Abstract
Despite significant progress in text generation models, a serious limitation is their tendency to produce text that is factually inconsistent with information in the input. Recent work has studied whether textual entailment systems can be used to identify factual errors; however, these sentence-level entailment models are trained to solve a different problem than generation filtering and they do not localize which part of a generation is non-factual. In this paper, we propose a new formulation of entailment that decomposes it at the level of dependency arcs. Rather than focusing on aggregate decisions, we instead ask whether the semantic relationship manifested by individual dependency arcs in the generated output is supported by the input. Human judgments on this task are difficult to obtain; we therefore propose a method to automatically create data based on existing entailment or paraphrase corpora. Experiments show that our dependency arc entailment model trained on this data can identify factual inconsistencies in paraphrasing and summarization better than sentence-level methods or those based on question generation, while additionally localizing the erroneous parts of the generation.
Introduction: The Challenge of Factuality
Modern language models are incredibly fluent, yet they frequently suffer from "hallucinations"—generating text that is grammatically correct but factually inconsistent with the source input. Traditional evaluation metrics like ROUGE or BERTScore often fail to detect these subtle factual slips because they focus primarily on surface-level similarity.
While Natural Language Inference (NLI) models have been explored as a solution, they typically operate at the sentence level. This approach has two major drawbacks: it lacks the granularity to pinpoint exactly where a sentence is wrong (localization), and there is often a task mismatch between general logical inference (e.g., MNLI) and the specific requirements of checking generation faithfulness.
The Dependency Arc Entailment (DAE) Framework
To address these issues, we introduce Dependency Arc Entailment (DAE). Instead of assigning a single truth value to an entire sentence, DAE decomposes the sentence into its constituent dependency arcs. Each arc represents a fundamental semantic relationship, such as a subject-verb or modifier-noun link.
By evaluating each arc individually, the model can identify which specific relationships are supported by the source text and which are "hallucinated." For example, if a model incorrectly swaps a subject and object, DAE can highlight those specific arcs as non-entailed.
Model Architecture and Training
The DAE model utilizes a pre-trained transformer encoder (such as BERT or ELECTRA). For every dependency arc in a generated sentence, the model extracts contextual embeddings for the head and child words, concatenates them with an embedding of the dependency label (e.g., nsubj, amod), and passes this combined representation through a classifier.
Automatic Data Creation
Obtaining human labels for millions of individual dependency arcs is impractical. We therefore developed a weak supervision strategy using the ParaNMT-50M dataset. Arcs from high-quality human paraphrases are treated as positive (entailed) examples. Negative examples are derived from lower-quality model generations where specific arcs do not appear in either the input or the "gold" paraphrase.
Key Results and Evaluation
Our experiments demonstrate that DAE significantly outperforms sentence-level baselines:
- Summarization Reranking: On the Falke et al. benchmark, DAE (using ELECTRA) achieved 83.6% accuracy, surpassing sentence-level BERT-MNLI (72.3%) and specialized Question Answering metrics like QAGS (72.1%). This performance nearly matches human-level accuracy (83.9%).
- Paraphrase Ranking: DAE performed on par with top NLI models (79.0%), proving its versatility across different generation tasks.
- Interpretability: Human evaluation confirmed that the automatically generated arc labels align with human judgment 82% of the time, validating the localization capabilities of the model.
Conclusion and Future Directions
DAE provides a surgical, interpretable method for ensuring factual consistency in generated text. While highly effective, current limitations include an inability to detect omitted information and challenges with complex semantics like negation or quantification. Future work will focus on integrating these semantic nuances and reducing reliance on the accuracy of underlying dependency parsers. Overall, this approach moves evaluation away from "vibe-based" scores toward precise, evidence-backed verification.
Summary updated