In goal-oriented natural language processing, success is often binary and easily verifiable. For instance, in a flight-booking system like the one studied by Li in AirDialogue: An Environment for Goal-Oriented Dialogue Research, an agent either satisfies the user's constraints or it does not. There is a mathematically defined ground truth.
However, when we transition to open discourse—such as open-domain dialogue, role-playing, social commentary, or computational argumentation—this ground truth vanishes. How do you objectively measure if a chatbot's response is "interesting" or if a legal argument is "convincing"? As Ashley et al. noted in their early work on adversarial legal reasoning (Compare and Contrast: A Test of Expertise), determining the relative quality of competing arguments is inherently subjective. This subjectivity creates a fundamental bottleneck: if human experts cannot agree on what makes a response or argument "good," developing automated evaluation functions becomes exceptionally difficult.
The Dimensions of Disagreement
To understand why evaluating open discourse is so challenging, we must look at where human judgments diverge. Subjectivity is not just random noise; it is driven by several systematic factors:
- Cultural and Contextual Norms: What constitutes an appropriate or empathetic response varies wildly across social groups. As Shang et al. point out in the Overview of the NTCIR-12 Short Text Conversation Task, conversational suitability depends heavily on cultural appropriateness. Similarly, Zhang et al. demonstrate in xDial-Eval: A Multilingual Open-Domain Dialogue Evaluation Benchmark that automated metrics focusing purely on objective coherence completely miss nuanced cultural expectations surrounding traits like humor or empathy.
- Domain Complexity: The level of disagreement among human raters changes depending on the topic. In a study on argument quality by Lauscher et al. (Creating a Domain-diverse Corpus for Theory-based Argument Quality Assessment), inter-annotator agreement plummeted when evaluating highly convoluted or subjective domains like political debates, whereas structured formats like product reviews yielded much higher consensus.
- Structure vs. Rhetoric: Identifying the structural components of an argument (e.g., premises and claims) is relatively objective. However, assessing its convincingness or rhetorical impact is deeply subjective. Gurevych emphasizes in Latest News in Computational Argumentation: Surfing on the Deep Learning Wave, Scuba Diving in the Abyss of Fundamental Questions that evaluating the perceived strength of a claim requires modeling its rhetorical impact within a specific context, a task distinct from simple structural parsing. This extends to delivery as well; Petukhova et al. (Multimodal Markers of Persuasive Speech: Designing a Virtual Debate Coach) highlight that nuances of how a debater presents themselves—such as confidence and assertiveness—remain difficult to quantify objectively.
Why Traditional Metrics Fail
When evaluating abstract qualities like "engagingness," "personality consistency," or "constructiveness," traditional automated metrics break down. As highlighted by Kumar et al. in AMUSED: A Multi-Stream Vector Representation Method for Use in Natural Dialogue, algorithmic benchmarks fail to capture whether a conversation is genuinely interesting, forcing researchers to rely on expensive and unscalable human evaluations.
Even when human evaluation is used, standard crowd-sourcing methodologies frequently suffer from a lack of calibration. See et al. (What makes a good conversation? How controllable attributes affect human judgments) observed high subjectivity and low calibration among human raters judging dialogue engagement. This issue is compounded when using absolute numeric scales (e.g., rating constructiveness from 1 to 10). As shown by Fujita et al. in Dataset Creation for Ranking Constructive News Comments, absolute rating scales vary so significantly between individual annotators that they fail to provide a reliable ground truth for training machine learning models.
Furthermore, open-domain conversations lack a unique structural mapping. The authors of Structured Attention for Unsupervised Dialogue Structure Induction note that real-world human conversations can be mapped in multiple valid ways, making it exceptionally difficult to establish standardized performance metrics for dialogue structure induction. This ambiguity extends to downstream tasks like summarization; Chen et al. (DialogSum Challenge: Summarizing Real-Life Scenario Dialogues) note that summarizing dialogues objectively becomes difficult when the conversation contains inherently subjective or unverified statements.
Algorithmic Solutions for Subjective Spaces
To build reliable models in the face of this inherent subjectivity, researchers have shifted from trying to eliminate disagreement to modeling it. Several key mechanisms have emerged to handle this transition.
Unanimity-Aware and Consensus Metrics
Instead of treating low inter-annotator agreement as a failure, some frameworks use it as a core signal. In the Overview of the NTCIR-13 Short Text Conversation Task, Shang et al. introduced evaluation frameworks that incorporate unanimity-aware metrics. By weighting evaluation scores based on the level of agreement among multiple human judges, the metric prioritizes responses that are universally recognized as high-quality. This strategy shifts the goal from finding a single "correct" answer to identifying universally acceptable responses.
Pairwise and Relative Ranking
Because humans are notoriously poor at assigning absolute scores but excellent at making relative comparisons, shifting the task framing can drastically improve reliability. Toledo et al. explored this in Automatic Argument Quality Assessment - New Datasets and Methods, examining whether argument quality is better captured through absolute individual scores or relative pairwise comparisons. Pairwise ranking mitigates individual annotator calibration bias, transforming a vague subjective judgment into a concrete choice ("Is argument A more convincing than argument B?"), which allows models to learn a more stable relative ordering of argumentative strength.
Multidimensional Taxonomies and Persona Constraints
When evaluating complex conversational agents, such as role-playing models, success cannot be measured by factual correctness. Tu et al. demonstrate in CharacterEval: A Chinese Benchmark for Role-Playing Conversational Agent Evaluation that role-playing success must be evaluated through subjective traits like emotional resonance and stylistic consistency.
To handle this, researchers break down abstract concepts into explicit, multidimensional taxonomies. For instance, Ferron et al. introduce a structured taxonomy of "engagingness" in MEEP: Is this Engaging? Prompting Large Language Models for Dialogue Evaluation in Multilingual Settings, allowing large language models to evaluate dialogue quality more precisely across different languages by focusing on specific behavioral dimensions rather than a single vague impression.
Human Preference Rewards
Finally, instead of optimizing for token-matching metrics like BLEU, modern architectures optimize directly for human preference distributions. In ViCo: Engaging Video Comment Generation with Human Preference Rewards, Sun et al. utilize a reward model trained on human engagement signals (such as "likes") to guide reinforcement learning. This allows the system to generate subjective, engaging social commentary without relying on a single, rigid ground-truth text, effectively letting the model learn the implicit distribution of human preference.
Key Foundational Work
Compare and Contrast: A Test of Expertise (1987) — contributed 2 concept elements to this direction.
A Grammatico-Statistical Approach to Discourse Partitioning (1994) — contributed 2 concept elements to this direction.