Paraphrasing Predicates from Written Language to Spoken Language Using the Web

Nobuhiro Kaji, Masashi Okamoto, Sadao KurohashiNAACL 2004

TL;DR

This paper tackles the challenge of making synthesized speech sound more natural by addressing the differences between written and spoken language. They propose a method to automatically identify and paraphrase "unsuitable expressions for spoken language" (UES) into "suitable expressions for spoken language" (SES) by leveraging large corpora collected from the web.

What Problem Does This Paper Address?

Speech synthesis applications, while improving, often produce unnatural-sounding speech. One significant reason for this is the inherent difference between how language is used in written text versus spoken conversation. Written language tends to employ more complex vocabulary and sentence structures, which can sound jarring when read aloud by a synthesizer. While some research has focused on improving the acoustic aspects of speech synthesis, this paper addresses the linguistic mismatch: transforming written text into expressions that are more natural and appropriate for spoken delivery. They specifically focus on "unsuitable expressions for spoken language" (UES) and aim to paraphrase them into "suitable expressions for spoken language" (SES).

Methods

The core of their method involves learning paraphrase pairs of the form 'UES → SES'. This is achieved by distinguishing UES and SES based on their occurrence probabilities in written and spoken language corpora, which are automatically collected from the web.

The process involves three main steps:

  1. Learning initial paraphrase pairs: They start by learning paraphrase pairs of predicates (verbs or adjectives) from a dictionary using a method proposed by Kaji et al. (2002).
  2. Collecting corpora: They automatically collect large written and spoken language corpora from the web. This is a crucial step, as existing Japanese spoken language corpora were found to be too small.
  3. Selecting UES → SES pairs: Using the collected corpora, they identify paraphrase pairs where the source expression is likely UES and the target expression is likely SES. This selection is based on comparing the occurrence probabilities of the source and target expressions in both the written and spoken corpora.

A key aspect of their corpus collection is the use of "interpersonal expressions" (familiarity and politeness markers) to classify web pages into written or spoken language categories. They also employ a Support Vector Machine (SVM) for binary classification of paraphrase pairs, using the occurrence probabilities as features.

Datasets

The paper introduces a method for collecting large-scale written and spoken language corpora from the Web.

  • Web Corpus: Initially, a large corpus of web pages (660,062 pages, 733M words) was collected.
  • Written Language Corpus: From the Web Corpus, 80,685 pages (77M words) were classified as written language.
  • Spoken Language Corpus: From the Web Corpus, 73,977 pages (113M words) were classified as spoken language.

These corpora are significantly larger than existing Japanese spoken language corpora. For evaluation, they used a dataset of 200 paraphrase pairs (70 positive, 130 negative) derived from the initial 5,836 learned pairs, which were manually annotated by two judges.

Key Insights and Ideas

  • Web as a Corpus Source: The paper demonstrates the feasibility and effectiveness of using the vastness of the World Wide Web to collect large, style-specific language corpora, overcoming the limitations of existing, smaller datasets.
  • Interpersonal Expressions as Classifiers: The insight that interpersonal expressions (familiarity and politeness markers) are strong indicators of spoken language is a novel approach to automatically classifying web pages by linguistic style.
  • Occurrence Probability as a Distinguishing Feature: The core idea that the difference in occurrence probability of an expression between written and spoken corpora can reliably distinguish between UES and SES is a powerful and intuitive principle.
  • Focus on Lexical Paraphrases: By focusing on lexical paraphrases (rather than syntactic ones), they target a specific type of linguistic mismatch that is crucial for natural-sounding speech synthesis.

Key Results

  • Corpus Collection Precision: The automatic collection method achieved a high average precision of 94% for classifying web pages into written and spoken language corpora.
  • Paraphrase Learning Accuracy: The SVM-based classifier achieved an accuracy of 76% in identifying suitable paraphrase pairs (UES → SES).
  • Feature Effectiveness: Comparing two feature sets, the one including occurrence probabilities in both written and spoken corpora (F-set1) performed slightly better (76% accuracy) than one using only spoken corpus probabilities (F-set2, 75% accuracy), highlighting the importance of the comparative analysis.
  • Qualitative Success: The paper presents examples of successfully classified paraphrase pairs, demonstrating the practical application of their method.

Limitations and Future Work

  • Corpus Bias: The authors acknowledge that the collected corpora can be biased by topic (e.g., a corpus containing many drama reviews might show high occurrence of words related to dramas). They suggest collecting from a larger web corpus to mitigate this.
  • Lexical Information: While their method is effective, they note that incorporating more explicit lexical information could further improve classification.
  • Broader Scope: The current work focuses on paraphrasing predicates. Future work includes examining other types of SES (slang, gendered speech) and extending paraphrasing to compound nouns and complex syntactic structures.
  • Iterative Improvement: They plan to iteratively improve the quality of their corpora.

Summary updated

BibTeX
@inproceedings{kaji2004paraphrasing,
  author = {Nobuhiro Kaji and Masashi Okamoto and Sadao Kurohashi},
  title = {{Paraphrasing Predicates from Written Language to Spoken Language Using the Web}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2004},
}
Figure