Toward Treating English Nominals Correctly

Richard W. Sproat, Mark Y. LibermanACL 1987

TL;DR

This paper presents a computational approach to assigning correct stress patterns to English nominal expressions, crucial for natural-sounding text-to-speech systems. By combining general English stress rules with specific knowledge about nominal types and their semantic relationships, the authors developed a program that significantly improves stress assignment accuracy over simpler methods.

What Problem Does This Paper Address?

Accurately predicting stress in English nominal expressions (like "rising prices" or "quarterback sneak") is a significant challenge for text-to-speech (TTS) systems. Unlike simple words, the stress in these multi-word phrases can depend heavily on their syntactic structure and, importantly, their meaning. This ambiguity means that even when the words are known, their stress pattern can be unpredictable, leading to unnatural-sounding speech. The paper aims to bridge this gap by developing an algorithm that can correctly assign stress to these complex nominals in unrestricted text, considering both syntactic and semantic factors.

Methods

The authors propose a two-pronged approach:

  • Parsing: While the paper's primary focus is stress assignment, it acknowledges the necessity of parsing complex nominals. They suggest that the knowledge used for stress assignment can also be beneficial for parsing, particularly by leveraging semantic relationships between words.
  • Stress Assignment: The core of their method involves a program that assigns stress contours to nominals. This program utilizes:
  • Idiosyncratic Knowledge: A lexicon of specific nominal types with known stress behaviors (e.g., "morning sickness" vs. "morning prayers").
  • General Knowledge: Rules derived from English stress patterns, including syntactic structure and semantic relationships.
  • Heuristic Patterns: A set of rules based on semantic, syntactic, and lexical properties of the nominal. These patterns are weighted, and the pattern with the most "votes" determines the stress. Examples include "MEASURE-PHRASE" (e.g., "pint jug" - right stress) and "LOCATION-TIME-OR-SUBSTANCE" (e.g., "kitchen towel" - right stress).
  • N-Ary Nominals: For nominals with more than two words, they employ a recursive stress assignment algorithm based on metrical grids, similar to established linguistic theories. This involves combining constituents and upgrading stress levels.
  • Rhythm Rule: A final check is performed for the applicability of the Rhythm Rule, which can shift stress in certain contexts.

Datasets

The paper mentions using the Brown Corpus (Francis and Kučera, 1982) to determine the frequency of complex nominals. For evaluation, they used 200 [N N] nominals from the Bell Labs News. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Lexicalization is not the sole determinant of stress: The authors debunk the myth that only lexicalized nominals are left-stressed. They provide examples like "dog annihilator" (left-stressed, not lexicalized) and "red herring" (right-stressed, must be lexicalized) to illustrate this.
  • Syntax and Semantics are Intertwined for Stress: The paper strongly emphasizes that predicting nominal stress requires understanding not just the word order but also the meaning relationships between the words. This is crucial for distinguishing between different types of modification.
  • Heuristics are essential for handling ambiguity: Since a purely syntactic approach is insufficient, the development of weighted heuristic patterns based on semantic and lexical properties is key to improving accuracy.
  • Parsing and Stress Assignment are mutually beneficial: The knowledge gained for stress assignment, particularly semantic relationships, can also inform the parsing process, suggesting a unified approach to handling complex nominals.

Key Results

The authors' program, which incorporates their heuristic patterns, achieved 80% correct stress assignment on a set of 200 [N N] nominals from the Bell Labs News. This significantly outperforms simpler methods:

  • Uniform rightward stress: 45% correct.
  • Leftward stress if N-N: 66% correct.

They classify failures into two categories: insufficient information (about 30 cases) and misanalysis (about 10 cases). The former suggests that expanding the set of heuristic patterns and word class knowledge would further improve performance.

Limitations and Future Work

  • Insufficient Information: A significant portion of the program's failures were due to a lack of information about word classes or semantic categories (e.g., chemical terms). This highlights the need to expand the program's knowledge base.
  • Manual Weight Assignment: The weights for the heuristic patterns are currently assigned "ad hoc" by hand. The authors plan to replace this with results from a statistical survey of nominal types.
  • Parsing Assumption: The stress assignment program assumes a parsed input, which is not a realistic assumption for a working TTS system. The paper acknowledges the need to integrate parsing capabilities.
  • Future Work: The authors suggest improving the program by increasing its set of heuristic patterns and its knowledge of word classes. They also aim to replace manual weight assignment with statistical data. They believe a 90-95% correct stress goal for [N N] nominals is plausible.

Summary updated

BibTeX
@inproceedings{sproat1987toward,
  author = {Richard W. Sproat and Mark Y. Liberman},
  title = {{Toward Treating English Nominals Correctly}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1987},
}

Loading related content...

Figure