Daejeon, July 27–31
Literary historiography often teaches conflict poetry through a narrowed “event hook”: a poem is anchored to a single canonical episode (or, conversely, dissolved into a generalized war-context), and the interpretive labor proceeds as if historical reference were either uniquely determinate or indefinitely diffuse. In Troubles writing (1960s-1990s Northern Ireland sectarianist conflict), this pedagogical habit is especially consequential because the archive of violence is dense, segmented, and unevenly remembered – what is culturally “available” for interpretation tends to default to a small set of headline events. Our project, “Retrieving Trauma”, asks whether AI-driven archival retrieval that is designed as an interpretable research instrument (rather than a generative co-author ) can expand the contextual field of close reading, surfacing subtle and cross-temporal historical resonances that are typically missed when scholarship (or teaching) begins from titles, pre-selected chronologies, or keyword search alone.
This work is positioned within three overlapping conversations in digital humanities and literary studies. First, it responds to the long arc from “distant reading” to post-distant interpretive practice (Moretti, 2000). The point for our project is not to abandon close reading, but to repurpose distance as contextual access, a way proposed to let a reader move between micro-interpretation and macro-archive without pretending that contextual discovery is magically produced by unaided scholarly memory. Second, it builds on DH’s now-explicit theorization of search as epistemic practice. Underwood argues that “[q]uantitative methods have been central to the humanities since scholars began relying on full-text search to map archives,” (Moretti, 2000). The point for our project is not to abandon close reading, but to repurpose distance as contextual access, a way proposed to let a reader move between micro-interpretation and macro-archive without pretending that contextual discovery is magically produced by unaided scholarly memory. Second, it builds on DH’s now-explicit theorization of search as epistemic practice. Underwood argues that “[q]uantitative methods have been central to the humanities since scholars began relying on full-text search to map archives,” (Underwood, 2014) while also noting how the intellectual implications of search are rendered opaque when algorithms are treated as mere tools. Our team take this as a methodological invitation: to treat retrieval not as infrastructure “outside” interpretation, but as a formalized part of interpretive method. Third, it engages trauma theory’s account of belatedness and incomprehension. Caruth characterizes traumatic experience as “experienced too soon, too unexpectedly, to be fully known,” and “not available to consciousness” in the moment (Underwood, 2014) while also noting how the intellectual implications of search are rendered opaque when algorithms are treated as mere tools. Our team take this as a methodological invitation: to treat retrieval not as infrastructure “outside” interpretation, but as a formalized part of interpretive method. Third, it engages trauma theory’s account of belatedness and incomprehension. Caruth characterizes traumatic experience as “experienced too soon, too unexpectedly, to be fully known,” and “not available to consciousness” in the moment (Caruth, 1996), only returning later through repetition. For literary studies, this suggests that conflict poems are not merely documentary correlates of discrete events but artifacts of cultural memory that encode what communities can and cannot know at given moments. For DH, it motivates a retrieval workflow that helps readers traverse the “knowing/not-knowing” gap by systematically reintroducing historically proximate and semantically resonant contexts. (Caruth, 1996), only returning later through repetition. For literary studies, this suggests that conflict poems are not merely documentary correlates of discrete events but artifacts of cultural memory that encode what communities can and cannot know at given moments. For DH, it motivates a retrieval workflow that helps readers traverse the “knowing/not-knowing” gap by systematically reintroducing historically proximate and semantically resonant contexts.
Technically, we draw on retrieval-augmented architectures but deliberately restrict the system to retrieval and inspection rather than generation. Retrieval-Augmented Generation (RAG) is typically defined as a pipeline combining parametric language modeling with non-parametric retrieval; Lewis et al. describe RAG as “models which combine pre-trained parametric and non-parametric memory.” (Lewis et al., 2020) In practice, however, DH faces a distinct constraint: interpretive accountability depends on traceable evidence, and generative paraphrase can obscure provenance. The RAG literature itself foregrounds provenance as an open problem (“providing provenance … remain open research problems”). For this reason, and because our archive partner’s governance concerns are heightened by any suggestion that texts will pass through an LLM, we implement a retrieval-first / non-generative workflow that returns ranked archival excerpts (with metadata) to the human reader, preserving inspection and citation as core scholarly acts.(Lewis et al., 2020) In practice, however, DH faces a distinct constraint: interpretive accountability depends on traceable evidence, and generative paraphrase can obscure provenance. The RAG literature itself foregrounds provenance as an open problem (“providing provenance … remain open research problems”). For this reason, and because our archive partner’s governance concerns are heightened by any suggestion that texts will pass through an LLM, we implement a retrieval-first / non-generative workflow that returns ranked archival excerpts (with metadata) to the human reader, preserving inspection and citation as core scholarly acts.
Data source and archival rationale
We use the Conflict Archive on the Internet (CAIN) as an authoritative, bounded archive suitable for a pilot that can still yield interpretively meaningful results. CAIN is hosted at Ulster University and provides “information and source material on ‘the Troubles’ (from 1968 to 2007) and politics in Northern Ireland.” (Melaugh, n.d.) CAIN’s Key Events section is explicitly structured for research and teaching: for each event, users can access “a selected reading list, a summary, a chronology, background information, and a list of source material.” (Melaugh, n.d.) CAIN’s Key Events section is explicitly structured for research and teaching: for each event, users can access “a selected reading list, a summary, a chronology, background information, and a list of source material.” ((CAIN Future Secured as a Live Archive, 2023) The archive’s longevity and civic-academic status are also relevant to our methodological claims about memory infrastructures. Ulster University describes CAIN as “freely available on-line since March 1997,” ((CAIN Future Secured as a Live Archive, 2023) with new materials added regularly. We therefore treat CAIN not merely as “data” but as curated historical representation whose structure shapes what becomes retrievable and hence interpretable.
Corpus construction, chunking, and embeddings
With a corpus of 188 Troubles poem, from CAIN, we scrape and normalize Key Event summaries (rather than interviews or long essays) to create an event-centric corpus aligned to the historiographic problem we address: readers frequently begin from events, but our intervention is to allow events to be rediscovered via poetic language, not merely recalled by name. We serialize content into JSON records with event metadata and apply chunking (segmenting longer summaries into smaller overlapping passages) to improve retrieval granularity; chunk boundaries are treated as interpretive design decisions because they shape what “counts” as a retrievable unit of context.
For semantic representation, we use an open-source embedding model. The model is trained via weakly supervised contrastive learning and is presented as “a family of state-of-the-art text embeddings” that can be used for “retrieval, clustering, and classification.” (Wang et al., 2022) In our implementation (selected for stability and accessibility), embeddings are fixed-length vectors, enabling deployment without paid APIs and supporting an AI democracy / lower buy-in ethos for DH tooling. We store vectors in a local vector database and retrieve the top-K most similar chunks (K=5) for each query. (Wang et al., 2022) In our implementation (selected for stability and accessibility), embeddings are fixed-length vectors, enabling deployment without paid APIs and supporting an AI democracy / lower buy-in ethos for DH tooling. We store vectors in a local vector database and retrieve the top-K most similar chunks (K=5) for each query.
Query design: from keyword search to “context search”
We operationalize Underwood’s call to theorize search by moving from direct keyword matching to what is known as semantic retrieval practice that resembles keyword search, expanding recall across paraphrase, vernacular drift, and indirect reference. Our prompt is intentionally simple and open-ended (e.g., “what historical events are related to this poem”), avoiding few-shot steering that would over-impose the researcher’s known event anchors and thereby reproduce the very historiographic narrowing we seek to challenge.
Evaluation protocol and interpretive case studies
We evaluate retrieval in two complementary ways.
Across poems and events, we further build an interactive network visualization that maps poem-event linkages (and their directionality) in the shared embedding space, enabling scholars to interrogate “hub” events that connect multiple poems, locations, and sentiments. In visualization practices, static network diagrams are often utilized, While they are often visually, they remain largely epistemically opaque. The interactivity of our network supports methodological transparency by allowing readers to inspect clusters and links directly. One striking outcome is that less-canonized events can emerge as connective hubs (even when not “the most famous”), inviting a reconsideration of cultural memory’s topology and the archive’s “hidden histories.” This directly supports our theoretical proposal of a trauma continuum: the archive-mediated semantics of conflict are not strictly linear, and later events can resonate backward as interpretive anchors, suggesting a partially circular historical representation rather than a one-way timeline.
Scholarly contribution and significance
We claim contributions at three levels:
By foregrounding a non-generative, open-source retrieval method over an Anglophone LLM “chat” paradigm, and by centering an archive of conflict memory whose structure mediates what can be known, the project expands DH methodological repertoire while remaining attentive to governance, accessibility, and the politics of remembrance in written forms.