<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1107">
  <Title>Semantic Lexicon Acquisition for Learning Natural Language Interfaces</Title>
  <Section position="3" start_page="0" end_page="57" type="intro">
    <SectionTitle>
2 Background
</SectionTitle>
    <Paragraph position="0"> The output produced by WOLFm can be used to assist a larger language acquisition system; in particular, it is currently used as part of the input to a parser acquisition system called CHILL (Constructive Heuristics Induction for Language Learning). CHILL uses inductive logic programming (Muggleton, 1992; Lavra~ and D~eroski, 1994) to learn a deterministic shift-reduce parser written in  Prolog. The input to CHILL is a corpus of sentences paired with semantic representations, the same input required by WOLFm. The parser learned is capable of mapping the sentences into their correct representations, as well as generalizing well to novel sentences.</Paragraph>
    <Paragraph position="1"> CHILL requires a lexicon as background knowledge in order to learn to parse into deeper semantic representations. By using WOLFIE, the lexicon can be provided automatically, easing the task of parser acquisition. Figure 1 illustrates the inputs and outputs of the complete system. The output of WOLFIE is a lexicon of (phrase, meaning} pairs; these aspects will be discussed more thoroughly in the following sections.</Paragraph>
    <Paragraph position="2"> One of the components of CHILL is 311 initial ow~rlygeneral parser, used to analyze the training data. This initial parser is specialized by the learner to generate only correct parses for the training examples. Given a correct lexicon, the overly-general parser should be able to parse all of the training examples.</Paragraph>
    <Paragraph position="3"> In this paper, we limit our discussion of CHmL to its ability to learn parsers that map natural-language questions directly into Prolog queries that can be executed to produce an answer (Zelle and Mooney, 1996). Following are two sample queries for a database on U.S. Geography paired with their corresponding Prolog query: What is the capital of the state with the biggest population? answer(C, (capital(S,C), largestCP, (state(S), population(S,P))))).</Paragraph>
    <Paragraph position="4"> What state is Texarkana located in? answer(S, (state (S), eq(C, cityid (texarkana, 3 ), loc (C, S))).</Paragraph>
    <Paragraph position="5"> Given a sufficient corpus of such sentence/representation pairs, CHILL iS able to learn a parser that correctly parses many novel sentences into logical queries.</Paragraph>
    <Paragraph position="6"> CHILL treats parser induction as a problem of learning rules to control the actions of the shift-reduce parser mentioned above. During parsing, the current context is contained in the contents of a stack and a buffer containing the remaining input. When parsing is complete, the stack contains the representation of the input sentence. There are three types of operators used by the parser to construct logical queries. One is the introduction onto the stack of a predicate needed in the sentence representation, due to the appearance a phrase at the front of the input buffer. The semantic lexicon associates phrases and their representations for use by this type of operator. A second type of operator unifies variables appearing in stack items. For example, in the first representation of a sample query given above, the first argument of answer is unified with the second argument of capital. Finally, a stack item may be embedded into the argument of another stack item, as is required for the first sentence/representation pair given above, to embed state(_) and population(_,_) into the second argument of largest.</Paragraph>
    <Paragraph position="7"> In sum, we concentrate on using machine learning methods to build a system for processing sentences in a narrow domain, but with the goal of obtaining deep semantic representations. This is in contrast to work in this general area that attempts to process broader corpora, but only obtains shallow representations as a result of processing.</Paragraph>
  </Section>
class="xml-element"></Paper>