Explainable and Decomposed Evaluation for Natural Language Generation

40 concepts39 papers2 proposals

Traditional automated metrics like BLEU or ROUGE relied on rigid n-gram overlaps to evaluate generated text. As generative models grew more sophisticated, the community shifted toward neural metrics (such as BERTScore) and Large Language Model (LLM) judges. As noted by Zhen Li et al. in Leveraging Large Language Models for NLG Evaluation: Advances and Challenges, while these modern evaluators align much better with human intuition, they introduce a major bottleneck: they function as opaque black boxes.

When an evaluator outputs a single scalar value (e.g., 0.65), it leaves developers in the dark. Is the score low because of a minor grammatical typo, or a catastrophic factual hallucination? Weizhe Yuan et al. point out in T5Score: Discriminative Fine-tuning of Generative Evaluation Metrics that this lack of interpretability severely limits the diagnostic utility of automated evaluation. Even highly accurate learnable metrics, as Wei Xu observes in LENS: A Learnable Evaluation Metric for Text Simplification, sacrifice the explicit transparency of traditional methods.

To bridge this interpretability gap, a new paradigm has emerged. Instead of treating evaluation as a single-step scoring problem, modern frameworks decompose the task into explicit sub-steps, generate natural language rationales, and map global scores back to fine-grained linguistic features.

Decomposition: Reducing Task Complexity

Direct prompting forces an LLM to simultaneously act as an error detector, a weight balancer, and a grader. This multi-task cognitive load often leads to erratic scoring. Hierarchical criteria decomposition solves this by breaking the evaluation process down into isolated, manageable sub-attributes.

[Opaque Direct Prompting]
Generated Text -------> [ LLM Judge ] -------> Single Scalar Score (e.g., 3/5)

[Decomposed Evaluation Framework]
                        /---> Sub-Criterion A (Fluency) ----> Score A \
Generated Text -------> ----> Sub-Criterion B (Accuracy) ---> Score B ---> [ White-Box Aggregator ] ---> Final Score
                        \---> Sub-Criterion C (Style) ------> Score C /

In DnA-Eval: Enhancing Large Language Model Evaluation through Decomposition and Aggregation, Minzhi Li et al. decouple evaluation into explicit criteria definition, individual scoring, and external aggregation. By offloading the final score calculation to an external tool rather than relying on the LLM's internal arithmetic, they significantly improve agreement with human labels. Similarly, Yuxuan Liu et al. introduce HD-Eval: Aligning Large Language Model Evaluators Through Hierarchical Criteria Decomposition, which uses a structured hierarchy of fine-grained criteria combined with "white-box" aggregators trained on human preferences.

Decomposition also allows metrics to adapt to highly specific or subjective tasks. For personalized long-form text generation, Hamed Zamani develops ExPerT: Effective and Explainable Evaluation of Personalized Long-Form Text Generation, which breaks text down into "atomic aspects" and evidence before scoring content and style independently. In open-ended question answering, Yongqi Fan et al. show in MinosEval: Distinguishing Factoid and Non-Factoid for Tailored Open-Ended QA Evaluation with LLMs that separating factoid extraction from creative instance ranking yields a much more reliable automated judge.

Emulating Human Experts via Error Notation

When human experts grade a text, they rarely guess a holistic score out of thin air; they mark specific errors in the margins and compute a penalty. Modern explainable metrics emulate this workflow by forcing models to perform explicit error notation before generating a final grade.

In machine translation, Qingyu Lu et al. propose Error Analysis Prompting Enables Human-Like Translation Evaluation in Large Language Models, a two-step method where the model explicitly identifies and categorizes translation errors before computing a score. Similarly, Patrick Fernandes et al. introduce AutoMQM in The Devil Is in the Errors: Leveraging Large Language Models for Fine-grained Machine Translation Evaluation, demonstrating that forcing an LLM to locate and categorize fine-grained errors provides developers with actionable feedback to debug translation pipelines.

This approach has proven vital across multimodal and high-stakes domains:

Granular Attribution and Verification Loops

To ensure that an evaluator's rationale is actually grounded in reality—and not just a hallucinated justification—researchers integrate explicit verification loops and game-theoretic attribution.

To evaluate factual consistency without reference texts, Leshem Choshen et al. design $Q^2$: Evaluating Factual Consistency in Knowledge-Grounded Dialogues via Question Generation and Question Answering. Instead of a black-box similarity check, $Q^2$ automatically generates questions from the model's response, answers them using the source material, and uses Natural Language Inference (NLI) to pinpoint exactly where a hallucination occurred.

                  /---> Generate Questions from Response
Generated Response ----> Answer Questions using Source Material ---> NLI Verification ---> Pinpoint Hallucination

Other approaches look inside the mathematical representations of the metrics themselves. To explain sentence-level metrics, Takumi Goto applies Shapley values from cooperative game theory in Improving Explainability of Sentence-level Metrics via Edit-level Attribution for Grammatical Error Correction, breaking down a global score into the exact mathematical contribution of each individual edit. Conversely, Steffen Eger et al. show in BMX: Boosting Natural Language Generation Metrics with Explainability that explainability techniques like LIME or SHAP can be inverted: by aggregating word-level importance scores back into a segment-level score, they actually make the underlying metric more accurate.

Closing the Loop: From Evaluation to Correction

The ultimate value of explainable evaluation is that it transforms metrics from passive scorekeepers into active participants in model improvement. When an evaluator provides a structured, self-explanatory rationale, that output can be directly fed back into the generative model for automated correction.

In Divide-Conquer-Reasoning for Consistency Evaluation and Automatic Improvement of Large Language Models, Wendi Cui et al. demonstrate that breaking paragraph-level consistency checks into sentence-to-paragraph checks yields diagnostic feedback capable of automatically fixing up to 90% of hallucinations in a single iteration. Similarly, Yongqiang Ma et al. introduce the concept of Revision Distance in From Model-centered to Human-Centered: Revision Distance as a Metric for Text Evaluation in LLMs-based Applications, using an LLM to generate structured, actionable edit suggestions.

By measuring quality through the lens of how to fix the text rather than an abstract distance formula, explainable metrics finally bridge the gap between model evaluation and model optimization.

Key Foundational Work

NNEval: Neural Network Based Evaluation Metric for Image Captioning (2018) — contributed 1 concept elements to this direction.
VizSeq: a visual analysis toolkit for text generation tasks (2019) — contributed 1 concept elements to this direction.
Train, Sort, Explain: Learning to Diagnose Translation Models (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_black_box_scoring_618fd50f,
  title = {black-box scoring},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/black-box-scoring/rdr_618fd50f39254ab48c94b7bb4038e5cd},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_black_box_scoring_618fd50f_v1,
  title = {black-box scoring},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/black-box-scoring/rdr_618fd50f39254ab48c94b7bb4038e5cd/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake