Using an On-Line Dictionary to Find Rhyming Words and Pronunciations for Unknown Words

Roy J. Byrd, Martin ChodorowACL 1985 · DOI

In the mid-1980s, researchers at IBM’s T.J. Watson Research Center developed WordSmith, an automated dictionary system designed to transcend the limitations of traditional alphabetical lookups. While standard dictionaries are excellent for finding definitions of known spellings, they struggle with phonetic tasks. WordSmith was engineered to address two specific linguistic challenges: identifying "true" rhymes based on phonetics and predicting the pronunciation of unknown words through computational analogy.

Beyond the Alphabetical List

Traditional dictionaries are organized by spelling, which is often a poor proxy for sound in English. For instance, "cough" and "dough" share similar endings but do not rhyme, whereas "write" and "right" are perfect rhymes despite their visual differences. WordSmith moves beyond these constraints by using phonetic encoding to find rhymes based on primary stress and sound similarity, rather than orthographic patterns.

Phonetic Rhyme Encoding: Stress as the Primary Driver

The system’s first major subsystem is a phonetic rhyme encoder. Instead of simply reversing the letters of a word, WordSmith converts entries into phonetic segments and reorders them to prioritize "rhyme-relevant" components.

The algorithm identifies the syllable with primary stress and then groups all subsequent vowels and consonants. By sorting these groups, the system ensures that words with identical stressed vowels and following sounds appear as neighbors. This approach correctly identifies "fervency" as a better rhyme for "urgency" than "agency," because it prioritizes the stressed vowel sound over the shared "-ency" suffix.

Pronunciation by Analogy: The "Brange" Case Study

To handle unknown words or "pseudowords," WordSmith implements a psycholinguistically-inspired "pronunciation by analogy" engine. This model mimics how humans often pronounce new words by comparing them to known "lexical neighbors."

The process involves a four-step pipeline:

  1. Segmentation: Breaking the unknown word into substrings.
  2. Matching: Finding the longest initial and final substrings in the dictionary.
  3. Mapping: Using PROLOG-based rules to map these substrings to phonetic segments.
  4. Combination: Merging these pieces into a candidate pronunciation.
Figure 2: Two possible phonetic mappings for the pseudoword "brange" based on different lexical neighbors.

As shown in Figure 2, this method often leads to multiple valid pronunciations. For the word "brange," the system might find neighbors like "bran" (suggesting a short a) or "range" (suggesting a long a). This "analogy conflict" captures real-world linguistic ambiguity; informal surveys conducted by the authors confirmed that humans are similarly split on such pronunciations.

Modeling Human Linguistic Ambiguity

By utilizing Webster's Seventh Collegiate Dictionary and its 120,000 pronunciations, WordSmith acts as an experimental tool for testing human perception. It allows for "phonetic equivalence classes," where different phonetic segments (like the French u and English oo) can be mapped to a single code to see if users perceive them as rhyming.

The system also demonstrates an ability to ignore irrelevant prefix syllables. It can correctly group complex words like "antiphonary" and "missionary" based on their trailing stressed patterns, effectively modeling how humans intuitively filter sound for rhyming.

Challenges and the Vowel Triangle

Despite its successes, the WordSmith system faces several hurdles. Assigning primary stress to a completely unknown word remains difficult, especially when neighbor words have conflicting stress patterns. Furthermore, the system currently focuses on word-initial and word-final matches, as indexing internal substrings is computationally expensive.

Figure 1: The phonetic vowel triangle, representing the relationship between different vowel sounds used in the system's encoding.

One specific area for future research is the "schwa" problem—the neutral "uh" sound. As illustrated by the vowel triangle in Figure 1, vowels exist in a continuous phonetic space. Determining exactly where the schwa fits within this triangle is crucial for refining the accuracy of rhyming results and phonetic mappings.

Conclusion

WordSmith represents a significant step toward making digital dictionaries more linguistically "aware." By prioritizing phonetic structure and modeling the human tendency toward analogy, it provides a more intuitive way to navigate the complexities of the English language.

Summary updated

BibTeX
@inproceedings{byrd1985using,
  author = {Roy J. Byrd and Martin Chodorow},
  title = {{Using an On-Line Dictionary to Find Rhyming Words and Pronunciations for Unknown Words}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1985},
  doi = {10.3115/981210.981244},
}
Figure