Augmented Transition Networks (1970) › Applications: Question Answering and Natural Language Interfaces
History of Language AIAll chapters Augmented Transition Networks (1970) › Applications: Question Answering and Natural Language InterfacesOne section · Sections in this chapter4 / 7 ↓
Reading tools Definitions & inline links
Choose your expertise level to adjust how many terms are explained. Beginners see more tooltips, experts see fewer to maintain reading flow. Hover over underlined terms for instant definitions.
Article links
Make inline references clickable
A well-known early application of ATNs was the LUNAR system, developed by Woods and colleagues at BBN. LUNAR answered questions about lunar rock samples from the Apollo missions. Users could ask questions in natural English such as "What is the average concentration of aluminum in high-alkali rocks?" The analysis supplied the information needed for a database query.
LUNAR's ATN parser handled the syntactic analysis of questions. The system included networks for different question types: yes-no questions, wh-questions, comparative questions, and others. Each network was designed to extract specific information needed for database queries. When parsing "What is the concentration of aluminum in high-alkali rocks?," the ATN would identify "aluminum" as the chemical element of interest, "high-alkali rocks" as the sample filter, and "concentration" as the property to retrieve.
The ATN formalism proved well-suited for this application because it could integrate syntactic and semantic processing. Registers could store syntactic categories along with semantic information extracted during parsing. When the parser encountered "high-alkali rocks," it could store both the syntactic structure (a noun phrase) and the semantic interpretation (a filter condition for the database query). This tight integration between parsing and interpretation allowed LUNAR to handle complex questions that required understanding both syntax and meaning.
The register mechanism in ATNs made it natural to integrate semantic processing with syntactic parsing. As the parser built up syntactic structures, it could simultaneously construct semantic representations. This integration was particularly valuable for question-answering systems like LUNAR, where the goal wasn't just to produce parse trees but to extract query-relevant information. The procedural nature of ATNs made it straightforward to add semantic actions alongside syntactic transitions.
Other researchers adopted ATNs for various natural language processing tasks. ATN parsers were developed for information retrieval systems, where users could express queries in natural language rather than Boolean logic. Machine translation systems used ATNs to parse source languages and generate target language structures. Text generation systems used inverted ATNs, where the networks guided generation rather than parsing, constructing sentences from semantic representations.
The formalism was not tied to one domain. A grammar author could replace the networks and attach interpretations appropriate to another language or application. That flexibility came from programming the arcs, not from a language-independent grammar supplied by the ATN formalism itself.
ATN grammars could also parse fragments such as "What about aluminum?" A surrounding dialogue system could supply the earlier question about silicon and interpret the fragment against that context. Registers held state during a parse; longer-lived dialogue context still had to be managed by the application.