“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
state in the database DB, and returns a plan starting with the play action. Next, the plan will contain a delete-message action if message number one is one of <old-msg-num-1> through <old-msg-num-n>, and a play-next action otherwise. This repeats until all 64 messages have been checked and deleted if they are old.
4.2 Formal Properties
We are now in a position to state the benefits of our NLIA reduction precisely. Abstractly, one can consider an NLIDB, N, as a function from English sentences to SQL statements. Similarly our translator, T (described above), is a function from SQL to planning problem specifications.5 Finally, a planner is a function from these problem specifications to action sequences. Since an NLIA takes an English sentence, ε, and generates action sequences for the appliance, one can summarize our reduction as follows:
NLIA(ε) ≡ P T N (ε)
Popescu et al. [20] define the conditions under which an SQL statement is a valid interpretation of an English sentence ε, but the definition is too complex to include in this paper. We borrow from [20] the far simpler definitions of soundness and completeness below.
Definition. An NLIDB is sound if any SQL it outputs is a valid interpretation of its input sentence ε. An NLIDB is complete if it returns all valid interpretations of ε.
Note that if an NLIDB is both sound and complete and it returns a single SQL statement in response to a user’s utterance, then it has unambiguously determined the user’s intent – subject to our assumptions, of course.
Definition. Let S be an SQL statement over a relational database DB. An appliance reaction R is consistent with S if S is a query and R answers the query, or if S is an update and R is a sequence of legal device commands that changes DB accordingly. An NLIA is sound if in response to input ε, its reaction is consistent with some valid interpretation of ε. An NLIA is complete if it makes a consistent reaction to a valid interpretation of ε, when one exists. There are a variety of formulations of automated planning [19], but we briefly summarize with the following.
Definition. A planner is sound if any plan it outputs will transform the initial situation into a world state where the goal holds. A planner is complete if it returns a plan when one exists.
We can now formally state the two central benefits of our reduction: Proposition 1 [Soundness]. Let N be a sound NLIDB, let P be a sound planner, and let T be the translation scheme described above. Then P T N is a sound NLIA.
Proposition 2 [Completeness]. Let N be a complete NLIDB, let P be a complete planner, and let T be the translation scheme described above. Then P T N is a complete NLIA. The proofs are omitted due to lack of space.
Another potential objection to our theory is that it does not guarantee the reliability of a full-blown conversational speech interface; speech recognition, in particular, is likely to result in errors. While this is clearly true, we see great value in having an NLIA that is guaranteed to reliable – this enables the interface designer to localize errors to other modules and to institute the appropriate recovery strategy.
Consider, by way of analogy, a sophisticated chess-playing program that combines mini-max search with alpha-beta pruning, a complex and tunable evaluation function, specialized hardware, etc. Suppose we prove that alpha-beta pruning is “reliable” in that it only prunes moves that the search procedure would eventually discard. Well, the reliability of alpha-beta pruning does not guarantee that the chess program always makes the best move. However, when the program makes a mistake, we know that it is definitely not due to alpha-beta pruning. Again, the guaranteed reliability of one module, enables the program’s designer to focus his attention on other modules.
5. THE “EXACT” IMPLEMENTATION
In order to test the theory developed in section 4, we built the EXACT natural language interface to a telephone using the Precise NLIDB [20] and Blackbox planner [13] as foundations. We handcrafted a database model for the Panasonic KXTC1040W from its user manual [1]. This model is used both as an input to the Precise NLIDB and as the source of state information for the planner. Finally, we created a set of actions that model the phone’s commands, as described in the user manual. This action set is also input to the planner.
The system takes an input sentence, converts it into a set of possible SQL statements using Precise, translates those into a set of goals, and looks for a plan to satisfy each goal. If there is more than one goal and at least one goal has a plan, then we have an ambiguous sentence, and EXACT needs to ask the user for help in disambiguating. If no goal has a plan, then the phone cannot support the function being asked for, so EXACT can tell the user as much. If there is exactly one goal and it has a plan, EXACT can simply carry out that plan. In our experiments, the last case was by far the most common.
The dataset on which we evaluated our system includes examples of impossible requests, but EXACT is well equipped to handle this problem: if a sentence does not map to an appropriate SQL statement, either because of unknown words or because there is no attribute-value pairing for the sentence, then we can say that the NLIDB cannot understand the sentence. On the other hand, if the sentence maps to an SQL statement, but the planner fails to find a plan for that goal, then since our planner is complete we can say that the appliance does not support this function.
Our interface inherits desirable qualities, like reliability and portability across many appliances, from the planner and the NLIDB, but we had to make extensions to both components as explained below.
4.3 Significance of the Theory
Of course, theoretical guarantees only apply in practice if their assumptions are satisfied (e.g., all the words in the sentence are known – see [20] for the complete enumeration). Our experimental results (section 6) provide some evidence that these assumptions are realistic.
5.1 The NLIDB Component
Precise is a highly portable NLIDB that guarantees soundness of its SQL interpretations for certain kinds of English sentences, called semantically tractable questions [20]. Precise automatically generates a lexicon based on the names of relations, attributes, and values in its input database. At its core, it reduces the problem of mapping a sentence to an SQL query to a graph-matching problem, which can be solved using the maxflow algorithm. Precise relies on the Charniak parser to extract attachment information from the
Given a fixed set of actions, a planning problem is an initial state / goal pair, thus T maps from SQL to the cross product of tuple specifications with itself.
5
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