[Paper Review] Sentence Compression in Spanish driven by Discourse Segmentation and Language Models

Alejandro Molina-Villegas, Juan‐Manuel Torres‐Moreno+3 more|arXiv (Cornell University)|Dec 14, 2012

Natural Language Processing Techniques21 references3 citations

TL;DR

This paper proposes a sentence compression method for Spanish that leverages discourse segmentation and language models to generate grammatically coherent, concise summaries. By segmenting sentences into Elementary Discourse Units (EDUs), generating compression candidates through EDU removal, and selecting the most probable candidate using a statistical language model, the approach achieves improved fluency and coherence in compressed sentences, demonstrating strong performance on Spanish text summarization tasks.

ABSTRACT

Previous works demonstrated that Automatic Text Summarization (ATS) by sentences extraction may be improved using sentence compression. In this work we present a sentence compressions approach guided by level-sentence discourse segmentation and probabilistic language models (LM). The results presented here show that the proposed solution is able to generate coherent summaries with grammatical compressed sentences. The approach is simple enough to be transposed into other languages.

Motivation & Objective

• To improve automatic text summarization in Spanish by enhancing sentence-level compression.
• To address grammaticality and coherence issues in sentence compression by integrating discourse structure.
• To reduce the solution space for compression by leveraging discourse segmentation into Elementary Discourse Units (EDUs).
• To evaluate compression candidates using statistical language models to select the most natural-sounding output.
• To develop a method transferable to other languages with limited discourse parsing resources.

Proposed method

• The method begins with discourse segmentation of input sentences into Elementary Discourse Units (EDUs) using a shallow discourse segmenter based on the Rhetorical Structure Theory.
• All possible compression candidates are generated by removing subsets of EDUs while preserving the original order, resulting in 2^k - 1 candidates for k EDUs.
• The original sentence is always included as a candidate (CC₀) to ensure at least one valid output exists.
• Each candidate is scored using a statistical language model that estimates the probability of the compressed sentence as a natural language sequence.
• The candidate with the highest language model probability is selected as the final compressed sentence.
• The approach is designed to be lightweight and adaptable to other languages with minimal linguistic resources.

Experimental results

Research questions

• RQ1 Can discourse segmentation improve the grammatical quality and coherence of sentence compression in Spanish?
• RQ2 How effective is using a language model to rank compression candidates compared to heuristic or rule-based methods?
• RQ3 To what extent does segmenting sentences into EDUs reduce the search space while preserving meaningful compression options?
• RQ4 Can this method generate fluent, natural-sounding compressed sentences even in short or simple sentences?
• RQ5 Is the approach generalizable to other languages with limited discourse parsing infrastructure?

Key findings

• The proposed method successfully generates grammatically correct and coherent compressed sentences by leveraging discourse structure and language modeling.
• Using discourse segmentation reduces the number of candidates from potentially large word-level combinations to a manageable set based on EDUs, improving efficiency.
• The language model effectively ranks candidates, favoring more natural and fluent compressed outputs.
• The approach maintains the original sentence as a valid candidate, ensuring robustness even when no meaningful compression is possible.
• The method demonstrates strong potential for adaptation to other languages due to its reliance on discourse segmentation and statistical language models.
• The results show that discourse-aware compression outperforms word-level compression in terms of fluency and grammaticality, particularly in complex sentences.