These are the most exciting AI research articles published this year. It integrates modern artificial intelligence (AI) and data science developments. It is chronologically ordered, and there is a link to a more comprehensive article.

Sentiment and Emotion-aware Multi-modal Complaint Identification

Complaining is a way for consumers to let an organization, product, or event know they don't like it. Customers usually write reviews about the products or services they buy on retail websites and different social media sites. These reviews may include complaints about the products or services. Organizations and online merchants need to be able to automatically find out when customers have problems with the products or services they buy. This information can help them meet customer needs, such as handling and addressing complaints. 

Images posted with reviews can help people figure out what's wrong, which shows how important it is to get input from many different sources. Also, the customer's feelings significantly affect how they express their complaint since emotions affect all speech acts. So, it's also essential to look into how emotion and mood affect automatic complaint identification. One of the critical parts of this work is the creation of a new dataset called Complaint, Emotion, and Sentiment Annotated Multi-modal Amazon Reviews Dataset (CESAMARD), a collection of reviews and images of products posted on Amazon's website. 

To show how useful the multi-modal dataset is, the researchers present an attention-based multi-modal, adversarial multi-task deep neural network model for complaint detection. The results of experiments show that the multi-modality and multi-tasking complaint identification methods work better than the single-task and single-modality methods.

TIGGER: Scalable Generative Modelling for Temporal Interaction Graphs

Recently, there has been a rise in the number of people who want to learn how to make graphs. Work on modelling static graphs has come a long way, but work on modelling temporal graphs is still in its early stages and has a lot of room for improvement:

  1. Existing generative models don't scale with the length of time or the number of nodes.
  2. The ways things are done now are transductive, which makes it hard to share knowledge.
  3. Existing models leak node identity information and can't change the size of the source graph because they depend on a one-to-one mapping between the source and the generated graph.

The researchers use a new TIGGER model to fill these gaps in this paper. TIGGER gets its power from a mix of temporal point processes and auto-regressive modelling. Researchers did a lot of tests on real datasets to show that TIGGER makes more accurate graphs and up to 3 orders of magnitude faster than the current best method.

Towards Building ASR Systems for the Next Billion Users

Recent speech and language technology methods pretrain huge, fine-tuned models for specific tasks. But the benefits of such big models are often only good for a few languages with many resources. In this work, the researchers make several contributions to building ASR systems for low-resource languages from the Indian subcontinent. 

  • First, they collect 17,000 hours of raw speech data in 40 Indian languages from various fields, such as education, news, technology, and finance. 
  • Second, the researchers used this raw speech data to train different wav2vec style models for 40 Indian languages. 
  • Third, they look at the already trained models to find critical features. 

For example, codebook vectors of phonemes that sound similar are shared across languages, representations across layers differ depending on the language family, and attention heads often focus on small local windows. Fourth, the researchers fine-tune this model for downstream ASR for nine languages and get state-of-the-art results on three public datasets, even for languages like Sinhala and Nepali, with very few resources. Furthermore, their work shows that multilingual pretraining is an excellent way to make ASR systems for the many different languages spoken on the Indian subcontinent.

Want to publish your content?

Publish an article and share your insights to the world.

ALSO EXPLORE

DISCLAIMER

The information provided on this page has been procured through secondary sources. In case you would like to suggest any update, please write to us at support.ai@mail.nasscom.in