Reading Guide: Advances in Soft Computing (MICAI 2021, Part II)

Advances in Soft Computing is the second of two volumes from MICAI 2021, the 20th Mexican International Conference on Artificial Intelligence, held in Mexico City in October 2021. Where the companion volume focuses on machine learning, vision and evolutionary algorithms, this one is about language and real-world applications: speech, text classification, machine translation, recommendation and robotics. If you are curious about applied NLP beyond the English-only mainstream, it is an especially rewarding read.

Get the book: View Advances in Soft Computing (MICAI 2021) on Amazon

What this book is (and who it's for)

The volume was edited by Ildar Batyrshin, Alexander Gelbukh and Grigori Sidorov and published by Springer in its Lecture Notes in Artificial Intelligence subseries (volume 13068). The full MICAI 2021 set contains 58 papers selected from 129 submissions; this second volume holds 28 of them across roughly 360 pages, organized into two sections: Natural Language Processing and Intelligent Applications and Robotics.

It is a good fit for:

• Graduate students in NLP or applied AI looking for project ideas and compact examples of experimental papers.
• Researchers and engineers working on low-resource languages, social-media text, speech or educational technology.
• Practitioners who want to see how transformers and classic methods were being applied to concrete problems in 2021.

You will be comfortable if you know the basics of machine learning (classification, evaluation metrics, train/test splits) and have heard of neural architectures such as RNNs, CNNs and transformers.

Before you start

A little background goes a long way. Refresh your memory on precision, recall and F1, on how word embeddings and pretrained language models work at a high level, and on what topic models like LDA do. If you want to reproduce results, a Python environment with common NLP libraries will let you experiment with similar pipelines on public datasets.

Keep a simple reading log: problem, data, method, metric, result. Many papers here are short and empirical, so a consistent log makes them easy to compare.

How to read it

Because this is a conference volume, read it by theme rather than in order. Here are suggested starting points for different readers, using papers listed in the volume's table of contents.

Track 1: Language understanding and social media

Start with Virality Prediction for News Tweets Using RoBERTa and Sentiment Analysis on Twitter About COVID-19 Vaccination in Mexico. Together they show how pretrained models and classic sentiment techniques handle noisy, short, real-world text. Follow with Hypernyms-Based Topic Discovery Using LDA to see how lexical knowledge can sharpen unsupervised topic models.

Track 2: Speech and multilingual NLP

Readers interested in speech should begin with Improving a Conversational Speech Recognition System Using Phonetic and Neural Transcript Correction and Text-Independent Speaker Identification Using Formants and Convolutional Neural Networks. For multilingual and low-resource work, read the comparative study of RNNs and transformers for Nahuatl neural machine translation, and Towards Multilingual Image Captioning Models that Can Read.

Track 3: Education and document analysis

Educators and EdTech builders should look at Supervised Machine Learning for Automatic Assessment of Free-Text Answers and Plagiarism Detection in Students' Answers Using FP-Growth Algorithm. STClass: A Method for Determining the Sensitivity of Documents is a useful bridge to information-security applications. For something unusual, there is a paper analyzing letter relationships in the Voynich Manuscript.

Track 4: Intelligent applications and robotics

Start with Question Answering for Visual Navigation in Human-Centered Environments, which links language with robot navigation. Then read A Tourist Recommendation System: A Study Case in Mexico and the comparative study on temporal-constrained scheduling under uncertainty for a taste of planning and recommendation work.

A first-pass strategy

Read the abstract and conclusion of every paper in your track, choose two or three for a full read, and skim the rest. Skip detailed hyperparameter tables on a first pass; come back to them only if you plan to replicate a result.

Key ideas to take away

• Pretrained transformers had become the default baseline for many text tasks, but classic methods such as LDA and association-rule mining still appear where interpretability or small data matter.
• Language diversity matters. Work on Spanish-language social media and on Nahuatl shows that progress in NLP depends on serving more than one language.
• Speech and text are converging. Correcting transcripts with neural models and identifying speakers with CNNs show how techniques move between modalities.
• Applied AI is about context. Recommendation, scheduling and navigation papers show that domain constraints often matter as much as the model.
• Interpretability is a recurring concern, especially for sensitive applications where decisions must be explained.

Practice and discussion questions

1. Pick one social-media paper. What makes tweets harder to analyze than news articles, and how did the authors deal with it?
2. Why might a team compare RNNs and transformers for a low-resource language instead of just choosing the newest architecture?
3. For automatic grading of free-text answers, what risks should an educator consider before deploying such a system?
4. How does adding hypernyms change what a topic model discovers? Can you think of a dataset where this would help?
5. Choose an application paper. Which assumptions about the environment or users would need to change to deploy it in your country?
6. Which paper would you most like to reproduce, and what data and tools would you need?
7. Where would interpretability be more valuable than a small gain in accuracy?

Pair it with

• The companion volume: MICAI 2021 Part I, Advances in Computational Intelligence, covers machine and deep learning, image processing and pattern recognition, and evolutionary and metaheuristic algorithms.
• Foundations in learning: Algorithms for Reinforcement Learning, also in our store, is a compact introduction to the learning methods behind many robotics and decision-making systems.
• Hands-on Python skills: practical Python and data-science skills make it much easier to experiment with the NLP pipelines described here. Browse the BooksPlusAi shelf for programming titles.

Final thoughts

This volume rewards selective, curious reading. Choose a track, read a handful of papers carefully, and you will come away with concrete ideas for projects in multilingual NLP, speech and applied AI, plus a better instinct for how research papers are built.

Start reading: Get Advances in Soft Computing (MICAI 2021, Part II) on Amazon

As an Amazon Associate, BooksPlusAi may earn from qualifying purchases.

Sources

• Springer: Advances in Soft Computing, MICAI 2021, Part II
• University of Wisconsin–Madison Libraries catalog record