<?xml version="1.0" standalone="yes"?>
<Paper uid="C82-2009">
  <Title>DEVELOPING THE COMMENTATOR, A COMPUTER SYSTEM SIMULATING VERBAL PRODUCTION</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
DEVELOPING THE COMMENTATOR, A COMPUTER SYSTEM SIMULATING
VERBAL PRODUCTION
</SectionTitle>
    <Paragraph position="0"> The project &amp;quot;COMMENTATOR&amp;quot; at the department of general linguistics at the university of Lund is intended to test ideas about language production. The system implemented in BASIC on the ABC80 micro-computer generates a scene on the monitor where two persons, Adam and Eve, move randomly around a gate. Not only the present positions of Adam and Eve are shown on the screen but even the positions before the last Jump. This setting is also used for eliciting comments from human subjects. The moves are generated randomly but the operator can choose the length of the Jumps. The initial placement of Adam and Eve can be determined by the operator, as well as the instructions for the machine concerning the focus of attention (Adam or Eve) and the primary goal of the actors (the gate or the other actor). On the operator's command the computer produces a set of written comments on the development of the scene. COM~ENTATOR is a research tool, and does not use any ready-made sentences describing forseable situations.</Paragraph>
    <Paragraph position="1"> The system works as follows: From the primary information (the coordinates of the gate and the two actors) some more complex values are derived (distances, relations &amp;quot;to left&amp;quot;, &amp;quot;to right&amp;quot; etc.). The conditions for using abstract predicates equivalent to &amp;quot;to left&amp;quot; etc. in the given situation are tested accoMin 6 to a question menu. This results in positive or negative abstract sentences. The abstract sentence constituents - 42 are ordered as sub.leers, predicates , and oh,leers. Connective elements, such as &amp;quot;also&amp;quot;, &amp;quot;either&amp;quot;, &amp;quot;yet&amp;quot;, are added If possible. These connect the last proposition to the previous ones.</Paragraph>
    <Paragraph position="2"> Proper namesp pronouns, or other NPs are chosen on the basis of reference relations to the preceding proposition. The abstract propositions are substituted by surface phrases and words. The assembled structure is printed. (For a more extensive description of the program and one version of t~e prepare itself see Sigurd 1980.) The text produced by COMMENTATOR may look like this: Eva ~r till h~ger om Adam. (Eve is to the right of Adam.) Hen ar till vanster om henne. (He Is to the left of her.) Hen ~r till vanster om porten cokeS. (He is to ~he left of the gate, too.) Hen n~zmar slg den. (He is approaching it.) Hen n~u~na~ sig Eva coke&amp;. (He is approaching ~re, too.) Hen ~r n~rmast porten dock. (Lit. She is closest to the gate, however. ) Hen ~r inte n~ra den. (She is not close to it.) Adam ~r lnte n~ra den heller. (Adam is not close to it, either.)</Paragraph>
  </Section>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
COMMIE
</SectionTitle>
    <Paragraph position="0"> COMMIE is a semantically and psychollngulstlcally more advanced program, which is intended to overcome certain shortcomlngs experienced with the original COMMENTATOR. This more complex pro~am generates a more sophisticated and more dynamic stimulus including more objects and persons on the scene and the vocabulary is about ten times larger. COMMIE is to produce human-like texts. One important task is to avoid unnecessary redundancies. The commentaries produced have to be based on relevant changes of the scene and they must avoid repetitions, including information implied by previous utterances. (g.g., &amp;quot;Adam is to the left of Eve&amp;quot; implies &amp;quot;Eve is to the right of Adam&amp;quot;, etc.) Not all propositions that ere not true in the given.moment deserve to be mentioned, either. Negated state- 43 ments are meaningful only as answers to expectations evoked by the pre~ou8 context. (E.g., &amp;quot;Bertil is chasing Eve. He is not ver~ near her yet, however.&amp;quot; but not l&amp;quot;Bertil is chasin K Eve. Neither of them is in the .upper riEht corner&amp;quot;.) All information about the world of the ~V screen is stored in a two-d~nensionePS array as mi~L-thouKhte - two-place predications that are either true or false at the given moment.</Paragraph>
    <Paragraph position="1"> Yorgettln8 is s~nulated by 8radu~lly letting the contents of the array pass a horizon of forgetting. In a future version the basic contents may be stored in &amp; lon8 term memory. The p~edPScations of the array serve 88 inputs for generati~ complex semantic structu~res, which are inputs to the production of commentaries proper. Referring in OOU~IE is to be based on the theory of FSP. By computiz~ &amp;quot;de~rees of ~Lveness&amp;quot;, it shoUld be able to cope with even such problems as the difference between thematic pronominal vs. thematic zero subjects in Sla3~o langusKes generally or the enclitio vs. full forms of oblique pronominal cases in West Slavic le~gue~es etc. (Cfe Bily 1981a, Chapter Three.) The deep structures are to be &amp;quot;univers8~&amp;quot; enough to allow generation of comments~es in several' lan~ae~es.</Paragraph>
    <Paragraph position="2"> COMMIE will also have more sophisticated conditions for the use of the two-place predicates. Instead of, e.g., purely geometrically defined conditions for &amp;quot;X is to the left of Y&amp;quot; to be true, a certain limit is put on the allowed vertical distance. The limit seems to be a function of the size of X and Y=, the size of the referential frame and the focus of attention. (Cf. Bily 1981b) Other versions in preparation A version produclnE spoken comments using VOTRAX speech synthesizer or more advanced technique ie being planned. As the COM~NTATOR has perfect control of the steps in the produation of comments, it should be able to produce better prosody than ordinary text-to-speech systems (e.g. Carlson et 81.</Paragraph>
    <Paragraph position="3"> - 44 1981). Contractive accents can be derived as the system knows if a comment is in contrast with a precidAng statement, Accents based on syntactic structure can be derived Am the syntactic structure of the comments if prefectly known, A system which can comment on asp arbitrary scene where persons and things appear, 8ctions and events take p~ace, would demand pattern recognition, which is a difficult problem. 0nly a smeA1 scale system is being planned. It requires a flexible question menu, where &amp;quot;relevant questions are produced successively, Beginning comments on such a scene would be something like this: &amp;quot;There is something at the left upper corner. Zt moves. Now something turns up at the right upper corner . It looks like a ship. Both objects move fast towards the centre.&amp;quot; Such comments suggest applications as the automatio radar 0ommen~ator, but so far such practical applAoations have been little considered. It is, however, envisaged that verbal comments on events produced by computers will be oonunon in the future world of robots.</Paragraph>
  </Section>
class="xml-element"></Paper>