adjunct status, unlike familiar voice alternations such as passive and anti-passive. In a language like Tagalog, there are several additional voice options (dative, instrumental, locative) but again these are all fully transitive; in none of them is the agent demoted to oblique or adjunct status.
Now obviously this is a problem for the original, classical form of the Lexical Mapping Theory (as developed by Bresnan & Kannerva 1989, Alsina & Mchombo 1993, among others). When I arrived at Stanford in the late 1980s, there were several different theories of linking being developed, and these facts seemed to be a problem for all of them. But for LFG, they were a problem ONLY for the linking theory. Once you allow a non-canonical linking pattern for these languages, the rest of the system functions pretty normally. Wayan Arka (2003) later developed a model of LMT, adapted from Alex Alsina’s model, that works for Balinese, and I am sure this could be further adapted for Tagalog. But having solved this problem does not force major changes to other aspects of the analysis.
I would like to contrast this with a very influential analysis of symmetric voice proposed within the Government and Binding framework by Guilfoyle, Hung and Travis (1992). They proposed that the agent is internal to VP at D- structure, specifically that it occupies [SPEC, VP]. Depending on the voice morphology of the verb, one argument will fail to get Case in its D-structure position and be forced to move into [SPEC, IP], the structural subject position. No theta roles get deleted or absorbed, so the agent is free to remain in [SPEC, VP] when it is not selected as subject. This is a very elegant model of non- demoting voice alternation. However, because the change of GFs is expressed in terms of phrase structure, it makes incorrect predictions about things like word order and long-distance extraction in Tagalog. When all information is represented in the same way, a change in any part of the system affects every part of the analysis. In contrast, the modular design of LFG allows us to address separately issues which are in fact independent of each other.
Wayan Arka (2003) also presented a fair bit of evidence for the claim that in Balinese Undergoer Voice clauses, in which the patient is the subject, the agent and verb form a very tight constituent which he labels V-bar. This is an interesting and typologically unusual claim. Similar claims have been made in other Western Malayo-Polynesian languages, e.g. Toba Batak (Schachter 1984) and West Coast Bajau (Miller 2007). In transformational theories that adopt the Uniform Theta Assignment Hypothesis (UTAH, Baker 1988), this hypothesis is virtually ruled out on theoretical grounds. Within LFG, however, because GFs and theta-roles are represented separately from constituent relations, it is simply an empirical issue: the analysis can follow the facts of the language.
Another thing that I appreciate about LFG is that there is a healthy respect for the degree to which languages may differ from each other. Universals that get