Visualization for Changes in Relationships between Historical Figures in Chronicles Masahiko ITOH Mina AKAISHI Institute of Industrial Science Faculty of Computer and Information Sciences The University of Tokyo, Japan Hosei University, Japan [email protected] PRESTO, Japan Science and Technology Agency [email protected] Abstract Our framework first provides a mechanism for extract- ing networks of historical figures from historical data and A huge number of historical documents that have been for visualizing time-varying changes in the structure of net- accumulated for a long time are currently being digitized. works along with a timeline in a 3D space. It also allows However, it is difficult for us to analyze and obtain insight users to extract networks of people with multiple aspects into the past from such documents. This paper proposes such as “battle” or “gift” and to compare the differences an interactive visualization system to extract networks of and similarities in their evolutions. Users can observe the historical figures from historical data and to show time- appearances of groups of people, the important persons in varying changes in their relationships. It enables users to these groups, and changes in the structure of the groups. explore changes in the structure of the network interactively. It is desirable to grasp not only structural changes but Moreover, it extracts characteristics of each relationship, also characteristics of relationships between people, such as such as hostile or friendly relations, and visualizes them on friendly or unfriendly relationships. We therefore provide the network. It enables us to understand changes in past so- mechanisms for extracting components of relationships and ciety better by exploring changes in relationships between visualizing them by coloring them different colors. This en- people. ables users to observe what kind of characteristics a group of people had. It also enables users to observe temporal changes in characteristics of relationships between two peo- 1 Introduction ple, e.g. the users can find out that two generals changed from being allies in 1561 to opponents in 1564. A huge number of historical documents that have been The contribution of our system is that it interactively en- accumulated for a long time are currently being digitized, ables users to explore temporal changes in networks of his- enabling us to search for and access such documents easily torical figures and in the details of relationships between and rapidly. However, it is difficult for us to analyze and figures in the networks. obtain insight into the past from such documents by sim- ply using statistical techniques. Exploring such data from various viewpoints requires an interactive environment. 2 Framework for Visualizing Changes in Re- In this paper, we propose an interactive visualization lationships between Historical Figures framework for exploring changes in relationships between historical figures that are extracted from historical docu- This paper proposes a framework for extracting networks ments. It enables us to support understanding of changes in of historical figures and characteristics of relationships be- past society through exploring changes in relationships be- tween such figures from historical documents and for visu- tween people, such as the appearance of groups of people, alizing and interactively exploring temporal changes in such or whether such groups had friendly or unfriendly relations. relationships (Figure 1). The proposed system allows us to check whether visualiza- Figure 1 shows an overview of our visualization frame- tion results accurately represent already-known knowledge work. A historical document database (Figure 1(a)) stores and explore novel phenomenon or patterns that are not al- records of events that have attribute values such as names of ready well-known. people, keywords, and dates related to the events. Users can select records related to a specific viewpoint in which they Viewpoint 1560 㻨㻱㼞㼍㻘㻌㻼㼑㼞㼟㼛㼚㼟㻘㻌㻷㼑㼥㼣㼛㼞㼐㼟㻪 a 1561 Represenng a person as a node, 㻔㻮㼍㼠㼠㼘㼑㻘㻌㻳㼕㼒㼠㻘㻌㼑㼠㼏㻕 b 1562 are interested by specifying ranges of years, lists of peo- connecon between people ple, and/or keywords. The system then generates a network c as an edge of historical figures from the selected records by extracting d e g f 1560 a 1561 co-occurring people in the same events (Figure 1(b)). If two b 1562 people have a high co-occurrence ratio, they have a connec- (b)Extracng a me sequenal h 䞉䞉䞉䞉䞉 c 䞉䞉䞉䞉䞉 䞉䞉䞉䞉䞉 network of people 䞉䞉䞉䞉䞉 tion. Users can extract networks related to specific people 䞉䞉䞉䞉䞉 c d e such as ODA Nobunaga, who was the best-known feudal (a) Historical Doc DB g lord in Japanese history, or TOKUGAWA Ieyasu, who was bale (d) Visualizaon of changes the founder and first shogun of the Tokugawa shogunate, by rebellion in relaonships inputting their names. They can also extract networks re- peace 㻨㻱㼞㼍㻪 lated to a battle in the warring states period in Japanese his- present Assigning cluster type to tory (called Sengoku Jidai) by inputting the range of years (c) Extracng clusters of keywords for colors of edge related to Sengoku Jidai and keywords such as “battle”, “at- represenng relaonships between people tack”, and/or “rebellion”. Moreover, they can extract and compare networks related to various viewpoints. Figure 1. Overview of our visualization frame• Our framework allows users to extract a network of a work for changes in relationships between specific viewpoint with a specified time window, e.g. every historical figures. year. It enables users to observe sequential changes in the network of people. The historical document database stores keywords sum- marizing events. We can extract information about why two 4 Extracting a Network of Historical Figures or more people have connections by using events and key- words related to them. However, keywords are too diverse Our framework enables us to extract networks of people for visualizing the reasons for connections. We therefore from a specific viewpoint. We first filter records out from extract clusters of keywords and assign colors to extracted a database by keywords, range of date, and/or persons for clusters (Figure 1(c)). specifying the viewpoint and then aggregate filtered records every year. We next extract the networks of people for every Our framework then visualizes a time-varying network year on the basis of co-occurring people in the same event. extracted in Figure 1(b) with assigning a person to a node More specifically, we define a connection between peo- of the network and a connection between people to an edge ple depending on person dependency. The following for- (Figure 1(d)). It also assigns colors to the edge that has mula defines the person dependency from one person to an- characteristics defined by co-occurring clusters of keywords other in a particular year y, which is an extension of term between two people (Figure 1(d)). This enables users to dependency described by Akaishi et al. [2, 1]. explore temporal changes in relationships between people. 0 0 recordsy(p \ p ) pdy(p; p ) = (1) recordsy(p) 3 Historical Document Database 0 recordsy(p \ p ) represents the number of records rep- resenting events in which persons p and p0 both appear We utilize the historical database Dai-Nihon Shiryo in year y. recordsy(p) represents the number of records 1 0 Database . The Dai-Nihon Shiryo is a collection of histor- in which only person p appears. If pdy(p; p ) ≥ α and 0 ical documents arranged chronologically and dating from pdy(p ; p) ≥ α, then we define a bidirectional edge between 0 0 0 the ninth to the seventeenth centuries. It has been published people p and p . If pdy(p; p ) ≥ δ and pdy(p ; p) ≥ µ, then 0 by the University of Tokyo since 1901. Currently, it is be- we define a edge from person p to person p 2. ing digitized and compiled into a database. Each record in Size and color of a node represent the importance of a the database represents one event. It includes the date of person. The size of node isP defined by an attraction power 00 the event, a list of people related to the event, a list of these of person p described by p002p pdy(p ; p), which means people’s official titles, a list of place names related to the the summation of other people’s person dependency on per- event, a list of keywords representing the event, and text as son p. The color of the node is defined by the ratio of the attributes. The current database consists of about 230,000 number of inlinks to outlinks of the node. We use the fol- records. lowing colors as default colors that can be modified by a visualization system mentioned at Section 6.1. 1http://wwwap.hi.u-tokyo.ac.jp/ships/db.html 2We use α = 1 and δ = µ = 0:8 in this paper. • #In Link > #Out Link : dark purple Cluster label Keywords Battle Ikko, Shin Buddhism, Follower, Bat- • #Out Link > #In Link : light purple tle, Rebellion, Crucifixion, Naval bat- • #In Link = #Out Link : gray tle, Navy,... Gift Whale meat, Gift, Persimmon, Wild The length of an edge represents the strength of a rela- goose, Garden lantern, Present,... tionship between two people. It is defined by the difference Allowance Loyalty, Allowance, Supply,... 0 0 in person dependencies pdy(p; p ) − pdy(p ; p). If person Grant Incense, Grant,... dependency is strong, the edge shortens. If one person one- Conferment Commission, Conferment sidedly depends on another, the edge lengthens. Succession Successor, Succession, Head of fam- ily 5 Extracting Characteristics of Relationships ... ... between Historical Figures Table 1. Example of keyword clusters. We first extract keywords in events related to two people to represent characteristics of their relationship. We then add such keywords as labels to their relationships and at- of keyword clustering described in Section 5.1.