Completion of Japanese Sentences by Inferring Function Words from Content Words

Koji Kakigahara, Teruaki AizawaCOLING 1988

TL;DR

This paper tackles the challenge of accurately translating spoken Japanese by automatically inferring crucial "function words" (like particles and auxiliary verbs) from recognized "content words" (nouns and verbs). By leveraging semantic features of nouns and verb valency patterns, their method significantly improves the generation of grammatically correct Japanese sentences, bridging a critical gap between speech recognition and machine translation.

What Problem Does This Paper Address?

Automatic Telephone Interpretation (ATI) systems aim to enable seamless communication between speakers of different languages over the phone. A key bottleneck in such systems is the interface between speech recognition (SR) and machine translation (MT). Spoken language, especially in noisy environments or with rapid speech, often results in errors from the SR module. These errors, particularly in recognizing short, grammatically essential "function words" (like particles and auxiliary verbs in Japanese), can be unacceptable for the MT module, leading to mistranslations or complete system failure. This paper focuses on a specific sub-problem: how to reliably infer these missing or incorrectly recognized function words in Japanese, given a phrase lattice of recognized content words from the SR output.

Methods

The core of the proposed method is a "selection-by-generation" approach. Instead of directly trying to select the correct function word from the SR output's candidates, the system first generates a plausible Japanese sentence by inferring the most appropriate function words based on the recognized content words. This generation process relies on two key pieces of information:

  1. Semantic Features of Nouns and Valency Patterns: Nouns are annotated with both general semantic features (e.g., abstract, concrete, location) and task-specific features. Predicates (verbs) are then associated with "valency patterns" that describe which types of nouns (with specific semantic features) can connect to them using particular function words. For example, a verb like 'okuru' (to send) might have patterns like N[concrete/document] 'wo' + V or N[time/process] 'madeni' + N[concrete/document] 'wo' + V.
  2. Syntactic Information: Basic syntactic rules of Japanese are also employed, such as how verb forms connect with other words or how nouns can be linked using the function word 'no' (similar to 'of' in English).

The inference process involves:

  • Using syntactic information to connect content words.
  • Applying valency patterns based on the semantic features of recognized nouns and the verb.
  • Considering possible connections between nouns using 'no'.
  • Finally, generating candidate sentences.

The paper emphasizes that the key step is this generation process, which then allows for a more robust selection of the most appropriate function words.

Datasets

The paper describes using dialogue sentences collected from a restricted task domain: inquiries about an international conference. This dataset was used to prepare valency patterns for 65 verbs. For evaluation, 70 test questions were prepared, also within the same interrogative sentence domain.

Key Insights and Ideas

  • Inferring Function Words by Generation: The most significant insight is the shift from direct selection to generation. By first constructing a plausible sentence structure, the system can more effectively determine the correct function words, even if they were poorly recognized by the SR.
  • Valency Patterns as the Core Knowledge: The paper highlights the power of valency patterns, enriched with semantic features, as a compact and effective way to represent the grammatical and semantic relationships between words in Japanese.
  • Task Domain Restriction is Crucial: The success of the method is heavily reliant on the restricted task domain. This allows for a more manageable set of semantic features and valency patterns, increasing the probability of correct inference.
  • Inferring Omitted Verbs is Possible: The research extends to inferring omitted verbs, suggesting that the semantic relationships captured by valency patterns can work in reverse, inferring missing content words as well.

Key Results

The paper presents experimental results in Table 2, showing the improvement in the "correct answer rate" over multiple trials.

  • At the first trial, using the prepared valency patterns, 64.3% of the generated sentences were correct.
  • With the addition of more valency patterns in subsequent trials, the correct answer rate increased significantly, reaching 95.7% by the fifth trial and 100% by the sixth trial.
  • The experiments focused on simple interrogative sentences, demonstrating that the valency pattern approach allows for easy production of complete sentences from a sequence of content words.

Limitations and Future Work

  • Reliance on Correct Content Words: The method assumes that the content words (nouns and verbs) have been correctly recognized by the speech recognition system. Errors in content word recognition are not directly addressed by this specific inference mechanism.
  • Restricted Task Domain: The effectiveness is demonstrated within a narrow domain (international conference inquiries). Generalizing this to broader, open-domain conversations would require a significantly larger and more complex set of semantic features and valency patterns.
  • Future Work: The authors are currently evaluating the effectiveness of their proposed procedures by applying them to actual speech recognition results, aiming to address the SR errors more directly. They also suggest that inferring omitted verbs is a promising avenue.

Summary updated

BibTeX
@inproceedings{kakigahara1988completion,
  author = {Koji Kakigahara and Teruaki Aizawa},
  title = {{Completion of Japanese Sentences by Inferring Function Words from Content Words}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1988},
}
Figure