Acquisition of Kowledge Data by Analyzing Natural Language

Yasuhito Tanaka, Sho YoshidaCOLING 1986

TL;DR

This paper tackles the challenge of acquiring knowledge data for natural language processing (NLP) systems, arguing that traditional methods relying solely on grammar and dictionaries are insufficient. The authors propose a phased approach to knowledge acquisition, emphasizing the need to define words by their characteristics and relationships, and explore methods for extracting this knowledge from text.

What Problem Does This Paper Address?

Modern NLP systems, particularly those involved in tasks like kana-to-kanji conversion and machine translation, struggle with ambiguity arising from homonyms and polysemous words. The authors contend that simply combining existing grammars and word dictionaries is inadequate to resolve these issues. They identify a fundamental gap: the lack of robust "knowledge data" that captures the nuanced meanings and relationships of words. This knowledge is crucial for building more sophisticated and accurate NLP systems.

Methods

The paper outlines a three-phase approach to understanding and acquiring knowledge:

  • Phase 1: Initial attempts to build NLP systems using only grammar and word dictionaries, which proved insufficient due to ambiguity.
  • Phase 2: Exploring limitations on vocabulary, word usage, fields, and sentence types to reduce ambiguity. While effective, this approach is deemed too restrictive for practical application.
  • Phase 3: The core proposal, which emphasizes defining words by their characteristics, coverage, and conditions. This is represented formally as $\omega_i \leftarrow \{P_{i1}, P_{i2}, \dots, P_{in}\}$, where $P_{ij}$ are limitations on word $\omega_i$. The authors suggest that while semantic markers and categories are common approaches, they focus on word-to-word relationships.

The paper then details three methods for acquiring this knowledge data:

  1. Extracting and segmenting Kanji strings: This involves mechanically extracting four-character Kanji strings from large text corpora (like JICST abstract files), identifying meaningful concept combinations, and classifying them into categories. This method is highlighted for its potential for easy expansion and mechanical processing.
  2. Word-to-word synchronism: Analyzing Japanese sentences to extract connections between independent words that appear in the same sentence, treating them as candidates for knowledge data based on mutual relation, frequency, and manual procedures.
  3. Syntax analysis: Analyzing sentences using word dictionaries and grammar to build syntax elements, extract connections, and thus derive knowledge data. This method is theoretically sound but faces practical challenges due to the combinatorial explosion of syntax trees.

Datasets

The paper mentions the use of "JICST abstract files" for extracting Kanji strings. It also refers to the examination of "78,000 four character kanji strings" and an ongoing attempt to reorganize "887,000 data, 200,000 in net." No specific, named datasets are detailed for training or evaluation in the traditional sense.

Key Insights and Ideas

  • The insufficiency of traditional NLP components: The paper strongly argues that grammar and dictionaries alone are not enough for robust NLP. A deeper understanding of word meaning and relationships is essential.
  • The importance of word context and conditions: Defining words not just by their dictionary entry but by their specific usage conditions and relationships is key to resolving ambiguity.
  • Phased evolution of NLP knowledge acquisition: The progression from simple dictionary-based approaches to more complex relational and contextual understanding reflects the historical development and challenges in the field.
  • Mechanical extraction as a viable path: The emphasis on mechanically extracting knowledge from large text corpora suggests a pragmatic approach to overcoming the manual labor involved in knowledge engineering.

Key Results

The paper does not present quantitative results in terms of accuracy or performance metrics for specific NLP tasks. Instead, it focuses on the conceptual framework and methods for acquiring knowledge data.

  • The authors claim that their mechanical extraction method (Kanji string analysis) has yielded "28,000 different kinds of knowledge data" from "32,000 types of kanji strings in net."
  • They highlight that while "Word-related Attributes" are well-established in machine-readable dictionaries, other attributes like thesaurus structures are less developed, and there's "almost nothing" for other attributes.
  • The paper suggests that a "sophisticated system would be unsuccessful if no basic knowledge data is available."

Limitations and Future Work

  • Practicality of Syntax Analysis: The authors acknowledge that syntax analysis for knowledge acquisition is "theoretically pertinent but not practical" due to the immense number of syntax trees.
  • Ambiguity Reduction: While methods are proposed, the paper doesn't quantify how much ambiguity is reduced.
  • Systematization Challenges: The paper notes that without automatic sub-categorization of long-unit terms in Japanese, a variety of knowledge data would result, making systematization troublesome.
  • Future Work: The authors mention ongoing efforts to reorganize a large dataset and explore applying their Kanji string extraction ideas to three and five-character Kanji strings. They also note that concepts related to word-to-word synchronism are "only in the planning stage" with "no news reporting that knowledge data is effectively available."

Summary updated

BibTeX
@inproceedings{tanaka1986acquisition,
  author = {Yasuhito Tanaka and Sho Yoshida},
  title = {{Acquisition of Kowledge Data by Analyzing Natural Language}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1986},
}
Figure