Pith. sign in

REVIEW 5 major objections 5 minor 36 references

CREFT: Sequential Multi-Agent LLM for Character Relation Extraction

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A sequential chain of specialized LLM agents produces more accurate and complete character networks from drama scripts than a single-agent pass.

desk verdict A useful multi-agent pipeline for character relation extraction with real completeness gains, but the accuracy claims rest on unvalidated similarity metrics and need a statistics pass. read the letter →

arxiv 2505.24553 v1 pith:WY5PN7TL submitted 2025-05-30 cs.CL cs.AI

Ye Eun Chun , Taeyoon Hwang , Seung-won Hwang , Byung-Hak Kim This is my paper · ORCID
classification cs.CLcs.AI
keywords characterrelationextractionmulti-agentLLMnarrativeanalysisknowledgedistillationgraphKoreandramascriptsPersonalizedPageRanksemanticsimilarity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper is trying to establish that breaking character-relation extraction into a sequence of specialized LLM agents, each handling one refinement task, produces more accurate and complete character graphs from long-form narratives than asking one LLM to do everything in a single pass. This would matter because script reviewers and production teams need to grasp who matters to whom in a story quickly, and existing single-pass methods miss implicit relations and leave alias duplicates. On a curated set of 15 Korean drama scripts, the sequential pipeline raises character recall by 18.8 percentage points and group-match F1 by 16.5 points over the single-agent baseline. The paper also positions the approach as a way to distill episodes 1 through 4 into a visual character relation structure before full scripts are available.

What carries the argument

The central object is the Character Relation Structure (CRS), a character graph whose nodes are characters and whose edges are explicit or implicit relations, enriched with roles and group assignments. The mechanism that carries the argument is a sequential refinement chain: a fine-tuned Korean LLM first converts script chunks into SPO triplets to form a base graph, then a chain of specialized agents performs PPR-based character selection, duplicate-node merging, explicit and implicit relation extraction, filtering of irrelevant characters, role identification, and grouping. Each agent consumes the output of the previous one, and the sequence is what distinguishes CREFT from a single comprehensive prompt.

What would settle it

Re-score a random sample of predicted roles, group names, and relations by human annotators and check whether the KURE-v1 cosine-similarity rankings agree with their correctness judgments; substantial disagreement would mean the reported gains are artifacts of the metric rather than real extraction improvements.

Watch

Extended reading notes

Core claim

The paper claims that its sequential multi-agent framework, CREFT, significantly outperforms a single-agent LLM baseline at building a Character Relation Structure (CRS) from the first four episodes of Korean drama scripts. On fifteen curated dramas, the multi-agent pipeline raises character recall from 54.1% to 72.9%, group-match F1 from 25.9% to 42.4%, and character-relation recall from 40.8% to 50.4%, with smaller gains on role, group-name, explicit-relation, and implicit-relation similarity. The same experiment also shows that the single-agent approach collapses entirely on some dramas, while the multi-agent approach is more consistent. A separate experiment finds that Personalized PageRank character selection is not uniformly better than a simple edge-count method; the better method depends on whether key characters appear often in the early episodes.

Load-bearing premise

The load-bearing premise is that cosine similarity from the KURE-v1 embedding model reliably measures whether a predicted role, group name, or relation matches the human-annotated ground truth; the paper does not validate that this metric tracks human judgment.

Editorial extensions

If this is right

  • If the result holds, a pipeline of chained agents can extract a usable character relation structure from only the first four episodes of a drama, without access to the full script.
  • Breaking one complex instruction into focused prompts appears to prevent failures like leaving aliases unmerged and missing implicit relations, which dominate single-agent errors.
  • Role identification is the easiest component, while group assignment remains the weakest part of the pipeline, so further work should target grouping.
  • Character selection by Personalized PageRank is not uniformly better than simple edge counting; narrative pacing determines which method wins.
  • Because each agent's output is a plain-text intermediate, the pipeline can be inspected or corrected at any stage during production review.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that the same sequential design could transfer to novels, films, or serialized podcasts, where alias resolution and implicit relationship tracking are equally central.
  • A testable extension is to vary the order or number of agents; the current design fixes one order, and the per-drama variance suggests some narratives might benefit from re-running the merge step after grouping.
  • The metric construction, which selects the highest-similarity pair when multiple roles or relations exist, may favor broad or verbose predictions; checking sensitivity to that matching rule would sharpen the comparison.
  • The confidentiality motivation for distilling GPT-4o instead of calling APIs directly suggests a production pattern for using frontier models on unreleased content through local distillation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes CREFT, a sequential multi-agent LLM framework for extracting character relation structures from Korean drama scripts. It first builds a base character graph via distilled SPO-triplet extraction and then refines it through specialized agents for character selection (using Personalized PageRank), duplicate-node merging, explicit/implicit relation extraction, filtering, role identification, and grouping. The evaluation compares the multi-agent pipeline against a single-agent prompt baseline on 15 dramas using seven metrics, and compares PPR-based character selection with a simple edge-count baseline. The main tables report average gains for the multi-agent approach on all seven metrics, including an 18.8-point absolute improvement in character recall and a 16.5-point improvement in group-match F1, but the appendix per-drama tables also show several reversals and the paper does not report statistical significance tests.

Significance. If the empirical comparison is valid, CREFT is a useful practical contribution to narrative analysis and script review, with two notable strengths: the secure knowledge-distillation route that avoids calling external APIs on confidential scripts, and the detailed prompt suite in the appendices. The exact-match metrics (character recall and group-match F1) provide independent, robust support for the completeness dimension of the central claim, and the evaluation is against human-annotated ground truth, so I do not see a circularity problem. However, the accuracy dimension of the central claim is not yet established: the semantic-similarity metrics are unvalidated and lenient, the per-drama results show large reversals, and no significance tests are reported. The paper also leaves open a possible training/evaluation overlap and does not show that the PPR component is beneficial.

major comments (5)
  1. [§4.2, Table 9] The scoring rule for the four semantic similarity metrics is too lenient to support the accuracy claims. The text states: 'If multiple roles or relations exist for the same character, we select the pair with the highest similarity score; any role or relation absent in the ground truth is excluded from evaluation.' As a result, a model that emits many candidate labels for a character or pair is scored by its single best match and is never penalized for extra labels not in the ground truth. This can inflate the reported multi-agent improvements (Role Sim. 76.7→84.8, Group Name Sim. 64.5→81.5, Explicit Rel. Sim. 70.2→74.0, Implicit Rel. Sim. 55.8→60.0) if the multi-agent pipeline produces more candidate outputs than the single-agent baseline. The paper provides no evidence that KURE-v1 cosine similarity tracks human semantic judgments, and no inter-annotator agreement is reported for the ground-truth annotations. Please report exact-match precision/recall/F1 for roles and relations in addition to the similarity scores, and validate KURE-v1 against human ratings or replace these metrics with stricter ones.
  2. [Table 1 vs. Appendix F] The abstract and Section 5.1 claim that CREFT 'significantly outperforms' the single-agent baseline, but no statistical test is reported. The per-drama tables in Appendix F show large reversals, e.g., Drama 1 Group Match F1: 35.7 (single) vs 0.0 (multi); Drama 12 Implicit Relation Sim.: 32.8 vs 0.0. With 15 dramas, a paired nonparametric test (e.g., Wilcoxon signed-rank) or bootstrap confidence intervals on the differences should be reported, and the per-drama standard deviations shown in the appendix should also appear in the main table. Without this, the word 'significant' is not supported by the evidence.
  3. [§3.1 and §4.1] The paper does not state whether the 15 evaluation dramas are disjoint from the 355 scripts used to create the distilled SPO-triplet training set. Both sets cover episodes 1–4 of Korean dramas aired in overlapping periods (1995–2023 vs. 2014–2023), so there is a potential data-leakage concern for the base character graph. Please state explicitly that the 15 dramas are held out from the distillation training data, or re-run the evaluation on a strictly held-out split. This is needed to establish that the reported absolute numbers are not inflated by training on the test scripts.
  4. [§5.1, Table 2] Experiment 2 does not support PPR as a beneficial component of CREFT. The edge-count baseline has slightly higher average precision (47.2 vs. 46.8), recall (79.2 vs. 78.5), and F1 (58.7 vs. 58.2), and the text reports that each method wins in five dramas. The conclusion that the optimal strategy depends on narrative pacing is post hoc and lacks a concrete decision rule. Either demonstrate a reliable benefit for PPR or present it as a neutral design choice rather than as a validated component of the framework.
  5. [Appendix D] The definition of Character-Relation Recall is ambiguous: 'Number of matched character pairs between ground truth and predictions / Total number of character pairs in ground truth' does not specify when a pair is considered matched. Is matching based on exact character names in the final CRS, on resolved identities after the merging step, on unordered pairs, or on a semantic match? Since this metric contributes to the completeness claim, please define it precisely.
minor comments (5)
  1. [Throughout] There are several typos and formatting issues: 'Charater Recall' in the Table 11 caption, 'relatioinships' in Table 6, 'to to leverage' in Appendix E, and 'The exact prompt 3' in §4.3 should be 'The exact prompts'. Please proofread the manuscript.
  2. [References] The KURE-v1 citation (Youngjoon et al., 2024) lacks a venue, technical report number, or URL, making it impossible for readers to verify the model or its properties. Please provide a complete citation or a link to the model card.
  3. [Figure 2] The letter labels in Figure 2 are difficult to read in the current rendering; please ensure that all node labels (E, O, L, P and E', O', L', P') are legible in the final version.
  4. [§3.2 and §6] Section 3.2 states that the approach 'begins by collecting user-provided information about main and sub-characters,' which makes the pipeline semi-automatic; the conclusion's claim that CREFT 'automates' character relationship extraction should be qualified to reflect this human input.
  5. [§4.3] The hyper-parameters (PPR threshold 0.02, importance scores 1 and 0.5, chunk size 512 characters) are fixed without sensitivity analysis. Please report robustness to these choices or justify them from data.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central multi-agent comparison is evaluated against human-annotated ground truth with fixed, symmetric metrics, and no derivation reduces to its own inputs.

full rationale

The paper's central claim, that CREFT outperforms single-agent LLM baselines, is tested against human-annotated ground truth for 15 Korean dramas. Both pipelines start from the same base character graph and use the same PPR-based character selection, so the comparison is controlled for those shared components. The evaluation metrics are applied symmetrically to both systems: exact-match metrics (character recall, group match F1, character-relation recall) and KURE-v1 cosine-similarity metrics for roles, group names, and relations. KURE-v1 is used only as a fixed scoring function, not as a fitted parameter or as a source of the predicted labels; it is not used to train or select the multi-agent outputs. The scoring rule that selects the highest-similarity pair and excludes predicted labels absent from the ground truth is lenient, and the paper does not validate KURE-v1 against human judgment, but this is an evaluation-validity concern rather than circularity: the ground truth remains external and human-defined, and the rule does not make the reported gains true by construction. The only apparent self-citation, Jeong et al. (2025) in Section 3.1, is used to justify the knowledge-graph construction approach for the base graph; it is not load-bearing for the multi-agent refinement claim, and the base graph is identical for both compared systems. The PPR threshold (0.02) is a hand-set hyperparameter, not fitted to the evaluation results, and Experiment 2 does not claim superiority. The paper does not rename a known result, import a uniqueness theorem, or smuggle an ansatz via citation. Overall, the derivation chain is self-contained with respect to the claims tested, and no step reduces to its own inputs.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests mostly on domain assumptions about the dataset, the metric, and the label space. The only hand-set numeric parameters are the PPR threshold and importance scores; the evaluation adds a semantic-similarity assumption that is untested. No new entities are introduced.

free parameters (3)
  • PPR threshold = 0.02
    Section 4.3 sets a threshold of 0.02 to filter low-relevance characters; sensitivity is not reported, and it may be tuned on the 15 evaluation dramas.
  • Main/sub character importance scores = 1 and 0.5
    Section 4.3 assigns importance 1 to main characters and 0.5 to sub-characters for Personalized PageRank; these values are chosen by hand with no ablation.
  • Chunk size for triplet extraction = 512 characters
    Section 3.1 divides scripts into 512-character segments; the size is arbitrary and not ablated.
assumptions (5)
  • domain assumption Key characters can be identified from episodes 1-4 via PPR seeded by user-provided main/sub characters
    Section 3.2 and footnote 1 state that producers only have access to early episodes, so this restricted scope is treated as sufficient for character selection.
  • domain assumption The Implicit Relationship List is exhaustive for meaningful character dynamics
    Appendix B.2 requires the model to select implicit relations only from a fixed list, constraining the label space and potentially missing valid relations outside the list.
  • domain assumption KURE-v1 cosine similarity reflects semantic accuracy for role, group, and relation labels
    Section 4.2 uses KURE-v1 for all similarity metrics without human validation or comparison to alternative embedding models.
  • domain assumption The human-annotated ground truth on 15 dramas is accurate and complete
    Section 4.1 describes expert annotation but reports no inter-annotator agreement and provides no dataset release for external verification.
  • standard math Personalized PageRank and cosine similarity are valid mathematical tools for this task
    Used in Sections 3.2 and 4.2; both are standard algorithms with well-known properties.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CREFT: Sequential Multi-Agent LLM for Character Relation Extraction." pith.science (2026). https://pith.science/paper/WY5PN7TL

@misc{pith2026250524553,
  author       = {Pith},
  title        = {Pith review of: CREFT: Sequential Multi-Agent LLM for Character Relation Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WY5PN7TL}},
  note         = {Machine review of arXiv:2505.24553}
}
read the original abstract

Understanding complex character relations is crucial for narrative analysis and efficient script evaluation, yet existing extraction methods often fail to handle long-form narratives with nuanced interactions. To address this challenge, we present CREFT, a novel sequential framework leveraging specialized Large Language Model (LLM) agents. First, CREFT builds a base character graph through knowledge distillation, then iteratively refines character composition, relation extraction, role identification, and group assignments. Experiments on a curated Korean drama dataset demonstrate that CREFT significantly outperforms single-agent LLM baselines in both accuracy and completeness. By systematically visualizing character networks, CREFT streamlines narrative comprehension and accelerates script review -- offering substantial benefits to the entertainment, publishing, and educational sectors.

Figures

Figures reproduced from arXiv: 2505.24553 by the authors.

Figure 1
Figure 1. Overview of the CREFT Framework. The base character graph is generated by extracting SPO triplet data from narrative texts and subsequently refined through specialized LLM agents. Each agent incrementally improves character composition, explicit/implicit relation extraction, role identification, and group assignments, significantly enhancing clarity and accuracy of narrative understanding for early-stage content eva… view at source ↗
Figure 2
Figure 2. CRSs for Drama 13. In (a), the single-agent approach fails to merge aliases (E, O, L, P), producing redundant nodes (E’, O’, L’, P’). In contrast, (b) shows the multi-agent approach accurately merging these aliases for a more concise representation. five individual dramas, while delivering equivalent results in the others. This suggests that the effec￾tiveness of each method may depend on specific narrative characte… view at source ↗
Figure 3
Figure 3. Step-by-step refinement of the CRS for Drama 13. Each subfigure illustrates the incremental changes at a distinct process stage (see Section 3.2 for details) [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison of Character Extraction Algorithm (PPR vs. Edge Count). [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 35 canonical work pages

  1. [1]

    List only same person name pairs

    **[Hee-su’s mother]-[Seon-ae / Seon-ae Eun]** - Must list two different name sets from <Character Name List>. List only same person name pairs. Do not repeat the same name as answer name pairs. Do not include any partial name of name set as another name. If there is no same character within <Character Name List>, just write as [Character1]-[No Same Person...

    work page

  2. [2]

    [Repeat for all characters from <Character Name List>.] </Response Format> <Output Example> <Character Name List Example> [Character A] [Character B] [Character C] [Character D] </Character Name List Example> Correct <Character Pairs> Example:

    work page

  3. [3]

    **[Lawyer Hong] - [Hong Ji-yoon]: Lawyer Hong and Hong Ji-yoon are the same person.** Incorrect <Character Pairs> Example:

    work page

  4. [4]

    If any specific relationships occur too frequently and disrupt diversity in the relationships list, select these for review

    Identify Overused or Abundant Relationships: - Task: Review the <Character Relationship List> again. If any specific relationships occur too frequently and disrupt diversity in the relationships list, select these for review. <Provided Information> <Character List>[Put your character list from previous step here]</Character List> <Character Identity List>...

    work page

  5. [5]

    **[Lee Ho]-[King] **Evidence: King represents Lee Ho’s title, and Lee Ho is the king’s real name.**

    work page

  6. [6]

    **[Hwang Gwi-in]-[No Same Person]

    work page

  7. [8]

    **[Cheong-ha] - [Cheong-ha]

    work page

  8. [9]

    Peer” for classmates) rather than the final one (e.g., “colleague

    **[Crown Prince] - [King] **Evidence: The Crown Prince and the King are the same person, and King represents a title. Remember: - List names in the same order of <Character Name List>. - Do not repeat the same name as answer name pairs. - *** In Korean, the name ‘Young-min Cha’ is composed of the last name ‘Cha’ and the first name ‘Young-min’. - If two ch...

    work page

Show all 36 references
  1. [10]

    Information not provided

    **Subject: [As in Initial Character Knowledge Graph]** **Object: [As in Initial Character Knowledge Graph]** **(Explicit) Who is Subject regarding to Object]:** [If there is a hierarchical relationship between Subject and Object, use a specific term (e.g., supervisor, subordin...

  2. [11]

    **Subject: [Repeat for the next pair]** **Object: [Repeat for the next pair]** **(Explicit) Who is Subject regarding to Object]:** [Repeat for the next pair] **Verification: [Repeat for the next pair]** **(Implicit) What emotions does Subject experience toward Object?:** [Repe...

  3. [12]

    Seung-won Han is Seung-tak Go’s older male cousin

    **Subject: Seung-tak Go / Seung-tak ** **Object: Seung-won / Han Seung-won ** **(Explicit) Who is Subject regarding to Object]: older male cousin** **(Explicit) Verification: [Incorrect] Seung-tak Go is not Seung-won Han’s older male cousin. Seung-won Han is Seung-tak Go’s old...

  4. [13]

    **(Implicit) What emotions does Subject experience toward Object?: wariness**

    **Subject: Seung-won / Han Seung-won** **Object: Seung-tak Go / Seung-tak** **(Explicit) Who is Subject regarding to Object]: older male cousin** **(Explicit) Verification: [Correct] Seung-won Han is Seung-tak Go’s older male cousin. **(Implicit) What emotions does Subject exp...

  5. [14]

    **(Implicit) What emotions does Subject experience toward Object?: wariness**

    **Subject: Hyeri / Hyeri Go** **Object: Sangwook / Sangwook Ju** **(Explicit) Who is Subject regarding to Object]: colleague** **(Explicit) Verification: [Incorrect] Hyeri is Sangwook’s subordinate. **(Implicit) What emotions does Subject experience toward Object?: wariness**

  6. [15]

    Information not provided

    **Subject: Sangwook / Sangwook Ju** **Object: Hyeri / Hyeri Go** **(Explicit) Who is Subject regarding to Object]: supervisor** **(Explicit) Verification: [Correct] Sangwook is Hyeri’s supervisor. **(Implicit) What emotions does Subject experience toward Object?: wariness** </...

  7. [16]

    Identify and select any names that seem to be a pronoun (e.g., ‘doctor’, ‘patient’, ‘professor’) not a specific character name

    Identify General Character Names: - Task: Review the <Character List> . Identify and select any names that seem to be a pronoun (e.g., ‘doctor’, ‘patient’, ‘professor’) not a specific character name. **If the last name exists it refers a specific person so do not include this ...

  8. [17]

    Identify and select any entries that seem to describe an action, reaction, or state rather than a stable identity or occupation

    Identify Inappropriate Character Identities: - Task: Review the<Character Identity List> . Identify and select any entries that seem to describe an action, reaction, or state rather than a stable identity or occupation

  9. [18]

    Identify and select entries that describe temporary actions or states instead of defining relationships

    Identify Inappropriate Character Relationships: - Task: Review the <Character Relationship List> . Identify and select entries that describe temporary actions or states instead of defining relationships

  10. [20]

    General Character List:

  11. [21]

    **Character: [Write the exact name only from the <Character List> .]** **Last Name: [True/False: Check if the name contains last name.]**

  12. [22]

    **[Continue to select all general character names from the <Character List> ]**

  13. [23]

    Inappropriate Character Identity List **[Write an exact word from the <Character Identity List> ]** **[Continue to select all inappropriate character identities]**

  14. [24]

    Inappropriate Character Relationship List **[Write an exact word from the <Character Relationship List> ]** **[Continue to select all inappropriate character relationships]**

  15. [25]

    Information not provided

    Abundant Relationship List **[Write an exact word from the <Character Relationship List> ]** **[Continue to select all abundant relationships]** If there is no abundant relationship, state “Information not provided”. </Response Format> Remember: - **If the last name exists it ...

  16. [26]

    Use an exact name set from <Character Name Sets> .]** **Role: [Main role(s) of the character: Write with specific details if available

    **Character: [Do not alter any name set provided in <Character Name Sets> . Use an exact name set from <Character Name Sets> .]** **Role: [Main role(s) of the character: Write with specific details if available. For example, instead of simply writing ‘prince’, use ‘first princ...

  17. [27]

    No Group

    [Repeat as above for all name sets from <Character Name Sets> ] <Examples> Correct: [lawyer Kim] - Role: prosecutor (based on information stating they work as a prosecutor) Incorrect: [lawyer Kim] - Role: lawyer (based solely on the character’s name) <Source Priority> - If the...

  18. [28]

    No Group

    **Character: [**Use the exact name set from <Character List>. ]** **Group: [Select one group from the Group List or state “No Group”.]** **Family: [If family members exists in <Character List>, answer as ‘Yes’ otherwise ‘No’.]** **Rationale: [Provide a reason why this characte...

  19. [29]

    affection- ate

    [Repeat as above for all name sets from <Character List>] Remember: - **IMPORTANT**: Group membership is exclusive—no character should appear in more than one group. - **Use the exact name set from <Character List>, [Young-min / Young-min Cha] is the full name set. - **If the ...

  20. [30]

    Relationships should not focus on temporal or incidental interactions but rather represent lasting or significant connections

    Your task is to identify and categorize explicit and implicit relationships for each character pair based on the <Character Information> and <Summary>. Relationships should not focus on temporal or incidental interactions but rather represent lasting or significant connections

  21. [31]

    Your task is to identify general character names

  22. [32]

    Your task is to identify one main role (e.g., lawyer, cardiothoracic surgeon, unemployed, king) for each character in <Character Name Sets> based on the <Character Information> and <Summary> provided

  23. [33]

    Your task is to identify and categorize essential groups of characters in <Character Role> based on family relationships, shared affiliations, professions, or significant roles, using the <Character Information>, <Summary>, and <Character List> provided

  24. [34]

    Your task is to to leverage diverse perspectives to arrive at the most robust and accurate character relationship chart for the K-drama, considering the factor (**Power dynamics and alliances**, **Family relationships**, **Romantic relationships**) <Key Directives for task #1>...

  25. [35]

    When Subject and Object are the same person: **Subject Character Name** : **Name1** (in Korean) **Object Character Name** : **Name2** (in Korean) **Explicit Relation**: **None** (When Subject and Object are the same person!) **Implicit Relation**: **None** (When Subject and Ob...

  26. [36]

    **Implicit Relation**: Choose a hidden or indirect relationship from <Implicit Relationship List>

    If Subject and Object are NOT the same person, AND both are in Main-Sub Character list: **Subject Character Name** : **Name1** (in Korean) **Object Character Name** : **Name2** (in Korean) **Explicit Relation**: Describe the superficial relationship (e.g., ‘Sibling’, ‘Mentor’,...

    2014

  27. [2010]

    In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 138–147, Uppsala, Sweden

    Extracting social networks from literary fic- tion. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 138–147, Uppsala, Sweden. Association for Compu- tational Linguistics. Yeonseok Jeong, Minsoo Kim, Seung won Hwang, and Byung-H...

    2025

  28. [2013]

    In Proceedings of the 51st Annual Meeting of the Associ- ation for Computational Linguistics, pages 352–361, Sofia, Bulgaria

    Learning latent personas of film characters. In Proceedings of the 51st Annual Meeting of the Associ- ation for Computational Linguistics, pages 352–361, Sofia, Bulgaria. Association for Computational Lin- guistics. Sabyasachee Baruah and Shrikanth Narayanan. 2024. Character a...

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.