TL;DR
This paper introduces BAQUIL, a system designed to understand the implicit meanings, or "connotations," within stories. It tackles the challenge of inferring unstated information by analyzing character descriptions and identifying inconsistencies, moving beyond simple denotation to capture deeper narrative layers.
What Problem Does This Paper Address?
Human language is rich with implicit meaning. Beyond the literal "denotation" of words, we constantly infer underlying intentions, emotions, and implications. For story-processing systems, this presents a significant challenge: how can a machine understand that a character "groaning" implies they are "suffering," or that a seemingly innocent statement might be a "lie" or a "joke"? This paper addresses the problem of inference control in story understanding, specifically focusing on how to manage and derive meaning from the subtle, often indirect, descriptions of characters and events. The authors highlight the need for systems to go beyond surface-level interpretation and to recursively explain inconsistencies that arise in character descriptions.
Methods
The core of the paper is the description of the BAQUIL system, which aims to capture "connotation" as a form of inference. BAQUIL operates on a hierarchical structure, with a Main Program orchestrating various modules. Key components include:
- Parser: Processes story input into case structures.
- Story Character Specialist: Analyzes character descriptions, performing comparisons and inferences.
- Dictionary Lookup: Utilizes a semantic network and lexical taxonomy.
- Connotation Rules: A set of metarules (M1-M3) and specific inference rules (R1-R8) that guide the system's reasoning. These rules are designed to identify and interpret inconsistencies or changes in character descriptions.
- Comparison Rules (C1-C7): These rules are crucial for detecting potential inferences by comparing predicates within a character's description. They look for negations, hyponymy, lexical exclusion, and other linguistic relationships.
The system's approach is characterized by:
- Recursive Explanation: BAQUIL can recursively explain inconsistencies, building layers of meaning.
- Metarules for Inference: M1-M3 provide a framework for when and how to trigger inferences, focusing on pairwise matching of predicates, the type of verb used (change or inconsistency), and leveraging lexical taxonomy.
- Lexical Taxonomy: The system relies on a structured lexicon that includes subcategories, supercategories, and property lists to understand relationships between words.
Datasets
The paper does not explicitly mention training or evaluation datasets in the traditional sense. Instead, it states that the examples and rules are illustrated using a set of 20 stories (ranging from 4 to 200 sentences) that were processed by the system at Essex between 1979-1980. The "Sample Session" section provides a concrete example derived from "The Big Red Joke Book" by Benton and Loomes.
Key Insights and Ideas
- Connotation as a Structured Inference Process: The paper proposes that connotation isn't just a fuzzy feeling but a structured inferential process that can be modeled computationally. This is a significant step beyond simple keyword matching or denotational understanding.
- The Power of Inconsistency: BAQUIL leverages inconsistencies in character descriptions as a primary driver for inferring deeper meaning. This suggests that contradictions are not just errors but rich sources of information.
- Recursive Nature of Meaning: The ability to recursively explain inconsistencies is highlighted as an essential feature, allowing for the accumulation of meaning through multiple layers of interpretation.
- Lexical Taxonomy as a Foundation for Inference: The paper emphasizes the critical role of a well-structured lexicon and its associated taxonomy in enabling sophisticated inferences about word relationships and semantic categories.
- Context-Dependent Interpretation: The system demonstrates how context (e.g., previous statements, character roles) influences the interpretation of utterances, as seen in the example of the doctor and nurse.
Key Results
The paper focuses more on the methodology and theoretical framework rather than presenting extensive quantitative results. However, the "Sample Session" provides a qualitative demonstration of BAQUIL's capabilities. The system successfully infers:
- That a child's statement "There is no God!" in response to a quote about God sending a cheese implies that the speaker (the child) believes the quoter (Krylov) has a "religious discourse."
- That the teacher's response "there is no cheese, either..." implies the child has an "inconsistent discourse."
- That the teacher's statement, when interpreted in the context of Krylov's discourse being "fiction," means the teacher is suggesting Krylov's discourse is fictional.
- That the child's statement, when linked to Krylov's discourse being fictional, leads to the inference that the teacher believes the child is mistaken.
The paper also notes that the sample session ran in approximately 1.8 seconds of CPU time on a PDP-10, using about 25K of core memory, indicating a relatively efficient processing of these inferential steps.
Limitations and Future Work
The authors acknowledge several limitations:
- Limited World Knowledge: BAQUIL is specialized in discourse attitudes (errors, lies, jokes) and lacks extensive general world knowledge, limiting its ability to handle more complex scenarios or characters with nuanced personalities.
- Ambiguity in Lexical Connections: The paper notes a difficulty in distinguishing between "necessary truths" (always true) and "sometimes true" truths in lexical connections, which can complicate inference.
- Handling of "Liar's Paradox": Rules R6 and R7 are described as "not quite satisfying" and needing refinement, particularly in handling situations like the Liar's Paradox.
- Topic Ambiguity: For rules like R3, a "topic" could be connected to multiple discourse types (e.g., "religious discourse" and "joke"), and the system currently has no mechanism to systematically prioritize these connections.
- Language Specificity: The system's precision is limited to French and English, and the authors acknowledge a lack of precision regarding how "Whorfian" the system would be with other languages.
Future work could involve integrating BAQUIL into larger AI systems, expanding its expertise to other types of indirect descriptions, and refining the handling of complex linguistic phenomena and world knowledge.
Summary updated