Europa Press completes a project using AI to streamline the media processing of audiovisual content

Lee la versió de este artículo en castellano

MADRID, 13 Ene. (EUROPA PRESS) -

   Europa Press has completed a project, financially supported by the Spanish Centre for the Development of Industrial Technology (CDTI), with the aim of using Artificial Intelligence to facilitate the analysis and news reporting of the audiovisual material that reaches its newsroom every day.

   As part of its R&D projects, the agency has created in collaboration with the Spanish company AyGLOO a prototype called "Smart News Analyzer (SNA)", capable of converting into text any news content that arrives in the newsroom, whether audio or video, classifying it by subject, finding moments of interest and processing it more efficiently.

HOW DOES THE TOOL WORK?

   To do this, the tool uses 5 different artificial intelligence techniques (NLP, Graph Theory, STS, NLG and XAI) that are applied to the audio and video content that arrives every day to media newsrooms in the form of interviews, press conferences, short statements, etc.

   By applying AI, this technological tool is capable of shortening reaction times by offering the journalist an almost real-time transcription of the audio, extracts of the different 'moments of interest' according to predefined interests by the user, as well as two types of summaries: one extractive and one abstractive, according to the terminology of the language models.

Front-end prototype of the tool

WHAT IS THE GOAL?

   The aim is to minimise the time between the production and reporting of news and to use technology to facilitate the daily work of newsrooms, offering journalists tools that can support them in the most repetitive tasks and in the distribution of audiovisual content.

   Aware of the importance of shedding light on the 'black box' that this type of models represent, especially in an environment such as journalistic information, the project also includes as an additional objective to provide the tool with an explainability module that gives journalists clues about possible biases in the decisions adopted by the AI.

TECHNICAL CHALLENGES

   One of the technical challenges has been the choice of a model to perform abstractive summarisation tasks, a decision that involves evaluating multiple technical and economic factors, including the performance of the model, the costs associated with its training and implementation, and the quality of the results obtained.

   After a thorough analysis, it was determined that using Openai was the most viable option rather than training a zero-based model on an Encoder-Decoder architecture or fitting smaller models such as LLaMA.

   Advanced language models, such as Openai, have demonstrated superior levels of accuracy and consistency in text generation tasks, including abstractive summarisation. While fitting a small model may be feasible in some cases, the quality and robustness offered by Openai surpasses the results expected from lighter models fitted in specific domains.

Search for latest linked news

   Another challenge has been to achieve a tool with a real practical utility in the day-to-day work of newsrooms. In this sense, taking into account that a person who is intervening live can give different news that is of interest to different sections of the newsroom, the need was established to develop a classifier of parts of speech by subject matter.

   Given that in the transcripts of press conferences, interviews, television programmes, etc., the information is not structured, it was necessary to develop a tool for separating paragraphs and encoding these sentences of the text in a format that understands the model ('embeddings') to be able to compare the result with the agency's corpus of news, and thus have enough context with which to carry out the classification by topic.

   In this sense, the use of the Europa Press database was incorporated to highlight the news events that represent current affairs so that the artificial intelligence algorithm can relate these current affairs contents with those described and identified within the text itself. In this way, the ranking positions of each of the identified phrases are statistically favoured.

Schematic diagram showing basic operation of the prototype

   In addition, user profiles had to be created with the particular interests of each journalist, which also becomes an embedding. The cosine similarity of the profile with each of the phrases leads to higher ranking positions for those phrases that have a meaning close to the user's profile description.

   As the project developed, new analyses were added, such as further processing to determine the way the person is expressing themselves. More controversial or emphatic expressions also move up in the ranking so that the journalist can evaluate them.

   Another of the new analyses incorporated in the last part of the project crosses the information extracted from the audios and videos with a database of verified facts and 'fake news' already disproved by certifying bodies so that, if so, it can influence the algorithm's decision when considering it a moment of interest.

NEXT STEPS

   Once the prototype has been completed, the next step will be to make the tool available to the editorial staff and partners in order to fine-tune its use and make it as practical as possible.