No source: created in electronic format.
The poster presents an ongoing PhD research project that applies Digital Humanities to Classics. The project is focussed on the extraction of information from modern scholarly texts (i.e. “secondary sources”), namely all the modern publications about ancient works written in Greek or Latin (being our so-called “primary sources”). The project addresses both the problem of extracting information fom scholarly texts in an automatic and scalable way, and that of providing users (i.e. scholars) with advanced and meaningful entry points to information rather than just search engine-like functionalities over an electronic corpus of texts.
A currently ongoing project such as the Million Book Library drew considerable attention to the characteristic features of the next generation of digital libraries and to the consequences of a change of scale on the practice and the results of the research itself (Crane 2006). This is a big chance for Humanities and Classics particularly given the extended “shelf-life” of humanities relative to scientific publication. However, once those secondary sources are digitized this does not mean that information contained within them will be immediately accessible. Issues that we need to address concern the accuracy of our electronic resources, such as encoding of Greek text, inaccurate OCR transcription due to low image quality, problem of missing pages (Boschetti 2009), as well as the scalability of ways to provide access to information, given that we cannot afford to correct manually every scanned page.
Secondary sources are without any doubt intrinsically valuable for Classicists,
as they shape the scholarly discourse of the discipline. Printed citations and
references – and even mentions of names or geographical places – can be
considered as being already a form of hypertext as they virtually create links
between texts. However in the currently available digital libraries – except for
the Perseus Digital Library
In particular, Classics scholars are interested in named entity mentions inside texts, as is reflected by the widespread use of different kinds of indexes and concordances in this field that basically organize in a systematic manner references to text passages when a given entity is mentioned. Translating the problem into computational terms, we are faced with the task of automatic entity extraction from a corpus of unstructured texts to develop a discipline-specific system of semantic information retrieval.
In addition to all the unstructured information being made available on the web, several projects in the Humanities have produced over the last decade an increasing amount of structured information that was then stored in a wide range of data formats (i.e. databases, XML fles, etc.). The approach we undertake is to reuse information contained within those structured data sources as training data for a supervised system that extracts semantic information fom an unstructured corpus of texts. A likely approach is suggested for instance by a research recently conducted by IBM to investigate the automatic creation of links between structured data sources (i.e. database containing product information) and unstructured texts (i.e. emails of complaint about products) (Bhide et al. 2008). More generally, the problem implied by our approach of determining when two bits of information refer to the same entity has been thoroughly explored in the AI (Artifcial Intelligence) field (Li et al. 2005).
The very first phase of the project is devoted to the task of building our corpus
of unstructured texts. So far we considered two corpora of texts: the papers
contained in the open archive Princeton/Stanford Working Papers in
Classics
In the second phase we integrate different structured data sources into a single
knowledge base. As far as this task is concerned, it is possible to observe at
least two main categories of lack of interoperability. The first category
consists of cases where entities that are similar from an ontological
perspective (e.g. a geographical place name) are encoded using different data
structures (i.e. the same place name could be encoded using elements belonging
to diferent XML dialects). The second category covers cases where chunks of
information common to more than one collection are described with different
degrees of depth and precision. For instance, the name “Alexandria” inside an
inscription is just marked up as a place name where instead a collection of
geographical data offers many details for the city of Alexandria such as
coordinates, orthographical variants of the name, denomination in diferent
languages etc. For this purpose we mean to apply high level ontologies to
aggregate information related to the same entity but spread over diferent data
sources. Among the most suitable ontology vocabularies are worth to be mentioned
FOAF,
At a further stage we are taking into account how to automatically extract information from the corpus. We are mainly interested in extracting: 1) named entities; 2) bibliographic references; 3) canonical references, namely references to ancient texts expressed in a concise form and characterized by a logical reference scheme (e.g. based on references to books or lines of a work instead of page numbers).
The very first step in named entities processing is the recognition and identification of named entities within texts. Once identified the named entities should be classified and then disambiguated on the basis of the context. This task will be accomplished through the comparison of semantic spaces using methods and algorithms developed in the field of Latent Semantic Analysis (LSA) (Rubenstein & Goodenough 1965; Sahlgren 2006). In particular the semantic spaces of all the contexts where a given named entity appear will be compared with each other in order to determine which resources are really referring to the same entity.
For the information extraction task we use mainly tools based on machine learning methods, such as Conditional Random Fields (CRF) or Support Vector Machine (SVM). Some of them are already available as open source sofware and just need to be trained to work with our data, while others are being specifcally developed such as a Canonical Reference Extractor (Romanello et al. 2009). Information contained in the knowledge base is meant to be used as training material for those sofware components.
This ongoing project is expected to prove a scalable approach to extract entity references from unstructured corpora of texts, such as OCRed materials. In addition to this, it will prove how to reuse several data sources of structured information to train text mining components and it will show to what extent those data sources can be made interoperable with each other. Finally we plan to evaluate with some experts how effective such an information retrieval system is in helping Classicists with their research.
Acknowledgement
This research project is supported by an
AHRC (Arts and Humanities Research Council) award.