Results of our study indicate that a dog may process independent items of verbal information provided in a single request and use them to organize sequentially her behavior. We showed that a complex utterance, made up by the combination of previously mastered terms, could control the dog's behavior such that a specified action would be performed on a specified object.
Previous work on language understanding by dogs has mostly centered on dogs' comprehension of verbal commands, taken as single communicative events and referring to whole situations and are not-syntactical in their form. In word recognition studies in which dogs are asked to retrieve one out of a set of objects, they, of course, both discriminate objects based on their labels and act in an appropriate way towards them (e.g. fetching), sometimes integrating information about object location 
. Fetching or approach (“go to” commands) are however a constant feature of the task 
and it is not possible, through the results of such studies, to evaluate the process by which action and object terms are separately taken into account and integrated into a successful performance. Such evaluation is obtained when, as in the present study, action and object are varied independently from one another.
It is here relevant, as in other experiments on complex communication, to rule out Clever Hans explanations. Sofia's performance could not be influenced by experimenter-produced cues about the object to be approached and the behavior to be performed towards it: turning back, after requests were voiced, she lost visual contact with the experimenter and had to rely exclusively on words.
Performance was not restricted to the specific training context and generalized to novel conditions: she was able to obey requests when non-semantic variables (visual access to the eyes or mouth of the experimenter 
) were lacking; when requests were emitted by an unfamiliar person; when the spatial location of objects was changed, and when testing was done outside the laboratory. She generalized correct performance to objects in the same category, e.g., balls differing in size, shape and color and to a new object with a new label (bear). Such versatility suggests the existence of a capacity to extract and process relevant verbal features and a relative independence from contextual parameters.
Could Sofia's performance be accounted for by assuming that she learned, one by one, the correct responses to each object-action pairings of words in the whole set? Several aspects of our results make this assumption implausible: on one hand, the very quick and correct transition of performance from sequential to simultaneous requests which indicates that previously acquired responses tendencies were combined without or with little further training; on the other hand, and maybe more significantly, the maintenance of level of responding (at least with a small number of objects) when requests shifted from object-action to action-object. Under the hypothesis of separate learning of each combination, the reversal of items would be expected to decrease correct performance. Lack of improvement throughout test sessions also constitutes evidence against such hypothesis.
It is interesting to note that Sofia's performance did not reach a hundred per cent correct score in any session and also that there was a consistent decrease in performance, both in sequential and simultaneous object-action phases, as the number of objects used in a session increased (). Both aspects may point to some constraint in Sofia's performance, maybe a difficulty in the discrimination of the spoken labels of objects; or a problem of memory which eventually grew higher when the number of objects presented increased.
Learning rate was higher with action terms than with object ones, a difference also obtained in an unpublished study in which eighteen dogs, submitted to training procedures similar to those used in the present experiment, were all shown to be able to acquire correct responses to action commands but failed, most of them, to master labels-objects associations 
. This intriguing difference which deserves a more thorough examination, may relate to the history of dog breeding, during which many breeds were developed for herding, tracking, etc., cooperative tasks in which commands for action (not for object discrimination) prevail.
When presented with new combinations (stick fetch and bottle point) of previously mastered action and object terms, Sofia did not reach a successful level of responding. This result, based on a restricted number of trials, should be confirmed by new observations. It may (taking it at face value) derive from Sofia's previous exclusive training with stick point and bottle fetch requests, and might indicate the possible prevalence of a simpler, stimulus-bound way of reacting in cases of invariable non-combinatorial training conditions.
Sofia's processing of two-item sentences probably involves working memory processes, as occurs in human sentence comprehension 
. In label training of dogs, responses happen in close temporal relationship to requests but do require the keeping of some information in memory (about what object is to be retrieved and where it is located). Performance under simultaneous requests depends on the storing of a more complex information (about object and
action), the items of which must be put into use at appropriate stages of performance. Object-action requests are not obeyed by simply following the order in which terms were dispensed in the request sentence (object first, action second). One item of information is used first (which is the appropriate object?
); the second one is used when near the object (what is the appropriate action to be performed?
). It is thus conceivable that terms are stored in a parallel manner, independently of their order of reception, and are retrieved when certain environmental conditions are met. To retrieve information, Sofia uses her knowledge of the sequential structure of the task (an object must be approached before any action can be executed), and demonstrates understanding of the general principle that some actions require an object to be executed upon.
Sofia's prompt and successful performance for the new bear fetch and bear point requests gives a strong indication that, going beyond the learning of specific stimulus-response relationships, she was able to combine an action (selected among alternatives) to an object (selected among alternatives) even in the case of an object never before responded to with pointing or fetching. In an unpublished follow up experiment with Sofia, she was trained to choose either of two identical objects, placed at right or left in the experimental room and to perform either pointing or fetching towards it. Requests were thus action-action requests (turn right or left – point or fetch), not action-object ones. Sofia's performance in this task was highly successful and provided a confirmation of the dog's capacity to take into account and combine information items of a different nature.
Attention to the order of terms has been demonstrated in nonhuman species, in contexts in which the structural difference of sentences is relevant. Dolphins Akeakamai and Phoenix, for instance, when requested to take the ball to the hoop
, pushed the ball until it got near the hoop and did the opposite, when requested to take the hoop to the ball
. They were also able to learn different sequential grammars (S-V-O, Phoenix, O-V-S, Akeakamai) 
. In Sofia's case, order of terms did not differentiate performance: Fetch key
and key fetch
were equivalent. Such equivalence may derive from training conditions which did not take order of items as a parameter. Further research may reveal to what extent dogs are able to discriminate the placement of terms in multiple-item requests.
Our results suggest that dogs share with “linguistic” animals 
the capacity to encode in memory at least two heterogeneous items of information to be used in subsequent directed performance, a capacity which, although far from being “an infinite use of finite means” 
as human grammars are, may have comparative relevance as a forerunner to syntactical functioning.