Corpus-based Generation of Numeral Classifier using Phrase Alignment

Michael Paul, Eiichiro Sumita, Seiichi YamamotoCOLING 2002

TL;DR

This paper tackles the challenge of translating between languages with and without numeral classifiers (like Japanese vs. English) by automatically learning to generate the missing classifiers. They achieve this by aligning phrases in bilingual corpora to extract implicit classifier information, offering a uniform method for various classifier constructions.

What Problem Does This Paper Address?

Many natural language processing (NLP) tasks, especially machine translation, struggle with acquiring knowledge that is obligatory in one language but not explicitly expressed in another. A prime example is numeral classifiers, which are grammatical components essential in languages like Japanese for categorizing objects being counted or quantified. English, on the other hand, typically does not use explicit classifiers, relying on context or prepositions. When translating from English to Japanese, this missing classifier information needs to be recovered. This paper addresses the problem of automatically generating these obligatory Japanese numeral classifiers from English input that lacks them.

Methods

The core of the proposed method is a corpus-based approach that leverages phrase alignment between bilingual sentence pairs. The process involves:

  1. Sentence Retrieval: Given an input English sentence, the system retrieves similar sentences from a bilingual training corpus using a dynamic programming (DP) matching algorithm. This algorithm considers the semantic distance between words to find the most relevant examples.
  2. Phrase Alignment: For the retrieved bilingual sentence pairs, a phrase alignment method is applied. This technique identifies equivalent partial word sequences (phrases) across languages, even if they have different syntactic structures. This is crucial for capturing relationships between English numerals and their implicit Japanese classifier counterparts.
  3. Numeral Classifier Assignment: Once phrases are aligned, the system maps the aligned Japanese numeral classifier expressions to the corresponding English numeral expressions. This mapping is guided by the DP matching results (insertions, deletions, substitutions). The most frequent classifier assignment for a given numeral expression is chosen.
  4. Classifier Type Generation: Heuristic rules, based on a predefined table of Japanese quantifier constructions, are used to determine the realization type of the numeral classifier within the sentence context.

Datasets

The authors used a bilingual corpus of Japanese and English utterances, specifically drawn from phrasebooks for tourists. This corpus was utterance-aligned and contained approximately 200K utterances. For evaluation, a test set of 300 utterances was randomly selected from a larger set of 6000 utterances. The research also utilized thesauri (Kadokawa Ruigo-shin-jiten) for calculating semantic distances between words during the alignment process.

Key Insights and Ideas

  • Implicit Knowledge Extraction: The paper demonstrates that valuable linguistic knowledge, like numeral classifiers, can be implicitly present in bilingual corpora and extracted through phrase alignment, even when not explicitly stated in the source language.
  • Uniform Approach: The phrase alignment method offers a unified way to handle various numeral classifier constructions, avoiding the need for language-specific rules for each type.
  • Contextual Importance: The study highlights the difficulty of assigning classifiers correctly for short utterances due to a lack of sufficient contextual information, suggesting that longer utterances or more sophisticated context modeling might be necessary.
  • Acceptable vs. Correct: The evaluation framework introduces a distinction between "correct" and "acceptable" classifier assignments, acknowledging that some variations might be semantically valid even if not the most precise.

Key Results

The proposed method achieved a precision of 84.2% and a recall of 42.0% for numeral classifier generation. The precision for exact-match samples (distance threshold of 0.0) reached 90%, which aligns with previous work. The recall was lower, primarily attributed to the limitations in retrieving similar examples for longer utterances and potential failures in the phrase alignment module. The evaluation showed that for test utterances with a distance threshold of less than 0.3, 331 numeral instances were analyzed, with 121 (36.6%) receiving a "Rank A" (correct) assignment and 18 (5.4%) receiving a "Rank B" (acceptable) assignment.

Limitations and Future Work

The authors acknowledge several limitations:

  • Database Size: The performance is heavily dependent on the size and coverage of the bilingual corpus. If no similar example exists within the given distance threshold, the method fails.
  • Long Utterances: The DP matching method struggles with longer utterances due to the lack of a preprocessing step to split them into smaller chunks, impacting retrieval.
  • Short Utterances: Very short utterances may lack sufficient context for reliable classifier disambiguation.
  • Parser Accuracy: The phrase alignment module's performance is limited by the accuracy of the underlying parsers.
  • Non-Numeral Use of Numerals: The alignment process can fail when numerals are used in non-numerical contexts (e.g., "one" as a determiner).

Future work suggested includes:

  • Improving recall by decomposing input utterances and retrieving examples based on phrase matching rather than complete sentences.
  • Incorporating better parsers to enhance phrase alignment.
  • Developing a word alignment method with higher recall to better handle non-numeral uses of numbers.
  • Addressing the lack of contextual knowledge and verifying the characteristics of numerals and quantified nouns to avoid incompatible assignments.

Summary updated

BibTeX
@inproceedings{paul2002corpus,
  author = {Michael Paul and Eiichiro Sumita and Seiichi Yamamoto},
  title = {{Corpus-based Generation of Numeral Classifier using Phrase Alignment}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2002},
}
Figure