<?xml version="1.0" standalone="yes"?>
<Paper uid="W97-1207">
  <Title>Semantic and Discourse Information for Text-to-Speech Intonation</Title>
  <Section position="5" start_page="48" end_page="49" type="metho">
    <SectionTitle>
3 The WordNet Lexical Database
</SectionTitle>
    <Paragraph position="0"> WordNet is a large on-fine Engfish lexical database, based on theories of human lexical memory and comprised of four part-of-speech categories: nouns, verbs, adjectives, and adverbs (Miller et al., 1993). Within each category, lexical meaning is represented by synonym sets (synsets) organized around semantic relationships. Polysemous words are represented by multiple synsets, one for each word sense. The release used in this work, WordNet 1.5, contains a total of 91,591 synsets and 168,135 word senses (Miller, 1995).</Paragraph>
    <Paragraph position="1"> Types of semantic relationships between synsets vary by category. The basic structure of each is discussed briefly below.</Paragraph>
    <Paragraph position="2"> 3.0.1 Nouns The nouns category is the largest and semantically richest of the four. It contains 60,557 synsets, grouped into 25 different topical hierarchies. Synsets in each hierarchy are organized using hypernymy/hyponymy  (IS-A) relationships. The noun hierarchies also include antonymy and three types of meronymy/holonymy relationships (PART-OF, MEMBER-OF, MADE-OF). Meronyms are typically defined at the level of basic concepts in the hierarchies.</Paragraph>
    <Paragraph position="3"> 3.0.2 Verbs  Verbs currently comprise 11,363 synsets in WordNet, divided into 15 categories based on semantic criteria. The primary semantic relationships for verbs in WordNet are lexical entailment (e.g. snoring ENTAILS sleeping) and hypernomy/hyponymy. Verb hierarchies also in- null clude troponymy (MANNER-OF) relationships, and to a. lesser extent, antonymy and causM relationships. Generally, verb hierarchies are much shallower with higher branching factors than noun hierarchies, but like nouns, verbs exhibit basic concept levels at which most troponyms are defined.</Paragraph>
  </Section>
  <Section position="6" start_page="49" end_page="49" type="metho">
    <SectionTitle>
3.0.3 Adjectives
</SectionTitle>
    <Paragraph position="0"> WordNet contains 16,428 synsets of adjectives divided into descriptive and relational types, and a small closed-class of reference-modifying adjectives. Descriptive adjectives are organized around antonymy, and relational adjectives according to the nouns to which they pertain.</Paragraph>
    <Paragraph position="1"> WordNet also encodes limitations on syntactic positions that specific adjectives can occupy.</Paragraph>
  </Section>
  <Section position="7" start_page="49" end_page="49" type="metho">
    <SectionTitle>
3.0.4 Adverbs
</SectionTitle>
    <Paragraph position="0"> Adverbs make up the smallest of the four categories, with a total of 3243 synsets. Adverbs are organized by antonymy and similarity relationships. null</Paragraph>
  </Section>
  <Section position="8" start_page="49" end_page="52" type="metho">
    <SectionTitle>
4 Implementation
</SectionTitle>
    <Paragraph position="0"> An overview of the system architecture is shown in Figure 2. Following (Cahn, 1994; Cahn, 1997), text files are first parsed by the NPtool noun phrase parser, which identifies noun phrases and tags each word with morphological, syntactic, and part-of-speech information (Voutilainen, 1993). The preliminary processing module then adds gender information for proper names, resolves ambiguous tags, and reformats the text for further processing. ~ Next, the previous mention, contrast, and theme modules assign pitch accents, phrase accents, and boundary tones, using WordNet to identify sets of synonyms and contrastive words. Finally, the annotated text is re-formatted for the TrueTalk speech synthesizer (Entropic Research Laboratory, 1995). Additional implementation details for the accent assignment modules are provided below.</Paragraph>
    <Paragraph position="1"> SGender resohition is performed via simple lookup using the CMU Artificial Intelligence Repository Name Corpus (l(antrowitz, 1994). Ambiguous parses are resolved using a set of heuristics derived from analysis of NPtool output.</Paragraph>
    <Section position="1" start_page="49" end_page="49" type="sub_section">
      <SectionTitle>
4.1 Givenness Identification
</SectionTitle>
      <Paragraph position="0"> The first of the three accent assignment modules assigns .pitch accents to words using the following given/new strategy: For each word W,  1. If W is a noun, verb, adjective, or adverb, and W C/ history(), and W C/ equiv(x), for any x E history( ): (a) tag W as a focused item (b) add W to history( ) (c) create equiv(W) 2. If W is a noun, verb, adjective, or adverb,  and W E equiv(x), tag W as inferable. 6 The history and equivalence lists are reset at each paragraph boundary. Matches are limited to words belonging to the same part-of-speech category, relying only on word roots. Equivalence (synonym) sets are created from semantic relationships for each WordNet category as follows:  1. Nouns: equiv(W) = union of hypernyms and synonyms for all synsets of W. The number of hypernym levels used for each sense is determined by searching for the existence of meronyms on the current level, climbing the hypernym tree until a level containing meronyms is found, or the root is reached. If no meronyms are found, then (1/4 x depth of W synset) levels are used. 7 2. Verbs: equiv(W) = union of hypernyms, synonyms, and entailments for all synsets of W. Only one level of hypernyms is included.S 3. Adjectives and adverbs: equiv(W) = syn- null focused, reflecting their status as not explicitly given. ~The present approach to identifying a &amp;quot;basic&amp;quot; concept level for nouns using meronymic relations is not the optimal solution. Many noun categories in WordNet do not include meronyms, and meronyms may exist at sev: eral levels within a hierarchy.</Paragraph>
      <Paragraph position="1"> 8 Because verb hierarchies have a much higher branching factor, considering more than one level is generally impractical.</Paragraph>
      <Paragraph position="2">  Equivalence lists are ordered and searched from most common to least common sense of a word. The current implementation is limited to single word matches; no WordNet entries consisting of multi-word descriptions are included in the equivalence list.</Paragraph>
    </Section>
    <Section position="2" start_page="49" end_page="51" type="sub_section">
      <SectionTitle>
4.2 Contrastive Stress Assignment
</SectionTitle>
      <Paragraph position="0"> The second accent assignment module compares each open-class word (nouns, verbs, adjectives, and adverbs) with other words previously realized in the text to identify possible contrastive relationships. The top-level algorithm for assigning contrastive stress is shown in pseudo-code in Figure 3.</Paragraph>
      <Paragraph position="1"> Sets of possible contrastive words for nouns and verbs are determined by the hypernym/hyponym relationships in WordNet as follows: null</Paragraph>
      <Paragraph position="3"> Identify the set of immediate hypernyms, hyper(W), corresponding to each sense of W (synsets containing W).</Paragraph>
      <Paragraph position="4"> For each h: h E hyper(W), identify the set of immediate hyponyms, hypo(h), such that W e hypo(h).</Paragraph>
      <Paragraph position="5"> The set of possible contrastive words is the union of hyponyms for each sense of W.</Paragraph>
      <Paragraph position="6"> The contrastive sets for adjectives and adverbs are simply the union of antonyms for all foreach word W1 { for each word W2 on the history list (from most to least recent) { for each A: A E contrast(W2) { if W1 equals A then { tag W1 for contrastive stress;</Paragraph>
      <Paragraph position="8"> if no contrast is found { add W1 to the history list;  ships are not used in WordNet for either class. All contrastive sets generated are ordered and searched from the most common to least common sense of a word. The present implementation is limited to single word searches. There are a number of shortcomings in the present implementation of contrastive stress assignment. The first is its failure to use textual information to facilitate identification of contrastive relationships. To rectify this situation, a search for keywords commonly used to indicate contrast (e.g. however, unlike, on- null the-other-hand), as well as explicit negation (not) must be incorporated. Identifying parallel phrasing may also be useful in identifying contrastive relationships not encoded in WordNet (namely for non-antonymic contrasts between adjectives and adverbs).</Paragraph>
    </Section>
    <Section position="3" start_page="51" end_page="52" type="sub_section">
      <SectionTitle>
4.3 Theme ~ Rheme Identification
</SectionTitle>
      <Paragraph position="0"> The modules described above determine the second tier of the information structure--that is, which items are eligible for focus based on their new or contrastive status. The theme/rheme identification module is responsible for determining the primary information structure delineation of theme and rheme. Based on an automatic segmentation of utterances or parts of utterances into theme and rheme, we can apply the mapping of tunes described in Section 2 to decide which pitch accents to assign and where to place phrasal and boundary tones.</Paragraph>
      <Paragraph position="1"> The automatic segmentation of utterances into theme and rheme is a difficult problem.</Paragraph>
      <Paragraph position="2"> Our preliminary approach is based on a number of heuristics, and generally performs quite well. Nonetheless, we expect this module to be substantially refined once we have concluded our empirical analysis of the Boston University radio news corpus (Ostendorf, Price, and Shattuck-Hufnagel, 1995).</Paragraph>
      <Paragraph position="3"> The theme/rheme identification algorithm begins by trying to identify propositional constituents within utterances. As noted in Section 2, a single utterance may contain several clauses corresponding to several semantic propositions. Propositional constituents are centered around verb occurrences. The algorithm looks for verb complexes--contiguous stretches of text containing verbs, adverbs and some prepositions. Utterances are then divided into propositional constituents such that each contains a single verb complex. The algorithm also considers multi-word clauses that are set apart by punctuation, such as utterance-initial prepositional phrases, as separate propositional constituents. 9 This segmentation scheme is sim9Note that we work with the part-of-speech output of NPtool rather than a complete parse tree. While this presents a number of diffictflties for dividing utterances ilar to Gussenhoven's division of utterances into focus domains (Gussenhoven, 1983).</Paragraph>
      <Paragraph position="4"> Once propositional constituents have been determined, the algorithm applies a number of heuristics to sub-divide each into theme and rheme. We consider two possible segmentation points: before the verb-complex and after the verb-complex. The heuristics are as follows, where PreV, V and PostV correspond to the pre-verbal material, the verb-complex material and the post-verbal material respectively.</Paragraph>
      <Paragraph position="5">  into propositional constituents, it allows us more freedom in sub-dividing those propositional constituents into theme and rheme. That is, our program can produce prosodic phrases, such as those shown in Figure 1, that are orthogonal to traditional syntactic structures.  8. If PostV contains focused items, but PreV and V do not:</Paragraph>
      <Paragraph position="7"> Note that these heuristics encode a preference tbr thematic phrases to precede rhematic phrases, but do not always dictate such an ordering. Also, note that the heuristics allow thematic phrases to sometimes contain focused items. This is in accordance with our observation in Section 2 that themes need not contain only background material.</Paragraph>
      <Paragraph position="8"> Based on the theme/rheme identification heuristics, we map L-t-H* accents onto focused items in themes and H* accents onto focused items in rhemes. L- phrasal tones are placed at theme and rheme boundaries. When theme or rheme phrases are also marked by punctuation, appropriate boundary tones and pauses are also inserted (e.g. H% for comma delimited phrases).</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>