TL;DR
This paper introduces a method for annotating Chinese text with "information structures" derived from the HowNet knowledge base. This approach aims to represent the meaning of texts more deeply by combining HowNet's semantic definitions with dependency relations, moving beyond traditional syntactic parsing towards a richer understanding of Chinese.
What Problem Does This Paper Address?
While significant progress has been made in building Chinese corpora with syntactic information (like word boundaries, part-of-speech tags, and syntactic structures), there's a notable gap in annotating these corpora with semantic information. Existing efforts have primarily focused on lexical-level semantics or syntactic parsing. This paper addresses the challenge of representing the deeper meaning of Chinese texts by proposing a method to annotate them with "information structures" that capture semantic relationships between concepts, going beyond surface-level syntax.
Methods
The core of the paper's methodology lies in leveraging the HowNet knowledge base. HowNet is a bilingual, general knowledge base that describes relationships between concepts and their attributes. The authors define an "information structure" as a combination of:
- HowNet Definitions: These provide semantic definitions for concepts, utilizing a set of basic semantic units called "sememes" organized hierarchically. HowNet defines various relation types, including hyponymy, synonymy, antonymy, meronymy, attribute-host, material-product, converse, dynamic role, and concept co-occurrence.
- Dependency Relations: These capture how concepts relate to each other in a sentence or phrase, similar to syntactic dependency parsing but applied at a semantic level. The paper details how these relations are identified and represented.
The annotation process involves:
- Tagging concepts with their most probable HowNet definitions.
- Determining the information structures based on these definitions and identified dependency relations.
- Recovering implicit information structures from surface structures using HowNet's internal relations and rules governing dynamic roles between event types.
The paper also highlights the development of new information structure types, particularly those arising from function words, to improve coverage at phrase and sentence levels.
Datasets
The authors report annotating a subset of the Sinica corpus (version 3.0), specifically 30,000 words from 103 newspaper texts covering the crime domain. This corpus was previously annotated with word boundaries and part-of-speech tags. The paper also references the work by Gan and Tham (1999) which used a corpus of 30,000 words annotated with HowNet's general knowledge structure.
Key Insights and Ideas
- Bridging Syntax and Semantics: The paper demonstrates a practical approach to bridge the gap between syntactic analysis and deeper semantic understanding by integrating a rich knowledge base (HowNet) with dependency relations.
- HowNet's Robustness: The bottom-up approach to defining sememes in HowNet, starting from Chinese characters and building up to concepts, is presented as a robust method for capturing a wide range of concepts, even new ones.
- Dynamic Roles for Deeper Meaning: The concept of "dynamic roles" in HowNet is crucial for capturing the nuanced relationships of actants in an event, going beyond simple agent-patient roles in traditional case grammar. This allows for a more precise semantic representation.
- Implicit Information Recovery: The strategy of recovering implicit information structures from surface structures is a key innovation, enabling a more complete understanding of text meaning, especially in languages like Chinese where information might be implied rather than explicitly stated.
Key Results
The paper reports on the annotation of a subset of the Sinica corpus with information structures. While specific quantitative results (like accuracy metrics) are not detailed in this excerpt, the paper indicates that the annotation process has been completed for the chosen subset and is under verification. They expect to release the corpus and annotation guidelines. The work extends previous efforts by incorporating dependency relations into the annotation, providing a more comprehensive semantic representation.
Limitations and Future Work
- Coverage of Information Structures: The authors acknowledge that the coverage of information structure types at the phrase and sentence levels was evaluated, and missing types were added. This suggests that the initial set of structures might not have been exhaustive.
- Implicit Information Recovery: The paper identifies the recovery of implicit information structures as a main issue to be resolved in the final stage of their strategy. This includes understanding how agents are implied (e.g., in a "suicide" event) or how means are implicit (e.g., "suicide by firing").
- Automatic Algorithms: The authors are working on developing automatic algorithms for disambiguating HowNet definitions and identifying information structures. This is a crucial next step for scalability.
- Future Knowledge Bases: The creation of two additional knowledge bases is planned as part of their future work, likely to support the implicit information recovery process.
Summary updated