CASE STUDY - A.I use case: Tools developed by INA to promote its audiovisual heritage

Following on from last year's discussions, this year's MIFC is back with another discussion on artificial intelligence, this time focusing on the application of tools to heritage archives.INA was represented by Pascal Flard and Xavier Lemarchand, facing journalist Perrine Quennesson, to discuss its new technological tools during a case study.

As explained by Xavier Lemarchand, INA's head of IA Coordination and Integration, the Institute has been making audiovisual archives available to professionals for use in fresh programs since 1975. Since 1992, they have been safeguarding the radio and TV archives of 192 channels, and since the early 2000s have been offering editorialized content in a heritage medium. Their catalog includes some 25 million hours of TV and radio legal deposit, 5.8 million hours of web legal deposit, and 2.5 million hours from professional archives. That's no less than 30 million hours that need to be archived, catalogued and processed. At a time when only 85.5% of data from global legal deposit is catalogued, without descriptors, captions or summaries, artificial intelligence (AI) is the tool of choice to ensure speed and reliability in cataloguing and data processing.

A variety of AI tools for a variety of applications

INA works with several AI tools, both internal and external, which are not at the same level of maturity. The main ones are : Whisper, TextRazor and InaSpeechSegmenter, which is the only internal component at present. Their first use is the transcription of archived content, which then provides useful data for research and catalog discoverability. This opens up new fields of analysis, research and exploitation. The role of humans is not called into question, but technology reduces the complete transcription of INA archives to just two years, compared with a hundred thousand years for humans. It should be noted that not all their tools are yet mature enough to be usable, despite their great value in use. Text generation, such as titles or summaries, image description and face and speaker identification, for example, are not yet mature enough, whereas named entity extraction and signature are highly developed.

The reliability of the tools proposed is also a crucial issue for INA, which does not hesitate to carry out field measurements to ensure the reliability of the results given. This involves assessing data quality before processing on an industrial scale, as well as monitoring changes in reliability to avoid cyclical drifts, and improving processing by using human analysis and annotation. The challenge is not simply to “correct” a result, but to improve a tool in the long term.

Limited application in heritage cinema

The tools used or developed by INA are designed to support media specialists or the curious who wish to represent media data, but for the moment they have not been designed for use in heritage cinema. However, they do open up avenues for further reflection. Pascal Flard, deputy head of the INA's “Media Management & Processing” department, presented the various fields reworked by AI in terms of audiovisual content restoration, whether the artificial intelligence is generative or “degenerative”, notably on the subjects of colorization and image resolution. Outpainting and lip-synchronized dubbing are not used, for ethical reasons.

Transcription tools could also be used for research into the representation of male or female characters in films. An experiment successfully tested on podcasts.

The main limitations of AI

The first limitation is sustainability. Indeed, the two INA representatives are quite clear: the reliability of AI diminishes over time. There is no question of storing images produced by AI for more than two years, as they will quickly become obsolete. Only the original elements are preserved over time. This eliminates the question of storage and the energy required to store them.

The durability of AI brings with it a second limitation, that of evolution over time. For example, the performance of the tools used by INA is undermined when new entities appear, such as new names or new faces. They can cause transcriptions to fail. Human validation is therefore necessary before data is published, even if errors are marginal: the confidence rate is 83%. In the event of errors, a warning is displayed to alert users that their work base may be in error.

The content to be reworked is also a limitation, making it difficult to apply to long sequences or 3D images, where artificial intelligence is successfully applied to 2D.

Finally, a fourth limitation is to be found in contextualization: AI does not yet have all the keys to contextualize elements, and can therefore propose erroneous results. This is the case, for example, with the word “Wagner”, associated with the composer Richard Wagner from 2022 onwards, when the media are talking about the Wagner group, a Russian paramilitary organization.

However, the two experts are adamant that the limits are being pushed back as time goes by and as knowledge of artificial intelligence deepens.

A few principles to respect

Reaffirming that AI is not a miracle tool, and that it requires human supervision to be used to best effect, Xavier Lemarchand and Pascal Flard underline the various principles that INA intends to respect in its use of artificial intelligence. Starting with respect for copyright, and the obligation to limit the risks of historical distortion. The production context is also important: INA only produces colorized images if there is a genuine editorial logic behind them. The viewer is informed of every modification made, and the original elements are carefully preserved. These principles are reminiscent of those already respected by professionals in the classic film industry.