TL;DR
This paper argues that current dialogue models, focused on topical coherence, fail to capture the nuances of social conversations. They introduce "new-topic utterances" (NTUs) as a key element of social talk and propose a new annotation strategy and classification for these utterances, paving the way for more sophisticated dialogue models that account for non-topical social coherence.
What Problem Does This Paper Address?
Traditional dialogue models excel at understanding task-oriented conversations where utterances are tightly linked by a shared goal and topic. However, these models struggle with social talk, or casual conversation, which is characterized by open-ended topics and less defined purposes. A significant challenge in modeling social talk is understanding how conversations remain coherent even when speakers introduce new topics. These "new-topic utterances" (NTUs) don't directly follow from the previous discourse but are crucial for maintaining social connection. Existing models often treat NTUs as incoherent, missing a vital aspect of natural human interaction. This paper aims to bridge this gap by identifying and characterizing these NTUs in social talk.
Methods
The authors propose a novel approach to identifying NTUs by first annotating content-based coherence relations between utterances. NTUs are then identified as those utterances that lack such a coherence relation to the prior discourse. This is a reverse approach to directly identifying NTUs. They leverage existing frameworks like the Penn Discourse Treebank (PDTB) and Cognitive approach to Coherence Relations (CCR), but enrich them with new annotations for social talk.
Key methodological steps include:
- Annotation Strategy: A pilot annotation study was conducted on social dialogues from the Disco-SPICE corpus.
- Enrichment: The corpus was enriched with pragmatic annotations (e.g., speech acts) and new discourse relations, particularly those that capture interactional aspects rather than just topical ones.
- NTU Identification: NTUs were identified heuristically as utterances that did not have a clear content-based coherence relation to preceding discourse.
- Pattern Identification: The interaction between identified NTUs and their discourse context was analyzed to identify patterns of "discourse topic changes" (DTCs) that signal non-topical coherence.
- Classification: Based on these patterns, a new classification of NTUs was proposed, introducing "sequence-based social intents" that go beyond traditional speech act taxonomies.
Datasets
- Disco-SPICE corpus: This corpus of naturally occurring social dialogues was used for the pilot annotation study. It was previously annotated with discourse relations from the PDTB and CCR frameworks.
- SPICE-Ireland corpus: The original pragmatic annotations for the Disco-SPICE corpus were derived from this corpus.
- Riou (2015) dataset: The authors plan to use this dataset, which includes 15-minute extracts from the Santa Barbara Corpus of Spoken American English, for future in-depth studies. This dataset is noted for its publicly accessible audio files, which are valuable for annotation.
Key Insights and Ideas
- NTUs are not always incoherent: The paper argues that NTUs are a fundamental part of social talk and contribute to its coherence, rather than disrupting it.
- Social talk has its own coherence: Beyond topical coherence, social conversations rely on "interactional coherence" or "talk-that-does," which is driven by social goals and relational work between speakers.
- New social intents: The identified patterns of NTUs suggest new social intents that are not captured by existing speech act theories. These intents relate to managing social focus and the relevance of social domains.
- Bayesian game-theoretic models are promising: The authors propose that a Bayesian game-theoretic framework is a suitable architecture for modeling social talk due to its ability to capture rational and interactive aspects.
Key Results
- 38 true NTUs identified: The pilot study successfully identified 38 instances of NTUs, representing approximately 50% of the initial candidates. These NTUs were characterized by discourse-new topics and focused entities.
- New DTC patterns: Several patterns of discourse topic changes (DTCs) associated with NTUs were identified, including "true interruption," "forgotten topic," "goodbye-initialized topic," and various "interlocutor-centric" and "speaker-centric" moves.
- Classification of 31 NTUs: A classification of 31 NTUs was proposed, categorizing them based on their motivation (social focus, social domain relevance, or both).
Limitations and Future Work
- Pilot study limitations: The current study is a pilot, and the annotation was performed by a single author. Future work will involve a full-blown annotation project with multiple annotators to ensure reliability.
- Generalizability of social intents: The authors acknowledge a caveat regarding the cross-cultural generalization of the identified social intents.
- Developing a robust game-theoretic model: The immediate future work involves developing an actionable Bayesian game-theoretic model for social talk. This will require decomposing the utility function and learning from existing work on goal-directedness in social interaction. The authors aim to address the challenge of capturing multiple sociolinguistic dimensions of discourse context in their models.
Summary updated