Using Language and Context in the Analysis of Text

Yigal Arens

Introduction

In the early 1980s, the field of natural language processing (NLP) was divided between two primary architectural philosophies. One school of thought advocated for a serial pipeline approach, where a syntactic parse was completed before applying semantic rules. The opposing school, championed by researchers such as Roger Schank, argued for "integrated parsing," where all forms of knowledge are applied simultaneously in a single stage.

Yigal Arens proposed a middle ground that moves beyond this binary debate. He argued that while syntax and semantics should not be strictly separated, a distinction must be made between linguistic knowledge (the structure and general meaning of phrases) and contextual knowledge (world knowledge and situational awareness). This paper introduces a two-tiered architecture designed to bridge this gap, utilizing a phrasal analyzer to map linguistic patterns to conceptual fragments which are then reconciled by a context-centered component.

The Dual-Level Architecture

The Language-Centered Component: PHRAN

The core of the proposed system is PHRAN (PHRasal ANalyzer). Unlike traditional parsers that operate on a word-by-word basis, PHRAN treats the phrase as the fundamental unit of analysis. The system utilizes a database of pattern-concept pairs:

  • Patterns: These range from literal strings (e.g., "in a jiffy") to class-based templates (e.g., "<person> <go> to <location>") and complex phrasal constructs.
  • Concepts: These are the semantic representations associated with specific patterns.

As PHRAN processes text from left to right, it maintains a buffer of "conceptual fragments." When a sequence of terms matches a known pattern in the database, the system replaces the sequence with its corresponding higher-level concept. This approach allows the system to handle idioms and complex phrasal units—such as "kick the bucket"—without requiring a convoluted syntactic breakdown of individual words that may not contribute to the idiomatic meaning.

The Context-Centered Component

The second tier of the architecture is a context-centered layer designed to reconcile the fragments produced by PHRAN with broader world knowledge. At the time of the paper's writing, this component was largely theoretical or simulated, but its intended functions included:

  • Reference Resolution: Determining that a pronoun like "she" refers to a previously mentioned entity (e.g., "Mary").
  • Disambiguation: Selecting the correct meaning of a polysemous word based on the environment (e.g., distinguishing between "glasses" as eyewear or drinking vessels based on whether the context involves an "old man" or a "bartender").
  • Inference: Filling in the gaps of a narrative, such as understanding that the sentence "Mary cut the salami" implies the use of a tool (a knife) and a specific physical action.

Methodology and Evaluation

Because the paper predates modern large-scale NLP benchmarks, the system was evaluated using a variety of targeted test cases:

  1. Hand-crafted Sentences: The author tested the system on a range of examples, from news-style snippets regarding oil exploration to idiomatic sentences like "Wills will drive Bill to The Big Apple."
  2. Cross-linguistic Application: To demonstrate that the pattern-concept approach is language-independent, the database was extended to include Spanish and Chinese sentences.
  3. Narrative Simulation: A four-sentence story about duck hunting was used to show how context influences interpretation. For instance, the phrase "She got the bird" was interpreted as "she killed the bird" within the hunting context, rather than the literal "she received a bird."

Key Results and Insights

The system demonstrated several key strengths during testing:

  • Efficiency: Running on a DEC KL10 system, an uncompiled version of PHRAN processed example sentences in 0.5 to 6 seconds of CPU time.
  • Longest-Match Heuristic: PHRAN successfully implemented a "longest-match" strategy. For example, in the sentence "sat the game out," the system correctly identified the phrasal verb "sit out" as a superior match over the simple verb "sit," despite the latter appearing first.
  • Declarative Knowledge: By keeping linguistic data separate from processing routines, the system proved highly extensible. New idioms or linguistic rules could be added to the database without modifying the underlying code.

Limitations and Future Work

Despite its successes, the model faced several challenges. The context-centered component remained a model rather than a fully realized software tool at the time of publication. Additionally, the author identified an "indexing problem": as the database of pattern-concept pairs grows, retrieving the correct pattern becomes a computational bottleneck. While a tree-based indexing system was suggested, it remained an area for further research.

Finally, the author noted that the current architecture is largely unidirectional, with PHRAN passing data to the context layer. A more robust future model would allow the context layer to influence PHRAN’s initial pattern-matching process, creating a more integrated flow between world knowledge and linguistic analysis.

Summary updated

BibTeX
@inproceedings{arens1981using,
  author = {Yigal Arens},
  title = {{Using Language and Context in the Analysis of Text}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1981},
}
Figure