A Two-step Approach to Sentence Compression of Spoken Utterances

Dong Wang, Xian Qian, Yang LiuACL 2012

Abstract: Spontaneous speech compression is hindered by disfluencies and the lack of a single "gold standard" for evaluation. We present a two-step framework: a local candidate generator using Conditional Random Fields (CRFs) and a global discriminative reranker. By incorporating speech-specific features and training on multiple human references, our system achieves significant performance gains over existing baselines on the ICSI Meeting Corpus.

1. Introduction

Sentence compression for formal text is a mature field, but spoken language remains challenging due to its inherent disfluencies (e.g., "uh," "um"), false starts, and repetitions. These elements make spontaneous speech more redundant and less structured than written text. Traditional word-by-word decision models often fail to maintain global sentence quality. This work addresses these challenges through a pipeline that combines local sequence labeling with global reranking, optimized using multiple human references.

2. Methodology

Our framework consists of a candidate generation phase followed by a reranking phase.

2.1 Candidate Generation

We frame compression as a sequence labeling task using CRFs. We utilize a "BIO" (Begin, Inside, Outside) labeling scheme to identify fragments for removal. Key to our approach is the introduction of speech-specific features, such as a "disfluency window" to flag nearby filled pauses and distance metrics to identify repetitions. This stage generates an n-best list of 25 candidate compressions.

2.2 Discriminative Reranking

A Maximum Entropy (MaxEnt) model is employed to rerank the n-best list based on global features. These include the log-likelihood of the sentence under a language model, the overall compression ratio, and the CRF's confidence score. To better model the variety in human compression, the reranker is trained against eight human references simultaneously.

3. Experimental Setup

We evaluated our model on the ICSI Meeting Corpus, using a subset of 2,860 sentences. Each sentence in this dataset is accompanied by eight human-generated compressions. Training for the CRF step utilized the "best" reference as determined by a separate panel of annotators.

4. Results and Discussion

The experimental results demonstrate the effectiveness of our proposed features and reranking strategy.

Performance comparison of various feature sets and reranking strategies.

As shown in the table, the "expanded features" significantly improved word-level accuracy (79.28%) and F1 score (67.37%) compared to the baseline (Liu and Liu, 2010). The reranking step further optimized performance; specifically, training with eight references yielded the highest BLEU score of 94.21. This suggests that multi-reference training helps the model learn a more robust representation of valid compressions. Notably, the "oracle potential" analysis revealed that a near-perfect compression exists in the n-best list 84.7% of the time, indicating that future improvements should focus on the reranking logic.

5. Conclusion

We have presented a two-step approach for compressing spontaneous spoken utterances. By integrating speech-specific local features with a global reranker trained on multiple references, we significantly outperform baseline methods. Future research will explore the integration of syntactic dependency information and downstream application in full meeting summarization systems.

Summary updated

BibTeX
@inproceedings{wang2012two,
  author = {Dong Wang and Xian Qian and Yang Liu},
  title = {{A Two-step Approach to Sentence Compression of Spoken Utterances}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2012},
}
Figure