TL;DR
This paper introduces IPSA, a comprehensive, structured corpus of Italian parliamentary debates spanning 175 years. By digitizing historical documents and meticulously tagging speakers, IPSA unlocks a rich historical and linguistic resource for computational analysis, addressing the long-standing inaccessibility of these valuable texts.
What Problem Does This Paper Address?
Parliamentary debates are a treasure trove of information, offering insights into political science, historical societal shifts, and language evolution. However, for many countries, including Italy, these historical records have remained largely inaccessible in their original, often paper-based, formats. This lack of digital accessibility hinders their use in computational linguistics, historical analysis, and the development of advanced AI models. The authors tackle this by creating IPSA, a structured, machine-readable corpus of Italian parliamentary debates from 1848 to the present, making this vast historical data available for research.
Methods
The creation of IPSA involved several key steps:
- Data Collection: Documents were sourced from the official websites of the Italian Parliament (Camera dei Deputati and Senato della Repubblica). For pre-1996 documents, which were primarily scanned PDFs, a multi-stage process was employed.
- Optical Character Recognition (OCR): Tesseract, an open-source OCR engine, was used to convert scanned PDF pages into digital text. This process was computationally intensive, requiring parallel processing on multiple machines to complete within a reasonable timeframe.
- OCR Error Correction: Initial attempts to automatically correct OCR errors using a spell-checker (SymSpell) proved counterproductive, leading to worse results. Therefore, the first version of IPSA is released without extensive OCR post-processing, acknowledging potential "silver" quality for older documents.
- Speaker Tagging: A crucial aspect of IPSA is the automatic tagging of speeches with the corresponding politician. This was achieved by leveraging the Linked Open Data (LOD) structures provided by the Italian Parliament (specifically the OCD ontology for the Chamber of Deputies). A fuzzy matching algorithm (fuzzywuzzy) was employed to handle variations in speaker identification across different legislative periods and potential OCR errors.
Datasets
The primary dataset introduced is IPSA (Italian Parliamentary Speeches and other Acts). It comprises:
- Parliamentary Debates: Spanning 175 years (1848-2022), containing approximately 1.2 billion tokens.
- Standing Committee Reports and Speeches: An additional 1.2 billion tokens.
- Law Proposals and Amendments: Also included.
The paper also references and compares IPSA to several existing parliamentary corpora, including:
- DutchParl
- GerParCor (German Parliament Corpus)
- Polish Parliamentary Corpus
- US Congressional Record
- CzechParl
- siParl (Slovenian parliamentary corpus)
- Aalto Finnish Parliament ASR Corpus
- ParlaMint (multilingual corpus)
Key Insights and Ideas
- Historical Data is Underutilized: The paper highlights that despite their immense value, historical parliamentary records are often inaccessible and thus underutilized for computational research.
- OCR Challenges for Historical Documents: The "silver" quality of OCR for older, scanned documents is a significant hurdle. Even advanced tools can struggle with historical printing errors, and automatic correction can sometimes degrade quality.
- Linked Open Data for Speaker Identification: Leveraging existing LOD structures (like the OCD ontology) is a powerful strategy for accurately tagging speakers, even with variations in historical formatting.
- Fuzzy Matching is Essential: The use of fuzzy string matching is critical for robust speaker identification, especially when dealing with OCR errors and diverse naming conventions across legislative periods.
- Trade-offs in Data Processing: The decision to release IPSA without extensive OCR correction for older documents reflects a pragmatic approach to balancing data completeness with quality assurance, acknowledging that some level of error is inherent in historical digitized text.
Key Results
- Corpus Size: IPSA contains approximately 2.4 billion tokens in total, with about 1.2 billion tokens from Italian parliamentary debates.
- Speaker Tagging Performance: The speaker tagging task achieved a global F1 score of 0.909. Precision was generally higher than recall, indicating that when a speech was identified, the politician was usually correctly tagged, but the identification of when a speech started was a greater challenge.
- OCR Error Trends: Analysis of OCR error rates (WER and CER) showed a general downward trend over time, indicating improved readability of more recent documents. However, specific legislative intervals showed peaks in errors, attributed to issues with document structure (e.g., line and paragraph separation).
- Data Format: Debates are released in text-only, XML (with speaker tags), and RDF formats. Committee reports and law proposals are provided in text-only format.
Limitations and Future Work
- OCR Quality for Older Documents: The paper acknowledges that documents predating 1996, processed via OCR, are prone to errors and are considered "silver" quality.
- Tagging Recall: The recall for the speaker tagging task is lower than precision, suggesting that identifying the start of new speeches remains a challenge.
- Future Work: The authors plan to implement state-of-the-art methods for improved OCR correction and tagging. Specifically, they aim to increase recall in tagging by splitting the task into speech identification and politician attribution. Exploring byte-level approaches for OCR error correction is also a future direction.
Summary updated