AI in Text Summarization

In today’s day and age, we are constantly exposed to disproportionately large amounts of information compared to the quantity we can consume naturally. For this reason, we commonly refer to this period as the era of Big Data. One source of information relevant to our everyday lives is literature. Articles, papers, websites, and books are written and published at a rate much faster than ever before, and this rate is only increasing. Unfortunately, we as humans can only read so quickly.

Thankfully, this is an area where AI and Machine Learning can shine. Text summarization is the process of condensing long documents into shorter summaries. This task falls under the category of Natural Language Processing (NLP), which is a very active area of research within the machine learning community. Other language generation tasks within the NLP sphere include machine translation, sentiment analysis, and text classification.


There are many use cases where text summarization can be applied. For example, when coupled with web scraping, we can combine many different news articles surrounding a particular event and summarize critical facts without having to read through all the fluff. This capability has clear benefits for businesses within the financial services, asset management, and healthcare industries, to name a few.

Another use case involves generating captions and titles for documents.  Taking this a step further, if we combine Natural Language Processing with the field of Computer Vision, we can create captions for images and videos. The possibilities seem endless.


When it comes to summarizing text, there are two main approaches: extractive and abstractive. Extractive summarization is where we pull keywords and sentences from the source document and use these for the summary. Although this may work for simple use cases, there are some drawbacks. For one, some of the sentences extracted may be out of context, and this has the potential to create confusing or misleading summaries.

Abstractive summarization is the process of creating new and unique sentences using language generation models. Large collections of text often referred to as a corpus are used to train these models to make them capable of such summaries. These models incorporate the external knowledge from this corpus when creating new abstracts and therefore can use words not found within the source text. We can potentially create high-quality summaries, competing with those made by actual humans, by using these techniques. The drawback is that abstractive models are a fair bit more complicated than extractive approaches.

At nova IQ, we are experimenting with different abstractive approaches such as the Compression-Paraphrase model1 and Google’s new Transformer Model 2.

Text summarization may still be a nascent area in artificial intelligence, but it will have applications in so many areas of business.

by John Conway -AI Engineer, nova IQ


1:           Y.-C. Chen and M. Bansal, “Fast abstractive summarization with reinforce-selected sentence rewriting,” in Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, 2018, pp. 675–686.

2:           A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems, 2017, pp. 6000–6010.