<?xml version="1.0" standalone="yes"?>
<Paper uid="C86-1111">
  <Title>A New Predictive Analyzer of English</Title>
  <Section position="3" start_page="0" end_page="471" type="metho">
    <SectionTitle>
2. Aspects of Predictions
</SectionTitle>
    <Paragraph position="0"> While reading or hearing English, we constantly predict or expect what may follow next. Such predictions can be classified into six types which we will describe below.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.1. Essential Predictions
</SectionTitle>
      <Paragraph position="0"> The simplest type of prediction, which forms the basis of the following discussions, is presented in this subsection. The characteristic of this type of prediction is that it is essential in forming an acceptable sentence structure.</Paragraph>
      <Paragraph position="1"> Phrase structure grammar rules, especially those in Greibach normal form, can naturally describe this kind of prediction: we can consider the terminal symbol (or the lexical category) on the right-hand side of a rule as the current word and the nonterminal symbols that follow the terminal symbol as new predictions \[3\]. For example, the following rule describes what we predict when we encounter a transitive verb at the beginning of a verb phrase.</Paragraph>
      <Paragraph position="3"> Note that the new prediction, NP, is essential to form a verb phrase. By adopting this kind of rules as a means of structural description of sentences, we can easily capture the structures by using the stack mechanism \[1\].</Paragraph>
      <Paragraph position="4"> In the following subsections, except for the last subsection, these rules and the mechanism are gradually reinforced in order to handle a newly introduced prediction type. The extended mechanism provides us with a simpler (yet still powerful) means for recognition of sentence structures than, for example, ATN framework \[4\]. Other factors that affect the predictive recognition process are discussed in the last subsection.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="470" type="sub_section">
      <SectionTitle>
2.2. Optional Predictions
</SectionTitle>
      <Paragraph position="0"> We now extend our recognition mechanism by introducing optional predictions. This type of prediction is needed to handle postpositional modifiers that are not essential to form a sentence.</Paragraph>
      <Paragraph position="1"> In the previous subsection, we saw that rules in Greibach normal form are suitable for expressing our predictive recognition process, but any rule should not predict too much. Consider the following rule that explains a possible structure of noun phrases.</Paragraph>
      <Paragraph position="2"> NP ~ article NP-ART ADJ_CLAUSE Concerning the correspondence with human language understanding process, however, the rule cannot be considered a good simulation of our understanding process: we predict a postpositional modifier, like an adjectival clause, not at the beginning of a noun phrase but at the beginning of the modifier. For our purpose, therefore, we must exclude this kind of rule that do not express our predictions properly.</Paragraph>
      <Paragraph position="3"> Optional predictions are used to capture these structures.</Paragraph>
      <Paragraph position="4"> Here, we also extend the rule description to keep the correspondence between the grammar rules and the recognizing mechanism: we introduce the shifting flag. The following rules are used to capture postpositional modifiers.</Paragraph>
      <Paragraph position="5">  CW CP SF NPr (1) art NP ~ t NP-ART (2) noun NP-ART ~ t *NP-N (3) rel_pro NP-N ~ nil ADJ_CLAUSE  The first rule, for example, can he interpreted as follows: IF the current word (CW) is an article and the current prediction (CP: the top element of the stack) is NP, THEN shift the current word pointer (since the shifting flag (SF) is t) and replace the current prediction by the new prediction (NPr).</Paragraph>
      <Paragraph position="6"> The shifting flag enables us to proceed two or more state changes while looking at a single word. By using these notations  and the rules we can specify the state changes of the stack as shown in Figure 2-1. The prediction NP-N, with a prefix '*' which shows it is optional, is interpreted as the state in wlfich a noun essential to form a noun phrase has already appeared and it may end there. It will be popped out from the stack or will be replaced by a new prediction according to the word that follows.</Paragraph>
      <Paragraph position="8"> We extend our model by introducing bunch predictions which enable us to predict a set of syntactic categories simultaneously. In the following subsection, we see that this kind of prediction is useful for handling coordinate conjunctions, too.</Paragraph>
      <Paragraph position="9"> Various kinds of syntactic units can follow the verb be in a verb phrase and we cam~ot selectively predict one of these possibilities when we are reading the words, such as am or were, etc. The bunch predictions we introduce enable us to cope with this kind of predictions.</Paragraph>
      <Paragraph position="10"> The following rule shows how to write a bunch prediction ill a rule.</Paragraph>
      <Paragraph position="11"> (be flit) (VP fat) ~ t \[bunch (NP) (ADJ-) ((VP ing))\] *VP_MOD When a bnnch prediction is pushed onto tile stack, it works as if it were a single prediction until it becomes the top of tile stack, and one of the constitnent of the bunch prediction is, then, chosen to be appropriate according to the word encountered.</Paragraph>
    </Section>
    <Section position="3" start_page="470" end_page="471" type="sub_section">
      <SectionTitle>
2.4. And Stack
</SectionTitle>
      <Paragraph position="0"> In this subsection, we introduce another stack called the and stack to handle coordinate conjunctions. The method described here resembles that in \[5\] or \[6\], but with the and stack we can handle them quite simply.</Paragraph>
      <Paragraph position="1"> The appearance of coordinate conjunctions are usually net predictable and it triggers a new kind of operation. Let us consider the following sentences.</Paragraph>
      <Paragraph position="2">  (1) Mary had a little lamb and a kitten.</Paragraph>
      <Paragraph position="3"> (2) Mary had a little lamb and washed him every day. (3) Mary had a little lamb and she was always with him.  Conventional phrase structm'e grammar rules like: S ~ S and S are not directly useful for predictive recognition of the sentences. The structure that follows and depends not on the word itself but on the proceeding syntactic units being constructed. In the above sentences, a noun phrase, a verb phrase, and a clause are being constructed before the word, and each of these categories reappears in each of the three sentences, respectively. By using the and stack, we can easily recognize these structures. Figure 2-2 shows the relationship between the prediction stack (the stack that holds predictions) and the and stack where unnecessary details are omitted. At stage (ii), the first prediction is replaced by two predictions NP and VP only by looking the first word Mary. The lower element of the and stack is dmnged to (VP S), which shows that while the VP of the prediction stack is being processed, we are constructing both VP and S. In the same way, the stacks change their states as shown in the figure and a list (NP</Paragraph>
      <Paragraph position="5"> word and. The only thing we have to do is that we make a bunch pl'ediction \[bunch (NP) (VP) (S)\] and replace *NP-N by the bunch prediction. By looking at the words that follow we can choose one of the constituent predictions of the bunch prediction and process the rest of the sentence.</Paragraph>
      <Paragraph position="6"> Note that the following sentence can also be i'eeognized by this strategy: Mary lookedJbr and Jound the unicorn.</Paragraph>
      <Paragraph position="7"> A list (NP VP S) has been built when we encounter tile conjunction, and VP is used to capture the structure of the rest of the sentence. null The following rule description is used to trigger the above explained operation: and ?P -, t (special and_stack) where ?P indicates that applicability of the rule does net depend on the current prediction.</Paragraph>
      <Paragraph position="8">  Some kinds of words trigger insertive sm~ctures which are usually not predicted, and cause a kind of suspension of construction of structures being built. Some adverbs, prepositional phrases and adverbial phrases and clauses are such structures. \]{ere are three examples, where we use a pair of quotes to distinguish insertire structures.</Paragraph>
      <Paragraph position="9"> (1) There are economic risks and &amp;quot;generally&amp;quot; a lack of available data. (2) He adapted &amp;quot;for linguists&amp;quot; an existing system of formalization. In order to express insertive sta'uctures, we use the following notation.</Paragraph>
      <Paragraph position="11"> These rules are applicable for ahnost all old predictions provided that the current word belongs to the CW part of the rules. In this case, however, the top element of the prediction stack will not be popped. The new prediction(s), if they exist, will be pushed onto the current prediction.</Paragraph>
      <Paragraph position="12"> For example, (2) will be processed as shown in Figure 2-3.</Paragraph>
      <Paragraph position="13"> At first, the object noun phrase, NP, of the verb &amp;quot;adapted&amp;quot; is predicted. The rule (A-2) is then applied and the recognition of NP is suspended until the prepositional phrase is recognized by the prediction PP.</Paragraph>
      <Paragraph position="14"> He for .for the the  2.6. NP stack We introduce yet another stack, the NP stack, to handle struc. tures where a noun phrase is missing, e.g. relative clauses, This approach is widely used, e.g. in \[7\]. The fact is that people do not handle these structures in a totally different way comparing with normal clauses. It seems that when we encounter a relative pro. noun, we push a noun phrase onto a kind of stack, which we call the NP stack, and pop it when it is needed to fill out the gap afterwards. null The following rule is used to simulate the above operations. rel pro ADJ CLAUSE ~ nil +S The prefix '+' of the new prediction indicates that we push a noun phrase onto the NP stack.</Paragraph>
      <Paragraph position="15"> 2.7. Looking ahead and Preference In this subsection we discuss necessity of looking ahead and preference among syntactic ambiguities that affect the predictive recognition process.</Paragraph>
      <Paragraph position="16"> Some sort of lookahead facility is necessary to reflect the delay in making syntactic structure of sentences. In sharp contrast with Marcus's deterministic parser \[2\], we only make use of a word as the unit of lookahead.</Paragraph>
      <Paragraph position="17"> In the middle of a sentence we usually do not look back to see what the preceding structure was in order to build up a dominating structure. In Marcus's parser, however, we can make a rule like: &amp;quot;IF the first element is NP and the second element is VP, THEN let NP and VP be sons of S,&amp;quot; where NP which was recognized some time before is referenced again. This framework seems to be too strong as a simulation of our internal process. The approach taken in this research is to permit more appropriate and generalized predictions as described in the previous subsections. In our experiment, we make use of lookahead by permitting backtracking within a limited range: once the analyzer reached the n-th word, it would not cancel the previous decision maOe when it was processing (n-k)-th word, where k is the length of lookahead. The necessary length of the lookahead is investigated in the experiment. null Currently, preference factors are treated in the following manner. The syntactic categories a word belongs to are linearly ordered. Grammer rules are divided into two groups, usual and unusual: the rules that trigger insertive structures with some other uncommon rules are included in the latter group. Although the strategies are not fixed, generally we try each syntactic category one by one according to the order induced, and the usual rules are tried before unusual ones.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="471" end_page="471" type="metho">
    <SectionTitle>
3. Experiment
</SectionTitle>
    <Paragraph position="0"> The mechanisms described in the previous section were tested by analyzing two kinds of articles. The articles used in the experiment were a manual of a computer software and an abstract article on world economics. At first, basic grammar rules were written and they were revised and reinforced by looking at the result of the previous analysis.</Paragraph>
    <Paragraph position="1"> The output of the analyzer is a kind of tree structure as shown in Figure 3-1.</Paragraph>
    <Paragraph position="3"> Finally foun the old man and woman w\[t~cope Figure 3-1. Tree structure constructed by the analyzer. Structures that are captured by optional predictions and predictions J made through the and stack or insertive rules are called pending structures. In the example, (1), (2) and (3) are pending structures: (i) is recognized as an insertive structure; (2) is captured tlu.ough the and stack; and (3) is captured by an optional prediction. As shown in the figure, we temporarily attach them to the preceding predictions. In this representation, the word woman is modified by the prepositional phrase with the telescope. We, however, can easily obtain other plausible sentence structures. For example, if we attach (3) to VP, we get a tree structure where the prepositional phrase modifies the verb phrase. In our experiment, a sentence is said to be successfully recognized if we can get an appropriate tree structure by moving pending structures (if necessary).</Paragraph>
    <Paragraph position="4"> The success rate and its relation with the length of lookahead was as follows. Of the 85 sentences from each article, 65 (manual) and 70 (abstract) of them were analyzed as desired by making use of looking two words ahead, the current and the next word, while only one additional success was reported on each article by looking one more word ahead.</Paragraph>
  </Section>
class="xml-element"></Paper>