<?xml version="1.0" standalone="yes"?>
<Paper uid="C92-4179">
  <Title>Ward, Nigel (to appear). A Connectionist Language Generator. Ablex. revised and extended version of A Flexil)le, Parallel Model of Natural Language Generation, Ph D. thems and Technical Report UCB</Title>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4. Implementation
</SectionTitle>
    <Paragraph position="0"> 1 have built a parser (Ward 1992) and a generalor (Ward to appear) which use participatory profiles. This section discusses the generator, not as a presentation of the best or only way to use profiles, but merely as a demoastration that case profiles are workable.</Paragraph>
    <Paragraph position="1"> FIG, a 'Flexible Incremental Generator', produces English and Japanese sentences starting from a meaning representation, using spreading activation in a knowledge net, work. One task of a generator is, given an input including some items with case profiles, to build a sentence whose syntactic form and function words reflect those e~e profiles.</Paragraph>
    <Paragraph position="2"> In FIG case features are implemented ms nodes in the associative network. They are linked to constructions and words, with appropriate weights. For example, the node responsible,, has a link to the node by,,, representing the word &amp;quot;by&amp;quot; , and this link fias weight +1.</Paragraph>
    <Paragraph position="3"> The participatory profiles of concepts in the inpnt are represented by links to nodes for case fi~atures, appropriately weighted. For example, the node for Mary may have a link with weight .5 to responsible, to represent a given mput.</Paragraph>
    <Paragraph position="4"> For such an input, when mary,, becomes activated, case features will become activated to the de gree appropriate for her profile. In tnrn byw and other prepositions will receive actiw~tion from these case features. The net effect is that the profile for a participant activates prepositions proportionally to their proximity m case space to that profile. (The measure of proximity computed is, to be precise, the dot product of the vector for the participant and the vector for the prototype.) The preposition whose pro totype is closest will receive the most activatiou, mid hence appear in the output. Like ease markers, coltstructknls receive activation from the profiles of participants, via case features. They thus become mobi lized to the extent that there is a participant with a profile matdring that of the construction. (Some case markers appear before the word they flag, others after, and so FIG has a distinction between activation fi'oni the profiles of concepts which remain to be expressed and activation from the profile of the concept just expressed.) Constituents which involve profiles also are linked to nodes for case features; from these activation flows to concepts, and so the concept whose participatory profile is closest to that activated by a constituent will receive the most activation. (Actually the case feature nodes used for activation flow from constituents to concepts are distinct from those used for activation flow fronl concelpts to ca.se markers and constructions. That is, each case feature is implemented a.s a pair of nodes; this is for technical reasons.) There are multipie profiles in any non-trivial conceptualization, and ACRES DE COL!NG-92, NANTES, 23-28 nOLq&amp;quot; 1992 1 1 4 0 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 it wouht seem that crosstalk aright be a problem, but this has not been the case in F\[G, primarily hecause generally there is one eoltsLruction an(l one (:oncepi with enough actiwttion to dominate.</Paragraph>
    <Paragraph position="5"> FI('~ originally expected deep (:as(: relations in its int)uts , and its grmnrnar and lexicon referred to those cases. One problem was that, ~s I extended 1,'lG's coverage of the two langllages, the nnlnher O\[ cil~ses kept growing and the grammar got uglier and uglier.</Paragraph>
    <Paragraph position="6"> In t)articular, there were lengthening lists of possible cases for constituents, for example there was at llst of fonr possible cruses to nse for subject. Switching to proliles solved these problems. Conversion wa.s relatively easy; other than the new references to profiles, the grammar did not need to be changed. Tim additional eomi)utation required is negligible.</Paragraph>
    <Paragraph position="7"> FIG currently uses 10 (:~use fi~atures: volitional, responsible, active, aflheted, direet-cause, partial cause, individuated, topic, object of-force, and touched; these replace the e~Lses agent, instrument, patient, experieneer, cause and percept. At this point the meanings of the cause features derive less from their flames than from the way they are related to the con structions of Jat)anese and English. This is hecause tbe numeric values for the t)rofiles, although originally chosen according to comlrlon sense arid with reference to the literature, tlaw~ had to i)e tuned in the course of making FIG able to generate sentences in both Inngnages for a largish mmlber of inputs. I ascribe no special significance to the particular set of profiles enrrentty in nse: they are specific to FIG's current gramtnar and implementation details.</Paragraph>
  </Section>
class="xml-element"></Paper>