<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2116">
  <Title>Sydney, July 2006. c(c)2006 Association for Computational Linguistics A Grammatical Approach to Understanding Textual Tables using Two-Dimensional SCFGs</Title>
  <Section position="4" start_page="0" end_page="906" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Natural language processing has historically tended to emphasize understanding of linear strings--sentences, paragraphs, discourse structure. The vast body of work that focuses on text  spoken language understanding. Yet real-life text is actually heavily dependent on visual layout and formatting, which compensate for cues normally found in spoken language but are absent in text. As Scott (2003) reiterated in the opening ACL'03 invited talk: &amp;quot;The overlay of graphics on text is in manywaysequivalenttotheoverlayofprosodyon speech... Just as prosody undoubtedly contributes to the meaning of utterances, so too does a text's graphical presentation contribute to its meaning.</Paragraph>
    <Paragraph position="1"> However... few natural language understanding systems use graphical presentational features to aid interpretation...&amp;quot; (Power et al., 2003).</Paragraph>
    <Paragraph position="2"> Nowhere is this more evident than in the wide-spread use of tables in real-world, unsimplified text documents. Tables have a comparable or greater complexity as other elements of text. Unfortunately, in mainstream NLP it is not uncommon for tables to be regarded as a somehow &amp;quot;degenerate&amp;quot; form of text, unworthy of the same degree of attention as the rest of the text. But as we will discuss, the degree of ambiguity in table understanding is at least as great as for many sense and attachment problems. Many of the same mechanisms used for understanding linear text are also required for table understanding. The same division of surface syntax and underlying semantics is found.</Paragraph>
    <Paragraph position="3"> Indeed, to perceive the limitations of existing table understanding models, we may distinguish severalvery differentlevels oftable analysistasks.</Paragraph>
    <Paragraph position="4"> In table classification, the table is classified into one of several coarse categories (in the extreme case, somemodelssimplypredictwhetherthepurpose of the table is for page layout versus tabular data). In table synactic recognition, the surface types of individual cells or block regions are labeled (e.g., as heading or data) but the underlying semantic relationships between the table elements remain unrecognized and usually highly ambiguous (i.e., no logical relations between the elements  in the table are assigned). In contrast, in table semantic interpretation, the exact logical relations between the elements in the table must be recognized (e.g., by associating the table and/or subregions thereof with precise table schemas in relational database style).</Paragraph>
    <Paragraph position="5"> Existingtableunderstandingworklargelyliesat the level of superficial table classification or syntactic recognition. Rarely, if ever, are precise logical relations assigned between the elements in the table. Ad hoc heuristic approaches tend to rule, rather than linguistic approaches.</Paragraph>
    <Paragraph position="6"> Ontheotherhand, inthelinguisticapproachadvocated by Scott (2003) and (Power et al., 2003), tables were not considered. The various physical presentation elements discussed included headings, captions, and bulleted lists--all of which exhibit numerous similarities to tabular elements.</Paragraph>
    <Paragraph position="7"> Possibly, tables were not considered because they are difficult to describe adequately within the expressivenessofcommonlinguisticformalismslike null CFGs.</Paragraph>
    <Paragraph position="8"> The work presented here aims to address this problem. Our model provides an enabling foundation toward a linguistic approach by first shifting to a two-dimensional CFG framework. This permits us to construct a grammar where all the rules are meaningfully discriminative, such that-unlike existing table understanding models--any analysis of a table includes a full parse tree that assigns precise data model labels to all its regions (including nested subregions) thereby specifying the logical relations between the table's elements.</Paragraph>
    <Paragraph position="9"> Additionally, probabilities on the production rules supportthresholding(orranking)ofthealternative candidate table interpretation hypotheses.</Paragraph>
    <Paragraph position="10"> As with many natural language phenomena, a full model of disambiguation must ultimately integrate lexical semantics. However, in this research  stepwefocusonthequestionofhowmuchsemantic interpretation can be performed on the basis of other features, in the absence of a lexical or ontological model. Just as syntax and morphology and prosody alone already permit much recognition and disambiguation of semantic roles and argumentstructuretobedoneforsentence, thesame can be done for tables. At the same time, we believefutureintegrationoflexicalsemanticswillbe null facilitated by the grammatical framework of our model.</Paragraph>
    <Paragraph position="11"> One way to think about this is that we wish to  model what you might be able to recognize from a &amp;quot;Martian&amp;quot; table such as that in Table 1. The non-Martian reader relies solely on knowledge of alphabetsandnumbers, canspotfontandformatting clues, and is familiar with the conventions (i.e., grammars) of tables in general.</Paragraph>
    <Paragraph position="12"> You might reasonably interpret this table as a collection of vertical records with an attributes header column (Pbje, Hoer, NQ, Ncowifl) on the left. You might additionally interpret it as a table that contains an record key header row (Kwe, Zxc, Amc) along with the attributes header column (Pbje, Hoer, NQ, Ncowifl). You might assign the latter interpretation a slightly higher probability, noticing the slightly longer form of Pbje compared to Kwe, Zxc, and Amc. On the other hand, even without reading English, you could reject the interpretation as a collection of horizontal records under the header attributes row (Pbje, Kwe, Zxc, Amc), since each row contains different forms and types, in a pattern that is consistent acrosscolumns. Otherinterpretationsarealsopossible, but unlikely given the regularity of the patterns. null Thus by analyzing the structure of a table, the reader would form a hypothesis about its data model, providing a semantic interpretation that allows the reader to extract information from the table. As can be seen from the restored original English version of the same example in Table 2, the most likely interpretation was predicted even without access to specific lexical knowledge. We aim to show that a fairly useful baseline level of semantic interpretation accuracy can already be achieved, even with relatively little lexical and ontological knowledge.</Paragraph>
    <Paragraph position="13"> We model these alternative hypotheses for the interpretation of ambiguous tables as competing parses. Just as with ordinary parsing and semantic interpretation, the reader often builds multiple competing interpretations of the same table.</Paragraph>
    <Paragraph position="14"> Note that many previous models do not even distinguish between the alternative possible interpretations in the Martian example. Existing mod- null els such as Hurst (2000) and Yang (2002) interpret tables with the same structural layout simply by assigning them same data model, which stops short of recognizing that it is necessary to rank multiple competing interpretations that entail different sets of logical relations.</Paragraph>
    <Paragraph position="15"> In contrast, our proposed model is capable of producing multiple competing parses indicating different semantic interpretations of tables having the same structural layout, by selecting specific data models for the table and its subregions.</Paragraph>
  </Section>
class="xml-element"></Paper>