Introduction
Researchers at Penn State have introduced ABCD, a novel graph-based framework designed to decompose complex sentences into a set of simple, "atomic" sentences. By framing decomposition as a graph-editing task—utilizing operations to Accept, Break, Copy, or Drop edges—the model preserves the semantic meaning of every tensed clause more effectively than traditional sequence-to-sequence approaches.
Problem Statement: The Challenge of Complex Sentences
In natural language processing, complex sentences often pose significant challenges for downstream tasks such as summarization, question answering, and argument mining. While humans can easily parse nested clauses and conjoined phrases, machines often struggle with these structures.
Existing solutions frequently fall short:
- Elementary Discourse Unit (EDU) segmentation often results in fragments that are not complete thoughts.
- Text Simplification may remove critical information to improve readability.
- Split-and-Rephrase (SPRP) tasks often generate sentences from non-clausal parts, losing the original predication structure.
The authors argue that tensed clauses (clauses with a verb indicating time) are the fundamental units of meaning. The goal of ABCD is to extract every tensed clause into its own independent, grammatically correct simple sentence without losing information or introducing "hallucinations."
Methods: Graph Editing for Sentence Decomposition
The core innovation of this framework is the Word Relation Graph (WRG). Instead of treating a sentence as a flat sequence of tokens, ABCD represents it as a directed graph where vertices are word tokens and edges represent linear adjacency (neighboring words) and grammatical dependencies (e.g., a subject linked to its verb).
The framework treats decomposition as a graph editing task. For every edge in the WRG, the model predicts one of four labels:
- Accept (A): Maintain the connection.
- Break (B): Sever the connection to initiate a new sentence.
- Copy (C): Duplicate a word (such as a shared subject) into a new sentence.
- Drop (D): Delete unnecessary elements, such as coordinating conjunctions ("and", "but").
The neural architecture employs a biLSTM or BERT encoder to generate word representations, followed by an Edge Self-Attention layer that identifies which grammatical relations are most critical for edit decisions. A final classifier assigns the ABCD labels, and a post-processor reconstructs the final simple sentences.
Datasets and Evaluation
The researchers evaluated ABCD on two distinct datasets:
- MinWiki: A subset of WikiSplit (Wikipedia edits) filtered to ensure a one-to-one mapping between source tensed clauses and output sentences.
- DeSSE (Decomposed Sentences from Students Essays): A newly introduced dataset of over 12,000 sentences from undergraduate essays. These sentences are linguistically diverse and "messier" than Wikipedia entries, containing varied discourse connectives and clausal arguments.
Key Insights
- Graph vs. Sequence: Traditional encoder-decoder models often suffer from "hallucination"—adding words not present in the source. Because ABCD operates by editing an existing graph of source words, it excels at preserving original meaning.
- The "Copy" Operation: A major hurdle in splitting sentences is "subject sharing" (e.g., "The dog barked and ran"). ABCD’s explicit Copy operation allows it to correctly generate "The dog barked" and "The dog ran," whereas simpler models might produce the fragment "ran."
- Linguistic Diversity: Models trained solely on Wikipedia (MinWiki) struggled with the varied structures in student writing (DeSSE), highlighting the necessity for diverse benchmarks in sentence decomposition.
Results
ABCD demonstrated superior performance in predicting the correct number of atomic sentences compared to strong encoder-decoder baselines.
As shown in the results table, the model achieved an overall F1 score of 0.82 on MinWiki and 0.67 on the more challenging DeSSE dataset. Qualitatively, ABCD was less likely to alter sentence meaning; for instance, it consistently maintained original vocabulary where baseline models might substitute words to simplify the text.
Limitations and Future Work
Despite its successes, the authors identified several areas for improvement:
- The "Break" Challenge: In the complex DeSSE dataset, the model occasionally failed to identify precise "Break" (B) points, sometimes defaulting to "Accepting" (A) the edge.
- Vocabulary Constraints: The current model cannot insert new words (e.g., pronouns like "it" or "they") to make rephrased clauses more natural.
- Dependency Sensitivity: Because the framework relies on initial dependency parsing, errors in the parser can propagate through the pipeline.
Future research will focus on expanding edit types to include word insertion and testing the framework's generalizability across different languages and grammars.
Summary updated