From Chunks to function-Argument Structure: A Similarity-Based Approach

Sandra Kübler, Erhard W. HinrichsACL 2001 · DOI

This paper, presented by Sandra Kübler and Erhard W. Hinrichs from the University of Tübingen, introduces a novel methodology to bridge the gap between "shallow" and "deep" natural language processing. By applying memory-based learning to the outputs of a chunk parser, the authors demonstrate a robust method for identifying functional roles—such as subject and object—within a sentence, achieving over 90% accuracy on spontaneous speech data.

The Challenge of Spontaneous Speech

In computational linguistics, researchers often face a trade-off between speed and depth. Chunk parsing is fast and robust; it identifies basic phrases (e.g., "the red car" or "should get") but ignores how they relate to each other in a larger sentence. Conversely, traditional parsers aim for a complete analysis but are often brittle, frequently failing when encountering the ungrammatical or disfluent nature of real-world spontaneous speech.

The "gap" identified by the authors is the lack of a system that can take robust, shallow chunks and reliably assign function-argument structure—the "who did what to whom" of a sentence. Without knowing which noun phrase is the subject and which is the object, a deeper semantic understanding of the text remains elusive.

The TüSBL Architecture

The authors introduce the TüSBL (Tübingen Similarity-Based Learning) architecture, which operates across three distinct levels:

  1. Part-of-Speech (POS) Tagging: The input string is tagged using a bigram tagger.
  2. Chunk Analysis: An adapted version of Abney’s CASS parser uses finite-state transducers to group words into non-recursive "chunks."
  3. Similarity-Based Tree Construction: This is the core innovation. Instead of using hand-coded grammar rules to combine chunks, the system utilizes an "instance base" of previously seen, fully annotated trees.
The core process_chunk algorithm for similarity-based tree construction.

When the system encounters a new sentence, it searches its memory for the most similar tree structure. If it finds a "direct hit" (a lexical match), it adopts that structure. If not, it performs a partial match on the lexical or POS level, "filling in the gaps" by adapting stored tree templates to fit the new input. This "lazy learning" approach allows the system to handle the idiosyncrasies of spoken language without needing to abstract them into complex, rigid rules.

Datasets and Representation

The system was evaluated using two semi-automatically constructed treebanks based on the VERBMOBIL project:

  • German Treebank: Approximately 67,000 sentences, using a "topological field" model with a rich set of 36 functional labels.
  • English Treebank: Approximately 35,000 sentences, modeled after Head-Driven Phrase Structure Grammar (HPSG) with 13 functional labels.
Example of an English syntactic tree with functional labels produced by the system.

Crucially, these datasets consist of transliterated spontaneous speech, which is significantly more challenging than the "clean" written text (like the Wall Street Journal) often used in parsing benchmarks.

Key Insights

  • Similarity over Rules: By treating tree construction as a similarity problem rather than rule-matching, the system gracefully handles "near misses" in the data.
  • Conservative Attachment: The algorithm follows a strategy of only attaching nodes if there is sufficient evidence in the instance base. This prioritizes precision over completeness, which is vital for downstream applications that rely on accurate data.
  • Handling Ambiguity: The system can correct POS tagging or chunking errors by looking at the broader context provided by the instance base. If a word was tagged as a preposition but the most similar stored tree suggests it is an adverb, the system can effectively "overrule" the lower-level tagger.

Results and Performance

Evaluation using ten-fold cross-validation yielded impressive results for labeled precision of functional labels (identifying subjects, objects, etc.):

  • German: 89.73% precision.
  • English: 90.40% precision.

While the precision for syntactic categories was lower (81.56% for German and 66.84% for English), the high accuracy of the functional labels validates the similarity-based approach. The system is also highly efficient, parsing 3,770 English sentences in just 106.5 seconds on contemporary hardware (Ultra Sparc 10).

Limitations and Future Work

While precision is high, recall remains lower (around 60%). The system often leaves constituents unattached if it lacks confidence, resulting in a "fragmentary" parse. Furthermore, the current version uses a basic "overlap metric" for similarity. The authors suggest that more sophisticated metrics, such as "gain ratio," could improve performance. Future work aims to improve generalization so the system can better handle unseen words by relying more heavily on structural patterns rather than lexical identity.

Summary updated

BibTeX
@inproceedings{kubler2001from,
  author = {Sandra Kübler and Erhard W. Hinrichs},
  title = {{From Chunks to function-Argument Structure: A Similarity-Based Approach}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2001},
  doi = {10.3115/1073012.1073057},
}
Figure