<?xml version="1.0" standalone="yes"?>
<Paper uid="H05-1105">
  <Title>Using the Web as an Implicit Training Set: Application to Structural Ambiguity Resolution</Title>
  <Section position="7" start_page="838" end_page="841" type="evalu">
    <SectionTitle>
3 Coordination
</SectionTitle>
    <Paragraph position="0"> Coordinating conjunctions (and, or, but, etc.) pose major challenges to parsers and their proper handling is essential for the understanding of the sentence. Consider the following &amp;quot;cooked&amp;quot; example: The Department of Chronic Diseases and Health Promotion leads and strengthens global efforts to prevent and control chronic diseases or disabilities and to promote health and quality of life.</Paragraph>
    <Paragraph position="1"> Conjunctions can link two words, two constituents (e.g., NPs), two clauses or even two sentences. Thus, the first challenge is to identify the boundaries of the conjuncts of each coordination.</Paragraph>
    <Paragraph position="2"> The next problem comes from the interaction of the coordinations with other constituents that attach to its conjuncts (most often prepositional phrases).</Paragraph>
    <Paragraph position="3"> In the example above we need to decide between [health and [quality of life]] and [[health and qual- null (1) &amp;quot;v n2 n1&amp;quot; 59.29 22.06 (2) &amp;quot;p n2 v n1&amp;quot; 57.79 71.58 (3) &amp;quot;n1 * p n2 v&amp;quot; 65.78 20.73 (4) &amp;quot;v p n2 n1&amp;quot; 81.05 8.75 (5) &amp;quot;v pronoun p n2&amp;quot; 75.30 30.40 (6) &amp;quot;be n1 p n2&amp;quot; 63.65 30.54 n1 is pronoun 98.48 3.04  v is to be 79.23 9.53 Surface features (summed) 73.13 9.26 Maj. vote, of - noun 85.01+-1.21 91.77 Maj. vote, of - noun, N/A - verb 83.63+-1.30 100.00  ity] of life]. From a semantic point of view, we need to determine whether the or in chronic diseases or disabilities really means or or is used as an and (Agarwal and Boggess, 1992). Finally, we need to choose between a non-elided and an elided reading: [[chronic diseases] or disabilities] vs. [chronic [diseases or disabilities]].</Paragraph>
    <Paragraph position="4"> Below we focus on a special case of the latter problem: noun compound (NC) coordination. Consider the NC car and truck production. Its real meaning is car production and truck production. However, due to the principle of economy of expression, the first instance of production has been compressed out by means of ellipsis. By contrast, in president and chief executive, president is simply linked to chief executive. There is also an all-way coordination, where the conjunct is part of the whole, as in Securities and Exchange Commission.</Paragraph>
    <Paragraph position="5"> More formally, we consider configurations of the kind n1 c n2 h, where n1 and n2 are nouns, c is a coordination (and or or) and h is the head noun4. The task is to decide whether there is an ellipsis or not, independently of the local context. Syntactically, this can be expressed by the following bracketings: [[n1 c n2] h] versus [n1 c [n2 h]]. (Collins' parser (Collins, 1997) always predicts a flat NP for such configurations.) In order to make the task more  realistic (from a parser's perspective), we ignore the option of all-way coordination and try to predict the bracketing in Penn Treebank (Marcus et al., 1994) for configurations of this kind. The Penn Treebank brackets NCs with ellipsis as, e.g.,</Paragraph>
    <Paragraph position="7"> The NPs with ellipsis are flat, while the others contain internal NPs. The all-way coordinations can appear bracketed either way and make the task harder.</Paragraph>
    <Section position="1" start_page="839" end_page="839" type="sub_section">
      <SectionTitle>
3.1 Related Work
</SectionTitle>
      <Paragraph position="0"> Coordination ambiguity is under-explored, despite being one of the three major sources of structural ambiguity (together with prepositional phrase attachment and noun compound bracketing), and belonging to the class of ambiguities for which the number of analyses is the number of binary trees over the corresponding nodes (Church and Patil, 1982), and despite the fact that conjunctions are among the most frequent words.</Paragraph>
      <Paragraph position="1"> Rus et al. (2002) present a deterministic rule-based approach for bracketing in context of coordinated NCs of the kind n1 c n2 h, as a necessary step towards logical form derivation. Their algorithm uses POS tagging, syntactic parses, semantic senses of the nouns (manually annotated), lookups in a semantic network (WordNet) and the type of the coordination conjunction to make a 3-way classification: ellipsis, no ellipsis and all-way coordination. Using a back-off sequence of 3 different heuristics, they achieve 83.52% precision (baseline 61.52%) on a set of 298 examples. When 3 additional context-dependent heuristics and 224 additional examples with local contexts are added, the precision jumps to 87.42% (baseline 52.35%), with 71.05% recall.</Paragraph>
      <Paragraph position="2"> Resnik (1999) disambiguates two kinds of patterns: n1 and n2 n3 and n1 n2 and n3 n4 (e.g., [food/n1 [handling/n2 and/c storage/n3] procedures/n4]). While there are two options for the former (all-way coordinations are not allowed), there are 5 valid bracketings for the latter. Following Kurohashi and Nagao (1992), Resnik makes decisions based on similarity of form (i.e., number agreement: P=53%, R=90.6%), similarity of meaning (P=66%, R=71.2%) and conceptual association</Paragraph>
    </Section>
    <Section position="2" start_page="839" end_page="840" type="sub_section">
      <SectionTitle>
Example Predicts P(%) R(%)
</SectionTitle>
      <Paragraph position="0"> (buy) and sell orders NO ellipsis 33.33 1.40 buy (and sell orders) NO ellipsis 70.00 4.67 buy: and sell orders NO ellipsis 0.00 0.00 buy; and sell orders NO ellipsis 66.67 2.80 buy. and sell orders NO ellipsis 68.57 8.18 buy[...] and sell orders NO ellipsis 49.00 46.73 buy- and sell orders ellipsis 77.27 5.14 buy and sell / orders ellipsis 50.54 21.73 (buy and sell) orders ellipsis 92.31 3.04 buy and sell (orders) ellipsis 90.91 2.57 buy and sell, orders ellipsis 92.86 13.08 buy and sell: orders ellipsis 93.75 3.74 buy and sell; orders ellipsis 100.00 1.87 buy and sell. orders ellipsis 93.33 7.01 buy and sell[...] orders ellipsis 85.19 18.93  and recall shown are across all examples, not just the buy and sell orders shown.</Paragraph>
      <Paragraph position="1"> (P=75.0%, R=69.3%). Using a decision tree to combine the three information sources, he achieves 80% precision (baseline 66%) at 100% recall for the 3noun coordinations. For the 4-noun coordinations the precision is 81.6% (baseline 44.9%), 85.4% recall. null Chantree et al. (2005) cover a large set of ambiguities, not limited to nouns. They allow the head word to be a noun, a verb or an adjective, and the modifier to be an adjective, a preposition, an adverb, etc. They extract distributional information from the British National Corpus and distributional similarities between words, similarly to (Resnik, 1999). In two different experiments they achieve P=88.2%, R=38.5% and P=80.8%, R=53.8% (baseline P=75%).</Paragraph>
      <Paragraph position="2"> Goldberg (1999) resolves the attachment of ambiguous coordinate phrases of the kind n1 p n2 c n3, e.g., box/n1 of/p chocolates/n2 and/c roses/n3. Using an adaptation of the algorithm proposed by Ratnaparkhi (1998) for PP-attachment, she achieves P=72% (baseline P=64%), R=100.00%.</Paragraph>
      <Paragraph position="3"> Agarwal and Boggess (1992) focus on the identification of the conjuncts of coordinate conjunctions. Using POS and case labels in a deterministic algorithm, they achieve P=81.6%. Kurohashi and Nagao (1992) work on the same problem for Japanese. Their algorithm looks for similar word sequences among with sentence simplification, and achieves a precision of 81.3%.</Paragraph>
    </Section>
    <Section position="3" start_page="840" end_page="840" type="sub_section">
      <SectionTitle>
3.2 Models and Features
</SectionTitle>
      <Paragraph position="0"> We use the following n-gram models: (i) #(n1,h) vs. #(n2,h) (ii) #(n1,h) vs. #(n1,c,n2) Model (i) compares how likely it is that n1 modifies h, as opposed to n2 modifying h. Model (ii) checks which association is stronger: between n1 and h, or between n1 and n2. Regardless of whether the coordination is or or and, we query for both and we add up the corresponding counts.</Paragraph>
      <Paragraph position="1">  The set of surface features is similar to the one we used for PP-attachment. These are brackets, slash, comma, colon, semicolon, dot, question mark, exclamation mark, and any character. There are two additional ellipsis-predicting features: a dash after n1 and a slash after n2, see Table 3.</Paragraph>
      <Paragraph position="2">  We use the following paraphrase patterns:  (1) n2 c n1 h (ellipsis) (2) n2 h c n1 (NO ellipsis) (3) n1 h c n2 h (ellipsis) (4) n2 h c n1 h (ellipsis)  If matched frequently enough, each of these patterns predicts the coordination decision indicated in parentheses. If found only infrequently or not found at all, the opposite decision is made. Pattern (1) switches the places of n1 and n2 in the coordinated NC. For example, bar and pie graph can easily become pie and bar graph, which favors ellipsis. Pattern (2) moves n2 and h together to the left of the coordination conjunction, and places n1 to the right. If this happens frequently enough, there is no ellipsis. Pattern (3) inserts the elided head h after n1 with the hope that if there is ellipsis, we will find the full phrase elsewhere in the data. Pattern (4) combines pattern (1) and pattern (3); it not only inserts h after n1 but also switches the places of n1 and n2. As shown in Table 4, we included four of the heuristics by Rus et al. (2002). Heuristic 1 predicts no coordination when n1 and n2 are the same, e.g., milk and milk products. Heuristics 2 and 3 perform a lookup in WordNet and we did not use them. Heuristics 4, 5 and 6 exploit the local context, namely the  adjectives modifying n1 and/or n2. Heuristic 4 predicts no ellipsis if both n1 and n2 are modified by adjectives. Heuristic 5 predicts ellipsis if the coordination is or and n1 is modified by an adjective, but n2 is not. Heuristic 6 predicts no ellipsis if n1 is not modified by an adjective, but n2 is. We used versions of heuristics 4, 5 and 6 that check for determiners rather than adjectives.</Paragraph>
      <Paragraph position="3"> Finally, we included the number agreement feature (Resnik, 1993): (a) if n1 and n2 match in number, but n1 and h do not, predict ellipsis; (b) if n1 and n2 do not match in number, but n1 and h do, predict no ellipsis; (c) otherwise leave undecided.</Paragraph>
    </Section>
    <Section position="4" start_page="840" end_page="841" type="sub_section">
      <SectionTitle>
3.3 Evaluation
</SectionTitle>
      <Paragraph position="0"> We evaluated the algorithms on a collection of 428 examples extracted from the Penn Treebank. On extraction, determiners and non-noun modifiers were allowed, but the program was only presented with the quadruple (n1, c, n2, h). As Table 4 shows, our overall performance of 80.61 is on par with other approaches, whose best scores fall into the low 80's for precision. (Direct comparison is not possible, as the tasks and datasets all differ.) As Table 4 shows, n-gram model (i) performs well, but n-gram model (ii) performs poorly, probably because the (n1,c,n2) contains three words, as opposed to two for the alternative (n1,h), and thus a priori is less likely to be observed.</Paragraph>
      <Paragraph position="1"> The surface features are less effective for resolving coordinations. As Table 3 shows, they are very good predictors of ellipsis, but are less reliable when  predicting NO ellipsis. We combine the bold rows of Table 4 in a majority vote, obtaining P=83.82%, R=80.84%. We assign all undecided cases to no ellipsis, yielding P=80.61%, R=100%.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>