Identrics - Best Content Management Solution
Company: Identrics
Company Description: Identrics creates bespoke automation and artificial intelligence solutions, delegating organisations’ menial data and text processing tasks to machines so that humans can focus on high-end, insight-oriented work. Our solutions range from data classification and indexing to content enrichment – such as named entity recognition, sentiment analysis, automated content recommendation and generation.
Nomination Category: Product & Service Categories - Business Technology Solutions
Nomination Sub Category: Content Management Solution
Nomination Title: Multilingual Abstractive Summarisation. A technology that “understands” the text and summarises it in “its own words”.
If this is a brand-new product, state the date on which it was released. If this is a new version of an existing product, state the date on which the update was released:
Our abstractive summarisation solution is a brand new product, which was released in May 2021.
Identrics offers predefined solutions based on a combination of web-based software and predefined machine learning models.
Recently, we launched our latest Multilingual Abstractive Summarisation technology. Text generation-related technologies are popular right now, especially when it comes to recreating an original document into a completely new one.
We worked to create a complicated set of technologies that allows us to automatically read documents, namely media articles in various different languages, and create an automated abstract in English. The cutting edge characteristic here is that the machine-generated summary is completely different from the original text. It has all the facts from the original document, but it is retold differently by the machine.
Our product transforms multilingual content from all over the world and summarises it in English.
Automated abstractive summaries are generated by using a combination of machine learning (ML) models. The algorithm “understands” the text and summarises it in “its own words”. Existing automated summarisation solutions work predominantly with “extractive” algorithms, which are quite reliable but limited in functionalities and involve copyright issues. We are tackling these problems and can generate near human-like shorter retellings of different types of texts. Our product transforms multilingual content from all over the world and summarises it in English.
The challenges most ML & artificial intelligence (AI) technologies face are related to the possible retelling errors made by the models during the improvisation when generating text. Identrics optimises its process by developing and implementing fact-checking algorithms as a step before delivering the final summary to the client.
Our process works in the following way:
- The AI/ML algorithm generates a summary of the original document
- Our fact-checking algorithm verifies the generated information
- After a successful verification, our client receives a bulletproof abstractive summary
Our service works in real-time. Technologies like this allow us to deliver real-time summaries and monitoring via chat servers, chatbots and more. We can automate a newsletter with this technology as well.
Our abstractive summarisation model is tuned to understand and describe the core idea of a news article or the essence of business, economic, crime or political events. The technology that we use can be applied to different domains too by simply inputting sample data.
To represent the original document in a shorter, unique text, we utilised different technologies. At the same time, we conducted and added a lot of our research and development efforts.
We created a fact-checking algorithm that looks for mistakes between the original document and the summary. In addition, we have enriched the technology with spell-checking algorithms as well. We also check textual contradictions in the corresponding sentences and the summary. Sometimes, language models can “reimagine” facts and therefore we also developed a hallucination-checking algorithm to meet the best level of relevance we can.
Taking into account the AI models’ limited size of input, we are also using a fact preserving technique when shortening long text, in order to stay consistent with the original factology.
We have tested our solution on two dimensions - quality of English language and factology. As a result of all the tests and changes we made during the research and development phase, we created a technology that provides relevant summaries that are accurate to the original documents. Ninety-five per cent of the abstracts have good style and English grammar. At the same time, over 70% of the summaries contain facts that are relevant to the original document.
We have a case study with a media intelligence company that wanted to optimise the work of its analysts and reduce the time they spent creating summaries of different articles, so they could focus on more important work.
We have implemented abstractive summarisation into their working process. This led to an average of 50% optimization of the time an analyst spent on an article and a 40% reduction in the production costs, reported by the company department that implemented our solution.
This technology has the potential to also optimise processes in publishing, risk and compliance, bioinformatics, libraries and universities.
Please find some relevant information that supports our nomination:
We presented our abstractive summarisation technology during two events organised by the world’s largest media intelligence association - FIBEP: Exploring the media intelligence business ecosystem interview series and FIBEP Tech Day 2021:
https://identrics.net/which-technologies-in-the-media-intelligence-industry-to-keep-an-eye-on-this-2021/
https://www.fibep.info/2021-techday
Promo material about Abstractive summarisation and Identrics: https://drive.google.com/file/d/1-1ot6DKMZFIsSef4elw-rY40_N0X_wcX/view?usp=sharing