
Natural Language Understanding using Temporal Action Logic Martin Magnusson Department of Computer and Information Science Linkoping¨ University, SE-581 83 Linkoping,¨ Sweden email: [email protected] www.phlai.com/nl1 Abstract performance that is required in real-world appli- We consider a logicist approach to nat- cations and that they, while realizing that back- ural language understanding based on the ground knowledge is important, tend to make its translation of a quasi-logical form into a role peripheral instead of a solid base on which to temporal logic, explicitly constructed for build upon. the representation of action and change, We describe a natural language understanding and the subsequent reasoning about this system based on a logicist knowledge representa- semantic structure in the context of a back- tion foundation that serves as a research platform ground knowledge theory using automated for experimentation with the interchange between theorem proving techniques. The ap- computational linguistics and knowledge repre- proach is substantiated through a proof-of- sentation and reasoning. The focus is the rep- concept question answering system imple- resentational and inferential adequacy of the un- mentation that uses a head-driven phrase derlying techniques, which have been selected for structure grammar developed in the Lin- generality and extensibility, rather than on imme- guistic Knowledge Builder to construct diate applicability or the similarity with human di- minimal recursion semantics structures alogue characteristics. The techniques are brought which are translated into a Temporal Ac- together in a simple and clear architecture that tion Logic where both the SNARK au- holds great potential for development and experi- tomated theorem prover and the Allegro mentation. A novel integration of natural language Prolog logic programming environment technology, knowledge representation technology, can be used for reasoning through an in- and automated reasoning technology in a proof- terchangeable compilation into first-order of-concept question answering system, with the logic or logic programs respectively. working title NL1, has been implemented and is available as open source1. 1 Introduction A complex and poorly understood area in com- 2 Temporal Action Logic putational linguistics is the integration and use of The Temporal Action Logic (TAL) is a non- background knowledge to aid parsing, interpreta- monotonic temporal logic developed specifically tion and understanding of natural language. There for reasoning about actions and dynamical do- is general agreement that background knowledge mains. The logic has its origin in the Fea- is needed, e.g. to select between ambiguous in- tures and Fluents framework developed by Sande- terpretations or to provide answers to questions, wall (1994) but was given a new characterization and that without at least a partial understanding in terms of circumscription by Doherty (1994). of the world a system can never hope to approach Many extensions since have turned TAL into a full natural language understanding. As artifi- very expressive language capable of represent- cial intelligence research moves closer to applica- ing, among other things, actions with durations, tions there is an increasing risk that too many nat- ural language projects concentrate on the robust 1http://www.phlai.com/nl1 ∀ ∧ per1 ¦ § © ¨ § £ ¡ ¡ ¢ t [P er(t, alive) P er(t, loaded)] ¡ ¢ £ § § acs1 [t1,t2] Load R((t1,t2] loaded) ¡ acs2 [t1,t2] Fire ([t1] loaded → £ ¤ ¥ R((t1,t2] ¬alive ∧ ¬loaded)) obs1 [0] ¬loaded ∧ alive occ1 ¤ ¡ ¨ § [1, 2] Load ¢ ¨ © § © ¡ occ2 [3, 4] Fire ¡ ¦ § ¨ ¨ § © © § © Figure 1: The Yale shooting scenario in TAL. ¥ ¡ © ¡ § © £ ¨ § context-dependent and non-deterministic actions, ¡ ¨ § ¡ concurrency, and action side-effects. It also pro- ¡ vides solutions to the frame, ramification and qual- ¦ ¤ ification problems. For a more detailed introduc- tion to TAL the reader is referred to (Doherty et al., 1998). Figure 2: An overview of the NL1 architecture. 2.1 TAL Narratives Domains are described in TAL using fluents that any fluent changes at any time points when the flu- represent properties of the world that change over ent was not occluded. The final step is the min- time. World laws governing the evolution of flu- imization of the Occlude predicate and the ac- ents are expressed in narratives, which are high- tion occurrences, expressing the default assump- level descriptions of observations, action schemas, tion that no unexpected fluent changes or spuri- and action occurrences. Narratives receive a se- ous actions occur unless explicitly specified. The mantics through a translation to the base language, minimization is accomplished through the circum- which is an order-sorted classical first-order logic scription of those parts of the translated theory that together with a circumscription policy described contain action schemas and action occurrences re- below. Figure 1 shows the well-known Yale shoot- spectively. ing scenario expressed as a TAL narrative. A per- sistence statement (labelled per1) constrains the 3 System Architecture fluents alive and loaded’s values to persist unless they are affected by some action. Action schemas NL1 carries on an interactive natural language text (acs1 and acs2) use the reassignment operator dialogue with the user, executing commands and R to make loaded true after performing the Load answering queries about a simulated blocksworld. action and to make both loaded and alive false af- An important emphasis of the system architec- ter performing the Fire action, but only if loaded ture, depicted in Figure 2, is the use of declarative was true when the action was initiated. An ob- knowledge structures in the hope of building a sys- servation statement (obs1) initializes loaded and tem that is both convenient to adapt to new usage alive to false and true respectively. Finally, the scenarios and that has a great ultimate potential for two occurrence statements (occ1 and occ2) de- extension while at the same time retaining the ba- scribe a world history in which the Load action is sic implementation components. performed between time points 1 and 2, and the 3.1 Natural Language Module Fire action is performed between 3 and 4. The natural language module is responsible for the 2.2 Occlusion parsing of input sentences and the construction of The key to the solution of the frame problem in a quasi-logical form. User interaction consists of TAL lies in the use of occlusion. When narra- natural language input and output through a mini- tives are translated into classical logic an Occlude mal user interface in the form of a text top-loop. predicate is introduced and constrained to be true Sentence are read from a prompt and passed as at those time points where fluents are forced to text strings to the Linguistic Knowledge Builder change their values due to reassignments. An ad- (LKB) component. The LKB chart parser uses a ditional no-change axiom is added that rules out Head-driven Phrase Structure Grammar (HPSG), based on the grammar in (Sag et al., 2003), to that the system is not limited to the relatively sim- parse the input text string and build a feature struc- ple blocksworld, even though it is used as an illus- ture representation. The grammar includes seman- trative example. tical relations, and the semantical part of the fea- ture structure representation constitutes a Minimal 3.3 Automated Reasoning Module Recursion Semantics (MRS) structure. Performing automated reasoning in the Tempo- One of the benefits of this set-up is that natural ral Action Logic is not trivial since it is a non- language generation can be achieved by running monotonic logic with a semantics based on cir- the chart parser “in reverse” using the same gram- cumscription. When Sandewall developed the mar that was used when parsing. The task of gen- basis of the formalism he was concerned more erating a response is then reduced to the task of with assessing its correctness rather than per- constructing a suitable MRS structure represent- forming automated reasoning. Later Doherty and ing an answer to the user’s request. Łukaszewicz (1994) showed how the semantics The HPSG grammar forms a declarative knowl- could be expressed using circumscription and, in edge source that is easily adapted to new vocab- (Doherty, 1994), that under certain restrictions ularies, by changing the lexicon of words and se- the circumscription policy could be reduced to mantical relations, and that has a great potential predicate completion (defined by Lifschitz (1994)) for extension, e.g., moving towards broad cover- which is computable and results in an equivalent age as in the English Resource Grammar project first-order characterization. (Copestake and Flickinger, 2000). These methods open up the possibility of rea- soning in TAL using a very flexible compilation 3.2 Knowledge Representation Module approach that fits in nicely with the TAL high- The knowledge representation module holds a rep- level narrative description to base language trans- resentation of both the basic world laws and the lation. The idea is to choose an existing domain- particulars of the current scenario, but also a his- independent automated reasoning paradigm and tory of the developments of the world during the develop a compilation step from the TAL base lan- ongoing dialogue. guage into a language compatible with that para- The world model, together with action specifi-
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages8 Page
-
File Size-