1994: Annotated Bibliography
Adele Goldberg
Abeille' Anne and Daniele Godard. French Word Order and the Word Phrase Distinction. Talk at Stanford Jan 10, 1997
Note polysemy is like the English ditransitive:
(1) Paul ache\`te tout a\` son fils

Paul brings everything to his son
Head adjacency of bare complements (lex+ complements):
Although adverbs can intervene between H and Bare Complements
(other complements have much freer word order)
H$<$ FLOAT-Quant$<$ Quant $<$ Ppart $<$ Vinf $<$ Bare N
If modified or conjoined, bare complements can appear after phrasal ones
She argued there was a generalization to Adjectives: prenominal adj can be postposed when modified or conjoined
Distinguished:
VP adverbs: manner adverbs, pas, plus
V adverbs: bien, beaucoup
Alternative explanation of adjectives: Head of modifier must be next to the head of phrase (Maling, Williams). [but tall enough of a basketball player violates this head-head constraint]


Ackerman, Farrell. 1987. Miscreant Morphemes: Phrasal Predicates in Ugric. UC Berkeley dissertation. Chapter 6.
Discussion of Sapir vs. Kroeber
FA proposes a hierarchy: XCOMP $>$ OBJ $>$ SUBJ, and suggests that incorporation will operate on the highest available GF on this hierarchy. (1987: 294). (XCOMPS here are supposed to be secondary predicates and directional/locational complements). Some exceptions to this generalization are also discussed


Akhtar, Nameera and Michael Tomasello. 1997 Young children's productivity with word order and verb morphology. Developmental Psychology 33 (6) 952-965.
Children do not begin with innate abstract categories or schemas.
Until 24 months, Tomasello's daughter tended to use verbs only conservatively (in the argument structures and morphology with which she had heard them used).
Cites paper by Olguin and Tomasello 1993: children's spontaneous utterances tended to replicate the surface structures with which they heard the verbs modeled.
[seems to assume patient as subject is incorrect (but it's only incorrect in transitive sentences: fine for intrans)]
Children rarely, on the other hand, make word order errors in production. But this would be expected since the children have heard those particular words before, so they do know where the putter and puttee go. (T's Verb Island Hypothesis).
They WERE productive with novel object labels (using them as patients when they had heard them as agents) and with plural morphology.
10 children were subjects 2;9 to 3;8. Four novel transitive actions were used.
Novel verbs were taught in one of 4 ways: 1) No arguments (Dacking!), 2) Agent argument only (Barney's chamming!), 3) Patient argument only (Dacking Cookie Monster) and 4) both agent and patient arguments (Barney's pudding Cookie Monster).
Methodology (!): each child was seen 8 times: 2/3 sessions/week. The experimenter also visited the children in their classrooms several times before and between sessions to become and remain familiar to the child.
MLU was determined by coding child's speech in 30-45 minutes of free play.
Each novel verb was modelled 10 times. The action was performed by the experimenter, the child or the two together.
Tried to elicit verbs by: “What's gonna happen how? What happened?”
Testing comprehension: “Can you make Mickie Mouse dack Ernie? Make MM dack E.”
Testing production: “What did Cookie Monster just do to Big Bird? What did he do?”
O\&T's finding was replicated: children expressed args in a given position significantly more often if they had heard an argument expressed in that position.
[additional arguments that were pronouns were NOT counted: because the referent was ambiguous and because “some pronouns are themselves marked for case in English.”
But 7 of the children used pronouns to mark arguments they had not heard marked]
There was one child who did consistently mark both Agent and Patient when she hadn't heard them before, but ALL of the other children either failed to code them, or coded them incorrectly half the time.
Children were more productive with -ing than -ed. T offers interesting motivating factors
for that.
Table 2: five children performed at chance in comprehension part (these had the lowest MLU): they did not productively use argument slots either.
The second study tried to encourage children to use full NPs and name arguments for verbs taught in the No Argument condition.
Children were taught to produce Agent V Patient structures with a familiar verb and one novel verb: “no, say it like this, ‘Cookie Monster's pushing Ernie.’ ”
For younger children (2-3), there was no difference in the number of correct and incorrect markings: there were very few of either.
One older child made 16 correct markings and 7 errors; one never used any characters names with verbs taught as No Argument; 8 older subjects marked consistently correctly
any arguments they expressed with No Argument verb.
All 10 of the older children (3 and up) were significantly better than chance at comprehension of arguments with verbs that were taught in the No Argument condition.
“The one relevant study that did employ novel verbs (Naigles 1990) unfortunately did not employ reversible sentences (see O\&T, Pinker 1994a for alternative explanations of this study).” pg 20

“...it is not until some time between 2;9 and 3;6 that children begin to demonstrate understanding and use of word order as a productive syntactic device.” pg 20
Final study contrasted -ed with -ing.
Matt Rispoli's 1991 Mosaic Nature of Grammatical Relations. is cited. [Bill Morris should read]
[Naigles reviewed (and rejected this article) on the grounds that the verb meanings taught involved indirect causation: therefore the transitive construction was not wholly appropriate!]


Our next speaker at the PDP/NLP series: Joe Allen: Univ. of Southern California Mar 11th. 1997
ABSTRACT
Argument Structures without Lexical Entries
Understanding how multiple sources of probabilistic information are combined during language acquisition is an important part of current research into language development such as speech segmentation and the discovery of grammatical categories (Kelly 1992; Jusczyk 1993; Saffran, Newport and Aslin in press, Morgan and Demuth 1996). Similarly, the capacity to represent information about the probabilistic co-occurrence of lexical, grammatical, semantic and pragmatic elements is a crucial component of several current models of sentence comprehension (McDonald, Pearlmutter and Seidenberg 1994; Trueswell, Tanenhaus and Kello 1993). The work presented here shows how probability sensitive mechanisms relevant to processing verb argument structures may also govern the acquisition of this knowledge, providing an additional link between the mechanisms of processing and those of acquisition.
I will be presenting results from a connectionist model of verb argument structure assignment that relies on probabilistic information regarding particular verbs, as well as semantic information shared among several verbs. The model illustrates how co-occurrence relations among different levels of linguistic representation may interact in acquisition by contributing to the discovery of language specific well formedness constraints and coherent semantic subclasses of verbs. Both semantic and observed conditions of use contribute to the formation of semantic subclasses, which in turn determine the set of constructions in which particular verbs appear.
Trained on utterances derived from the CHILDES database, the model shows syntactic boostrapping behavior, in that a verb's emergent semantic representation is partly a function of the utterance types in which that verb was used during training. The model is able both to
assign appropriate thematic roles to arguments of utterances it has not been exposed to and to reject utterances for which no coherent assignment can be made. For example, the model assigns an appropriate thematic role to the second post verbal NP in He kicked me the ball,
although the model has never seen the verb kick used this way before. The model also correctly fails to assign thematic roles to arguments in unacceptable configurations such as *The paper cut. Constraints against assigning thematic roles in certain constructions arise from the interaction of lexical and semantic cues to the role assignment problem.
The acquisition task as modeled here is the discovery of a set of lexical, grammatical and semantic cues that provide computational solutions to the language comprehension problem. Understanding the interaction among the cues relevant to the role assignment problem provides insight into why languages might exhibit coherent semantic subclasses such as those detailed in Levin (1993).
*******
FROM MARYELLEN"S TALK at CMU:
Joe Allen's network determined that productivity depends on:

(1) semantics of the verb
(2) frequency of the verb
(3) frequency of the verb in this construction
(4) number of different constructions and frequencies of those constructions
(5) set of events of the verb (?)
(6) semantic neighborhood of the verb
(7) set of constructions the other verbs appear in
(8) number and frequency of neighboring verbs
(9) combined frequency of verbs used in similar ways
Barsalou, Lawrence W. 1983. Ad hoc categories. Memory and Cognition. 11 (3) 211-227.
Ad hoc categories:
(1) a. things to sell at a garage sale

b. things to take on a camping trip

c. possible costumes to wear to a Halloween party

d. places to look for antique desks
Central properties of Common Categories
Graded structure: some instances are better examples of a category than others. (robin vs penguin : bird)
Well established category representation: taxonomically organized
Comparison-Network Model:
I. Similarity comparison process
Tversky (1977): two concepts become more similar as the number of properties shared by them increases and the number of distinctive (nonshared) properties decreases. [not really going to do anything]
assumptions:
1) as instances become more similar to a category concept, they become more typical of the category
2) as noninstances become less similar to a category concept, they become more typical of the category's complement [??]
3) the similarity of unclear cases to a category concept is close to the minimum amount necessary for category membership.
II. Spreading activation
Concepts and properties as nodes, associations between concepts and properties as pathways that carry spreading activation.
Strength of association varies continuously and increases as a function of how frequently and recently an association has been active in working memory. Can also be asymmetrical
Common categories have well-established concept-to-instance associations, and well established instance-to-concept associations.
How Ad Hoc categories relate to the Comparison-Network Model
If graded structure results from concepts bearing different amounts of similarity to a category concept and if concepts vary in similarity from the category concepts of ad hoc categories, then graded structure is expected.
Or, ad hoc categories could be represented as lists without internal structure.
Ad hoc categories by hypothesis do not have well-established category representations in memory (“people rarely, if ever, think of them.”)
Experiment 1:
(2) a. things to inventory at a department store

b. ways to make friends

c. things that conquerors take as plunder

d. ways to escape being killed by the Mafia

e. things that babies do

f. times to write a term paper

g. things that could fall on your head.
Subjects were given 8 sets of 6 items and asked whether each item belonged to some category and then to rank order all six items for how good a member each was of the category.
In 4 sets, 3 were from category and 3 weren't.
In other 4 sets, 2 were from category, 2 clearly weren't, and 2 were unclear.
Ways to Make Friends (3-3):
(3) a. join a car playing club

b. get convicted for murder

c. don’t take a bath more often than once a month

d. go back to school

e. have a garage sale

f. get convicted for robbery
Ways to Escape Being Killed by the Mafia (2-2-2)
(4) a. change your identity and move to the mountains of South America

b. move to the remote reaches of Wyoming

c. stay where you're presently living in Los Vegas

d. move to Reno

e. move to the mountains of Mexico

f. change where you're living in Las Vegas
Finding: subjects generally agree on the graded structure of these categories, are uncertain where to draw the category boundary.
The model predicts that typicality should be very similar for ad hoc and common categories because he assumes 1) the same similarity comparison process constructs graded structure for both category types and 2) how well established in memory a category is should not affect the process.
Later (pg 224) B notes that ad hoc categories are not related by averages of their category instances [in any sense], but instead are structured by dimensions that are relevant to the goals that are relevant to the category.
The model does predict that subjects should show less agreement when generating category members for ad hoc than for common categories since common categories have more well established concept to instance associations to their typical instances. Subject's search of ad hoc categories should be more random, showing less agreement.
In fact, subjects were found to agree on typicality ratings the same for both types of categories. “This suggests that subjects use the same similarity comparison process to construct graded structure for both category types.” pg 218.
If instance to concept associations are not as well established for ad hoc categories, then subjects should have trouble labeling the categories without context. If contexts prime the category concepts, then the concept should be easier to identify.
Finding: subjects generated concepts for common categories 100% of the time; for ad hoc categories without context, 87% of the time, and for ad hoc categories with context, 97% of the time (subjects were variable in what concepts they discovered).

“It may be optimal that perceiving an entity does not activate all the ad hoc categories to
which it belongs. Seeing a chair and having categories such as ‘emergency firewood,’ ‘fits in the trunk of a car,’ ‘used to prop doors open’ come to mind would be highly distracting...” 223.
Interestingly, subjects who were given contexts were more variable in their categorizations of common category item sets than if they were given no context. Category such as ``fruit" may often be incorporated into ad hoc categories relevant to current purposes (fruit for desert, fruit to pick from trees, fruit available in the fall).
Barselou also notes that what began as ad hoc categories become processed more frequently and begin to become common categories. After a number of garage sales, or after a number of camping trips, things to sell at a garage sale or things to take on a camping
trip become well-established categories.
Bates, Elizabeth and Brian MacWhinney. 1982. Functionalist Approaches to Grammar. In Eric Wanner and Lila R. Gleitman (eds.) Language Acquisition: The State of the Art. New York: Cambridge University Press. 175-218.
Nice paper to read for course on Functionalism/Formalism.