Computational Social Science Literature DailyReading Project

Citation: Szcześny, A. (2025). Leveraging positional bias of LLM in-context learning with class-few-shot and Maj-Min alternating ordering. Computational Science – ICCS 2025.

Local files: Vault Taiwan_Framework; markdown szczesnyLeveragingPositionalBias2025.md.

Core claim: Class-balanced few-shot examples plus Maj-Min alternating ordering outperform standard random few-shot, with roughly five percentage points gain on minority-class F1.

1. Mechanism: Two separable effects: (1) class balance alone (random class-few-shot) beats standard few-shot; (2) ordering adds further gain on top.
2. Positional bias: Recency bias dominates overall accuracy, so Maj-Min Alt (majority first, minority last) works best globally.
3. For minority-class recall specifically, Maj-Min Seq (minority examples clustered near the end) outperforms alternating; the recency window matters more when detecting a rare category.

Citation: Uma, A. N. (2021). Learning from disagreement: A survey. Journal of Artificial Intelligence Research.

Local files: Vault Taiwan_Framework; markdown umaLearningDisagreementSurvey2021.md.

Core argument: Taxonomy mapping onto annotation failure modes: annotation scheme problems (ill-defined or overlapping classes), ambiguity (multiple valid interpretations of the same text), and item difficulty (genuinely unclear cases). The diagnostic question is which source dominates a label’s error pattern; fixes differ.

1. Concentrated confusion label: When two labels confuse each other more than anything else, disagreement is structured and reflects conceptual proximity rather than vague definitions.
2. Soft labels: Uncertain assignments are treated as signal marking a real ambiguity zone at the label boundary.

Citation: Gao, W., Wang, Y., & Li, Y. (2026). Ca-MCF: Category-level multi-label causal feature selection. arXiv:2602.12961.

Local files: Vault Taiwan_Framework; markdown Category-level Multi-label Causal Feature selection.md.

Core contribution: Addresses “label blocking”: a dominant label suppresses causal signal for features tied to rarer categories. Remedy is label-category flattening—decompose each multi-class label into binary category nodes, then run causal discovery (Markov blanket) at that granularity.

1. Metrics: SCSMI (feature–target category association); DCSMI (category–category association across labels).
2. Theorem 1 intuition: A blocked feature is recovered if its SCSMI with the target exceeds the DCSMI between the blocking category and the target (the feature explains the target better than the dominant category does).
3. Ablations: Phase 3 (blocked-feature recovery) drives most gains, especially on sparse high-dimensional data; Phases 1–2 stabilize structure; Phase 4 trims redundancy via cross-dimensional SCSMI.

Citation: Törnberg, P. (2024). Best practices for text annotation with large language models. arXiv:2402.05129.

Local files: Vault Taiwan_Framework; markdown tornbergBestPracticesText2024.md.

Overview: Normative best-practices piece; nine recommendations with unequal depth.

1. Prompt stability: Run multiple prompt paraphrases on a subset and measure Krippendorff’s α across runs; high sensitivity to minor wording signals the prompt is not constraining interpretation enough.
2. “I don’t know” option: An explicit uncertainty escape reduces stochastic forced-choice answers when no label fits well.
3. Validation set size: Minimum about 20–30 samples per category for basic confidence; imbalanced categories need more.
4. Failure examination: Characterize how failures occur, not only aggregate F1, because systematic misinterpretation can bias downstream use while overall metrics look fine.

Citation: Bibal, A. (2025). Automating annotation guideline improvements using LLMs: A case study. Proceedings of ACL.

Local files: Vault Taiwan_Framework; markdown bibalAutomatingAnnotationGuideline2025.md.

1. Across iterations, the dominant LLM move is negative constraints (“Do not mark X unless…”, “Do not mark generic terms without a specific name”) rather than restructuring the positive definition.
2. Core definitional scope stays largely stable; gains come from explicit boundary marking on the negative side.
3. LLMs often agree on hard boundary cases (e.g. Twitter handles) that systematically divide human annotators—suggesting LLM-identified disagreement may reflect structurally different confusion than human IAA alone.

Citation: Halterman, A., & Keith, K. A. (2024). Codebook LLMs: Adapting political science codebooks for LLM use and adapting LLMs to follow codebooks. arXiv:2407.10747.

Local files: Vault Taiwan_Framework; markdown haltermanCodebookLLMsAdapting2024.md.

Core argument: Contrasts the universal label assumption (LLM applies its pretrained concept of a label) with the codebook-construct assumption (LLM applies the researcher’s operationalization). Social scientists should enforce the latter and test compliance empirically.

1. Empirical finding: Evidence that Mistral behaves as if under the universal-label assumption—matching label names to surface lexical cues rather than reasoning from definitions.
2. Lexical overlap heuristic: Primary failure mode; e.g. the word “rally” in the text triggers RALLY though the codebook match is DEMONSTRATION.
3. Error taxonomy (A–F): Semantics/reasoning mistakes (E) dominate BFRS and CCC errors; non-compliance/hallucination (D) dominates Manifestos.
4. Codebook format: Label → Definition → Clarification → Negative clarification → Positive example → Negative example, targeting observed boundary confusion between adjacent categories.

Citation: Henning, S., Beluch, W., Fraser, A., & Friedrich, A. (2022). A survey of methods for addressing class imbalance in deep-learning based natural language processing. arXiv survey.

Local files: Vault Taiwan_Framework; markdown henningSurveyMethodsAddressing2022.md.

Scope: Taxonomy of imbalance interventions for DL-based NLP: re-sampling, data augmentation, loss functions, staged learning, model design; value is the decision framework and cross-method comparison.

1. Ranking loss: Maximizes the true-positive score while suppressing the top competing incorrect label; explicitly ignores the catch-all class’s own score rather than penalizing it.
2. LDAM: Label-distribution-aware margin loss pushes minority instances further from the boundary by scaling margins inversely to class frequency.
3. Distribution-balanced (DB) loss: Addresses logit suppression in BCE (minority labels mostly negatives) and instance-level class-weight imbalance; negative-tolerant regularization sharply drops penalty once a negative logit crosses a threshold, avoiding over-learning “don’t predict rare label.”

Citation: Dunivin (2025). Scaling hermeneutics: a guide to qualitative coding with LLMs for reflexive content analysis. EPJ Data Science.

Local files: Markdown notes in PhD Vault Taiwan_Framework; PDF in D:/下载/Document/Projects/Taiwan_Election_Discourse (e.g. Dunivin 2025 PDF).

Core argument: Hybrid workflow (computational grounded theory + LLMs). Codebooks built for human hermeneutics often fail on LLMs because models lack implicit sociological/historical context shared by human coders.

1. LLMs need more precise, literal definitions; translate implicit assumptions into explicit prompt constraints.
2. Ask the model for a 2–3 sentence justification before the code label to improve reliability and surface systematic errors.
3. Mandatory/prohibitory constraints and explicit scope reduce boundary drift; literal code labels outperform abstract ones.
4. Evaluating codes independently (“per code”) outperforms presenting the full codebook at once (higher κ in the paper’s comparison).

Citation: Than, N., Fan, L., Law, T., Nelson, L. K., & McCall, L. (2025). Updating “The Future of Coding”: Qualitative Coding with Generative Large Language Models. Sociological Methods & Research.

Local files: Vault Taiwan_Framework; PDF in D:/下载/Document/Projects/Taiwan_Election_Discourse (McCall 2025 PDF).

Core argument: Generative LLMs shift text-as-data work from programming to dialogue, approximating iterative inductive/deductive qualitative coding.

1. Sequential binary steps (e.g. relevance then A vs. B) can match supervised-ML-level accuracy on complex concepts.
2. LLM-expanded definitions from a human codebook (7–10 bullets) work at least as well as hand-polished definitions.
3. Zero-shot with clear definitions can beat few-shot on long texts (examples add noise or blow the context budget).
4. One-sentence qualitative justification alongside the label helps audit model reasoning.

Citation: Billig, M., & Skey, M. (2025). Banal Nationalism, 30 Years On—A Review. Nations and Nationalism.

Local files: Vault Taiwan_Framework; PDF in D:/下载/Document/Projects/Taiwan_Election_Discourse (Billig & Skey 2026 PDF filename on disk).

1. Banal nationalism: hot nationalist outbursts rest on everyday, unnoticed flagging of the nation in the established state.
2. Global dimension: pushback to pure “borderless” cosmopolitanism—international order, trade accounting, and climate governance still presuppose nation-states; nationalism is also an international doctrine.
3. Sewell: surface struggles over resources vs. deep common frame that naturalizes the nation-state.
4. Process: banalization is not static; banal ↔ hot cycles (e.g. crises heating populist nationalism).
5. Entitlement: what is routine for a dominant group can feel threatening to marginalized groups; everyday vs. banal should not be reduced to top-down vs. bottom-up only.

Citation: Ho, J. C.-t. (2025). Computational Methods and Nationalism Studies. State of Nationalism.

Local files: Vault Taiwan_Framework; related PDFs under D:/下载/Document/Projects/Taiwan_Election_Discourse.

Evolution (CSS in nationalism): keywords → topic models (LDA) → embeddings → supervised / few-shot (incl. LLMs) → network views of identity and discourse.

1. Risks: algorithmic bias from web-scale training data; LLMs may yield generic, “apolitical” labels vs. rich qualitative coding (Rogers & Zhang, 2024)—relevant to detecting implicit/banal nationalism.
2. Practice: theory-first operationalization on repurposed “found” data; mix inductive (e.g. LDA) with deductive supervision; always validate against human gold standards and use interpretability tools for error analysis.

Citation: Bozdağ, U., Tóth, T., & Demeter, M. (2025). When Articulating Populist Dichotomies Is Paramount: Exploring the Effects of Explicit and Implicit Populist Styles on User Engagement in Turkish Election Tweets. Media and Social Identities.

Local files: Vault Taiwan_Framework; PDF in D:/下载/Document/Projects/Taiwan_Election_Discourse (Bozdağ et al. 2025).

Operationalization (PPCS refinement): explicit populism names both people and elite/other in one unit; implicit people-centrist (IPP) centers the people/general will without naming the foe; implicit antagonist (IPA) attacks elite/others without naming “the people.”

1. Explicit dichotomization predicts strongest engagement (favorites, RTs, replies).
2. IPP: positive for RTs/replies, not favorites; IPA: null on engagement in their models.
3. Domestic topics such as terrorism were heavily framed via EP/IPA as threats to public safety.

Citation: Tóth, T. (2021). Fractured implicitness. Why implicit populism matters? An International Journal of Pure Communication Inquiry.

Local files: Vault Taiwan_Framework; PDF in D:/下载/Document/Projects/Taiwan_Election_Discourse.

1. EP: same coding unit refers to people and a harmful other, stressing harm to the volonté générale (Mudde, 2004).
2. IP: stresses either popular sovereignty or threat to the general will without the full dyad in one unit (Hameleers, 2018).
3. IP1 (people-centrism): people/sovereignty/demands vs. antagonist interests, foe absent from unit.
4. IP2 (antagonism): enemy (person, group, vague circle, ideology, austerity, “danger”) without evoking the masses in the same unit.
5. Practical suggestion: code EP and IP at narrow and extended unit levels within the same content.

Citation: Olsson, F. (2024). Implicit Nationalism: The Missing Link in the Study of Nationalist Attitudes? Sociological Research for a Dynamic World.

Local files: Vault Taiwan_Framework; PDF in D:/下载/Document/Projects/Taiwan_Election_Discourse (Olsson 2024).

Aim: Bridge banal nationalism and implicit-cognition work; implicit measures may capture more automatic, less conscious nationalism (cf. Skey 2017b). Events/discourse/symbols can raise implicit more than explicit nationalist attitudes.

1. Implicit associations: attitude = object–sentiment link; stereotype = object–trait link (Greenwald & Krieger, 2006); overlap common (Kurdi et al., 2019).
2. Implicit nationalism: positive association between sentiment/trait/attribute and nation; includes symbols and representatives, not only “nation” as word.
3. Anthem and “unwaved” flag alike can prime nationalist associations; superiority and tolerance framings can both strengthen positive nation links.

Citation: Gerber, A. S., & Green, D. P. (2012). Field Experiments: Design, Analysis, and Interpretation. New York: W. W. Norton & Company. Publisher

Random assignment and unbiased inference

1. Random assignment constructs a treated group and a control group that match in expectation on everything that precedes treatment, which is the standard way experiments confront the missing-counterfactual structure of causal inference.
2. Expected potential outcomes:
   1. Under randomization, E[Y_i(1) | D_i = 1] = E[Y_i(1) | D_i = 0].
   2. The average treatment effect can be written as the contrast of observed group means: ATE = E[Y_i(1) | D_i = 1] − E[Y_i(0) | D_i = 0].
   3. When treatment is assigned with equal probability for all subjects, potential outcomes are independent of the treatment actually received.
3. Unbiased estimator: an estimator is unbiased if its expectation equals the parameter of interest (here the ATE).
4. Difference in sample means (cf. text eq. 2.14): consider E[(1/m)Σ_i=1^mY_i − (1/(N − m))Σ_i=m+1^NY_i] — the mean outcome among the m treated units minus the mean among the N − m untreated units.
   1. By linearity of expectation, the expectation of the difference equals the difference of the expectations of the two group means.
   2. Under random assignment this chain simplifies to E[Y_i(1) | D_i = 1] − E[Y_i(0) | D_i = 0] = E[Y_i(1)] − E[Y_i(0)] = E[τ_i] = ATE.

Citation: Morgan, S. L., & Winship, C. (2015). Counterfactuals and Causal Inference: Methods and Principles for Social Research (2nd ed.). Cambridge University Press. DOI

2.3 Treatment groups and observed outcomes

1. Causal exposure variable D can be realized along (at least) three pathways: the individual’s decision; planned random allocation; or a decision by an outside actor.
2. If more than two causal states matter, the book shifts to language of alternative treatments and corresponding treatment groups, moving away from default “control state / control group” labels.
3. Outcome data we observe contain only part of the information that would be needed to compute individual-level causal effects for everyone. (Text figure links exposure D to observed outcomes — see PDF.)

2.4 The average treatment effect

1. Aggregate causal effects are averages of individual-level effects.
2. Average treatment effect (ATE): it is crucial to define clearly the population of individuals the analysis targets.

Document: Full text: Counterfactuals and Causal Inference (PDF in local library / Zotero).

Citation: Morgan, S. L., & Winship, C. (2015). Counterfactuals and Causal Inference: Methods and Principles for Social Research (2nd ed.). Cambridge University Press. DOI

2.1 Defining the causal states

1. Binary cause: treatment and control
   1. Fine articulation: much recent literature stresses finer articulation of causal states to reduce the limitations of a blunt binary.
   2. Nominal states from constitutive features (Catholic-school example):
      1. Nominal states: enrolled in a Catholic school vs. enrolled in a public school.
      2. Outcome: learning.
   3. Value of mounting causal analysis on a base that combines a metaphysics of causal powers with one of counterfactual dependence.
2. Local and reasonable
   1. Why not default to “fundamental-cause” talk: often non-productive for the analytic job.
   2. If a facile ceteris paribus assumption excuses the analyst from discussing other contrasts that almost certainly co-occur, the proposed causal states may be too improbable or ill-defined to support causal analysis built on them.

2.2 Potential outcomes and individual-level treatment effects

1. Notation: y_i¹ = potential outcome in the treatment state for individual i; y_i⁰ = potential outcome in the control state.
2. Individual-level causal effect: δ_i = y_i¹ − y_i⁰.
3. The same idea can be framed as a difference between expectations of individual-specific random variables, E[Y_i¹] − E[Y_i⁰], which is useful when units self-select into treatment states.

Citation: Holland, P. W. (1986). Statistics and Causal Inference. Journal of the American Statistical Association, 81(396), 945–960. https://doi.org/10.2307/2289064

1. Binary clarification — causes of effects vs. effects of causes: Holland proposes “no causation without manipulation” and places the treatment itself in the primary, prior role.
2. Associational inference vs. causal inference:
   1. In causal inference, the role of time is more central than in associational inference.
   2. Associational inference amounts to descriptive statistics.
3. Rubin’s model:
   1. Distinguish A(u) (a property or characteristic of unit u) from S(u) (the treatment / cause).
   2. Because time matters, variables split into pre-exposure (fixed before exposure to the cause) and post-exposure (fixed after).
   3. Basic causal contrast: Y_t(u) − Y_c(u).
   4. Fundamental problem of causal inference: Y_t(u) and Y_c(u) cannot both be observed on the same unit, so the effect of t on u is not directly observable.
      1. Even when manipulation seems physically intervenable, knowability of the underlying counterfactual remains hypothetical rather than directly observed.
      2. Solution 1: Rely on homogeneity or invariance assumptions.
      3. Solution 2: Average causal effect T of t (vs. c) over U: E(Y_t − Y_c) = T.
4. Some special cases:
   1. Temporal stability and causal transience.
   2. Unit homogeneity: cannot be proven; aim to make it plausible.
   3. Independence:
      1. Large samples help.
      2. Contrast E(Y_t) vs. E(Y | S = t): whether the unit is in the treated vs. untreated group matters for identification.
      3. Prima facie causal effect of t: only under independence can the associational contrast support full causal interpretation.
   4. Constant effect:
      1. ATE applies at the individual level.
      2. Independence is still needed for the prima facie effect to equal the average causal effect T.
   5. Observational studies:
      1. Independence is typically not assumed to hold.
      2. Strong ignorability (strong unconfoundedness / overlap-type assumptions).
      3. Use covariate adjustment to approximate randomization.

Citation: Bjarnøe, C., Adams, J. & Boydstun, A. (2023). British Journal of Political Science.

1. Background & Core Question: Political parties face strategic decisions during campaigns: should they rely on “issue-based” appeals (e.g., debates on economy or immigration) or “valence-based” appeals (emphasizing leadership character, integrity, and competence)? This study investigates how parties choose between these two strategies when presenting themselves versus attacking their opponents across 10 Western European democracies.
2. Key Findings:
   • “Our issues versus their valence” effect: Parties tend to emphasize concrete policy issues when presenting themselves, but heavily focus on valence (character and competence) when attacking rivals. Valence attacks are highly effective at winning votes; emphasizing their own policy issues helps the winning party claim an “issue mandate” post-election.
   • Extremist party issue focus effect: Compared to moderate parties, parties with more extreme ideologies place a significantly stronger emphasis on policy issues rather than valence in their self-presentations.
   • Prime ministerial valence focus effect: The incumbent prime minister’s party heavily emphasizes valence in its self-presentation, as prime ministers are the “public face” of the government and possess unique opportunities to showcase leadership and governing competence.
3. Conclusion: Campaign strategies are highly asymmetrical; parties apply entirely different standards to themselves and their opponents. Understanding these strategic choices is crucial for comprehending modern democratic elections and the dynamics of political polarization.

Citation: Boydstun, A. E., Bevan, S. & Thomas III, H. F. (2014). Policy Studies Journal.

1. Background & Core Question: Traditional studies of political attention often narrow their focus to a single issue (e.g., media coverage of the economy). Ignoring the broader agenda can lead to flawed assumptions. The authors argue that scholars must measure “attention diversity”—the extent to which attention is concentrated or widely distributed across various items (e.g., different issues on a policy agenda or different frames within a debate). The field lacked a standard measurement approach.
2. Key Findings:
   • Attention diversity is an independent and vital variable that reveals the openness and accessibility of policy agendas to various problems and perspectives.
   • The study evaluated four widely used diversity metrics: the inverse Herfindahl Index, Shannon’s H (entropy), and their normalized versions.
   • Through testing with simulated data and three real-world datasets, Shannon’s H and its normalized form were found to be superior measurement tools.
3. Conclusion: When measuring the distribution of attention in agendas or policy debates, Shannon’s H effectively minimizes the risk of spurious findings. The choice between raw and normalized Shannon’s H should depend on whether the variance in the total number of possible items is substantively meaningful for the specific research context.

Citation: Feezell, J. T., Glazier, R. A. & Boydstun, A. E. (2021). Politics, Groups, and Identities.

1. Background & Core Question: Iyengar’s (1991) classic research shows that “episodic” frames (specific individual experiences) lead to individual responsibility attribution, while “thematic” frames (broader social contexts) lead to government/societal attribution. This paper investigates whether these framing effects change when the identity of the target population portrayed in the news changes (e.g., “American” vs. “Muslim American”).
2. Key Findings:
   • In the context of Poverty: The experiment replicated Iyengar’s findings. Participants reacted similarly regardless of whether the subject was “Muslim American” or “American.” Episodic frame led to individual attribution; thematic frame led to societal attribution. Identity did not alter the direction of responsibility.
   • In the context of Religious Radicalism: The classic framing effects disappeared. Regardless of frame (episodic or thematic), audiences consistently attributed responsibility to the radicalized individuals rather than government or society.
   • The deep impact of identity: When the news subject was identified as “Muslim American” in the radicalism issue, respondents were significantly more likely to cite “religion” as the substantive underlying cause of the individual’s responsibility.
3. Conclusion: While classic framing theories demonstrate robustness in certain contexts, they are complicated by the interacting effects of target group identity and issue type. A group’s identity deeply influences how the public forms internal attributions about their behavior, especially in specific policy contexts.

Citation: Jeffries, L. (2008). Journal of Literary Semantics (JLSE). Open in Zotero

1. TL;DR: This article examines whether reader-involvement in a poem can be at least partly explained by stylistic features. By comparing Peter Sansom’s “Mittens” and Mebdh McGuckian’s “Pain tells you what to wear,” the study uses deictic shift theory and blending theory to explain how different linguistic choices draw readers into the story world or alienate them.
2. Theory & Concepts:
   • Deictic Positioning: “Thus, it is thought that the way that deictic expressions work in narrative is that they cause the reader to take up the deictic positioning of the character, place and time which are indicated by the textual triggers. Unlike everyday interaction, then, where the first person pronoun I typically references the person speaking, the use of first person in a narrative both identifies the narrator and also provides a perspective for the reader to enter the text world.” (Jeffries, 2008, p. 71).
   • Time References: “These alternating time references may cause the readers of this poem, then, to psychologically place themselves repeatedly in the present and then the past of the narrator’s deictic field.” (Jeffries, 2008, p. 74).
   • Merging of Narrative Perspective: “The incongruities of the text, whereby the narrator appears to be at once a small boy and an adult man, emerge more strongly in the second – and final – stanza and create less a sense of deictic shifting between two different ages of the first person narrator than a blending of the two into a new, rather unusual, character who is an adult but is being treated like a child.” (Jeffries, 2008, p. 74). This creates an effect where both past and present identities are simultaneously primed, challenging normal decompression.
   • Merging of Historic Context and Present: “In purely textual terms, then, there is no way of knowing in these first five lines, though the first stanza context may incline us towards the uncomfortable reading that this is genuinely happening in the present to the adult narrator, whilst the disjunction with what we know about the world may incline us to revise the deictic field to one in the past (i.e. a flashback), but described in the historic present tense.” (Jeffries, 2008, p. 75).
   • Voiding the Deictic Field & Generalized Narration: Some poetic openings stop chronological progression by failing to instantiate a specific WHO, WHERE, or WHEN, effectively voiding the deictic field. McGuckian’s poem uses a generalized second-person narrative: “The use of generalized second person pronouns, superficially hints at an addressee, whose referent could be the reader if this were so, but is more likely to refer to some generic third person.” This keeps the reader estranged and at arm’s length.
   • Priming and Prominence: The psychological prominence of textual triggers dictates the depth of reader involvement. “These two poems support such a notion since the deictic fields of ‘Mittens’, though unusual and at times apparently contradictory, are nevertheless strongly primed, whereas those of ‘Pain tells you what to wear’ are less clear and thus less fully primed.”
3. The Role of Pronouns:
   • First-person “I” (Mittens): Provides an entry perspective for the reader to enter the text world; facilitates identification with the narrator and allows readers to interpret the poem in personally relevant ways; juxtaposing two “ages” of the narrator demonstrates that identity is not singular across a lifetime.
   • Second-person “you” (Pain tells you what to wear): Iterative descriptions induce a generic third-person reading rather than true address; with no overt “I” to identify with, conceptual blending between reader and narrator remains incomplete, creating alienation; the narrator effectively hides behind a generalized pronoun that fails to establish a clear deictic centre.

Citation: Bamman, D., Lewke, O. & Mansoor, A. (2020). Proceedings of the Twelfth Language Resources and Evaluation Conference. ACL Anthology · Open in Zotero

1. TL;DR: The paper introduces a literary coreference dataset and shows that English fiction contains distinctive long-distance, bursty entity patterns; models trained on in-domain literary data perform substantially better than those trained on standard news benchmarks.
2. Theoretical Frame: Computational Literary Studies / coreference resolution domain adaptation.
3. Key Concepts:
   • Literary coreference: Differs from standard coreference because literary narratives track entities across long spans, shifting narrative perspectives, and character development over the course of a novel.
   • Near-identity: A relation shaped by neutralization/compression and refocusing/decompression, which respectively minimize or maximize perceived differences between discourse entities.
4. Data & Method: 210,532 tokens from 100 English-language fiction works in LitBank (1719-1922), annotated for ACE-style entity categories and used to evaluate a BERT-based neural coreference system across training domains.
5. Key Findings:
   1. Most mentions in literary texts are characters/people (83.1%), and pronouns make up 54.3% of mentions.
   2. Long-range entities display bursty narrative behavior, while pronouns have a median distance of only two entities to their nearest antecedent.
   3. A model trained on literary in-domain data achieves an F-score six points higher than a model trained on OntoNotes.
6. Takeaways:
   • Reusability: The annotation distinctions around generic vs. specific mentions and identity revelation are directly useful for narrative text analysis, especially mystery or character-centered texts.
   • Limitation: The dataset only annotates six ACE entity categories rather than unrestricted coreference.
   • Direct application: Essential for arguing that domain-specific data matters in NLP, especially when adapting entity tracking to literature, culture, or communication research.
7. Key Quote: “While a model trained on in-domain literary data achieves an F-score six points higher than one trained on OntoNotes, it performs equivalently to a model trained on PreCo”.

Citation: Fang, A. C., Lo, F., Chinn, C. K., Bel, N., Hinrichs, E., Osenova, P. & Simov, K. (2009). Proceedings of the Workshop on Adaptation of Language Resources and Technology to New Domains. Link

1. Research Objective: Establish a computational framework for analyzing creative language in classical Chinese poetry using NLP and corpus analysis, focusing on extraction, classification, and structural analysis of poetic “imageries” to investigate inter-poet and intra-poet stylistic differences.
2. Ontology of Imageries: Built on a complete collection of Tang dynasty poems; words segmented and indexed into semantic classes (synsets) across six main categories: human, affair, time, space, object, and miscellany.
3. Four Structural Levels of Imagery:
   • Primary imageries: Head nouns with imagery potential (e.g., “winter”)
   • Complex imageries: Primary imageries modified by a premodifier or determiner (e.g., “harsh winter”)
   • Extended imageries: Complex imageries serving clausal functions or acting as predicates
   • Textual imageries: A system of extended imageries designed to articulate artistic conception
4. Syntactic Parsing: Java-based phrase structure grammar (PSG) parser generates syntactic trees for POS-tagged poetic lines, enabling automatic identification of imagery units.
5. Case Study — Su Shi vs. Liu Yong:
   • Grammatical: Liu Yong: higher verb proportion (16.2%), fewer adjectives (11.9%) → colloquial/vocal style. Su Shi: fewer verbs (10.2%), more adjectives (16.5%) → dense, formal, scholarly style.
   • Syntactic: Su Shi favored SVO constructions (more subjects, objects, adverbials, complements) typical of formal prose. Liu Yong showed higher predicate and premodifier proportions, consistent with casual style.
6. Conclusion: Structured sub-categorization of imageries successfully links linguistic analysis (lexical, grammatical, syntactic) to literary evaluations of style; the framework proves effective for automatic analysis of classical Chinese poetry.

Citation: Kaplan, D. M. & Blei, D. M. (2007). Seventh IEEE International Conference on Data Mining (ICDM 2007). DOI

1. Research Question: Computationally capture a comprehensive scope of poetic style; poetry as a domain has “gone largely unexplored” compared to prose.
2. Place in the Literature: Moves beyond predominant bag-of-words and diction-based approaches to text analysis.
3. Operationalization: Maps poem text to a high-dimensional vector via multi-layered metrics:
   • Orthographic: Word count, line count, stanza count, average line/word length, lines per stanza; frequencies of the most frequent noun, adjective, and verb as proxies for repetition
   • Syntactic: Frequencies of contractions and parts of speech aggregated at different levels of specificity
   • Phonemic: Alliteration and formal rhyme scheme
4. Visualization: PCA projection of high-dimensional poem vectors onto two dimensions, preserving greatest variance to depict relative poem similarity.
5. Validity Risks:
   • POS tagger trained on Wall Street Journal corpus — domain mismatch with poetic language
   • PCA dimensionality reduction may obscure non-linear stylistic patterns
6. Comments: Pioneering early work on computational poetic style analysis. However, the feature set remains surface-level and does not fully capture the concise construction of narratives.

Citation: Kao, J. T. & Jurafsky, D. (2015). Linguistic Issues in Language Technology. Link

1. Research Questions:
   • How do literary movements transform the ideals of poetic beauty?
   • How do changes in aesthetic standards impact different levels of expertise (elite vs. amateur)?
   • Can we measure these shifts using computational methods?
2. Place in the Literature: Builds on Martindale (1990)’s work on elite corpus measurement; extends to analyzing changes surrounding Imagism and measuring the trickle-down effect from elite to mass aesthetics.
3. Key Concepts & Operationalization:
   • Concrete imagery: Average concreteness rating of all words → Concreteness score
   • Emotional language: Average valence and arousal ratings from affective norms database → Valence & Arousal scores
   • Sound devices: Identity rhyme, perfect rhyme, slant rhyme, alliteration, consonance, and assonance
   • Diction: Summed word frequencies divided by poem length → WordFreq score
   • Exactness: Unique word types / total word instances → Type-Token Ratio
4. Research Design: Three comparative analyses — Imagists vs. 19th-century professionals; contemporary vs. 19th-century professionals; contemporary professionals vs. amateurs.
5. Comments: Effective framework for measuring stylistic features across expertise levels and historical periods. The elite-to-masses comparison is compelling. However, most measurements remain limited to word-level statistics.

1. Social and intellectual factors shaping the political economy of communication:
   • Social driving forces:
   1. Media have been organized along industrial lines, with most media labor working for wages; yet media conglomerates are now so powerful that they can control the circuit of accumulation without retaining the risks of full ownership—flexible accumulation has consolidated global media power
   2. However, political economy is ill-equipped to examine behaviors organized around the household; likewise, a new vocabulary used to discuss audiences challenges political economy—audiences participate in the accumulation process through listening, reading, watching, and purchasing media content
   3. The process of transnationalization: also constitutes the demand for a New World Information and Communication Order (NWICO), while “globalization” frequently and vaguely refers to transnationalization
      • E.g., the Non-Aligned Movement — national self-determination — universal access to communication media, control over production and distribution, and communication as a fundamental human right — examining the role of communication in the U.S.-led postwar restructuring of global capitalism
      • New technologies expand audience control and undermine national sovereignty — embedded imagined American-style imagery — charting the growth and power of transnational communication corporations
      • Information and communication technologies possess the power to restructure the global division of labor
   4. The emergence of the information society:
      • A radical rupture of time, space, and social relations
      • Dan Schiller, Digital Capitalism
      • Open source and hacker networks challenge ownership: a contradiction between the desire for free access to information and capitalism’s drive to harness information solely for generating surplus value
2. The communication industry is fundamentally incompatible with the traditional classical economic model

1. Two Guiding Questions:
   • Have the repertoires of representation around “difference” and “otherness” changed, or do earlier traces from colonial and racialized narratives remain embedded in contemporary culture?
   • Is there a possibility for an effective politics of representation that can challenge and transform dominant stereotypical portrayals?
2. Persistence of Difference and Power:
   • Even with new modes of visual and cultural production, many symbolic codes, visual motifs, and narrative frames of the “racialized Other” still echo colonial-era patterns
   • Old and new forms often overlap: modern advertising, film, and journalism may adopt new aesthetics, yet still rely on deep-seated binaries (civilized/primitive, modern/backward, white/non-white)
3. Representation as a Site of Power Struggle:
   • The “spectacle of the Other” emphasizes that seeing and being seen are part of a broader power relation
   • Mainstream media and cultural industries do not merely “depict” the Other; they define the Other—this power to represent is central to cultural hegemony
   • Representation is not a one-way process; it is constantly negotiated among producers, texts, and audiences, though dominant discourses generally prevail
4. Counter-Representation Strategies:
   • Reversing the stereotype: Flipping a negative stereotype into its opposite, re-signifying it in a positive or empowering way
   • Positive imagery: Presenting affirmative, authentic, and diverse representations to offset prevailing negative portrayals
   • Contesting from within: Working within mainstream forms and platforms to insert new narratives that disrupt established codes
   • Hall warns that each approach has limits—reversal can still lock thinking into binary oppositions, while positive imagery may gloss over structural inequalities

1. How to define communication:
   • The rhetoric of conversation should provide the standards of science just as the logic of inquiry does; understanding occurs when two or more persons exchange observations, ideas, and express them in a language that does not merely reveal reality but contributes to constructing it.
   • Two definitional approaches: information transmission versus the construction of meaning
   • Mathematically grounded communication science: Shannon and Weaver — the transmission of information from a communicator and encoder to a decoder
   • Sociological perspective: The process of constructing meaning
   • Market specialists: An interactive process between two or more parties, in which meaning is exchanged through the intentional use of symbols
   • Implications for political economy: Communication studies should view communication systems as integral to the fundamental economic, political, and cultural processes of society: the framework of capitalism’s essential elements — capital accumulation and labor wages

1. Worker-Peasant-Soldier Students’ Attitude: They considered the loss of collective consciousness in new college students as an indicator of the reduction of political rights and participation in the New Era. This phenomenon illustrates a huge generation gap; the traditional habitus no longer existed, replaced by the alternation of ideology with individual self-consciousness.
2. Middle School Changes: The rebirth of the keypoint classes, demonstrating that a brand-new hierarchy of educational systems had emerged, producing the education elite.
3. Party Membership Transformation: Shorn of its ideological meaning, party membership retained its instrumental value as a political credential and networking tool, attracting ambitious university students who aspired to public service and leadership positions.
4. Loss of Inspirational Core: The loss of the inspirational core of communist party membership reduced it to a political credential and networking tool.
5. Core of New Tsinghua University: The transfer of class power, the flow and change of the knowledge hierarchy.

Language Standardization Process:

1. Standardized Prompt: “Please summarize this news article in 7–10 English sentences. Article: [insert article text]”
2. Claims Extraction: Prompt GPT-4o to extract the “descriptive, normative, causal, and classificatory claims” and “people, places, things, and events” included in each summary
3. Bias Mitigation: How to deal with the bias of the LLMs

Two-Step Candidate Generation Process:

1. Bi-encoder to Cross-encoder: Using the cutoff to filter
2. Efficient Pairwise Comparison: Reducing computational complexity through hierarchical filtering

1. Marx’s conception of the commodity: all commodities are nothing but a definite quantity of congealed labour-time
2. A critique within communication studies: orthodox Marxism overemphasizes the instrumentality and productivity of labor, reducing human beings to objects of production or mere factors of production — thus minimizing the other material practices of labor
3. The political economy of communication proposes to demonstrate how communication and culture constitute material practices, how labor and language are mutually constitutive, and how communication and information form part of the same social activity — the dialectical relationship of the social construction of meaning
4. The conservative critique of political economy:
   1. Conservatives hold that the growth of wealth originates from an organic order that leads people to respect tradition, which in turn clarifies social roles and the moral foundations compelling their fulfillment
   2. At its core: opposition to empowering the masses; the belief that the essence of economics is the maximization of pleasure — minimizing the sacrifice of what is unwanted while obtaining what is most desired (Jevons, 1965)
   3. Classical economics, under its preoccupation with equilibrium states, focuses on consensual change while neglecting the traditional interest in history — rendering it unable to analyze periods of transformation
   4. The market provides classical and neoclassical economics with their moral justification
5. Several variants of political economy:
   1. Neoconservatism: Government as the primary beneficiary of regulation — hence the push for deregulation; essentially releasing politics to preserve scientific rigor
   2. Institutional economics
   3. Marxist political economy: Critique of corporate concentration, exploration of state intervention, and response to transformations in production technology, industrial organization, and world markets
   4. Feminist political economy: Incorporating domestic labor into the system and analysis of market exchange; lacking a concrete calculative framework
   5. Environmental political economy: The transformation of the moral horizon
6. Why political economy is needed:
   1. Economics favors describing statics — everything is resolved in equilibrium, with interest limited to describing incremental change within a given set of institutions
   2. It is unable to incorporate major socioeconomic determinants into its analysis
   3. It tends to treat the market as a natural product of individual interaction, thereby exacerbating social divisions of class, race, and gender
   4. There is an urgent need to disentangle economic orthodoxy from its rhetorical system and to study it as part of a system of power

Research Purpose: To study how ideas and narratives diffuse across text, we must first develop a method to identify whether texts share the same information and narratives, rather than the same broad themes or exact features.

Methods: Using large language models to distill texts to their core ideas and then compare the similarity of claims rather than of words, phrases, or sentences. Sorting by SBERT and claimed by GPT-4o.

Research Questions: The spread of Russian claims about the development of a Ukrainian bioweapons program in U.S. mainstream and fringe news websites.

1. Narrative diffusion: A process that creates narrative similarity as an empirical relic
2. When conducting or using one word, you should do a related literature review of this specific word
3. Gold standard approach: Providing infinite resources to identify narrative similarity
4. LLM-SBERT: Using the key claims and subjects to measure the similarity
5. How to minimize the scale of large pairwise comparisons:
   1. Use LLMs to distill the core claims and subjects
   2. Using the pretrained BERT-based model to generate a semantic similarity
   3. Using human annotators to label pairs identified as positive cases

Questions: How to deal with LLMs bias?

1. Social relations, particularly power relations, interactively constitute the production, distribution, and consumption of resources; this requires the political economy of communication to examine the constantly shifting modes of control along the lines of production, distribution, and consumption — or how activists use new media to resist the concentration of power. However, within the communication industry, producers, distributors, and consumers cannot be precisely defined: audiences cannot be straightforwardly identified as consumers, for in consuming media products they also produce symbolic value.
2. An alternative definition of political economy: the study of control and survival in social life.

1. The fundamental analytical framework of the political economy of communication: commodification, spatialization, and structuration
2. The audience commodity thesis: Workers sell their labor power (attention) in return for the media content they receive; this extends the scope of labor, expands the traditional logic of the commodity, and may serve as an account of the labor transformation process under capitalism
3. Labor (Braverman, 1974): A conceptual unity of imagination and creative capacity, design and execution; in the process of commodification, capital functions to separate conception from execution, to separate skill from the unskilled performance of tasks, thereby forming a managerial class, while new configurations of skill and power in production are combined to restructure the labor process
4. From commodification to spatialization: The emergence of information systems enables information to transform space — redistributing industrial resources
5. The tradition of the political economy of communication treats spatialization as the institutional extension of corporate power within the communication industry

1. A totalitarian society has two distinguishing characteristics:
   1. The nature of the tie between the totalitarian party and its active adherents
   2. Social atomization: “the obliteration of social ties that are not directly harnessed to the party’s aims. Totalitarian societies recognize no legitimate distinction between private and public spheres.”

Questions: Does social atomization only exist in a totalistic society? What about the rise of media and internet platforms?

1. He insisted that the modern socialist society became less enforced. Where the totalitarian image places its emphasis on the disincentives and psychological states created by fear and inbred caution, the neo-traditional image emphasizes the meshing of economic and political power on the structured incentives offered by the party.
2. How the social structure works: The result is a highly institutionalized network of patron-client relations that is maintained by the party and is integral to its rule: a clientelist system in which public loyalty to the party and its ideology is mingled with personal loyalties between party branch officials and their clients.
3. How it’s maintained: The neo-traditional image stresses the social network, not the group, as its main structural concept.

Previous Problems: As harmful content fostered many problems, and social platforms conducted content moderation, many users held low trust towards the restrictions. However, many actors tried to use community guidelines to restrict users, but created a vast difference. No research has put attention on users’ situation.

Research Questions: Focusing on the length, readability, and semantic complexity of community guidelines.

1. Platform Typology: Chat platforms, creator platforms, forum platforms, social network platforms
2. Alt Tech Platforms: Provide alternatives to Silicon Valley-controlled platforms like Gab
3. Platform Governance Archive (PGA): GitHub repository makes historical versions available
4. Readability Measurement: Flesch-Kincaid Grade Score relies on average sentence length and average word length in a given text