[Paper Review] SocAoG: Incremental Graph Parsing for Social Relation Inference in Dialogues

Liang Qiu, Liang Yuan+6 more|arXiv (Cornell University)|Jun 2, 2021

Topic ModelingComputer Science54 references12 citations

TL;DR

This paper proposes SocAoG, an incremental graph parsing framework that models social relations and individual attributes jointly using an And-or Graph to ensure relational consistency. It employs an α–β–γ inference strategy with MCMC sampling to dynamically update relations and attributes in real time as new dialogue utterances arrive, achieving state-of-the-art performance on DialogRE and MovieGraph datasets by outperforming prior methods in relation inference accuracy.

ABSTRACT

Inferring social relations from dialogues is vital for building emotionally intelligent robots to interpret human language better and act accordingly. We model the social network as an And-or Graph, named SocAoG, for the consistency of relations among a group and leveraging attributes as inference cues. Moreover, we formulate a sequential structure prediction task, and propose an $\alpha$-$\beta$-$\gamma$ strategy to incrementally parse SocAoG for the dynamic inference upon any incoming utterance: (i) an $\alpha$ process predicting attributes and relations conditioned on the semantics of dialogues, (ii) a $\beta$ process updating the social relations based on related attributes, and (iii) a $\gamma$ process updating individual's attributes based on interpersonal social relations. Empirical results on DialogRE and MovieGraph show that our model infers social relations more accurately than the state-of-the-art methods. Moreover, the ablation study shows the three processes complement each other, and the case study demonstrates the dynamic relational inference.

Motivation & Objective

• To address the lack of relational consistency modeling in existing dialogue-based relation extraction methods.
• To integrate personal attributes (e.g., gender, profession) as inference cues for improved relation prediction.
• To enable dynamic, incremental inference of social relations upon each incoming utterance in a dialogue.
• To model the joint evolution of attributes and relations in a socially coherent manner using structured graphical representations.
• To provide interpretable reasoning paths through incremental graph updates that reflect human-like social inference.

Proposed method

• SocAoG models social relations and attributes as a hierarchical And-or Graph (AoG), encoding group-level consistency and contextual constraints.
• The α process performs discriminative prediction of attributes and relations based on utterance semantics using contextual embeddings.
• The β process updates social relations by enforcing consistency with predicted attributes and existing relational structure.
• The γ process refines individual attributes based on interpersonal relations, ensuring attribute-relation coherence.
• An MCMC sampling strategy is used to jointly infer posterior distributions over attributes and relations through iterative α–β–γ updates.
• The framework supports incremental parsing, enabling real-time belief updates with each new utterance in a dialogue.

Experimental results

Research questions

• RQ1 Can joint modeling of attributes and relations improve social relation inference in dialogues compared to isolated relation extraction?
• RQ2 How effective is incremental graph parsing with α–β–γ processes in tracking evolving social relations during dialogue?
• RQ3 To what extent do attribute cues (e.g., gender, profession) enhance relational inference accuracy?
• RQ4 How do the three processes (α, β, γ) complement each other in maintaining consistency and improving performance?
• RQ5 Can the model reveal interpretable reasoning paths through dynamic belief updates in social networks?

Key findings

• Our model achieves state-of-the-art F1 scores of 69.1 (σ=0.5) on the DialogRE dataset, outperforming prior methods.
• The ablation study confirms that all three processes (α, β, γ) contribute significantly, with α providing the primary predictive capacity.
• The α–β–γ strategy enables dynamic inference, with MCMC acceptance rates stabilizing quickly, indicating convergence after each utterance.
• Case studies show that the model correctly infers indirect relations (e.g., S1 and S2 as parents via S5’s utterance) and updates attributes like age and profession.
• The model successfully resolves contradictions in beliefs—e.g., detecting 'Mr. and Mrs. Geller' in turn 7 triggers a belief update from 'friends' to 'parents'.
• The convergence curve peaks at key information turns, demonstrating sensitivity to critical relational cues in dialogue.