Equitable Metadata for Diverse Voices: Sustainable Computational Poetry Analysis with HathiTrust Extracted Features

2.1 HathiTrust Extracted Features

The HathiTrust Research Center (HTRC), established in 2011, supports large-scale computational analysis of works in the HathiTrust Digital Library (HTDL). The Center develops infrastructure, tools, and datasets that support massive text and data mining of the HTDL corpus for educational and non-profit research. Through these services, HTRC provides researchers with computational access to vast digital collections while adhering to copyright laws and non-consumptive research practices.

The HathiTrust Extracted Features (EF) dataset provides a set of structured features derived from OCR-processed pages. These features include volume-level metadata, page-level metadata, token counts, and part-of-speech (POS) tagged tokens generated using the Stanford NLP library[10, 11]. Files are distributed in JSON-LD format. The EF dataset contains more than 18 million volumes, 6.8 billion pages, and 3.2 trillion tokens, including both public-domain and in-copyright items.

EF enables broad-scale computational analyses, such as lexical, stylistic, and genre-based studies, across the HTDL collections without violating copyright restrictions. Previous studies have used EF-based features for tasks such as word similarity[12], topic modeling[13], and literary genre identification[14]. Despite its usability, the linguistic features in EF are derived from OCR-processed text, which may contain OCR errors. These errors can influence subsequent tokenization and POS tagging. In addition, general-purpose POS taggers, such as the Stanford NLP parser, are typically trained on modern prose and may perform less reliably on non-standard, domain-specific texts, including poetry. These inherent limitations should be considered when interpreting token-level results.

In October 2024, HathiTrust Executive Director Mike Furlough announced that HathiTrust will cease funding and data access for the HTRC at the end of 2026. Since the announcement, HTRC has been preparing a transition and sunsetting process. Unlike the discontinued access to Data Capsule, access to HTDL, Bookworm, and derived datasets, including EF, will remain available. As a result, EF is expected to remain an important resource for computational analysis of the HathiTrust collections after the retirement of the HTRC infrastructure.

2.2 Poets from Historically Underrepresented Groups (PHUG) Dataset

In this study, we build on the PHUG dataset introduced by Kang and Choi[9], a curated collection of American poetry by poets from historically underrepresented groups in the HTDL. The dataset comprises 4,723 poems from 120 collections by African American (AA), Asian American (APA-AA), Pacific Islander (APA-PA), Latin American (LXA), and Native American (NA) poets, with a detailed breakdown provided in Appendix Table Table A1. Most importantly, it provides poem-level boundary annotations, including the start and end page numbers of each poem. This structural metadata enables large-scale computational poetry analysis through retrieving poem records in HTRC EF.

To construct the derived metadata representation, we cross-referenced the PHUG dataset with EF. Initially, we used EF v2.0, yielding 92% coverage, with 8 volumes and 375 poems missing. By upgrading to EF v2.5, retrieval coverage reached 100%; all 4,723 poems across 120 volumes were matched to their corresponding EF data. This shows the high reliability and robustness of the EF dataset.

We also observed meaningful qualitative improvements in token extraction in v2.5, resulting from HTRC's recently updated OCR pipeline. When comparing the exclusive words extracted from both versions, we found that EF v2.5 more accurately recovers orthographically complex forms, especially words with apostrophes and diacritics. For instance, words that appeared in EF v2.0 in simplified or degraded forms such as Wen, shiyazhf, Hooulu, tia, que, and Dine are correctly captured in EF v2.5 as W'en, Shiy'azh'i, Hoʻoulu, tía, Qué, and Diné. These examples demonstrate how improved OCR processing preserves critical orthographic markers that the earlier algorithms often missed. Such enhanced token recognition allows a more robust linguistic analysis across diverse and non-standard language forms. The matching extracted features and analysis code are publicly available via Zenodo and GitHub, respectively.¹

3.1 Method

To identify unique linguistic characteristics and culturally specific words for each poet group, we conducted an exclusive word analysis using token frequency from EF. We chose this approach because general POS taggers perform poorly on poetry, and we preserved original capitalization to identify culturally significant proper nouns and capture deliberate stylistic emphasis. We recognized that a strict zero-tolerance criterion (a threshold of 0) was too restrictive and would eliminate highly distinctive terms that appeared only rarely in other groups.

To address this, we empirically tested a cross-group overlap threshold, increasing it incrementally. We observed no significant changes beyond a threshold of 2; a threshold of 3 yielded few new words, and those added tended to have relatively low frequencies. Therefore, we set the final threshold at 2. This slightly relaxed approach helped retain important, high-frequency words that are distinctive in a particular group, such as Dat and dere in the AA group, without compromising overall cross-group distinctiveness. We visualized the top 50 exclusive words per group with a threshold of 2 in Figure 1. Frequency counts and ranked word lists are provided in Table Table B1 in the Appendix.

3.2 Results and Discussion

In the African American (AA) group, the most frequent exclusive words represent African American English (AAE) variants and African American Vernacular English (AAVE). For example, 'dat', which indicates a common pronunciation of 'that' in AAE[15], appears 154 times and 'wid' referring to 'with' in AAVE has 124 counts, along with 99 counts of 'dey' meaning 'they' in AAVE. 'fo (Fo)' meaning 'for', reflects a common characteristic of AAE known as r-dropping, in which the final r/ is not pronounced[15].

Asian American (APA-AA) group has the fewest exclusive word counts, indicating that the group is less homogeneous than the other groups. There are more words referring to various cultures and locations than other groups. 'RICE (rice)' appears 28 times across the corpus, which holds cultural, agricultural, and culinary significance in Asian countries[16, 17]. 'Korean' (16 counts), 'Shanghai' (13), 'Armenian' (8), 'Kashmir' (7), 'Baghdad' (7), 'Turkish' (7), 'Delhi' (6), 'Nepal' (6) and 'Armenia' (6) demonstrate the diversity of ethnicity and nationality within the same group. Some popular surnames for specific ethnic groups, such as 'Chang' (16), 'Wong' (9), and 'Zhang', (8) are also included in the APA-AA exclusive word list. Names are important indicators for inferential ethnic classification, which overcomes the limitation of an aggregate Asian American category[18].

Pacific Islander (APA-PA) group's exclusive words display the most linguistic distinctiveness. The words 'Waiolola' and 'Waiololi', both of which count 43 times, are respectively a female and male figure from The Kumulipo, the creation chant in Hawaiian religion, written by Liliuokalani, Queen of Hawaii. The former translates as 'broad stream' and the latter as 'narrow stream'[19]. However, direct translations do not deliver the significance of the mythic figures throughout the narrative poem. These examples indicate that preserving linguistic originality in multilingual poetry is important because direct translations can hinder multiple interpretations of original words and their musicality in the poem[19].

In Latin American (LXA) group, ethnic personal words frequently appear: 'Enrique' (80 counts), 'Delmira' (73), 'Manuel' (64), María (48), 'André' (41), 'Eugenia' (35), 'Renata' (24), 'Lisbeth' (16), 'Luisa' (14), 'Ugarte' (14), and 'Agustini' (12)[20]. This shows a balanced mix of male and female names, but there are more words referring to women: 'Mamá' (98), 'nena' (40), 'Ella' (17), 'ella' (14), and 'tía' (13), respectively translating as 'mom', 'little girl', 'she/her', and 'aunt'[21]. Location words, such as 'Havana' (30), 'Montevideo' (16), 'Andes' (16), 'Minas' (10), and 'Juarez' (10), demonstrate that LXA poets have different national backgrounds, including Cuba, Uruguay, Brazil, and Mexico. Interestingly, unlike APA-AA group, in which there are words referring to ethnicity and country names, LXA group has more words of city names.

Native American (NA) group has a lot of words related to NA history and culture. The most frequent exclusive word 'reservation' with 87 counts and the third most frequent word 'rez' with 45 counts both refer to a designated land area managed by a Native American tribe. In addition, popular tribe names appear in the word list: 'Navajo' (51 counts), 'Comanche' (35), 'Diné' (35), 'Cherokee' (23), 'Choctaw' (20), 'Mohawk' (17), 'Kiowa' (15), 'Sioux' (13), 'Choctaws' (13), 'Stoney' (12), and 'Navajos' (11). Some words, such as 'Directions' (22)[22], 'powwow' (21; a Native American ritual), 'frybread' (18; traditional Native American food), INDIAN (Indian) (17), 'Ceremony' (14), and 'tepee' (11; a traditional Native American tent) represent Native American culture and history[23–27]. Many counts of community-related words, such as 'Grandpa' (21), 'pishno' (21; meaning 'we', 'us', 'ours')[28], and 'Granny' (16), as well as nature- and landscape-related words, including 'Raven' (27), 'Plains' (24), 'Rushmore' (20), and 'shiprock' (11), indicate NA group's iconic interests in family and the natural environment[29–31].