A Grammatico-Statistical Approach to Discourse Partitioning

Tadashi Nomoto, Yoshihiko NittaCOLING 1994

TL;DR

This paper introduces a novel method for automatically segmenting unstructured text into coherent discourse units, called "discourse segments." By combining linguistic insights about Japanese discourse structure with statistical techniques from information retrieval, the authors demonstrate a successful approach to recovering article boundaries from a corpus of Japanese newspaper articles.

What Problem Does This Paper Address?

The paper tackles the challenge of discourse partitioning, which is the task of dividing a continuous text into meaningful, coherent segments. Existing approaches often treat text as a simple sequence of words or sentences, overlooking potentially important linguistic structures. The authors argue that discourse is better understood as a series of "discourse segments" – minimal coherent units that are often demarcated by specific linguistic cues, such as the Japanese topic particle "wa." This segmentation is crucial for various natural language processing tasks, including information retrieval, summarization, and question answering, where understanding the thematic structure of a document is key.

Methods

The proposed method is a hybrid of linguistic and statistical approaches.

  • Linguistic Motivation: The core idea is based on the theory of "discourse segments" (Nomoto and Nitta, 1993). These segments are conceptualized as topic-comment structures, where the topic is marked by the particle "wa" in Japanese. A key linguistic observation is that anaphoric references within a discourse segment tend to be resolved internally, not crossing segment boundaries. The authors define a discourse segment as a block of sentences bounded by text breaks or "wa"-marked elements.
  • Statistical Approach: The paper leverages techniques from information retrieval, specifically the vector space model and tf-idf weighting.
  • Each discourse segment is represented as a term vector, where terms are nominal occurrences within the segment.
  • tf-idf (term frequency-inverse document frequency) is used to weight the importance of terms. This weighting scheme favors terms that are frequent within a segment but rare across the entire corpus, thus highlighting distinctive vocabulary.
  • Coherence between adjacent segments is measured using the Dice coefficient, a standard similarity measure in information retrieval. This coefficient quantifies the overlap between the term vectors of two segments, weighted by their tf-idf scores.
  • Segmentation Process: The partitioning is performed in two main steps:
  • Collect all nominal occurrences from the corpus.
  • Divide the corpus into disjoint discourse segments.
  • Compare adjacent blocks of discourse segments using the coherence measure.
  • Assign a coherence/similarity value to each pair.
  • Analyze the resulting "coherence curve" to identify valleys (low coherence) as potential discourse breaks.

Datasets

The experiments were conducted on a four-week collection of editorials from Nihon Keizai Shimbun, a Japanese economics newspaper. This corpus contained approximately 1111 sentences, around 10,000 nouns, and 556 discourse segments. The segmentation of nouns was performed semi-automatically using the JUMAN morphological analyzer.

Key Insights and Ideas

  • Discourse Segments as Fundamental Units: The paper strongly advocates for viewing discourse as composed of "discourse segments" rather than arbitrary blocks of text. This linguistic framing provides a more principled basis for segmentation.
  • Internal Cohesion of Segments: The observation that anaphoric references are typically resolved within discourse segments is a crucial insight, suggesting that these segments represent self-contained thematic units.
  • tf-idf for Discourse Coherence: Applying tf-idf, a standard IR technique, to measure the similarity between discourse segments proves effective. It highlights the distinctiveness of vocabulary within segments, which is indicative of thematic shifts.
  • Coherence Curves for Boundary Detection: The coherence curve, generated by pairwise segment comparisons, effectively reveals potential discourse boundaries as local minima. This visual representation aids in understanding the thematic flow of the text.
  • Linguistic Cues as Structural Markers: The paper emphasizes that linguistic cues, like the "wa" particle and text breaks, are not just stylistic but can serve as actual structural markers for discourse segments.

Key Results

  • The proposed method demonstrated significant success in recovering article boundaries from the Japanese newspaper corpus.
  • The Dice coefficient analysis on the Nikkei corpus showed a strong correspondence between the coherence curve and the global discourse structure, with article boundaries often coinciding with major minima on the graph.
  • Out of 27 local minima identified, 7 were found to be incorrect, resulting in an error rate of approximately 25%. Similar results were obtained with Jaccard and cosine coefficients.
  • Experiments with varying block sizes for segmentation indicated that a block size of 10 segments yielded good results.
  • The paper also compared its approach to the Vocabulary Management Profile (VMP) method (Youmans, 1991), finding the VMP less effective in detecting significant patterns in the tested corpus, possibly due to its sensitivity to text repetition.
  • Recall and precision analysis showed that segmentations based on discourse segments (labeled 'G') achieved the highest average recall and precision compared to segmentations based on fixed word block sizes.

Limitations and Future Work

  • Evaluation Difficulty: The authors acknowledge that evaluating discourse segmentation is inherently problematic due to the subjective nature of human judgment on discourse structure. They relied on orthographic structures (article/paragraph breaks) as a proxy for evaluation.
  • Corpus Size Impact: The paper suggests that larger corpus sizes might lead to better results, as the document frequency (df) factor in tf-idf becomes more meaningful.
  • Granularity Control: While the method is effective at detecting article-level boundaries, achieving finer-grained segmentation poses challenges due to the lack of reliable evaluation metrics for smaller units.
  • Future Work: The authors suggest exploring alternative weighting policies, such as weighted inverse document frequency (widf), and further investigating the impact of corpus size and the development of more robust evaluation metrics for finer-grained discourse segmentation. They also note the need to relate their findings more directly to rhetorical analyses of discourse.

Summary updated

BibTeX
@inproceedings{nomoto1994grammatico,
  author = {Tadashi Nomoto and Yoshihiko Nitta},
  title = {{A Grammatico-Statistical Approach to Discourse Partitioning}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1994},
}
Figure