Computational Linguistics and Intelligent Text Processing: Proceedings of the 8th International Conference, CICLing 2007, Mexico City(invited paper), A. Gelbukh (Ed.), pp. 311-324, Springer, Berlin, Germany, February 2007.Learning for Semantic ParsingRaymond J. MooneyDepartment of Computer Sciences, University of Texas at Austin1 University Station C0500, Austin, TX 78712-0233, [email protected]. Semantic parsing is the task of mapping a natural languagesentence into a complete, formal meaning representation. Over the pastdecade, we have developed a number of machine learning methods for in-ducing semantic parsers by training on a corpus of sentences paired withtheir meaning representations in a specified formal language. We havedemonstrated these methods on the automated construction of natural-language interfaces to databases and robot command languages. Thispaper reviews our prior work on this topic and discusses directions forfuture research.1 IntroductionSemantic parsing is the task of mapping a natural language (NL) sentence intoa complete, formal meaning representation (MR) or logical form. A meaningrepresentation language (MRL) is a formal unambiguous language that allowsfor automated inference and processing, such as first-order predicate logic. Inparticular, our research has focused on applications in which the MRL is “ex-ecutable” and can be directly used by another program to perform some tasksuch as answering questions from a database or controlling the actions of a realor simulated robot. This distinguishes the task from related tasks such as se-mantic role labeling [8] and other forms of “shallow” semantic parsing which donot generate complete, formal representations.Over the past decade, we have developed a number of systems for learningparsers that map NL sentences to a pre-specified MRL [44, 35, 37, 24, 17, 39, 23].Given a training corpus of sentences annotated with their correct semantic in-terpretation in a given MRL, the goal of these systems is to induce an efficientand accurate semantic parser that can map novel sentences into this MRL. Someof the systems require extra training input in addition to (NL, MR) pairs, suchas syntactic parse trees or semantically annotated parse trees.In this paper, we first describe the applications we have explored and theircorresponding MRLs, and then review the parsing and learning systems that wehave already developed for these applications, along with experimental resultson their performance. We then discuss important areas for future research inlearning for semantic parsing.2 Sample Applications and their MRLsWe have previously considered two MRLs for performing useful, complex tasks.The first is a database query language, primarily using a sample database onU.S. geography. The second MRL is a coaching language for robotic soccer de-veloped for the RoboCup Coach Competition, in which AI researchers competeto provide effective instructions to a coachable team of agents in a simulatedsoccer domain [9].When exploring NL interfaces for databases, the MRL we have primarilyused is a logical query language based on Prolog. We have primarily focused onqueries to a small database on U.S. geography. This domain, Geoquery, wasoriginally chosen to test corpus-based semantic parsing due to the availability ofa hand-built natural-language interface, Geobase, supplied with Turbo Prolog2.0 [3]. The language consists of Prolog queries augmented with several meta-predicates [44]. Below is a sample query with its English gloss:answer(A,count(B,(state(B),const(C,riverid(mississippi)),traverse(C,B)),A))“How many states does the Mississippi run through?”The same query language has also been used to build NLI’s for databases ofrestaurants and CS-job openings, including a component that translates ourlogical queries to standard SQL database queries [36, 35]. The resulting formalqueries can be executed to generate answers to the corresponding questions.RoboCup (www.robocup.org) is an international AI research initiative usingrobotic soccer as its primary domain. In the Coach Competition, teams of agentscompete on a simulated soccer field and receive advice from a team coach in aformal language called CLang. In CLang, tactics and behaviors are expressedin terms of if-then rules. As described in [9], its grammar consists of 37 non-terminal symbols and 133 productions. Below is a sample rule with its Englishgloss:((bpos (penalty-area our)) (do (player-except our {4}) (pos (half our))))“If the ball is in our penalty area, all our players except player 4 should stay in ourhalf.”The robots in the simulator can interpret the CLang instructions which thenstrongly affect their behavior while playing the game. The semantic parsers wehave developed for this MRL were part of a larger research project on advice-taking reinforcement learners that can accept advice stated in natural language[25].3 Systems for Learning Semantic ParsersOur earliest system for learning semantic parsers called Chill [44, 35] uses In-ductive Logic Programming (ILP) [26] to learn a deterministic parser written inProlog. In our more recent work, we have developed three different approacheshas2VP−bownerplayer the ballNN−player CD−unum NP−nullNN−nullVB−bownerS−bownerNP−playerDT−nullPRP$−teamourplayer the ballN3−bowner(_)N7−player(our,2)N2−null null nullN4−player(_,_) N5−teamourN6−unum2N1−bowner(_)hasN8−bowner(player(our,2))Fig. 1. The SAPT and its Compositional MR Construction for a CLang Sentence.to learning statistical semantic parsers that are more robust and scale more ef-fectively to larger training sets. Each exploits a different advanced technologyin statistical natural language processing. Scissor [17, 18] adds detailed seman-tics to a state-of-the-art statistical syntactic parser (i.e. the Collins parser [12]),Wasp [39] adapts statistical machine translation methods to map from NL toMRL, and Krisp [23] uses Support Vector Machines (SVM’s) [13] with a subse-quence kernel specialized for text learning [27]. We briefly review each of thesesystems below. A version of our Geoquery data has also been used to eval-uate a system for learning semantic parsers using probabilistic CombinatorialCategorial Grammars (CCG) [45].3.1 ScissorScissor (Semantic Composition that Integrates Syntax and Semantics to getOptimal Representations) [17, 18] learns a statistical parser that generates asemantically augmented
View Full Document
Unlocking...