Bringing the Dictionary to the User: The FOKS System

Slaven Bilac, Timothy Baldwin, Hozumi TanakaCOLING 2002 · OpenReview

Abstract

Slaven Bilac, Timothy Baldwin, Hozumi Tanaka. COLING 2002: The 19th International Conference on Computational Linguistics.

2002.

Introduction

For learners of Japanese, the writing system presents a significant barrier to vocabulary acquisition. Unlike alphabetic languages, Kanji characters (ideograms) often have multiple, context-dependent readings. For instance, the character 生 can be read as sei, shou, nama, i(kiru), u(mu), and more. When these characters combine into compounds (jukugo), phonological shifts like rendaku (sequential voicing) or onbin (euphonic changes) further complicate pronunciation.

Traditional dictionaries—both paper and electronic—usually require the user to provide the exact correct reading or navigate cumbersome radical/stroke-count indices. If a learner guesses the reading incorrectly, they hit a dead end. The Forgiving Online Kanji Search (FOKS) system addresses this by acting as a "smart" dictionary that maps "expected" (but potentially incorrect) readings to actual dictionary entries using a probabilistic model.

Methodology

FOKS treats dictionary lookup as a probabilistic inference problem. Instead of strict string matching, it employs a generative model to determine the likelihood that a learner would produce a specific reading for a given dictionary entry. The system uses a Naive Bayes model based on three main components:

  1. Morphophonemic Alignment: Calculating the base probability of a specific Kanji taking a specific reading ($P(r|k)$).
  2. Phonological Rules: Modeling common transformations like gemination and sequential voicing, assigning probabilities to these shifts.
  3. Corpus-Based Frequency: Using the EDR Japanese corpus to determine the prior probability of a word ($F(s)$), boosting common words and penalizing rare ones.

The final "plausibility grade" for a word given a user's input reading is calculated as: $$Grade(s|r) = P(r|s) \times F(s)$$ By pre-calculating these scores for the entire EDICT dictionary, the system provides near-instantaneous results through a web interface without runtime overhead.

Key Insights: The Impact of Query Length

A core insight of the research is that learner errors are predictable, often based on partial knowledge of Kanji (e.g., using a common on-yomi reading where a kun-yomi is required). The authors found that the system's precision improves significantly as the length of the query increases.

Average number of results returned depending on length of reading

As shown in the figure above, for queries of 4 characters or more, the system typically returns 3 or fewer candidates. This makes it highly efficient for complex compounds, where the probability of a random match is low.

Key Results

The system was evaluated against a "conventional" dictionary search using a test set of 139 real-world errors collected from native speakers and learners. The results demonstrate a significant improvement in error resilience.

Performance comparison between conventional search and the FOKS system

On the test set, FOKS successfully retrieved the intended word 34 times, compared to only 10 times for the conventional system. This represents a 35.8% reduction in the error rate for dictionary lookups (calculated against the 77 target words actually present in the dictionary). For successful "intelligent" searches, the desired entry had a mean rank of 4.71, meaning the user usually found their word in the top five results even when their input was wrong.

Limitations and Future Work

While effective, the current model assumes Kanji readings in a compound are independent, which is linguistically imperfect. The authors acknowledge that future versions could address graphic similarity (confusing Kanji that look alike) and semantic similarity (confusing Kanji with similar meanings). Additionally, the model could be adapted to a specific learner's proficiency level or common mistake patterns to further refine ranking accuracy.

Summary updated

BibTeX
@inproceedings{bilac2002bringing,
  author = {Slaven Bilac and Timothy Baldwin and Hozumi Tanaka},
  title = {{Bringing the Dictionary to the User: The FOKS System}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2002},
  url = {https://openreview.net/forum?id=tma33VJFcrF},
}
Figure