Transparent generosity. Introducing the impresso interface for the exploration of semantically enriched historical newspapers

Figures & data

Table 1. impresso’s user requirements gathering process. Overview of the involvement and contributions of stakeholders in history and libraries.

Stakeholders	Frequency	Contribution	Exposure to the app	Interaction
Two historians in the impresso core team	Day-to-day	Development of user requirements for interface design; data modeling	Constant involvement in the development	Personal Slack Skype
Historians and librarians in the consortium	6-monthly exchanges	Initial user requirements, feedback, and validation	Presentations and independent testing	Personal Email
Associated researchers	Intermittent	Feature requests, feedback, validation	Independent testing based on personal research interests, partial participation to presentations	Slack Skype Email Personal
Individual sessions	Intermittent	Feature requests, feedback, validation	Presentations, testing during sessions	Personal
University teaching	Intermittent	Feature requests, feedback, validation	Presentations and testing, increasingly toward the end of the project, on advanced versions.	Personal
Principal investigators	Intermittent; annual	Feedback and validation	Presentations	Personal Consortium meetings
impresso app users	Irregular	Feedback on UX and errors	Usage	Slack, Email, Feedback app

Table 2. Retrieve relevant content: Search, filter, and discover.

Inputs: ideas and requirements			Outputs: interface components	ID
Historians	Designers	NLP
Retrieve relevant content; Identify sets of relevant keywords; Deal with OCR mistakes	Autofill based on matching enrichments; Documentation of search query; Results overview via metadata facets	Word embeddings, named entities, language detection, topics, and content item type detection; Result ranking, fuzzy match, exact match	Search pill; “Find similar” search component	#R1
Identify persons in the text	View distributions in context with other enrichments	Named entity detection and disambiguation	Filter by entity	#R2
Identify and (de)select semantically coherent articles from search		Group articles by topic models with relevance score	Filter by topic	#R3
Support further content segmentation		Lexical preprocessing and content segmentation	Filter by language, content type, content length	#R4
Collect articles in bulk and curate queries manually, control the collection of articles	Include collections in article search and as filters		Create collection from a query; Add selected articles to one or more collections; Focus on exploration of collections	#R5
Search and explore images	Add image search to article viewer	Image similarity detection and text-based search	Detect image reuse and find similar images	#R6
Trace content across the corpus		Text reuse detection	Text reuse explorer and text reuse indicators in article view	#R7
Avoid the “Filter bubble effect”	Recommend articles based on a collection	Recommender system	Customizable article recommender in the collection space	#R8

Table 3. Understand what is there: Aggregate, document, and contextualize.

Inputs: ideas and requirements			Outputs: interface components	ID
Historians	Designers	NLP
Understand the composition of the corpus	Visualise print runs and missing data	Collect missing data during ingestion	Newspaper page	#U1
History of query and collection building	Transform queries to readable text		Human readable summary	#U2
Read article in page/issue context	Transcription and facsimile view with semantic enrichments as marginalia Searchable table of contents	Articles recommendations based on topics	Article viewer	#U3
	Summary overviews of the distribution of semantic enrichments using small multiples		Overview pages for entities, newspapers, collections etc.	#U4

Table 4. Compare and assess significance.

Inputs: ideas and requirements			Outputs: interface components	ID
Historians	Designers	NLP
High-level overview of queries and collections;	Compare search queries and represent proportions		Inspect & Compare	#C1
Trace semantic shifts	N-gram visualization and comparison	N-gram frequencies	Multiple Unigram view (n-grams)	#C2

Table 5. Externalize: Export, link, and embed.

Inputs: ideas and requirements			Outputs: interface components	ID
Historians	Designers	NLP
Share and process content outside the interface	Embed content as iframe outside the interface; Reproduction of queries Data export to csv		Export function for article metadata, NLP metadata, article contents; Stable URL, shareable among users; Widget for external referencing	#E1

Table 6. Create transparency through documentation and educational materials.

Inputs: ideas and requirements			Outputs: interface components	ID
Historians	Designers	NLP
OCR quality assessment per article	Help users to assess and act on OCR quality	Calculate a measure for OCR quality	OCR QA search filter (forthcoming)	#T1
Convey the idiosyncrasies of enriched newspapers	Videos and animations		Tutorials Ranke2 lecture with quiz	#T2
Support reasoning based on semantic enrichments; Understand the interface functionalities; Understand choices made by the project and their impact	Visual language for educational materials Integration of educ. materials in the interface	Data processing documentation (Co-)production of educ. materials	Blog FAQ Community Calls impresso Challenges impresso YouTube channel	#T3

Figure 1. Search pill with suggestions for the keyword “atom” (blue) including entity mentions (orange), topics (green) linked named entities (yellow), user-created collections (blue).

Figure 2. A Query for “atom” and the linked entity “Otto Hahn” with suggested keywords based on word embeddings (left), the human readable summary of the executed query (centre top, highlighted) and snippet previews of search results (centre).

Figure 3. Image similarity ranking reveals image reuse between 1984 and 1996 (images blurred for copyright reasons).

Figure 4. Searching for instances of text reuse in the nuclear power collection revealed a pro-nuclear PR campaign published between May and November 1977 in Swiss French-language newspapers.

Figure 5. Experimental recommender system with option to assign weight on co-occurring entities, temporal proximity, topics, and text reuse.

Figure 6. Newspaper page with information on publication periods, available issues, and indication of missing pages.

Figure 7. Distribution of articles over time for the query on nuclear power content.

Figure 8. Preview of the distribution of a linked named entity in the corpus.

Figure 9. Overview of the distribution of a linked named entity in the corpus and contexts in which it appears.

Figure 10. Facsimile view of a newspaper page with marginalia and searchable table of content on the left.

Figure 11. Using the Inspect component to contrast the distribution of “bikini” in French and German language content.

Figure 12. Using the Compare component to assess the proportion of content related to nuclear power and nuclear weapons in German-language newspapers.

Figure 13. N-Gram frequencies of “atom”, “atomenergie” and “kernenergie”.

Table 7. Overview of the integration of main semantic enrichments in impresso Search page, Inspect & Compare and availability of dedicated components for exploration.

	Search	I&C	Dedicated component
Topics	✔	✔	✔
Text reuse			✔
Named entities	✔	✔	✔
Image search			✔
Word embeddings	✔	✔
Language	✔	✔
n-grams			✔
Content item type	✔	✔
OCR quality	✔

Figure 14. Example of an i-button with link to corresponding FAQ article.