Source-linked AI summary
Les Misérables Dialogue: A film-based network representation of social communication
Theodora Moldovan, Georgios Panayiotou, Davide Vega
TL;DR
The classic Les Misérables co-occurrence network and many communication networks omit important details of who communicates, what is said, and when. The authors reconstruct dialogue from the 1998 film as a temporal text network and add ten-dimensional social annotations. The resulting networks are disassortative, with the reported tendency weaker when all recipients are included.
Problem
Co-occurrence and sender–receiver edges can omit what was said, when it was said, and who was addressed or also received a message.
Source: §1 · pp. 1–2Method
The authors manually annotate the 1998 film’s utterances as temporal text networks and add classifier scores for ten social dimensions to support multilayer representations.
Source: §1 · p. 2, §2.3 · p. 6Results
Both communication networks are disassortative, most strongly for LMD-flata (r = −0.40); including all recipients weakens this tendency (r = −0.29).
Source: §3.1 · p. 8Takeaways & Limitations
The temporal text networks and multiplex representations provide a richer alternative to the canonical co-occurrence benchmark and a framework for studying communication in narrative media.
Source: §4 · p. 11Methods for analysing feature-rich networks can be computationally demanding, motivating more scalable approaches.
Source: §4 · p. 11
Abstract
from arXiv · showhide
In this work, we revisit the classic Les Misérables benchmark network from the perspective of social communication. Using the 1998 film adaptation of the novel, we construct the publicly available Les Misérables Dialogue temporal text network, which explicitly distinguishes speakers, addressees, and other recipients of each utterance, along with its content and timing. Using the ten social dimensions model, we also derive multilayer network representations of the network, in which layers correspond to distinct relational dimensions. Beyond the network dataset, this work offers a general methodology for building similarly rich communication datasets from similar narrative sources.
1 Introduction
The canonical Les Misérables network captures chapter co-occurrence, while communication networks can omit who speaks, who is addressed, what is said, and when. This work reconstructs the 1998 film’s dialogue as a richly annotated temporal text network and derives representations for studying distinct social dimensions and relationships over time.
1 Introduction
- The canonical network records chapter co-occurrence, while communication edges can omit utterance content, timing, intended addressees, and other recipients.These abstractions capture only part of social communication.Source: §1 · pp. 1–2
- The authors manually reconstruct the 1998 film’s dialogue as a temporal text network, annotating each utterance with its speaker, addressee, other recipients, content, and timing.Film cues such as gaze, body orientation, proximity, and tone provide evidence for identifying participants.Source: §1 · p. 2
- Each utterance receives scores for ten social dimensions, which are used to construct multilayer networks with one layer per dimension.The dimensions include forms such as trust, conflict, support, power, and intimacy.Source: §1 · p. 2
- The authors derive network projections for structural, temporal, textual, and multilayer analyses, comparing their structure with the canonical network and examining relationships as they evolve through the film.The paper presents the dataset and modeling approach as a framework for studying communication in narrative media.Source: §1 · p. 2
2 Les Mis´erables: Revisited
The section builds a communication network from the 1998 film dialogue, preserving utterance roles, content, and timing. It also derives character-level and multilayer representations for analysis.
2 Les Mis´erables: Revisited
- The film-derived network represents directed communication from speakers to intended addressees or other characters who receive an utterance.Visible participants and audiovisual cues support annotation of secondary recipients and interaction duration.Source: §2 · p. 3
2.1 Data collection
- The dataset contains 1,686 verbal utterances with text, speaker, addressee(s), recipient(s), and subtitle-based start and end times.Annotators used audiovisual cues in ambiguous cases, but did not encode those cues themselves as messages or edges.Source: §2.1 · pp. 3–4
- The annotations distinguish the utterance’s intended addressee from its broader set of recipients, which includes the addressee and other attending characters able to hear it.Addressees are always a subset of recipients; most utterances have one addressee, while some have none or address a group.Source: §2.1 · pp. 3–4
2.2 Annotation and validation
- Majority agreement exceeded 98% for every annotation component, while unanimous agreement remained above 90%.The validation annotators reviewed flagged utterances, discussed them, and reconciled the annotations.Source: Table 1 · p. 5, §2.2 · p. 5
2.3 Temporal text network construction
- The LMD temporal text network represents actors, messages, message content, and timed producer and recipient edges.Reception edges connect messages either to intended addressees only or to all annotated recipients; message attributes also include ten social-dimension scores from the classifier.Source: §2.3 · pp. 5–6
2.4 Derived network representations
- The derived character-level networks vary by recipient definition, whether interaction frequency is retained, and whether social dimensions are included.The multiplex projections use ten social-dimension layers and thresholds of 0.5 or 0.8.Source: §2.4 · p. 6, Multiplex projections · p. 6
Flattened projections
- LMD-flata links speaker–addressee pairs, while LMD-flatr also includes speaker–recipient pairs, so its edges include all those in LMD-flata.Source: Flattened projections · p. 6
Multiplex projections
- Multiplex projections add a layer-specific edge when a message’s social-dimension score meets the selected threshold.A message can contribute to multiple layers, while repeated qualifying messages between the same pair collapse to one edge per layer.Source: Multiplex projections · pp. 6–7
3 Analysis
The analysis compares communication-network projections with the canonical benchmark and examines how social-dimension patterns vary across relationships and over time. Including all recipients changes network structure, while the character pairs show distinct dimension profiles.
3.1 Network structure
- Approximately twice as many edges in LMD-flatr as in LMD-flata accompany average degree rising from 3.25 to 5.95 and clustering from 0.09 to 0.29.LMD-flatr also places all characters in one connected component; additional recipient ties close triangles in multi-party scenes.Source: §3.1 · pp. 7–8
- Both communication networks are disassortative, most strongly for LMD-flata (r = −0.40), while recipient inclusion weakens this pattern (r = −0.29).The canonical co-occurrence network is less disassortative (r = −0.17).Source: §3.1 · pp. 8–9
- Trust is the most prominent layer, followed by romance, support, and conflict; similarity has one edge and identity has none.Most layers remain sparse and weakly clustered; trust and romance have the largest connected components.Source: §3.1 · p. 9
3.2 Temporal dynamics of social dimensions
- At θ = 0.8, Cosette and Valjean exchange 230 messages, with trust prominent in both directions and rising conflict from Cosette to Valjean.Valjean-to-Cosette messages are also marked by romance scores, which the paper cautions indicate intimacy or affection in language, not a romantic relationship.Source: §3.2 · pp. 9–10
- Javert and Valjean exchange 224 messages dominated by conflict, especially from Javert to Valjean; other dimensions are also generally more prominent in that direction.The paper relates this pattern to Javert's pursuit of Valjean.Source: §3.2 · p. 10
- Fantine and Valjean exchange 89 messages, with trust prominent from Fantine to Valjean and their interaction plateauing around the film's first hour.The plateau reflects the comparatively short duration of their relationship following Fantine's death.Source: §3.2 · p. 10
4 Conclusion
The paper introduces a method for constructing feature-rich communication networks from film dialogue and applies it to Les Misérables. Its temporal text networks and multiplex representations offer a dataset and framework for studying social communication in narrative media.
4 Conclusion
- The film-dialogue method produces a richer alternative to the canonical co-occurrence benchmark and a framework for studying narrative-media communication at multiple abstraction levels.The conclusion identifies temporal analyses, higher-order representations, and comparison of social-dimension structures as future directions.Source: §4 · pp. 10–11
- The authors note that existing methods for analysing feature-rich networks can be computationally demanding, motivating more scalable approaches.The paper hopes the dataset can serve as a benchmark and template for constructing similarly detailed networks from other films and narrative sources.Source: §4 · p. 11
Data availability
- The generated network datasets are publicly available in the project repository.The conclusion also describes the repository's HIF-format higher-order representation.Source: §4 · p. 11, Data availability · p. 12