While the use of computational methods in corpus-level analysis has become ubiquitous in recent years, the optimal integration of these tools into digital archiving and corpus presentation remains an open question. Massive archives such as JSTOR, Gale, and Google Books have experimented with computational tools for traversing and exploring underlying documents in their collections.[3] Meanwhile, progress in machine learning (ML) and natural language processing (NLP) continues to advance the state-of-the-art for the underlying computational tools and methods these archives employ. However, there is a critical disjuncture between the datasets these latter tools and methods are trained upon and the real-world needs and expectations of digital archives [Jo and Gebru 2020] [Birhane et al 2022]. The application of computational methods to archival research thus risks both misrepresenting the findings of models and techniques, on the one hand, and misconstruing the exigencies of the underlying archive on the other. These concerns have resulted in a trend toward out-of-the-box computational tools that are not well integrated with the underlying archive and often obscure rather than explain interesting phenomena. Computational tools in archival analysis are thus frequently designed, implemented, and presented as abstractions of the archive. As such, these tools rarely serve as more than a heuristic method of inquiry, failing to meet the higher expectations of either computational humanities scholarship or humanistic textual inquiry.

We designed Project Quintessence with these concerns in mind. By bringing together a team of scholars intimately acquainted both with computational analysis in NLP and with the texts and historical context of our archive, this project has attempted to build a corpus exploration tool from the ground up that fully integrates computational tools and methods with the specific archive in question. For our corpus, we chose the Early English Books Online Text Creation Partnership (EEBO-TCP) archive, a collection of 60,331 texts transcribed and tagged with TEI markup by the TCP team. Each model we trained and technique we incorporated was designed to take advantage of this highly idiosyncratic archive. Rather than decoupling textual analysis from the texts themselves, we incorporated these tools into the very architecture of the site. The result is a tool that allows one to dynamically interact with the archive in a way that is not possible with any one method independently, and which we believe presents a new model for archival research and analysis. By presenting multiple, complementary tools side-by-side for use by domain experts, we allow for the tools to be checked against each other and also compared with the underlying archive. We believe that this framework provides a stable site for the continual improvement of knowledge about the archive and domain alike.

In this paper, we first contextualize Project Quintessence within a longer trajectory of digital archival research tools, and then describe how the various elements of the project cohere by employing our concept of dimensionality. We continue this discussion by outlining the project pipeline, including model training parameters and other considerations for each of the models and tools included. Finally, we explore some potential future directions for our project and for database visualization frameworks in general. We argue that an attention to textual dimensionality allows for an interdependent database framework that augments research at any scale of analysis and has the capacity to both to confirm the doxa of pre-existing research while also providing surprising results that enrich the body of knowledge in the field. We believe that our web tool accomplishes all of these charges.

Project Quintessence is built using well-established, highly nimble modeling techniques; primarily Latent Dirichlet Analysis (LDA) Topic Modeling and Word2Vec Word Embeddings. By leveraging dynamic visualization libraries such as Plotly, we were able to extract a high level of archival information from each of our models. Each dynamic visualization on the site can accommodate any number of research questions or approaches. In addition, by synthesizing approaches from multiple techniques on the same archive, we were able to harness the affordances of each method rather than attempting to answer all questions with a single approach. We describe below how these methods differ from more contemporary techniques, such as Structural Topic Models (STM), Dynamic Topic Models (DTM), Document Embeddings, Contextual Embeddings (CoVe), Transformer Embeddings (such as BERT), and Stacked Embeddings (such as Flair); most of which we used at different stages of our design process to supplement our findings and enrich our analysis. Overall, the trend toward large language models and massive amounts of masked perimeters has introduced elements of bias and structuring assumptions which do not hold true for all datasets, disguising crucial inferences the model makes based on its training data rather than the object being interpreted [Bender et al 2021] [Chun 2008]. By incorporating historical lexicographic tools like MorphAdorner into our text cleaning pipeline and generating many smaller language models on specific cross-sections of the archive, we were able to preserve germane distinctions within the archive while still providing a mode of ingress into its analysis.

The use of visualizations in representing digital archives has become more nuanced with the recent increase in critical attention to these methods. Most of the early literature on data visualization in the Digital Humanities has remarked on the late adoption of visualization techniques by researchers in the humanities, a trend that appears to be justified when examining naive approaches to what visualizations can and do tell us of the underlying data, and the extent to which that lens is mediated and controlled by its designers. However, as media studies and digital humanities scholarship have produced compelling critiques of the unmediated view of visualization, alongside compelling design objectives for researchers and engineers, dynamic visualizations are increasingly seen as a site for emancipatory politics and critical inquiry.

To that end, we have framed our project’s intervention as one of harnessing new technological advances in ways that opens rather than closes archives to countervailing narratives and critique of canon. In our design practices, we asked how digital methods might appropriately interact with digital textual archives. Michael Witmore has aptly described this relationship in terms of addressability, and describes archives of texts as digital objects that are, “massively addressable at different levels of scale” [Whitmore 2010]. We suggest that the oscillations of scale afforded by dynamic exploration tools could be best considered by thinking in terms of dimensionality. As described in David J. Staley’s influential account, the capacity for visualization techniques to augment the affordances of text can be understood of as an increase in dimensionality [Staley 2015]. Staley frames visualization as a way of breaking out of the implied linear one-dimensionality of linguistic syntax. However, we can expand upon this account. As Peter Gärdenfors has suggested, treating semantic meaning as a conceptual space with abstract, dimensional relations allows for complex semantic representation and analysis [Gärdenfors 2014]. Word embeddings, for instance, are commonly employed to locate clusters of semantically similar words by assigning spatial coordinates to those words based on hundreds of dimensions of training data. Following Gärdenfors, we do not treat dimensionality rigidly as a geometrical principle, contrasted with hierarchical or ordinal relationships, but instead as a higher-level abstract relationality of terms. By allowing ourselves to depart from three-dimensional thinking, dynamic visualizations are freed from their metaphorical relationship with abstract, Cartesian space, and instead function as tools for making tangible the many relationships between words with startling complexity.

Over the course of our project, we have increasingly become informed by this dimensional thinking. Describing syntactic, semantic, and thematic relationships in texts as dimensional relationships insists on the reality of those relationships. By virtue of the relational character of meaning-making, individual uses of words are necessarily and inevitably put into relationship with other words. While any computational attempt at measuring or visualizing those relationships is necessarily an approximation of varying validity, these computational tools are nevertheless measuring real relationships. Thus, we argue that maximizing researcher access to textual dimensionality should be a primary objective for database design and visualization. This emphasis on dimensionality has been influenced by many earlier digital projects that we will discuss in three groupings; large scale digital archives, small-scale digital projects, and pre-existing work on the EEBO-TCP corpus. Our objective was to design Project Quintessence as a coherent visualization tool that incorporated lessons from each of these categories.

Recent work in the digital humanities has rightly critiqued the over-reliance on spatial metaphors of analysis through terms like “close-reading” or “distant-reading” [Da 2019] [Bode 2018]. Project Quintessence fully incorporates these critiques into its design. As such, conventional notions of scale collapse entirely when using our web tool for textual analysis: at the same time as a researcher is engaging in a close reading of a text, they can inspect the semantic associations of any word in the document with other documents written by that author, in that decade, or printed at that print location. At the same time as one collects a personal archive of documents using our subsetting tool, one can also inspect the thematic topic distribution of that archive, and compare it to the distribution of topics in the corpus as a whole. These authorial, temporal, spatial, and thematic contexts allow for a renegotiation of how scale operates in textual analysis. Rather than having to choose between the closeness of a “human scale” or the distance of a “computational scale,” Quintessence instead encourages a multiscalar view of the corpus by emphasizing different units of analysis — word, document, or topic — each providing a unique approach to exploring the entire archive.

In summary, Project Quintessence de-emphasizes discussion of scale that predominate in digital archival contexts, favoring instead a focus on dimensionality . Prompted by the computational tools incorporated into our project, we describe token-, document-, and topic-level relationships in the archive as real and meaningful discursive relationships that can be explored through quantitative and inferential analysis despite not being reducible to the latter. The goal of computational modeling and data visualization can and should be the expansion of textual dimensionality in productive and meaningful ways. Moreover, we believe these visualizations should prioritize utility and interpretability for actual domain experts. That is why we chose tools and designed the site based on our own research priorities and those of our advisory team. In conversations with early users of Project Quintessence, we have observed that the site becomes more useful the better one understands the EEBO archive; the inverse trend of many out-of-the-box methods.

We hope in this paper and through the Quintessence site to have contributed to the design of future archival research frameworks through an emphasis on dimensionality and dynamic exploration that interfaces with texts. As with any project at this scale, limitations on resources and time have necessarily influenced the scope of the site itself. Nevertheless, continued advancements in textual analysis tools present many exciting opportunities for future enhancements.

Our “Document Viewer” prototype (Figure 13) perhaps best represents the future direction of our project. We are designing this tool to be fully integrated into other models and visualizations on our site as a dynamic overlay to any text that is accessed from our database. Through an automatically-generated element tagging method, topic, part-of-speech, and named entity information will be available as one is examining the text itself. The Document Viewer figures this relational data for tokens in the corpus as latent dimensions within the texts themselves that can be studied and visualized. As with every other tool on our site, one can seamlessly navigate between this textual view and the models discussed above, which we believe mirrors the recursive nature of the archival research process.

As mentioned above, a project to produce a ground-up bidirectional encoding or contextual embedding model for the EEBO-TCP would enable a trove of new tools and methods. Such a project might help to integrate top-down topic modeling approaches with bottom-up embeddings, like in the Neural Topic Modeling approaches of Wang et al. and Gupta et al., or phrase- and document-focused comparison tools like phrase vectorization, Gensim’s Doc2Vec, or Kusner et al.’s Word Mover’s Distance [Wang et al 2018] [Gupta, Chaudhary, and Schütze 2021] [Kusner 2015]. These approaches could radically alter researcher interactions with texts if vector data were built into document indexing through a program like Apache Lucene. A researcher could plausibly search or access dynamic maps of phrases, documents, topics and more based on semantic similarity. A vector-based approach has also been fruitfully paired with image and sound tagging, and our analysis of the latter could be enhanced by reinserting these media into their textual contexts.

The normalized tagging necessary to produce such an embedding model could be used for network analysis of named entities, enabling search and categorization by metadata categories such as author, publisher, and location, and even diegetic textual categories like character, historical figure, location, or chronology. A synthesis between diachronic embeddings and diachronic topic modeling, such as Blei et al.’s Dynamic Topic Model (DTM) or Roberts et al.’s Structural Topic Model (STM) could help enrich our understanding of the complexity of discursive change beyond a simple word-based or topic-based approach. We have also experimented with the use of Gerrish et al.’s Document Influence Model (DIM) to map the persistence and change of topics over time linked to the publication of specific texts. This modeling technique would only be enhanced by integrating embeddings with topic modeling, and could assist with recovery projects and questions of canonicity and biases in scholarly attention.

While the above by no means exhausts the potential future directions of the Quintessence framework, we hope it paints a picture of the rich potential for future archival research that harnesses textual dimensionality. As we extend the Quintessence framework beyond the EEBO-TCP archive to inform database design on other archives and in other domains, we hope to reverse the trend toward massive and opaque automated textual modeling and encourage critical and scholarly apprehension of an ever-growing archive of digitized historical material.