Book review for Language, Vol. 72, No. 2, 1966, pp. 417--421:
Head-driven Phrase Structure Grammar. By Carl Pollard and Ivan A. Sag. Chicago & London: The University of Chicago Press. 1994. Pp. xi, 440.
Reviewed by Takao Gunji, Osaka University*
This book was originally intended as to follow Information-based syntax and semantics volume 1: Fundamentals (Pollardand Sag 1987), which laid out the overall organization of the theory of Head-driven Phrase Structure Grammar (HPSG). Detailed analyses of specific languages such as English were meant to be presented in the second volume. In fact, the theory has evolved to a somewhat different form in the current book, and I will distinguish them as HPSG-I and HPSG-II, respectively, in this review.
HPSG is developed as a nonderivational grammatical framework, i.e., a constraint-based grammar, with a highly hierarchical organization of the features. With regard to feature organization, one of the most prominent aspects of HPSG-II, as compared with HPSG-I, is its emphasis on a closer integration of semantic and syntactic feature structure. In particular, while the LOCAL feature in HPSG-I was located under SYNTAX, which entails that any local information is strictly syntactic, in HPSG-II, the CONTENT (semantic) feature is located under LOCAL, making it possible to treat semantic information also as local. Furthermore, the BINDING feature in HPSG-I, which was located under SYNTAX, is renamed as the NONLOCAL feature in HPSG-II and appears outside of the CATEGORY (syntactic) feature. In this way, what appears under LOCAL in HPSG-II is an integration of both syntactic and semantic (as well as pragmatic (CONTEXT)) information. Furthermore, the LOCAL feature is bundled together with the NONLOCAL feature under the new SYNSEM feature. This contrast is graphically represented below:
HPSG-I: + PHONOLOGY + | SYNTAX + LOCAL + HEAD ++| | | + SUBCAT +|| | + BINDING +|| sign+ SEMANTICS +
HPSG-II: + PHONOLOGY + | SYNSEM + LOCAL + CATEGORY + HEAD + ++| | | | + SUBCAT + ||| | | | CONTENT + INDEX +||| | | | + RESTRICTION +||| | | + CONTEXT +|| sign+ + NONLOCAL ++
Now that the value of SYNSEM contains both syntactic and semantic information but excludes phonological information, the value of SUBCAT (for subcategorization) can now be specified more appropriately as a list of SYNSEM information, as phonological information is usually not used for subcategorization. In short, the new bipartite geometry makes it clear that subcategorization is concerned only with syntax and semantics.
Based on the revised feature geometry, the authors discuss specific phenomena in natural language (mostly English), ranging from agreement, complementation, unbounded dependencies, relativization, binding, and control to quantification in Chs. 2--8. As the book was written in the late 1980s and early 1990s, P&S carefully compare their theory with the representative analyses then, mostly in government-binding (GB) theory.
As it is not possible to discuss here every innovative proposal in this book, I would like to highlight only a few. Concerning agreement, P&S make a distinction among three types of agreement: index agreement, syntactic agreement, and pragmatic agreement. Among these, index agreement is based on semantic, rather than syntactic, concepts. One of the important proposals in HPSG-II is that the INDEX feature, which has information about person, number, and gender, among others, is located under CONTENT (cf. above) and is thus treated as part of semantics. Moreover, since the architecture of feature-based theory allows underspecification, agreement can be made nondirectional. These characterizations of agreement solve many mysteries and eliminates redundancies (e.g., German case concord) in directional theories of agreement.
Even though the use of indices (in general) has been common in generative grammar, few, if any, explicit characterizations have been given. In HPSG, the index is defined as the value of the INDEX feature; this value is much more restrictive than the value of SYNSEM as is evident in the feature geometry shown above. This distinction is used in expressing the difference between traditional `raising' and `EQUI' types of verbs: the former involves sharing the value of SYNSEM, while the latter only the value of INDEX, excluding syntactic information such as case. This is typically how HPSG eliminates the empty categories hypothesized in GB; NP trace and PRO are eliminated in this case.
The chapter on control also clarifies what has been mentioned but never made explicit in previous studies in GB. The control theory presented in Ch. 7 is very specific; it proposes toe divide verbs into three types according to their semantic nature: influence, commitment, and orientation types. Verbs of the influence type such as persuade show object control, formally defined as coindexing (in the above sense) between the bearer of the influenced role (typically the direct object) and the subject of the complement VP. Similarly, the bearer of the committor (typically the subject) is involved in coindexing for verbs of the commitment type such as promise, and that of experiencer (typically the subject) for the orientation type such as want. As the control theory is stated in terms of semantics rather than syntax, whether the controller is realized as the subject or the object is contingent on the subcategorization properties of the respective types of verbs. This theory of control, coupled with P&S's analyses of passivization and binding, correctly predicts and integrates various generalizations that have been presented in the literature.
Another major innovation is P&S's treatment of binding. While pointing out inadequacies of an account that relies on such configurational concepts as c-command in GB, P&S provide a nonconfigurational theory of binding. Perhaps as a service for readers familiar with configurational analyses, they formalize their theory in terms of obliqueness command (o-command) and state it in the form of the familiar three principles of GB. Even though the wording is deliberately similar, the concept of o-command is defined based on the obliqueness order among coarguments, and this theory dispenses with any configurational concept. Moreover, P&S give persuasive arguments to prove the superiority of obliqueness-based analysis of binding.
What is common in P&S's treatment of agreement, control, and binding, as just seen, is their sense of balance between syntax and nsemantics. In presenting their theories, they not only criticize purely syntactic analyses of these phenomena but also point out inadequacies of purely semantic treatments. The status of the INDEX feature typically exhibits this subtle decision. Even though the value of INDEX is part of semantic information, it is not the semantic referent itself; it is an abstract grammatical entity with its own raison d'être. For example, the interaction between the value of INDEX and the elements in SUBCAT in agreement and binding cannot be adequately explained in purely referential theory. As the name of their first book indicates, and the feature SYNSEM in HPSG-II symbolizes, HPSG is aimed at an integral theory of syntax and semantics, this is quite persuasively demonstrated in their actual analyses of representative phenomena.
Some of the contents in the book have been further modified by the researchers in HPSG during the course of writing the earlier chapters, resulting in the emergence of what is called HPSG-III. The last chapter, Ch. 9, reflects this transition. It is dedicated to discussion of many possible revisions of the theory presented in the preceding chapters and can be regarded as a point of departure for a further revision of the theory of HPSG. Specifically, P&S propose to make a drastic change in the status of SUBCAT. Adopting the proposal by Borsley (1987), separate lists are now assigned for the subject and other arguments; the subject resides in the SUBJ list and others in the COMPS list. This rearrangement of the argument lists helps make the treatment of the idiosyncratic behaviors of the subject in many languages straightforward. Moreover, various possibilities of combinations of values of SUBJ and COMPS open the way to treatments of unaccusativity and other related phenomena, as briefly mentioned in a footnote (379).
The SUBCAT feature was first introduced by GPSG (Gazdar et al. 1985) as a classification feature (subcategorization in its original sense) of lexical categories. Then, in HPSG (I and II), it was made to incorporate information about whether a phrase is saturated or not, which was partially expressed by the bar feature in GPSG. Now, in HPSG-III, the subcategorization information, which is purely lexical, and the constituency information, which is phrasal, are again separated; the former is encoded in SUBCAT (now taken to be a lexical feature),1 while the latter is in SUBJ and COMPS.
In my view, the new organization of valence features may have further consequences for linearization. HPSG follows the tradition first proposed in GPSG in treating the information about word order in a component separate from that for constituency. Thus, the schemata that define the phrase structure (ID schemata) are essentially order-free and determine only immediate dominance. In this respect, unlike traditional phrase structure approach, including various versions of X'-theory, the constituency determined by ID-schemata doesn't define the local order among constituents; the valence features (SUBJ and COMPS) only determine constituency and are not responsible for linear order. Explanation of word order variation in such languages as German, Dutch, and Japanese would, then, become straightforward in a purely nonderivational framework, if the idea of linearization proposed by Dowty (1996) and Reape (1996) is adopted, as exhibited by Kathol (1995), among others. In such a treatment, the terminal yield of a phrase structure tree is no longer assumed to define the phonological shape of the sentence the tree is to represent.
Even though it may sound contradictory to some that a grammatical framework allowing this kind of view toward language has `phrase structure' in its name, P&S's book, particularly the final chapter, gives one a chance to appreciate the flexibility of constraint-based grammar.