“I have always wished that my computer would be as easy to use as my telephone. My wish has come true. I no longer know how to use my telephone.” – Bjarne Stroustrop (originator of C++) As household appliances grow in complexity and sophistication, they
sentence to help constrain the matching problem. Finally, Precise generates the complete set of possible SQL interpretations of the sentence that are consistent with its lexicon and its parser.
Precise was originally built to support only questions, which translate to SELECT queries in SQL. Since an NLIA has to respond to requests for changing the state of the appliance, which translate naturally to SQL UPDATE statements, we had to extend Precise appropriately. There is no room to explain the technical details of Precise, but the essence is an extension of Precise’s graph matching algorithm to keep track of two attributes, a “pre-attribute” and a “post-attribute.” The pre-attributes contain values in the database state before an UPDATE, and the post-attributes contain the new, updated values. Any values that are set by an UPDATE statement are matched with post-attributes, and all other values are matched with pre-attributes. In order to understand an input sentence, the system needs to classify it as an UPDATE or a SELECT, and proceed appropriately. Precise’s other modules, including its tokenizer, lexicon, and parser, remain unchanged.
physical processes (in this case the process of a message being played) that evolve over time. We do not need to model the processes explicitly, since (by assumption) the networked device notifies the agent of events such as incoming messages. Thus execution of the play command does not terminate with the device playing; it simply stops later of its own accord. By modeling the device in this fashion, we are able to use the Blackbox planner [13], which operates by compiling PDDL planning problems into a set of propositional clauses that is satisfiable exactly when an n-step plan exists. A fast satisfiability algorithm is used; if an assignment is found, a reverse compilation phase generates the plan; otherwise, the plan length is incremented.
6. EXPERIMENTAL EVALUATION
To test our system, we gathered a total of 72 sentences from seven graduate students and faculty at the University of Washington. The people who provided us with data were given a concise list of the features available on the Panasonic phone and were asked to write down sentences in English that they might use to invoke those features.
Of the 72 sentences gathered, we used 61 in our experiments. The eleven sentences we excluded did not describe a goal or an action, but rather they implied it. For example, we excluded the sentence, “It’s too loud.” This sentence does not describe a goal state or an action to achieve a goal. Instead, it describes the current state and implies the goal. Such sentences clearly demonstrate the importance of processing speech acts, but are beyond the current capabilities of EXACT.

5.2 The Planning Component
We have already explained how we prevent unwarranted side effects and how we finesse the problem of information gathering. But the domain of household appliances also has a surprising amount of temporal complexity.
First, user commands may involve events occurring at specific future times (e.g., “record Star Trek”). While one could use a temporal planner to handle these goals, we instead rely on the temporal capabilities of the device itself. Since it is possible now to set a VCR to program later, EXACT can handle this goal with a classical planner. If the VCR did not support this type of operation, EXACT would explain that the goal was unachievable.
Second, numerous actions are durative; execution occurs over an interval of time (e.g., “play all messages”). Although it might seem natural to model the temporal aspects of these actions explicitly (e.g., in PDDL 2.1 level 36), we tried this and discovered problems (see below).
Finally, different commands naturally translate into goals with different temporal annotations, but the nature of the temporal mapping is complex and subtle. For example, consider the command “Play all messages.” Note that the user (presumably) doesn’t want the answering machine to play the messages forever; once is enough. Thus if one models `play’ as a durative action (which transiently plays a message), one cannot model the goal as one of achievement. In a durative model, the goal has an implicit temporal annotation that it must be true at some point during execution, even if it is not true at the end of all execution. There are two problems with such a model. First, few planners support such expressiveness; indeed PDDL 2.1 level 3 does not even allow one to express the goal. But the deeper problem is related to disambiguating natural language. Consider the commands “Play all messages” (which does not require playing to be true at the end of the plan) and “Turn on answering machine” (which does). We could think of no principled way to distinguish between these goals; clearly the user would be very upset if EXACT responded to the last command by turning answering on and then off again!
Our solution is to use a classical atomic model of time, implicitly recognizing that exogenous events may subsequently change the device state. Philosophically, this is consistent with PDDL 2.1 level 5 in which all actions are instantaneous, but some may initiate
Figure 3: EXACT’s performance in our experiment.
Figure 3 shows EXACT’s performance on our dataset. Recall is the fraction of the sentences in the dataset where EXACT put forth an interpretation.7 We refer to these sentences as tractable sentences. On intractable sentences, EXACT indicates it cannot understand the request and asks for a paraphrase.
Precision is the fraction of the tractable sentences in the dataset that EXACT interpreted correctly (including the interpretation that the sentence was an impossible request, where appropriate). To measure precision and recall on our data, we labeled each sentence with one or more valid interpretations.8 As the black bars in the figure show, we achieve 100% precision on our dataset,
Remember that to maximize reliability, EXACT does not try to “guess” the correct interpretation when it is not sure. 8
EXACT’s interpretation is correct if it is a member of the set of valid interpretations.
7
6
doc.guandang.net
A Reliable Natural Langu... 暂无评价 8页 免费 Application of Natural L.... Developing a Natural Language Interface to Complex Data GARY G. HENDRIX, ...
A Natural Language Inter... 暂无评价 4页 免费 A Reliable Natural Langu....Interface to Data management (EUFID) system, a n a t u r a l - l ...
Many of the seniors language has reached a level...interface circuits,to highlingt the target of ...3.In the applicationof household appliances It ...
A visualization of small high-level language ...interface circuits, highlight the object of ... household appliances, medical equipment, aerospace,...
reliable news 新闻/通讯稿 news release 头条新闻 ...interface 无线掌上导航器 wireless palm pilot 蜂窝...household electrical appliances 局域网 Intranet 网站...
Many high-level language has reached the level ...interface circuits, highlight the object of ...In Household Appliances can be said that the ...
Many high-level language has reached the level ..., navigation systems, and household appliances. ...interface circuits, highlight the object of ...
code language, and as so low why is the use?..., navigation systems, and household appliances. ...interface circuits, highlight the object of ...
Many high-level language has reached the level ...interface circuits, highlight the object of ...In Household Appliances can be said that the ...
Many high-level language has reached the level ...interface circuits, highlight the object of ...In Household Appliances Can be said that the ...
热门文档