Natural language processing, a sub-discipline of artificial intelligence has received much attention in recent years. There is a paradigm shift in natural language processing with the introduction of deep-learning based techniques. Extracting features manually from text data is difficult and time consuming. Supervised deep learning-based techniques are well-known for their automatic feature extraction capabilities. But these systems suffer from interpretability issues and high computational complexity. Moreover in order to train the systems a huge amount of labeled data is required. NLP includes many research problems like summarization, dialogue systems, machine translation, question answering where generating labeled data is a challenge.
Nowadays, there is a trend in developing unsupervised deep learning-based architectures. Most of the research articles are only based on utilizing deep learning. Researchers are also developing unsupervised optimization techniques that are widely used in solving different problems of real-life. Evolutionary algorithms (EAs) are a type of meta-heuristic optimization technique which are well-known for determining near-optimal solutions in a limited amount of time. They are based on the concepts of the natural evolution process. In the recent world, a lot of research is also going on developing new evolutionary algorithms (EAs) and their applications in various domains.
EAs have received considerable attention from academics, researchers, and domain workers in solving the problems of NLP because these algorithms can be used without availability of large training data as opposed to deep learning techniques. In solving different problems in the domain of NLP, there is a requirement of simultaneously optimizing several objective functions. For example in summarization systems, conflicting objectives like coverage, anti-redundancy, readability, cohesion etc. are required to be simultaneously optimized by the search capability of some optimization techniques. Concepts of multiobjective optimization are useful in such cases where multiple conflicting objective functions are simultaneously optimized. Moreover, for determining the appropriate parameters/architectures for deep learning based models for solving different NLP problems, optimization based approaches are widely used. Feature selection, classifier ensemble based approaches can also be successfully solved using optimization frameworks. Several quality measures like accuracy, precision, recall, F-score can be simultaneously optimized for selecting the best combination of classifiers for the purpose of ensemble. Similarly in the domain of feature selection, filter and wrapper based approaches can be designed by simultaneously optimizing several feature quality measures by utilizing the search capability of EAs. But there are still many open problems which are related to applications of EAs in solving NLP problems: 1) stability of EAs in solving different real-life
problems; in general EAs are randomized algorithms. Their performance varies from generation to generation. How to generate stable solutions using EAs? 2) EAs produce several solutions on the final population. How to select a single solution for reporting? 3) As the researchers are also investigating the effect of adding multimodal information in solving different NLP problems, in this scenario, how to develop an evolutionary framework handling multimodal NLP data in an efficient way? 4) How to adopt/develop an evolutionary framework when it is applied on multi/cross-lingual data?? Answering these questions is a prerequisite for widespread deployment of evolutionary algorithms in NLP application.
This special session aims to bring together the current research progress on developing new EAs for solving different NLP problems. The articles demonstrating the applications of existing EAs in solving various real-life problems are also welcome. We however do NOT encourage the submission that only focuses on new theories and algorithms, without demonstrating their application on NLP data. High quality articles based on unsupervised deep-learning techniques are also welcome. As per our knowledge, there is no previous special session held anywhere as most of the NLP community focuses on using deep learning-based methods.
Articles focusing on the following topics (but not limited to) related to applications of EAs in solving different problems of NLP are invited for this special session.
1) Architecture selection for deep learning based techniques using evolutionary algorithms 2) Review paper showing comparison between deep and evolutionary techniques
3) Computer aided Machine Translation 4) Evolutionary Computing for mono-lingual and cross-lingual NLP task 5) Summarization 6) Natural language inference 7) Evolutionary algorithms for Textual entailment 8) Text Classification/Clustering 9) Entity linking 10) Named Entities Recognition
11) Knowledge extraction and information analysis 12) Natural language generation 13) Dialogue management 14) Slogan generation 15) Visual Question-Answering 16) Sentiment Analysis
Because of the wide scope of NLP, some important topics that fit in the scope of the special session may not be listed above. Therefore, if you are unsure whether your work would fit, we encourage you to get in touch with any organizer. All papers must comply with the basic requirements of IEEE SSCI 2021. The review process will comply with the standard review process of the IEEE SSCI. Each paper will receive at least three reviews from experts in the field. As per our knowledge, there is no previous special session held anywhere as most of the NLP community focuses on using deep learning-based methods.
Organizers: Sriparna Saha, Naveen Saini, Jose G Moreno