Guest Editors: Marta R. Costa-jussÓ, Cristina Espa˝a i Bonet, Pascale Fung, Noah A. Smith
Semantic representations at different levels — word, sentence, paragraph — are central to many solutions to Natural Language Processing (NLP) tasks. Text annotation, information retrieval, sentiment analysis, text summarisation, question answering, and machine translation solutions have achieved significant improvements using semantic representations of words. The number of tasks has also grown with the popularisation of unsupervised word embeddings, a fast and efficient way of estimating continuous representations of words. Now, representations derived from deep learning methods are giving a new boost to the field.
But in an increasingly globalised world, multilingual and cross-language applications are needed. Extensions to monolingual representations of words such as multilingual Brown clusters, multilingual or interlingual word embeddings, multilingual topic models, and cross-lingual semantic parsers have been successfully applied, though performance lags behind their monolingual counterparts. On the other hand, multilinguality should be able to overcome one of the main limitations of standard representations: multiple senses of a word are conflated into a single vector. In addition, representations of sentences and paragraphs are arising through the encoder-decoder approach (in the fields of natural language inference, machine translation, and text summarization, among others) and the extension to multilingual inputs is equally or even more interesting for NLP applications.
Abstract semantic representations attract interest both in academia and industry, as shown by the wide and varied publications from universities and companies on this topic. Achieving improvements in this direction will change the conception of translation (from pairwise to interlingua), enable greater sharing among different NLP applications, and overcome low-resource limitations through zero-shot learning, among other advantages. This research direction will not only increase the performance of current architectures in cross-language settings, but will also lead to novel language-independent architectures and tasks for NLP.
This call aims to motivate research on multilingual and interlingual representations that, in the context of any NLP task, go beyond projections from one language into another one and that target crosslinguality. The special topic proposed aims to put together a compact and openly accessible volume, which presents high quality research works that cover an overview of this multidisciplinary field as well as most recent advances.
Topics include, but are not limited to:
Papers should be submitted according to the Computational Linguistics style: http://cljournal.org/. As in regular submissions to the journal, paper submissions should be made through the CL electronic submission system. In Step 1 of the submission process, please select 'Special Issue: Multilingual and Interlingual Semantic Representations' under the 'Journal Section' heading. Please note that papers submitted to a special issue undergo the same reviewing process as regular papers.