In Search of News Past

Impresso 1

2017-2020

Impresso 2

2023-2027

Developing new approaches and interfaces for the exploration and critical analysis of historical media archives.

Swiss and Luxembourg newspapers

Cutting-edge platform

New collections from across Europe

Radio broadcasts

Datalab

An interdisciplinary team spanning history, web development, linguistics, and digital humanities

A network of partner institutions that provide digitized newspaper and radio collections

Press from 1732 to 2018

Radio programs in audio format with speech-to-text

Transcripts of radio news broadcasts

Impresso user plans

Impresso web app

Multiple filters

date range

language

keywords

source

length

article type

and many more...

Advanced features

Named entities profiles

Corpus overview

Ngrams

Text reuse

Browse newspapers and radio broadcasts at the same time

Example of future integration of audio files into the Impresso app

Impresso app

Currently under development

To preserve a corpus, it can be saved as a collection.

All articles are now tagged so they stand out in search results.

Each collection has its own page, complete with some descriptive statistics.

A collection can be exported as a CSV file.

Once exported, the CSV file can be downloaded and opened in a spreadsheet program.

It includes all the data that your plan grants you access to.

The Datalab contains instructions for programmatically accessing the Impresso API. But for the general public, it’s the notebooks that are of particular interest. 

We have also developed additional notebooks that are currently pending. We also invite colleagues outside the project to develop their own notebooks, which we could add here.

Descriptive statistics for a collection

Network of entities

Map of locations

Semantic proximity

The goal is to spark users’ curiosity, make computational methods accessible to them, and encourage them to explore code without needing to master it.

But what exactly is a notebook?

A Jupyter Notebook is an interactive, (often) web-based tool that lets you combine live code, visual graphics, mathematical equations, and narrative text into a single shareable document.

1. A notebook contains lines of code 

2. It also includes comments explaining what we're doing

3. Each cell of code can be executed to see the result immediately 

But then, who develops these tools? And to meet what needs?

We can try to anticipate the needs of historians and students, but we can also identify them collectively. 

For example, we invited colleagues to a workshop to develop analytical scenarios together

(University of Luxembourg, fall 2024).

The process begins, therefore, with a face-to-face meeting between people from different professions, to “bridge the gap” in a practical way.

For example, one group came up with the idea of creating a small tool that would allow users to perform exploratory data analysis (EDA)—enabling them to generate a brief statistical report based on a collection of Impresso articles.

This drawing is literally the result of that group work session.

We have established a Datalab Editorial Board, which develops guidelines to standardize the notebooks and manage their development.

Development and internal/external review processes

Internal review

External review

Drafting

The entire process is a transparent collaborative space that brings together historians, digital humanists, and developers.

This is the result of this collaborative process: a notebook to "inspect" a collection of articles

This notebook was created for educational purposes, to generate basic statistics based on an Impresso collection. 

We open it in Google Colab, an interface that lets us work online with this Jupyter notebook

And we upload the CSV file of the collection we created during our previous exploration of radio coverage in the press.

Each line of code is commented to guide the user

Here, we load the dataset

This is literally where the two languages meet

In this example, we use the exact collection we created earlier

If there are any questions, Gemini helps you to understand or fix the code

One first way to bridge the gap?

The notebook guides us through the process of creating our first chart.

In this case, a histogram that shows how many articles in our collection come from which newspapers.

Here, our collection includes articles containing the keyword “radio” from three Swiss newspapers during the 1950s.

The notebook then generates a timeline of articles, organized by newspaper.

We can see that the number of articles in La Sentinelle (LSE) mentioning the radio increased toward the end of the decade

The notebook then guides us through the process of generating other exploratory statistics.

For example, an analysis of article length

The length of articles varies from one newspaper to another: on average, articles about radio in La Liberté (LLE) are twice as long as those in Le Journal de Genève (JDG).

This second notebook helps highlight the main named entities that appear together in the articles

No need to create a collection beforehand; the notebook will connect to the Impresso data on its own.

So we can run a query directly in the notebook. For example, to find the 100 people who appear most frequently in articles containing the word “radio” in the three newspapers covered by our previous analyses for the same period.

It generates a network that shows how often the most frequently mentioned people appear in the same press articles.

This network can then be exported and viewed more clearly in software such as Gephi

In the case of the "radio" keyword, the entities are very strongly influenced by the music programming. Composers (in white) are much more prominent than political figures (in blue) or media personalities (in red).

This third notebook will retrieve the geographic coordinates of the locations mentioned in the articles and generate a map.

Just like in the previous notebook, we run the query directly here.

Next, the top 100 results are geolocated by querying Wikidata.

It generates a map showing the cities (in black) and countries (in red) mentioned in the news articles.

In the case of a corpus built around the keyword “radio,” the results aren’t very interesting: the locations are generally the cities where the broadcasts or concerts take place, with a few capital cities occasionally mentioned in news stories. This approach is much more interesting when working with a more general-purpose corpus.

Incorporating these tools into historians' toolkit

EPFL MA Course 2025-2026

Examples of student works, made with the Impresso datalb

How can historians really learn from the kind of tools computer scientists use and develop instead of merely outsourcing specific tasks?

How do you NOT do what you did, simply differently or at greater scale, but do something that's genuinely new?

And conversely, how do historians avoid being relegated to "consultant" roles, providing a sort of historiographical guarantee?

How do you push research forward and not simply adapt it to new technology?

Embedding algorithms are trained, in similar fashion to LLMs, to map the position of words in relation to each other across billions of texts, and, based on these patterns, they generate vector representations for those texts: this is what we mean when we say that these algorithms embed words in their semantic context.

Here: war and peace appear more often in the same texts than peace and tomato—therefore the vectors of war and peace are closer than those of peace and tomato or war and tomato

Embedding algorithms are trained, in similar fashion to LLMs, to map the position of words in relation to each other across billions of texts, and, based on these patterns, they generate vector representations for those texts: this is what we mean when we say that these algorithms embed words in their semantic context.

Here: war and peace appear more often in the same texts than peace and tomato—therefore the vectors of war and peace are closer than those of peace and tomato or war and tomato

Words are transformed into multidimensional vectors

A semantic map, reducing the complexity of the vectors to see their proximity

Integrate embeddings into the interface so that if we search for "bank" we get results about credit institutions rather than river banks

Animation: Coenen and Pearce, “Understanding UMAP”, https://pair-code.github.io/understanding-umap/ 

How can we interpret a UMAP, a cloud of dots that is a “flattened” version of our original multidimensional object (which is itself too complex to grasp)?

Visualizing vector representations of articles in 2D reduces the dimensionality of the semantic distance between them (from hundreds of dimensions to just 2).

Interpreting semantic proximity maps (UMAPs)

And a tool to compare newspaper biases?

At first glance, we see that the colors (the newspapers) are not always evenly distributed within the groups of dots. This means that the newspapers do not cover the same areas, or not with the same intensity.

Back to our 60,000 articles published in 1950 in three newspapers.

A tool to compare newspaper biases?

We therefore need to explore the map to understand what is in these clusters.

Accident

Korean War

International relations

Job advertisement

Weather

A tool to compare newspaper biases?

The editorial line of the newspapers becomes quite clear with this overall view.

Classified ads

Religion

Music, culture

Economic news

Crimes, accidents

Can embeddings be transposed into historical research pipeline?

We worked on different scenarii to apply the tech on historical data

Fusing computer scientists' knowledge of the tech with historical lines of questionning

Pictures of sources taken in archives

A collection retrieved from Impresso

How to visualise the semantic similarities between them?

Transcribe the sources in CSV

Download the collection in CSV

Use a semantic proximity map to visualise them in 2D

Embed the two datasets

      Original press release

       Article in the press

A tool for exploring large press corpora

An interactive interface to explore a collection, zoom in, select and get back to the original texts

A tool for comparing press and radio

A semantic map showing that radio and the press do not cover the same news stories

Impresso also organizes workshops and international conferences

coming soon!

Impresso Seminar

Building bridges with existing communities

An example of how the tools developed by Impresso can be used to address a media history question

Radio Nations in Prangins in the 1930s

United Nations Archives Geneva (P132-01-007)

Newspaper stand at Zurich train station, 1962

ETH-Bibliothek Zürich (Com_C17-073-047-001-005)

Welcome to the Impresso web app

Before starting a search, log in to access more content and facsimiles.

Let's search for "radio" in article contents

This very broad search returns over a million results

But what are we looking at?

Total number of articles

Proportion (%)

For example: Is the result biased by the nature of the corpus, since some newspapers are only available during certain periods?

But what are we looking at?

Or: Does this keyword mainly return results in a specific language, a specific country, or from specific news outlets?

And is that even the right keyword?

This is a classic problem in searching large text corpora: keywords are an imperfect way to filter

- Is the keyword polysemous?

Radio = the device, the institution, the broadcast?

- Does it appear in any other word?

Radioactivity, Radiography, ...

- Is the concept we're looking for gradually being replaced by another one?

- Are there different ways to spell it?

(in the case of radio, no problem)

Radio -> Podcast

- Does this word often cause OCR issues?

radio, r4dio, radid, adio, 'radio, ...

- Should it be paired with other keywords?

Radio AND Broadcast, Radio OR Antenna

By selecting a date range (in this case, a decade), we reduced the number of results by a factor of 10.

Let's narrow down the search using filters

Here, three French-language Swiss newspapers from different regions and with different political leanings

Remove articles that are less than 100 words long

Separate articles from advertisements

Keep only what is featured on the front page

Search filters

Search summary

Content item

The facsimile is visible only if you have access rights

Each result is available as a facsimile and/or transcript

Article in context, along with the entire issue

Block-by-block transcription, for close reading

Now that we’ve explored our collection, how is radio depicted in print media?

Category 1.1 Advertisements

At first glance, this may not be what we’re looking for, but radio is widely present in the press as a commodity.

Radio can also serve as a selling point for other products, such as cars.

Category 1.1 Advertisements

If we compare an opinion newspaper with a mainstream newspaper, the difference in the proportion of advertisements is clear.

Journal de Genève

8 % of "radio" articles are ads

L'Express (Feuille d'Avis de Neuchâtel)

31 % of "radio" articles are ads

Category 1.1 Advertisements

Category 1.2

Job advertisements

Radio isn't just a device, it's also an employer. Job listings in the press allow for documentation of this professional reality.

La Liberté 1956

La Liberté 1978

Category 2

Radio program

From the very beginning, the radio needed the press to publicize its weekly schedule.

Journal de Genève 1952

Feuille d'avis de Neuchâtel 1930

It is worth noting that even the dissemination of this very practical information creates synergies among the media

Category 2

Radio program

The opposite is also true: newspapers often include the schedules of many stations because they know this is an essential service for many of their readers. 

Gazette de Lausanne 1986

Category 3

Radio as a source of information

More interesting than advertisements or radio programs, radio can also be cited as a source of information in news articles.

Feuille d'avis de Neuchâtel 1940

A search of the Impresso corpus shows that this was fairly rare in Swiss newspapers during the early decades (1920–1950), but it became more common thereafter.

Category 3

Radio as a source of information

An interesting observation: it is very often radio stations from communist countries that are mentioned in Swiss newspapers

Feuille d'avis de Neuchâtel 1950

Feuille d'avis de Neuchâtel 1960

The Swiss media do not have correspondents in these countries and therefore rely on reports from government radio stations

Category 3

Radio as a source of information

As the second half of the 20th century progresses, radio is cited more and more often as a source of information. It is increasingly becoming an integral part of the media landscape.

Feuille d'avis de Neuchâtel 1973

Category 4

Radio as a subject in the news

The “richest” occurrences in this corpus involve references to radio as the subject of a news article.

Feuille d'avis de Neuchâtel 1940

Here, the press is complaining about the poor language used by radio hosts!

As soon as radio became widely accessible, it became a topic of discussion in its own right for the press. For example, newspapers began to review the programs themselves.

Category 4

Radio as a subject in the news

Feuille d'avis de Neuchâtel 1960

"Miroir du monde" is the flagship international news program on the Swiss French-speaking radio, coming soon in the Impresso app!

The press is becoming a platform for reacting to what is said on the radio, a forum for discussing the new profession of radio journalism.

Feuille d'avis de Neuchâtel 1970

The "radio critique" section

Category 4

Radio as a subject in the news

Feuille d'avis de Neuchâtel 1940

The press is a valuable source for exploring the institutional history of radio.

Feuille d'avis de Neuchâtel 1970

Feuille d'avis de Neuchâtel  1980

Category 4

Radio as a subject in the news

Feuille d'avis de Neuchâtel 1940

The press also shows us when and how radio becomes part of people’s daily lives.

Feuille d'avis de Neuchâtel 1960

Category 4

Radio as a subject in the news

L'Impartial 1926

Finally, the press is also a source for studying the history of radio technology

Feuille d'avis de Neuchâtel 1940

Feuille d'avis de Neuchâtel 1960

Category 4

Radio as a subject in the news

Feuille d'avis de Neuchâtel 1960

Worth noting: While radio and TV magazines would be a major source for writing about the people behind the radio, the press is also very valuable.

Feuille d'avis de Neuchâtel 1960

These two articles, published in the same issue on March 30, 1960, offer two unique perspectives on the people behind the microphone:

On the left, a popular radio host “reveals” himself in a photo… for an advertisement for a brand of razors! On the right, a report states that a radio employee has been convicted for refusing to serve in the Swiss army due to his pacifist and religious beliefs.

This leads us down an interesting path: can we study the people who appear in the selected articles? 

In the Impresso app, “named entities” (people, places, organizations) mentioned in articles are automatically identified. 

This leads us down an interesting path: can we study the people who appear in the selected articles? 

In a collection centered on the keyword “radio,” three categories of people emerge: composers, heads of state and prominent politicians, and… figures from the radio world.

In our case, two individuals emerge: the journalist René Payot (who moved from print media to radio), and the satirist Jack Rollan (who did the opposite).

Entities are linked to the Wikidata entries for the relevant individuals, if such entries exist.

The platform allows users to search for all instances of a named entity

This provides insight into the timeline of René Payot’s career, a journalist for the Journal de Genève who went on to become a radio commentator on international news during World War II, before hosting widely listened-to international news programs for two decades.

But very soon, there will also be radio collections available right in Impresso! 

1. Radio in text format

Radio typescripts

We have received the transcripts from Swissinfo, the Swiss international radio station, covering the period of World War II, in French and German.

But very soon, there will also be radio collections available right in Impresso! 

2. Radio in audio format

Radio broadcasts

We are making major changes to the platform to enable it to support audio.

We have collections from RTS, radio stations in French-speaking Switzerland, and the INA, which archives French public radio, just waiting to be released.

The key challenge is not only to integrate the audio, but also to ensure that the transcript can be used by users – for example, to find similar news articles.

In short, the Impresso project is a series of challenges: 

I.

The data

Multiple sources, multiple formats, massive data, and numerous legal issues

II.

The design

Make all of this accessible and useful to the public

III.

The analysis

Enrich data to create added value and develop relevant analysis tools

IV.

Writing history

How should all of this be interpreted in the context of historical research?

Conducting transmedia and transnational historical research using semantically enriched historical media sources

Impresso API

to process other sources

via notebooks in a datalab

and compare them to our data

Enriching and connecting historical sources by transforming noisy and heterogeneous sources into semantically enriched and structured data, ultimately connected in a common vector space.

Welcome to the Impresso datalab

Descriptive statistics for a collection

Network of entities

Map of locations

Semantic proximity

First, you need to log in to the datalab to be then able create a temporary personal token.

Let's analyze our collection about "radio" as a UMAP

This notebook, which is not yet available on the Datalab, retrieves the embeddings already calculated by Impresso or calculates them again if they are not provided by the API.

We've tweaked the collection's filters slightly to include advertisements and short articles that aren't on the front page, which results in a larger and more diverse corpus. This corpus of 4,000 articles is processed in about 20 minutes.

A map for exploring a corpus of press articles about radio

Each dot represents an article from 1950 that contains the keyword “radio” in the three selected newspapers.

This UMAP “map” is a simplification of the semantic distance between all these articles; it makes significant compromises to ensure it is readable in 2D.

This doesn't mean that if two articles are close to each other, they are the most similar; it means that they probably belong to the same semantic register (that they are about the same thing). That's why we'll focus mainly on the groups rather than on individual situations.

What can we see on this map?

In fact, this seems like a very good way to broadly categorize the "poor" or "rich" uses of the term radio in our corpus (as we saw more qualitatively in the previous chapter).

Focus on international news: Why is radio mentioned?

Since all the articles contain the word “radio,” this does not accurately reflect media coverage in 1950; it skews the results in favor of regions where news reaches Swiss newspapers via official (often communist) radio stations.

Now, we want to use embeddings to compare two collections.

For example, to study how radio news programs select information compared to the print media.

Here, we have selected a week's worth of news bulletins from Radio Suisse Internationale in June 1945.

To make them comparable to news articles, these bulletins must first be broken down into segments of information.

This is the text that was read on the air. There’s one every day in French, German, English, and sometimes Spanish.

One of the major news stories this week is the diplomatic progress made at the San Francisco conference among the “Big Five” (establishment of the future UN).

45 radio news

1674 press articles

Let's compare the news coverage in the press and on the radio for the week of June 4–10, 1945

All the articles (no ads) from the Journal de Genève, Gazette de Lausanne, La Sentinelle, Feuille d'Avis de Neuchâtel, L'Impartial and La Liberté.

News coverage in the press and on the radio

June 4–10, 1945

Area of the semantic map that is not covered by the radio bulletins

What is the order of priority in Radio Suisse Internationale's news bulletins?

June 4–10, 1945