TL;DR
This paper explores using FORTRAN, a mathematically-oriented programming language, for non-numerical data processing, specifically for automatic translation. It introduces a suite of FORTRAN subroutines designed to handle linguistic data at various granularities, from individual bits to entire machine words and strings, demonstrating how these can be leveraged for complex tasks like dictionary lookup and grammatical analysis.
What Problem Does This Paper Address?
The proliferation of problem-oriented programming languages, while beneficial for users, creates fragmentation. It makes it difficult for users to share experiences and can lead to programs needing complete rewrites if a different machine with fewer compilers is used. The authors question whether the advantages of these specialized languages can be achieved without multiplying their number, proposing that a widely used, mathematically-focused language like FORTRAN, augmented with specific subroutines, could handle non-numerical tasks like automatic translation efficiently.
Methods
The paper details a set of FORTRAN subroutines designed to process linguistic data. These subroutines operate on different levels of data granularity:
- Bits:
PAKR(packs grammatical information into bits of a machine word) andKBIT(identifies individual bits) are used for encoding and checking grammatical features.LBITallows for flipping bits, useful for marking or unmarking information. - Single Characters:
IFFP(finds the first position of a character),SEPO(seeks positions of characters), andICPC(compares a character against a table) are used for character-level analysis, including identifying word boundaries. - Machine-words:
CPWO(compares machine words, ignoring the sign bit) is crucial for non-numerical comparisons. - Units larger than machine-words:
CPST(compares strings of a defined length) andSHA2(shifts strings within an array) are used for more complex string manipulations.ICSS(composes strings) is used to move strings between arrays. - Higher-level operations:
ADT(changes word order) addresses linguistic requirements like adjective placement. A significantLOOKUPsubroutine implements a logarithmic search for dictionary lookups, efficiently finding translations and associated grammatical information.
The core idea is to build complex linguistic processing capabilities by combining these fundamental building blocks.
Datasets
Not discussed in the paper. The paper focuses on the methodology and subroutines rather than specific datasets for training or evaluation. It mentions a "Hebrew-English translation project" and a "dictionary" but does not detail their composition or size.
Key Insights and Ideas
- FORTRAN's Versatility: The paper strongly suggests that FORTRAN, often perceived as purely for numerical computation, can be effectively adapted for complex non-numerical tasks like linguistic processing through well-designed subroutines.
- Modular Linguistic Processing: By breaking down linguistic tasks into bit, character, word, and string operations, the authors demonstrate a modular approach to building sophisticated translation systems.
- Efficient Dictionary Lookup: The logarithmic search implemented in the
LOOKUPsubroutine is a key innovation, allowing for rapid retrieval of translations and grammatical information, a critical component for any translation system. - Handling Grammatical Nuances: The subroutines for bit manipulation (
PAKR,KBIT,LBIT) highlight the importance of encoding and processing fine-grained grammatical information, which is essential for accurate translation. - Bridging Language Structures: The
ADTsubroutine specifically addresses the challenge of reordering words to conform to English syntax when translating from languages like Hebrew where grammatical structures differ (e.g., adjective placement).
Key Results
The paper doesn't present quantitative results in terms of translation accuracy or performance metrics. Instead, it focuses on demonstrating the feasibility and methodology of using FORTRAN and its subroutines for automatic translation. The "key results" are the successful design and description of these subroutines and their conceptual application to a Hebrew-English translation project. The efficiency of the dictionary lookup is highlighted, stating that it requires "never more than ten times to find a word."
Limitations and Future Work
- FORTRAN as a Platform: While advocating for FORTRAN's applicability, the authors acknowledge that it might not be the ultimate solution. They suggest that languages like PL/1 or future, yet-to-be-created languages might handle these tasks more efficiently.
- Need for Broader Applicability: The paper's ultimate goal is to promote the idea of universally applicable problem-oriented languages, where progress in one field (like programming) is accessible to others, and programs can be more easily shared across different computing environments.
- Further Subroutines: The authors mention that "additional subroutines that facilitate information retrieval and other linguistic data-processing" exist beyond those described, implying that the presented set is not exhaustive.
- Refinement Needed: The paper notes that the indefinite article selection routine "will, of course, still have to be refined for the case when the prefix 'un' is followed by a vowel."
Summary updated