Helping Machines to Understand Human Language Better

Natural language is ambiguous. Human beings can easily disambiguate the meaning of linguistic units such as words, phrases and sentences as they read or listen. However it is difficult for machines to do the same without explicit representations of syntax, semantic and discourse.

As a researcher in Natural Language Processing (NLP) at NTU, Professor Joty Shafiq is seeking to address this by giving explicit representations to different aspects of natural language such as its syntax (constituency and dependency parsing), semantic structures (named entities, semantic roles) and discourse structures (co-reference, coherence). NLP tools can then be built to parse natural language in terms of these representations in order to help machines to understand human language better.

Previously, researchers used statistical machine-learning methods to build such NLP tools. However, these methods required a lot of human effort and expense, and were not generalisable. They were also limited in their performance.

Applying neural network models
However like other fields of AI, NLP has gone through a deep-learning  tsunami with researchers now applying neural network models that require less human effort because the model learns automatically from the data.

With the new NLP tools, Prof Joty is also looking to develop better end-user applications for machine translation, text summarisation, dialogue systems and sentiment analysis.  

For example, with the parsing tools, named entities can be used to build knowledge graphs that are used for search; and constituency and dependency trees can be used to understand syntactic dependencies in order to build sentiment analysis tools, and so on.

Alternatively, applications can also be built end-to-end using with neural (deep learning) methods that rely minimally on the NLP tools.

Diving into machine translation and multilingual and multimodal processing
One of the areas that Prof Joty is working on is the development of a new deep-learning architecture for machine translation – a huge industry that is worth about US$100 billion a year. Other aspects of machine translation that his lab is focusing on include data augmentation through diversification, unsupervised and semi-supervised neural machine translation to support low-resource languages like Malay, Nepali, Hindi, Sinhala and Tamil, word translation (cross-lingual embeddings) and discourse-based machine translation.

Another significant part of Prof Joty’s current research is in multilingual and multimodal processing. His team is interested in multilingual parsing tools (e.g. multilingual Named Entity Recognition) as well as multilingual applications in question answering, summarisation and natural language inference. He is also collaborating with the Vision group at NTU to build multimodal applications including image and video captioning models.

Recently, Prof Joty has also been looking into the security of AI systems as well as the robustness of NLP models to ensure that NLP tools and applications do not exhibit algorithmic bias and discriminate on the basis of factors such as gender, race or speaker.

“Language is a great creation and working to make machines understand language and generate language is fascinating,” said Prof Joty. “NLP methods directly impact humanity and commerce, and as NLP technologies reach human parity, we as researchers need to also think about its social impact.”

For more information on Prof Joty and his research work, please click here.