Russian Stress Prediction using Maximum Entropy Ranking

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 879–883, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Russian Stress Prediction using Maximum Entropy Ranking Keith Hall Richard Sproat Google, Inc New York, NY, USA {kbhall,rws}@google.com Abstract We explore a model of stress prediction in Russian using a combination of lo- cal contextual features and linguistically- motivated features associated with the word’s stem and suffix. We frame this as a ranking problem, where the objec- tive is to rank the pronunciation with the correct stress above those with incorrect stress. We train our models using a simple Maximum Entropy ranking framework al- lowing for efficient prediction. An empir- ical evaluation shows that a model com- bining the local contextual features and the linguistically-motivated non-local fea- tures performs best in identifying both primary and secondary stress. 1 Introduction In many languages, one component of accu- rate word pronunciation prediction is predict- ing the placement of lexical stress. While in some languages (e.g. Spanish) the lexical stress system is relatively simple, in others (e.g. En- glish, Russian) stress prediction is quite compli- cated. Much as with other work on pronuncia- tion prediction, previous work on stress assign- ment has fallen into two camps, namely systems based on linguistically motivated rules (Church, 1985, for example) and more recently data- driven techniques where the models are derived directly from labeled training data (Dou et al., 2009). In this work, we present a machine- learned system for predicting Russian stress which incorporates both data-driven contextual features as well as linguistically-motivated word features. 2 Previous Work on Stress Prediction Pronunciation prediction, of which stress pre- diction is a part, is important for many speech applications including automatic speech recog- nition, text-to-speech synthesis, and translit- eration for, say, machine translation. While there is by now a sizable literature on pro- nunciation prediction from spelling (often termed “grapheme-to-phoneme” conversion), work that specifically focuses on stress predic- tion is more limited. One of the best-known early pieces of work is (Church, 1985), which uses morphological rules and stress pattern templates to predict stress in novel words. An- other early piece of work is (Williams, 1987). The work we present here is closer in spirit to data-driven approaches such as (Webster, 2004; Pearson et al., 2000) and particularly (Dou et al., 2009), whose features we use in the work described below. 3 Russian Stress Patterns Russian stress preserves many features of Indo- European accenting patterns (Halle, 1997). In order to know the stress of a morphologically complex word consisting of a stem plus a suf- fix, one needs to know if the stem has an accent, and if so on what syllable; and similarly for the suffix. For words where the stem is accented, 879 acc unacc postacc Dat Sg гор'оху г'ороду корол'ю gor’oxu g’orodu korolj’u Dat Pl гор'охам город'ам корол'ям gor’oxam gorod’am korolj’am ’pea’ ’town’ ’king’ Table 1: Examples of accented, unaccented and postaccented nouns in Russian, for dative singular and plural …

Content truncated to 500 words