Bi-particle Adverbs, PoS-Tagging and the Recognition of German Separable Prefix Verbs

Martin Volk, Simon Clematide, Johannes Graën, Phillip StröbelKONVENS 2016 · OpenReview

Introduction

German verbs often present a unique challenge for Natural Language Processing (NLP) due to their "separable" nature. A verb like nachweisen (to prove) can split across a sentence, with the prefix nach appearing at the very end. Standard Part-of-Speech (PoS) taggers, which typically analyze words in a local window, often fail to re-attach these prefixes to their stems, leading to incorrect lemmatization and a loss of semantic meaning.

This paper introduces a high-precision algorithm designed to solve this re-attachment problem while specifically addressing a common "trap": bi-particle adverbs. These are fixed phrases like nach wie vor (still) that are frequently mistaken for verb prefixes, causing cascading errors in syntactic parsing.

The "Adverb Trap" and Structural Impact

A major contribution of this research is the identification of bi-particle adverbs as a primary source of error in German NLP. Phrases such as ab und zu (sometimes) or nach wie vor contain words that look identical to common verb prefixes.

Syntactic structure of a sentence containing both a bi-particle adverb ("nach wie vor") and a separable verb ("festhalten").

As shown in Figure 1, mistagging a component of an adverb (like nach) as a verb prefix doesn't just affect the verb; it can break the entire syntactic parse. If a tagger incorrectly identifies a particle as a preposition, the parser may attempt to find a non-existent noun phrase to attach to it, resulting in a nonsensical sentence structure. In the Text+Berg corpus, the authors found that nach in the phrase nach wie vor was mistagged as a verb prefix 55% of the time.

Methodology: A Multi-Step Re-attachment Algorithm

The authors propose a robust algorithm for large-scale corpus annotation consisting of five key steps:

  1. Initial Tagging: A standard PoS tagger (e.g., TreeTagger) identifies potential verb stems and separated prefixes (tagged as PTKVZ).
  2. Adverb Correction: Before re-attachment, the system uses rule-based logic to identify bi-particle adverb patterns (e.g., "particle + und/wie + particle") and corrects their tags.
  3. The Search: For every valid separated prefix, the algorithm searches backward in the sentence for the most recent finite full verb or imperative.
  4. Validation: To prevent "hallucinating" verbs, the combined prefix and verb are checked against a list of 8,500 known German separable verbs and the GerTwol morphological analyzer.
  5. Special Case Handling: The logic includes specific rules for auxiliary and modal verbs that function as full verbs when paired with a prefix (e.g., innehaben).

Insights from Large-Scale Corpora

The researchers utilized several major German resources, including the TIGER and TuBa-D/Z treebanks, to benchmark their system. By applying their tools to the Text+Berg corpus (22.5 million tokens of alpine literature), they were able to track how the usage of these problematic adverbs has evolved over 150 years.

Google Ngram Viewer data showing the frequency trends of common bi-particle adverbs in German from 1860 to 2008.

Using the Google N-gram viewer (Figure 4), the authors demonstrated a significant semantic shift: while some adverbs like nach wie vor have surged in popularity, others like nach und nach have declined. This highlights the necessity for NLP tools to remain updated with evolving idiomatic usage to maintain accuracy.

Key Results and Performance

The algorithm achieved an impressive 96.8% to 97% precision when tested against the gold-standard TuBa-D/Z treebank. Key factors in this success included:

  • Dictionary Expansion: Supplementing the verb list with a morphological analyzer added 468 rare verbs, boosting precision.
  • Error Mitigation: The specialized treatment of bi-particle adverbs nearly eliminated errors related to the prefix nach in specialized domains like banking news.

Limitations and Future Work

Despite its high precision, the algorithm faces challenges in three specific scenarios:

  1. Topicalization: When a prefix is moved to the front of a sentence for emphasis, the backward-search logic fails.
  2. Nested Clauses: Sub-clauses appearing between a verb and its prefix can lead to incorrect attachments.
  3. Coordinated Prefixes: The system currently does not handle verbs with multiple coordinated prefixes (e.g., "the value increases instead of decreases").

Future research will focus on integrating these re-attachment techniques into machine translation and cross-lingual word alignment to improve the handling of complex German verb structures in multilingual contexts.

Summary updated

BibTeX
@inproceedings{volk2016bi,
  author = {Martin Volk and Simon Clematide and Johannes Graën and Phillip Ströbel},
  title = {{Bi-particle Adverbs, PoS-Tagging and the Recognition of German Separable Prefix Verbs}},
  booktitle = {KONVENS},
  year = {2016},
  url = {https://openreview.net/forum?id=0TkXt0kUdH},
}
Figure