TL;DR
This paper introduces MorphoClass, a system designed to identify and classify unknown German words by inferring their stem and morphological properties. It tackles the challenge of handling out-of-vocabulary words in German text by leveraging global information, morphological rules, and external knowledge sources.
What Problem Does This Paper Address?
In natural language processing, dealing with words not present in a predefined lexicon (unknown words or out-of-vocabulary words) is a persistent challenge. For morphologically rich languages like German, this problem is compounded because a single base word can have many inflected forms. Identifying these unknown words and determining their grammatical properties (like gender and inflection rules) is crucial for tasks such as machine translation, information retrieval, and text analysis. Existing systems often rely on extensive lexicons or complex linguistic rules, leaving a gap in efficiently handling truly novel or rare words.
Methods
MorphoClass tackles this by breaking down the problem into several stages:
- Unknown Word Identification: It first identifies potential unknown words by exploiting the German linguistic property that nouns are always capitalized. It then uses statistics like total frequency, capitalized frequency, and start-of-sentence frequency to hypothesize if a word type could be an unknown noun.
- Stem Generation: For each potential unknown noun, MorphoClass generates all possible stems by reversing common German inflections, considering umlaut and ß alternations. This process aims to group together different inflected forms of the same underlying word.
- Stem Coverage Checking and Refinement: It checks if the generated stems can be covered by at least one morphological class. If not, it attempts to refine the stem coverage.
- Morphological Stem Analysis: Each stem is then analyzed using a lexicon-based approach and suffix-based morphology. It checks if the stem is already known (rejecting it if so), if it's a compound word (by splitting it into known parts), or if its class can be guessed from its ending using Mikheev-like rules.
- Word Type Clusterisation: Finally, it selects a set of stems to cover all identified word types, prioritizing stems that are compounds, then those identified by ending-guessing rules, and finally longer stems.
Datasets
The paper mentions training the ending-guessing rules model over "8,5 MB of raw text." For evaluation, the system was tested on an "85 KB German literature text: Erzählungen by Franz Kafka." This text contained 3510 different word forms, including 493 unknown nouns. The evaluation was performed manually on a quarter of the identified stems.
Key Insights and Ideas
- Leveraging Capitalization: The simple rule of capitalizing German nouns proves to be a surprisingly effective heuristic for initial identification of unknown nouns.
- Stem Generation as a Grouping Mechanism: Generating all possible stems from inflected forms is a clever way to group related words and infer their common root, even if the root itself is unknown.
- Cascading Analysis: The system's approach of first trying compound splitting and then falling back to ending-guessing rules is a practical way to balance precision and coverage, as compound analysis is generally more reliable.
- Focus on "Unknown" Stems: By explicitly rejecting stems that are already in a known stem lexicon, the system ensures it's focusing on truly novel word forms.
Key Results
The evaluation on Franz Kafka's text yielded promising results:
- Coverage: The system achieved an overall coverage of 88.99%, meaning it was able to assign a morphological class to nearly 89% of the identified unknown noun stems.
- Precision:
- Compound splitting achieved a high precision of 93.62% with a coverage of 43.12%.
- Ending-guessing rules provided a precision of 56% (precision1) to 70% (precision3) with a coverage of 45.87%.
- The overall system precision was reported as 74.23% (precision1), 76.08% (precision2), and 81.44% (precision3).
- Manual Evaluation Breakdown: Out of 120 evaluated stems:
- 72 were correctly classified with a single class (YES).
- 12 had a set of classes assigned (SET).
- 7 had a correct class but not all correct ones (PART).
- 18 were assigned a wrong single class (WRONG).
- 11 were skipped (SKIP).
Limitations and Future Work
- Context Independence: The current system relies on very simple rules and does not exploit any contextual information from the surrounding text. This is a significant limitation for disambiguation.
- Single Inflected Form Assumption: The paper notes that many unknown nouns might appear with only one inflected form, which simplifies the problem but might not reflect real-world scenarios with more complex word usage.
- Future Work: The authors suggest that a similar approach could be applied to other inflectional languages and other open-class parts of speech like adjectives, verbs, and adverbs, with most of the steps being adaptable. They also hint at the possibility of incorporating more sophisticated linguistic rules.
Summary updated