Born-digital document collections contain vast amounts of historical facts and knowledge. However, manual assessment of these large text collections is infeasible. In this paper, we demonstrate a retrieval system, DIGITALHISTORIAN, that analyzes these document collections using semantic annotations in the form of temporal expressions and named entities linked to a knowledge graph. For queries about entities or events DIGITALHISTORIAN utilizes state-of-the-art methods to understand and analyze temporal expressions in the content of documents. It understands uncertainty in temporal expressions and uses them to mine interesting time intervals for keyword queries. These time intervals are further used for re-ranking and diversifying documents, so that the ranked list of documents portray a historic overview of the query. Further, to contextualize the interesting time intervals, we use frequently occurring named entities and display them in informative visualizations. DIGITALHISTORIAN is designed to help scholars in digital humanities explore large document collections quickly without having any prior knowledge about interesting time intervals or entities for given keyword query. Large born-digital document collections cannot be analyzed by manual human effort. Nowadays, however, they can be automatically annotated with temporal expressions and named entities. Time has been found to be a very important part of generic Web queries; recent studies [9] estimate that around 17.1% of them are implicitly time-sensitive in nature. In the context of digital humanities, these figures can be expected to be much higher. The desiderata that we believe many commercial search engines do not currently meet and which will be useful for scholars in digital humanities are: . the ability to automatically suggest interesting time intervals for history-oriented queries; . the ability to diversify or re-rank documents using temporal expressions; . the ability to establish relationships between the time intervals of interest for query and other evidences in text such as named entities; . the ability to visually analyze the different relationships established between the annotations in documents.
Organization. The rest of the paper is organized as follows. In Section 2, we give a brief description of the key methods applied for mining interesting time intervals, and how these are used for search result re-ranking and diversification. In Section 3, we describe the technical building blocks of DIGITALHISTORIAN. In Section 4, we describe how DIGTALHISTORIAN can be utilized to explore document collections. We put our system in context to related systems in Section 5 and summarize the contributions in Section 6. 2
The underlying methods for temporal search in DIGITALHISTORIAN are derived from our prior
research [
Understanding Time. Temporal expressions in documents can be highly uncertain, for example
1990s. For such temporal expressions it is unclear how the time interval should be constructed for
further analysis. In order to represent such ambiguities in time, we use the time model proposed by
Berberich et al. [
Time Intervals of Interest to Queries. Given a history-oriented query such as george w bush, our
approach [
Re-ranking Documents Using Time. The ranking of the initial set of pseudo-relevant documents
can be refined by using one of the interesting time intervals for query expansion. Consider for example
the query george w bush reformulated with the time interval [2000,2004]. The documents which
contain temporal expressions that can generate the time interval in the query more frequently and also
have a higher textual relevance to the query will be promoted in the rankings [
Diversifying Documents Using Time. The time intervals of interest can be considered to reflect
different temporal aspects underlying the query. The initial pseudo-relevant set of documents can then
be diversified so as to contain at least one document relevant to the different temporal aspects [
Counting Frequent Named Entities. The time intervals of interest can further be used as a basis for aggregating different annotations in text. In particular, we aggregate the occurrence of unique named entities. E.g., given the query george w bush and the time interval [2000,2004], we obtain the aggregate counts of co-occurring named entities such as al gore. 3
The key building blocks of DIGITALHISTORIAN are: a document collection, semantic annotators, an information retrieval framework, a visualization engine, and the graphical user interface. We describe each of them briefly in the following paragraphs.
Document Collection. We used the The New York Times Annotated Corpus,1 a collection of news articles published in The New York Times between 1987 to 2007. It comprises of roughly two million news articles. As metadata, we used only the publication dates.
Semantic Annotators. For annotating temporal expressions we utilized the HEIDELTIME temporal
tagger [
Information Retrieval Framework. All the documents along with their annotations were indexed with the ELASTICSEARCH2 framework. As a baseline retrieval model for pseudo-relevant documents we used the Okapi BM25 method implemented in ELASTICSEARCH. All our methods for temporal search and aggregation were implemented in the JAVA language.
Visualization and GUI. For generating visualizations, we used the BRUNEL VISUALIZATION3 API. The entire graphical user interface was programmed using JAVA’s SWING API. 4
As outlined in the following, there are two key use cases that we would like to demonstrate with DIGITALHISTORIAN. We also describe how the users will be able to interact with DIGITALHISTORIAN using illustrations for the different use cases.
Exploring Search Results. The foremost task that we address with DIGITALHISTORIAN is that of exploring the document collection using the interesting time intervals identified for the query. The main view (Figure 1) of the DIGITALHISTORIAN addresses this by providing the user with a search field to issue keyword queries. Subsequent to the search operation, various interesting time intervals are displayed in a list on the left hand-side of the interface. The list of time intervals are ordered by their interestingness, i.e., how frequently they are generated by the temporal expressions in document contents for the given query. A diversified set of documents is shown in the main display which gives a temporal overview of documents for the query. Each document in the list is depicted by its headline, its URL, a snippet from its contents and the normalized temporal expressions in its contents. Furthermore, the users can double-click the various time intervals in the list to expand the query, so as to obtain more documents concerning it. Unlike many commercial search engines, all of this is done automatically, without imposing any sliders or check-boxes to manually specify relevant time intervals.
Analytical Visualizations. We further construct informative visualizations by contextualizing the interesting time intervals with co-occurring named entities. Currently, there are two analytical visualizations available. Both of them show the frequency of various named entities that occur in different time intervals. The first visualization is a chord diagram (Figure 2a), where an arc is drawn between a time interval and a corresponding named entity that occurs in that time interval. The thickness of the arc is in proportion to the frequency of the named entity in that time interval. The user can also hover over to each individual chord in the graph to see the time interval and the entity it connects and the corresponding aggregate count. The second visualization is a heatmap diagram (Figure 2b), where on the x-axis the different time points and on the y-axis different named entities are plotted. The intensity of the cell in the heatmap shows the frequency of that named entity in that time point. The user can further drill up from years to decades and drill down from years to days by scrolling on the time axis to inspect the different time intervals with their respective frequency of named entities. 1https://catalog.ldc.upenn.edu/LDC2008T19 2https://www.elastic.co/ 3https://github.com/Brunel-Visualization/Brunel Gupta, Stro¨ tgen, and Berberich (a) In the chord diagram, we can see the key entities by large chords corresponding to iraq, george w bush, and iran. Iraq was involved in multiple conflicts which can be seen in the highlighted boxes, i.e., the American invasion of Iraq in 2003 and the Iran-Iraq war of 1980-1988. The most (b) In the heatmap diagram, each cell shows the frequency of the named frequent entities are highlighted in the legend entity in the corresponding time point.
There exists few demonstrations that utilize named entities and temporal expressions for search and
analysis of documents. None of them put any emphasis on the understanding and analysis of temporal
expressions in document contents as we have done in this paper. Some of the challenges that we
addressed, have been described in detail by Tahmesebi et al. from the digital humanities perspective [
In this paper, we demonstrated DIGITALHISTORIAN, a system that is able to analyze temporal expressions in document content to generate interesting time intervals which are subsequently used to re-rank and diversify documents to give a historic overview for the issued query. It also offers capabilities to analyze frequent named entities in the YAGO knowledge graph for informative visualizations. DIGITALHISTORIAN thus provides scholars in digital humanities an informative and innovative way of exploring semantically annotated document collections.