AIRI Institute Develops AI to Identify Tolkien Characters by Textual Clues
Researchers at the AIRI Institute for Artificial Intelligence have trained an AI system to recognize characters from John Ronald Reuel Tolkien’s works by analyzing the characters’ lines and the descriptions surrounding them. The goal is to determine how characters relate to one another, based on their dialogue and the narrative context. This finding was shared with socialbites.ca in an AIRI briefing.
For the training dataset, the team drew on Tolkien’s authored corpus, including posthumously edited notes overseen by Tolkien’s son. The collection comprises material from History of Middle-earth, The Lord of the Rings, and The Hobbit. The training approach involved labeling key passages and character references to teach the model how to map textual cues to character identities and inter-character relationships.
Using markers to segment the source material, researchers extracted 156,482 sentences and compiled a list consisting of 518 character names, 15 race indicators, and relevant biographical facts in a process described as manual tuning. After refining this list, the dataset used for analysis reached 880 items, providing a robust basis for the AI to learn from textual patterns and character dynamics.
In subsequent steps, the team outlined an automated analysis workflow for examining artifacts. The sequence of procedures demonstrated in the work can be applied to solve a range of practical tasks. For instance, it could be used to interpret regulatory documents or to summarize legal texts in plain language that is accessible to a general audience. This approach helps transform dense material into clearer, more comprehensible content.
The technology holds potential for training speech-enabled assistants and translation systems. It also promises efficiency gains for users who need to search through large volumes of text, saving time while improving search relevance and comprehension. The described methods may support easier information retrieval, especially in extensive document collections. [Citation: AIRI Institute] [Citation: Socialbites.ca]