ABCD: A Graph Framework to Convert Complex Sentences to a Covering Set of Simple Sentences

Yanjun Gao, Ting-Hao Kenneth Huang, Rebecca J. Passonneau

Introduction

Breaking down complex sentences into simple, atomic clauses is a foundational step for downstream NLP tasks like summarization, argument mining, and question answering. However, existing methods often fall short. Discourse segmentation (e.g., EDUs) chops sentences into spans that aren't always grammatically complete. Text simplification often deletes important propositions. Split-and-Rephrase (SPRP) models, typically based on sequence-to-sequence (seq2seq) architectures, are notorious for hallucinating new words or failing to preserve the exact meaning of the original sentence.

The ABCD framework addresses these issues by mapping every tensed clause in a source sentence to an independent, grammatically complete simple sentence, without losing original propositions or hallucinating new ones.

Methods: The ABCD Framework

To ensure strict fidelity to the source text, the authors formulate sentence splitting as a graph edit problem rather than a generation task.

Graph Construction and Operations

An input sentence is first converted into a Word Relation Graph (WRG). In this graph, words are nodes, and edges represent either linear adjacency (neighboring words) or grammatical dependencies derived from a dependency parser.

Word Relation Graph (WRG) and ABCD operations for sentence decomposition.

The model examines every edge triple (source node, target node, edge label) and predicts one of four "ABCD" edits:

  • Accept (A): Keep the edge to maintain connectivity.
  • Break (B): Sever the edge between words to create a sentence boundary.
  • Copy (C): Duplicate a target word into a new connected component (e.g., sharing a subject across conjoined verb phrases).
  • Drop (D): Remove the word entirely (e.g., deleting discourse connectives like "and").

Neural Architecture

The system uses a biLSTM or BERT encoder to generate sentence representations. These are processed through a multi-head edge self-attention layer, followed by a classifier (MLP or bilinear) that predicts the specific ABCD operation for each edge.

Neural architecture of the ABCD framework featuring edge self-attention.

Finally, a deterministic graph segmenter applies the predicted edits, identifies connected components, and arranges the words back into linear simple sentences.

Datasets

The framework was evaluated on two primary datasets:

  1. DeSSE (Decomposed Sentences from Students Essays): A new dataset of 12K training examples from undergraduate social science essays. It features complex linguistic phenomena like relative clauses and is significantly more challenging than standard Wikipedia text.
  2. MinWiki: A filtered subset of MinWikiSplit (Wikipedia revision histories), containing 18K training examples. It is structurally simpler than DeSSE.

Key Results and Insights

ABCD demonstrates that graph-based editing is superior to seq2seq generation for maintaining strict fidelity. By restricting the model to editing a graph of original words, it avoids the "hallucination" problem common in encoder-decoder models.

  • Performance on DeSSE: ABCD significantly outperformed baselines on the more challenging DeSSE dataset, achieving a 53.42 BLEU4 score (vs. 45.91 for the Copy512 seq2seq baseline) and correctly predicting the number of output sentences in 53.29% of cases (vs. 36.20% for the baseline).
  • Performance on MinWiki: ABCD remained highly competitive, achieving 75.80 BLEU4 and predicting the correct sentence count in 78.61% of cases.

Interestingly, a biLSTM trained from scratch with positional encodings often outperformed fine-tuned BERT, suggesting that structural and positional awareness is more critical for this task than general semantic pre-training.

Limitations and Future Work

The model currently struggles with the "Break" operation due to a lack of broader sequential context in the edge-triple classification. Additionally, because it only edits existing words, it cannot perform natural language "smoothing" like replacing a copied subject with a pronoun. Future work involves adding an INSERT operation and extending the framework to other languages.

Summary updated

BibTeX
@inproceedings{gao2021abcd,
  author = {Yanjun Gao and Ting-Hao Kenneth Huang and Rebecca J. Passonneau},
  title = {{ABCD: A Graph Framework to Convert Complex Sentences to a Covering Set of Simple Sentences}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
  doi = {10.18653/v1/2021.acl-long.303},
}
Figure