Ella: Hybrid model (Gaspito)

The market for natural language generation by artificial intelligence (AI) has developed by leaps and bounds. However, lack of source transparency or generated false statements cloud the euphoria – especially in the media and content industry. The hybrid model of Ella Media AG now combines quickly achievable linguistic diversity with factual accuracy.

Efficiently create automated content with minimal instructions: innovative language models are currently creating everyday contact points with AI for everyone and can change content production fundamentally. But these impressive tools also bring problems with them. Because ChatGPT & Co. do not deliver quality-assured output: their models aggregate texts according to patterns that are based on their diverse training data. They do not guarantee factual accuracy. Whether a piece of information is in the right context or has a reliable source remains unclear – human expertise is needed to identify possible misstatements. Especially for journalism, factual accuracy is absolutely necessary. 

 

The best of both worlds

For its AI software solutions, the media tech company Ella has therefore developed the hybrid model “Gaspito”. Specially created for the journalistic environment and the German language, it combines the strengths of the language model GPT-3 with those of the specially developed language model Maskito. In turn, the weaknesses of the models balance each other out. Maskito is characterzsed by a high degree of factual accuracy and content proximity to the original text. It is supplemented by a large software package that can be used to check facts, for example. However, due to the closeness of the content to the original, the model is often unable to make major changes to the text – which is precisely what users want. GPT-3, on the other hand, modifies original texts to a high degree and creates high-quality output quickly and effectively. But it has two disadvantages: firstly, it does not create variety in all parts of the text equally, and secondly, the factual accuracy and source transparency of the output often do not meet journalistic standards of quality.

By combining the two models, the hybrid language model Gaspito generates creative content – and achieves better text quality as well as a high degree of factual certainty with great linguistic diversity for unique content.

 

Quality has top priority

For an optimal result, Gaspito works in two stages: users first select the desired sources via the AI Assistant tools. This ensures source transparency and high-quality original content. In the first step, GPT-3 generates a new, linguistically heavily modified text. 

 

In the second step, Maskito modifies the sections of the text that GPT-3 has not previously modified, using various methods. In contrast to ChatGPT, for example, the hybrid model then checks the generated text for factual correctness. Additional software and verification mechanisms compare the text with the original. If original quotations, proper names or data defined by developers as facts do not appear in the generated text, the hybrid model corrects this. In addition, Ella’s specially developed fact-checking model ensures the factual accuracy of the text generated by Maskito. It also learns, recognises and evaluates implicit facts, connections or changes in comparison to the original text. Both this assessment as well as the direct comparison, together with other factors, flow into the overall evaluation of the text. If certain parts are insufficient, Maskito changes or replaces them until the result meets Ella’s high quality standards.

 

Specialization in a market of generalists

There are now a number of generative AI tools that offer more than mere paraphrasing. But they are often generalists with broad training data. Ella, on the other hand, specializes in the media and content industry. “Our media-first solutions allow us to tailor the process of automated copywriting to the specific needs and requirements of the industry. This is also helped by our pilot customers with whom we test our tools and models,” says Nasrin Saef, Head of Machine Learning at Ella Lab. “The integration of GPT-3 opens up new possibilities. Right now we are working on many new features, such as a bullet-to-text feature, a summary feature and a headline generator.”

“Our goal is to revolutionize the media and content industry through automated, high-quality text creation and processing,” explains Michael Keusgen, CEO of Ella. “Through comprehensive automated and human quality checks, we guarantee a high quality in accordance with journalistic standards, both in terms of training data and the content generated.” Users have full control over their texts with regard to variation and origin of the content and can also compare the generated content with the original. They get exactly the automated solution they need for their daily work and which actually relieves them.

“We are currently evaluating where GPT-4 adds value for which use cases and where we want to use it,” says Ella’s Deep Learning Architect, Michael Janz.

 

About Ella

Ella Media AG was founded in 2020 by CEO Michael Keusgen in Switzerland. The MediaTech company specializes in AI-based text creation and processing. Ella’s AI-based language models and software products aim to automatically create high-quality content for news and media as well as creative stories for the entertainment industry. With the help of AI, Ella wants to revolutionize the media and content industry and relieve newsrooms and content creators through effective assistance. Ella is committed to high ethical and quality standards, based on many years of experience in machine learning and natural language processing as well as a deep knowledge of the media industry. 

Ella Media AG has over 80 employees. In addition to its headquarters in Switzerland (Canton of Zug), Ella operates further locations in the USA (Miami) and Luxembourg (Luxembourg). Research is carried out at the Ella Lab in Germany (Cologne and Berlin). More information can be found at https://ella2.inthemaking.be/.

 

+++

Press contact:

Fink & Fuchs AG
Patrick Rothwell
ella@finkfuchs.de
phone: 0611 -74131 – 16