[Paper Review] Neural Data-to-Text Generation via Jointly Learning the Segmentation and Correspondence

Xiaoyu Shen, Ernie Chang+3 more|arXiv (Cornell University)|May 3, 2020

Topic Modeling65 references10 citations

TL;DR

This paper proposes a neural data-to-text generation model that jointly learns segmentation and data-to-text correspondence as latent variables, enabling interpretable, low-hallucination outputs with reduced computational cost. By modeling generation segment-by-segment and explicitly aligning each segment to a data record, the model achieves state-of-the-art performance on E2E and WebNLG benchmarks with zero factual hallucinations and minimal information missing or repetition.

ABSTRACT

The neural attention model has achieved great success in data-to-text generation tasks. Though usually excelling at producing fluent text, it suffers from the problem of information missing, repetition and "hallucination". Due to the black-box nature of the neural attention architecture, avoiding these problems in a systematic way is non-trivial. To address this concern, we propose to explicitly segment target text into fragment units and align them with their data correspondences. The segmentation and correspondence are jointly learned as latent variables without any human annotations. We further impose a soft statistical constraint to regularize the segmental granularity. The resulting architecture maintains the same expressive power as neural attention models, while being able to generate fully interpretable outputs with several times less computational cost. On both E2E and WebNLG benchmarks, we show the proposed model consistently outperforms its neural attention counterparts.

Motivation & Objective

• Address the lack of interpretability and high hallucination rates in neural attention-based data-to-text models.
• Overcome the black-box nature of attention mechanisms that makes error detection and control difficult.
• Reduce computational cost and improve structural control in text generation without relying on human-annotated segmentation or alignment.
• Enable explicit, verifiable correspondence between generated text segments and input data records for improved reliability.
• Maintain high fluency and diversity while minimizing information missing and repetition through end-to-end training with latent structure induction.

Proposed method

• Model text generation as a sequence of segments, where each segment is generated by attending only to a single selected data record.
• Jointly learn segmentation and correspondence as latent variables during training, without human annotations.
• Use dynamic programming to efficiently marginalize over all possible segmentations and alignments, maximizing likelihood in polynomial time.
• Integrate an autoregressive neural network to model dependencies between segments while preserving tractable inference.
• Apply posterior regularization with a soft statistical constraint to control segmental granularity and prevent overly fine or coarse splits.
• Enable constrained decoding by monitoring the data record associated with each segment, allowing easy integration of heuristic rules or user constraints.

Experimental results

Research questions

• RQ1 Can joint learning of segmentation and correspondence improve interpretability and reduce hallucination in neural data-to-text generation?
• RQ2 How does the proposed model compare to standard neural attention models in terms of fluency, factual consistency, and computational efficiency?
• RQ3 To what extent can the model’s explicit structure enable better control over output content and reduce repetition and information missing?
• RQ4 Can the model achieve performance comparable to rule-based systems in factual correctness while maintaining the fluency of neural models?
• RQ5 How effective are decoding constraints in reducing repetition and missing facts when applied to the model’s interpretable segment-level structure?

Key findings

• The proposed model achieves zero factual hallucinations on the E2E benchmark, outperforming all neural baselines and matching the factual accuracy of rule-based systems.
• On E2E, the model reduces information missing to zero and repetition to only 2 cases with the (+RM) constraint, significantly outperforming standard seq2seq models.
• On WebNLG, the model maintains high automatic metric scores while achieving superior information adequacy and diversity, with only 5 missing facts under the (+RM) constraint.
• The model reduces repetition from 19 to 2 cases on WebNLG using the (+R) decoding constraint, demonstrating the effectiveness of structure-aware control.
• The model is several times more computationally efficient than standard attention-based models due to localized attention per segment rather than full input attention.
• Visualization of attention maps confirms that the model’s attention is sharply focused on individual data records, unlike the diffuse attention in standard seq2seq models that leads to hallucination.