Abstract
The progressive digitization of historical archives provides new, often domain-specific, textual resources that report on facts and events that have happened in the past; among these, memoirs are a very common type of primary source. In this article, we present an approach for extracting information from Italian historical war memoirs and turning it into structured knowledge. This is based on the semantic notions of events, participants, and roles. We evaluate quantitatively each of the key steps of our approach and provide a graph-based representation of the extracted knowledge, which allows to move between a Close and a Distant Reading of the collection.
Introduction
The "Distant Reading" paradigm in Digital Humanities often relies on coarse tools like topic modeling or simple keyword searches. While these provide a bird's-eye view of a collection, they often lack the granular context historians require. For a researcher studying the Italian Resistance, knowing that "arrests" is a common topic is insufficient; they need to identify exactly who was arrested, by which militia, in which city, and be able to jump directly to the source text.
Historical texts present unique challenges: they are often the product of Optical Character Recognition (OCR), contain domain-specific jargon, and feature significant named entity ambiguity (e.g., distinguishing between a city and a partisan's nom de guerre). Furthermore, standard linguistic resources like FrameNet are often incomplete for Italian, complicating standard Semantic Role Labeling (SRL).
Methodology: The Information Extraction Pipeline
The researchers developed a multi-stage Information Extraction (IE) pipeline tailored for Italian war memoirs.
Preprocessing and Dependency Parsing
The system uses the TINT pipeline (an Italian-specific suite based on Stanford CoreNLP) for tokenization and dependency parsing. This step is crucial for identifying the syntactic structure of sentences, which serves as the foundation for semantic role assignment.
Named Entity Extraction & Disambiguation
To handle ambiguity, the system employs a two-fold approach:
- Cross-class ambiguity: A k-Nearest Neighbor (k-NN) classifier determines if a word is a Person, Location, or Organization based on context.
- Intra-class ambiguity: An iterative process uses co-occurrence scores to resolve specific entities (e.g., which "Marco" is being mentioned) based on other entities in the sentence.
Event & Role Labeling
The core of the system is a custom event-predicate dictionary. It maps lexical units (verbs or nouns) to 88 event classes (e.g., DEPARTING, KILLING, ARREST). It then assigns semantic roles (e.g., Agent, Victim, Source) based on the syntactic dependencies identified during preprocessing.
The authors found that a relatively small dictionary of "lexical patterns" can cover a high percentage of event mentions. By focusing on the subject-verb-object relationship, they could accurately assign roles even when the vocabulary varied.
Datasets
The study utilizes four distinct subcorpora:
- Memoirs: 25 digitized books by Italian partisans (~1.5 million words).
- Biographies: 189 short records from Wikisource used for lexical expansion.
- Wiki-Articles: 1,748 Italian Wikipedia articles related to the Resistance.
- Memoirs-test: A "gold standard" set of 3 books (112,000 words) for evaluation.
Key Results
- Entity Disambiguation: The system achieved high performance for Locations (0.93 F1) and Organizations (0.92 F1). Person disambiguation reached an 0.74 F1.
- Semantic Tagging: An embedding-based classifier reached 75.4% accuracy for identifying the semantic type of event participants.
- Event Extraction: On the test corpus, the system achieved 0.78 Precision but lower Recall (0.50). The lower recall is attributed to the metaphorical or "diaristic" language common in memoirs. Precision jumped to 0.89 when the system identified at least two matching arguments for an event.
Conclusion and Future Work
This research demonstrates that structured IE can bridge the gap between macro-statistical trends and specific textual evidence. By representing memoirs as a graph, researchers can toggle between "Distant Reading" and "Close Reading."
Future work will address current limitations, including improving handling of non-literal language, implementing anaphora resolution to link pronouns back to entities, and developing event coreference to collapse multiple mentions of the same historical fact across different books.
Summary updated