Submission Instructions

Articles submitted to this special issue must adhere to the NLE journal guidelines 
(see section "Manuscript requirements" for the journal latex style).
We encourage authors to keep their submissions below 30 pages.
Send your manuscript in pdf attached to an email addressed to JNLE-SIO@disi.unitn.it
    - with subject filed: JNLE-SIO and
    - including names of the authors and title of the submission in the body

An alternative way to submit to JNLE-SIO is to submit a paper to TextGraph 6 and being selected
for contributing to JNLE. See the website:

http://www.textgraphs.org/ws11/index.html

The selected workshop papers must be extended to journal papers by following the indications of
both the TextGraph 6 reviewers and the JNLE-SIO editors. These upgraded versions have to be
submitted to JNLE-SIO no later than August 28, 2011 for the second round of review of JNLE-SIO.

Last Update: March 12, 2011