TL;DR
This paper introduces the Integrated Partial Parser (IPP), a system designed to understand natural language text by modeling a reader's purpose. Unlike previous parsers that focused on exhaustive syntactic analysis, IPP prioritizes understanding based on the reader's goals, allowing it to efficiently process information relevant to its domain (international violence and terrorism) and skip irrelevant details.
What Problem Does This Paper Address?
Traditional natural language understanding systems often treat all input text with equal importance, leading to inefficient processing. This paper argues that human readers don't process information uniformly; their goals and interests dictate what they pay attention to, how deeply they process it, and what they retain. The challenge is to build a computational model that mimics this purposeful reading, moving beyond purely syntactic parsing to a more goal-directed, semantic understanding. Previous systems like SAM and PAM aimed for complete understanding of every detail, making them fragile and computationally expensive, while FRUMP, though robust, could miss important information by being too top-down.
Methods
IPP employs a "reading with a purpose" approach, focusing on understanding relevant to its pre-defined interests (international violence and terrorism). Key methodological aspects include:
- Integrated Partial Parsing: IPP maps lexical input directly to high-level semantic representations, bypassing the need for a preliminary, strictly syntactic parse tree. This allows high-level predictions to influence low-level processing.
- Top-Down Processing with Interests: The system makes predictions about what information is likely to be relevant based on its interests. It actively seeks information that aligns with these interests and can skip over irrelevant parts.
- Script-Based Representation: Stories are represented using scripts (e.g., $KIDNAP, $AMBUSH) and tokens (representing concepts like "airliner" or "kidnapper"), similar to SAM and FRUMP, but with a focus on integrating them into a coherent, purposeful understanding.
- Bottom-Up Processing for Unexpected Input: While top-down processing is primary, IPP can also incorporate unexpected but interesting information, integrating it with its existing understanding.
- Word Classification: Words are classified based on their contribution to understanding: those that must be processed immediately (interesting, structure-building words), those that can be saved for later processing, and those that can be skipped entirely.
- Pragmatic Heuristics: IPP uses heuristics, such as assuming the first actor in a story is important, to guide its processing and make efficient guesses.
Datasets
Not discussed in the paper. The paper uses examples from the Boston Globe for illustration.
Key Insights and Ideas
- Purpose is Paramount: A reader's purpose is the most critical factor in determining how text is processed and understood. This insight is central to IPP's design.
- Efficiency Through Relevance: By focusing on pre-defined interests, IPP can achieve efficient understanding by selectively processing information, mirroring how humans skim and focus.
- Direct Semantic Parsing: Bypassing intermediate syntactic representations can be more effective for achieving deep understanding, especially when the goal is not a complete syntactic parse.
- Robustness via Pragmatism: Heuristics and the ability to skip irrelevant information contribute to robustness, allowing the system to handle real-world text that is often complex or imperfect.
- Interest as a Control Mechanism: The system's "interest" level guides processing depth and can even lead to stopping processing a story if it becomes uninteresting, similar to human behavior.
Key Results
The paper demonstrates IPP's ability to process newspaper stories about international violence and terrorism. While quantitative results are not explicitly detailed, the paper provides detailed examples of IPP processing specific sentences and constructing story representations. The qualitative findings highlight:
- IPP successfully identifies key events and actors in stories.
- It infers script activations (e.g., $TERRORISM, $SHOOT, $AMBUSH) based on lexical input and predictions.
- It builds semantic representations that capture the core meaning and relevant details of a story, as judged by its purpose.
- IPP can process complex sentences and achieve understanding in approximately 2 CPU seconds on a DECsystem 20/50.
- The system's representation for a story about a shooting incident in Ireland successfully captures the script, actors, victim, and location, demonstrating its ability to extract relevant information.
Limitations and Future Work
- Limited Domain: IPP's current domain of interest is restricted to international violence and terrorism.
- Incomplete Understanding: Due to its focus on purpose, IPP may not fully understand stories that require understanding of goals and plans, or other higher-level representations not currently supported.
- Potential for Errors: The reliance on pragmatic heuristics and top-down predictions means there's a possibility of making incorrect inferences.
Future work planned includes:
- Expanding the vocabulary and increasing the program's expertise in its current domain.
- Adding support for higher-level knowledge structures like goals and plans.
- Expanding the domain of interest to encompass a wider range of topics.
Summary updated