Digital Humanities Lab: Text analysis and the humanities

Main ideas, concepts, and lessons

1) Text analysis in the humanities (overview and motivation)

• Digital humanities increasingly uses computational methods to uncover meaning, patterns, and trends in large text-based datasets.
• Relevant textual sources include:
  • historical newspapers
  • social media collections
  • correspondence
  • collections of academic publications
• Text analysis can support tasks such as:
  • topic discovery (what texts are about)
  • sentiment/emotion detection
  • trend/pattern detection across large populations
  • event prediction and tracking shifts in belief or social movements (mentioned as a capability)

---

2) Ashley Dennis-Henderson: Analyzing World War I war diaries

Core goal (Masters → PhD)

Use mathematical/computational techniques to analyze Australian World War I diaries to understand:

• what soldiers wrote about over time (topic change)
• how they felt (sentiment/emotional tone)

Dataset

• Source: State Library of New South Wales digitized WWI collection
• Total documents: 966
• Labeled diaries: 557 diaries
• Size: over 9 million words

Why this matters

• Traditional close reading was limited to ~20 diaries due to time constraints.
• Computational approaches enable analysis at much larger scale.

Methods and workflow (Masters)

Data cleaning

• Ensure text/data are in a computer-processable format.
• Fix inconsistent formatting so the computer does not miss content.

Date extraction

• Extract entry dates to map analysis across time during the war.

Topic analysis

Word frequencies

• Identify top frequent words while excluding function words (e.g., and, to, or because).
• Interpretation:
  • Frequent “middle” words often reflect everyday terms (times of day, weather, eating, locations, travel, letter writing).
  • Additional war-related terms add another layer (e.g., wounded, guns, trench).
  • Identify homonyms/context-dependent terms (words that are general but become war-specific in context; e.g., horse as “Light Horse”).

Topic modelling (LDA)

• Concept: each document is a mixture of topics; each topic is a list of words.
• Analogy: a “cat/dog” sentence might produce topics such as animals + emotions.
• Parameter choice:
  • 10 topics were best for the dataset.
• Outputs:
  • Topics were manually interpreted/labeled (e.g., “war at sea,” “Egypt,” “Gallipoli,” “trench” variants, “armistice,” “home,” “everyday life”).
• Temporal validation:
  • Topic proportions over time matched expected historical sequence:
    • “war at sea” early
    • “Egypt” training period
    • peaks for “Gallipoli”
    • trench periods
    • “armistice” then “home”

Sentiment analysis

• Dictionary-based approach:
  • classify text as positive / negative / neutral using sentiment dictionaries (words are pre-rated).
• Ashley used 8 different dictionaries to handle issues such as:
  • dictionary suitability depending on domain
  • word meaning shifts over time
• Findings from sentiment + topics:
  • General sentiment followed expected patterns:
    • more positive during training/pleasant contexts (e.g., Egypt)
    • drops around Gallipoli
    • rises after the armistice
  • Notable exception:
    • the “White Christmas” topic had more negative sentiment because diaries described:
      • harsh trench conditions (mud, slush, lack of food/rations, poor access to help)
      • limited “Christmas” references

Key lesson from the diary study

• Contrary to intuition that diaries would focus mainly on battles, over 60% of writing was about everyday life activities.

PhD research questions (future work)

• Accuracy of sentiment analysis on historic data

  • Question: how reliable are sentiment dictionaries/techniques for historical language?
  • Motivation: dictionary performance depends on:
    • when the dictionary was created
    • which domain it was built for
    • changes in meaning over time
  • Aim: an accurate and easy-to-apply sentiment method for historic corpora.

• Comparing diary vs newspaper portrayal of WWI

  • Use topic modelling + sentiment analysis to test hypotheses such as:
    • Was WWI portrayed differently in newspapers than in diaries?
    • Were newspaper narratives “more positive” or “more negative”?
  • Use methods to contribute evidence to open historical questions (example: why people enlisted?).

---

3) Rachel A. Ankeny: Organism selection in science using quantitative textual methods

Core problem (philosophy of science / science policy / practice)

• A claim exists that model organisms (e.g., mice, fruit flies) have “swamped” attention away from other organisms.
• Rachel argues this is an empirical question with insufficient evidence, especially about:
  • whether swamping occurs
  • how it changed over time
  • why scientists choose organisms
  • how diversity is valued in practice

Research focus

• Project: ARC project “Organisms in us” (Rachel leads)
• Primary questions:
  • Has model organism research excluded other organisms?
  • How has this changed since the 1970s?
  • What factors influence these patterns?
  • What is the value of biological diversity, and how is it enacted?

Method (quantitative sampling + coding of article content)

Journal selection and sampling strategy

• Select journals representative of various scientific fields.
• Sample in five-year intervals.
• For each sampled year:
  • review all items published
  • keep research articles only
  • exclude non-research content

Coding unit

• Focus on discussion of research articles that:
  • use more than one organism

Manual coding (with student/undergraduate assistance)

• Articles are accessed automatically, but coding requires heavy human effort.
• For each article, coders determine:
  • what the article is about
  • which organisms it focuses on
• Data entry into large Excel spreadsheets

Coding difficulties (explicit pitfalls)

• Misspellings can break coding because organism names are often Latin.
• Organism names change over time (same organism may appear under different names).
• Some articles don’t clearly state study organisms (coders infer from domain knowledge).

Findings (high-level)

• Contrary to the “swamping” narrative:
  • the data did not show blanket privileging of model organisms across all contexts.
• Differences by field:
  • Ecology- and diversity-oriented fields:
    • more diverse organism use
  • Fields aimed at translating knowledge to humans:
    • more dominance of model organisms
    • less organism diversity

“Presumption of generality”

• Classic idea: findings in model organisms generalize widely (e.g., “what is true for E. coli is true for bacteria and elephants”).
• Rachel frames this as a problematic/untested assumption that may influence organism choice.

Additional analytic approaches (conceptual)

• Map/statistical approaches relating organism use to the tree of life.
• Rank abundance curve analysis:
  • examine changes in diversity and dominance across organisms over time.

Conclusions and next steps

• Preliminary conclusions:

  • Biological diversity is necessary/valued when the object of study is diversity itself (e.g., ecology or multi-species comparisons).
  • Diversity is more limited where studies are designed to translate toward humans and where presumption of generality dominates.

• Moving beyond counts:

  • Past work gathered qualitative criteria scientists cite when choosing organisms (oral histories, publications, interviews, biographies).
  • Next steps aim to:
    • combine qualitative explanations with large-scale quantitative trend evidence
    • refine theme coding from article content
    • consider machine learning, but remain cautious due to organism-name identification challenges
    • apply social-science textual analysis techniques

• Emphasized caution:

  • Even with digitization and online access, historical texts can be hard to computationally harvest due to:
    • old writing norms
    • older language assumptions
    • incomplete metadata (some articles don’t state what organisms they study)

---

4) Martin Schweinberger (pre-recorded) + Michael Haugh: Infrastructure for computational humanities (LADAL + ATAP)

Core message: build infrastructure + training for computational text analysis

• Humanities disciplines risk falling behind computational capability needs unless computational skills are integrated into curricula.
• Benefits of computational humanities (as stated):
  • enables work on big data and multi-modal resources
  • supports creation of real-world applications (language learning, machine translation, speech recognition)
  • improves transparency and research stability/reproducibility
  • addresses limited educational resources for HASS researchers by providing broadly applicable, transparent resources for applying computational methods to humanities questions

Three main problems to address

• Researchers and students have very different computational skill levels.
• Different disciplines require different tools and framings.
• Need broad, reusable resources that serve multiple audiences.

Training and infrastructure design (three components)

• Generic training via libraries
  • e.g., introductions to R, data visualization fundamentals
• Specialized training via LADAL
  • basic to advanced methods
  • learn tools via a “read the method + copy/paste code” workflow
• ATAP cloud environment
  • interactive, cloud-based analysis independent of local computer setup
  • adaptive: uses existing data or allows users to upload their own datasets

---

LADAL (Language Technology and Data Analytics Lab)

Institutional role

• An e-research support infrastructure maintained by the UQ School of Languages and Cultures.

What LADAL enables

• Digital tool and data management skills, including:
  • file naming conventions
  • project folder organization
  • avoiding data loss
• Computational methods + basic programming skills

Tutorial topics

• Data extraction from text:
  • extracting occurrences of words/phrases
  • transforming and processing text data
• Data visualization, including:
  • word clouds
  • geospatial mapping
  • interactive web apps
• Text analytics / NLP and statistics:
  • sentiment analysis
  • topic modelling
  • network analysis
  • classification and machine learning
  • mixed-effects regression modelling (for research relevance)

How LADAL supports learning

• self-guided online study materials + code examples
• face-to-face consultations
• workshops

Quality aims

• implement procedures with best practices
• make research reproducible and replicable
• improve data management for more efficient workflows
• build computational skill foundations for humanities researchers

---

ATAP (Australian Text Analytics Platform)

Role

• Integrated collaborative cloud-based environment connecting:
  • users
  • providers of text analytics

Key features

• Perform analyses in the cloud (not dependent on personal computer setup)
• Adaptive support:
  • use already-available datasets
  • upload user datasets
• Provides training resources (not only infrastructure)

---

What “text analytics” means (definition + comparison)

Definition

• Using computational means to extract information from unstructured textual data.
• Includes:
  • NLP
  • data analytics
  • machine learning

Common prominent methods

• topic modelling
• sentiment analysis
• network analysis

Difference from traditional discourse/content analysis

• Text analytics:
  • works on much larger corpora (tens/hundreds/thousands of texts)
  • relies more on computational processing than manual annotation
  • uses machine learning/statistics rather than manual post-editing
  • emphasizes data management + reproducibility strongly
• Traditional approaches:
  • often analyze a single or very small number of works manually

---

Case study example: COVID discourse in the Australian Twittersphere (LADAL/ATAP example)

Researchers mentioned

• Michael Haugh and Sam Hames (case study collaborators)

Goal

• Understand emotional shifts and discourse development around COVID-19 in Australian Twitter discourse.

Method summary

• Problem addressed:
  • discourse researchers often use qualitative analyses on small datasets
  • big-data analysts may treat discourse as one undifferentiated mass
• Approach:
  • break discourse into phases/sub-discourses using:
    • keyword frequency to identify phases
    • clustering to discover sub-discourse structure

Results

• Discourse evolved into:
  • 7 phases
  • 5 main sub-discourses:
    1. medical discourse
    2. international (abroad) discourse
    3. domestic restrictions/home situation discourse
    4. discourse about spread/contagion
    5. economy-related discourse
• Emotional findings:
  • very negative when COVID entered Australia
  • sentiment recovered toward more positive attitudes afterward
  • the “spread” topic stayed notably negative
• Key methodological takeaway:
  • emotional shifts should be analyzed separately by topic (not averaged across the whole corpus)

---

Speaker Q&A / discussion highlights

• Ashley on tools
  • Uses R primarily; also Python exists.
  • Uses packages for topic modelling and sentiment dictionaries.
• Ashley on history vs math
  • History helps interpret quantitative results in context.
  • Math provides logical structure; history helps address ambiguity and multiple perspectives.
• Rachel on peer review/editorial influence in organism selection data
  • Only published results are included; rejected work is not represented.
  • She argues journals/peer review still matter because they shape what is representative of fields.
  • She cannot make a “negative case” for what was not published, but can make a positive case for what is in the record.
• Michael on how to get involved with LADAL/ATAP
  • Weekly webinar series; recordings uploaded (including on LADAL’s YouTube channel).
  • Outreach is expanding via a newly hired communications/engagement role.
• Advice for historians with no computational background
  • Collaborate with someone who can do the technical work.
  • Learn enough “language” of the methods to communicate effectively (e.g., understand topic modelling concepts).
  • Attend webinars and use resources; “dabble” to think critically even if not implementing personally.

---

Speakers / sources featured (as named in subtitles)

1. Ashley Dennis-Henderson (presenter; PhD student/Applied Mathematics & History background)
2. Rachel A. Ankeny (Professor of History and Philosophy, University of Adelaide; presenter)
3. Alexis (moderator/organizer; frequently introduced speakers and managed Q&A)
4. Martin Schweinberger (pre-recorded; UQ School of Languages and Cultures; LADAL/ATAP co-presenter)
5. Michael Haugh (present; Professor of Linguistics and Applied Linguistics; LADAL/ATAP co-presenter)
6. Karen (Producer)
7. Jennifer (asked a question in the chat; about Ashley’s background/interdisciplinary aspect)
8. Matilda (asked a question in the chat; about peer review/editorial influence in Rachel’s organism selection analysis)
9. Sam Hames (co-authored case study mentioned in Martin/Michael’s section)
10. Mark McPeek (ecological scientist collaborator mentioned in Rachel’s project)
11. Aesha Duong (undergraduate research assistant mentioned in Rachel’s project)
12. University of Adelaide Library Digital Humanities Lab (series host described)
13. Australian Text Analysis Platform (ATAP) (described infrastructure)
14. Language Technology and Data Analysis Laboratory (LADAL) (described infrastructure)
15. Australian Research Council (ARC) (funding agency mentioned for Rachel’s project)
16. Medical Research Future Fund (mentioned as part of Rachel’s team’s recent grant context)
17. State Library of New South Wales (source of the WWI diaries used in Ashley’s study)
18. Clarivate / Web of Science / indexing systems (mentioned in context of how articles are selected/indexed)
19. Big Data and Society (journal where the COVID Twitter discourse case study is published)