A Computational Model for Arguments Understanding

Stephane GuezCOLING 1990

TL;DR

This paper proposes a computational model for understanding arguments in natural language dialogues by integrating linguistic, conceptual, and discourse-level analyses. It argues that traditional AI approaches over-rely on conceptual knowledge, neglecting crucial structural cues from language itself that can disambiguate meaning and infer intent.

What Problem Does This Paper Address?

Understanding arguments in natural language is a complex task. While previous AI research has focused on the conceptual aspects of arguments (i.e., the logical relationships between propositions and domain knowledge), this paper highlights that argumentation is also a deeply linguistic phenomenon. Language provides specific structures and "clue words" (like "but," "so," "besides") that don't just convey information but actively orient how utterances should be interpreted as arguments. The challenge is to build a model that can leverage these linguistic cues, alongside conceptual and discourse structure, to understand arguments, especially when conceptual knowledge is incomplete or ambiguous. The authors argue that a purely conceptual approach is insufficient because it assumes complete prior knowledge of speakers' beliefs, which is often not the case in real-world dialogues.

Methods

The paper proposes a multi-level computational model for argument understanding, consisting of several interconnected modules:

  • Conceptual Base: Stores domain-specific conceptual relations, common sense knowledge, and individual speaker beliefs.
  • Relation Finder: Derives appropriate conceptual relations from the knowledge base.
  • Base of Linguistic Constraints: Describes the argumentative role of linguistic operators and connectives (e.g., "but," "besides"). This is a key novel contribution, moving beyond purely conceptual analysis.
  • Context Analyzer: Tracks the dialogue's topic and argumentative orientation, incrementally building the discourse structure.
  • Argumentative Analyzer: Computes the argumentative orientation of utterances by integrating constraints from the linguistic and conceptual levels.
  • Learning Module: Addresses gaps in conceptual knowledge by forming hypotheses about new relations, which are then checked for plausibility and consistency.

The core idea is to integrate these levels, ensuring that interpretations are coherent across all of them. The model uses first-order logic to represent constraints and relations. A novel aspect is the explicit representation and use of "argumentative operators" and their associated linguistic constraints, which can disambiguate meaning and infer intent even when conceptual knowledge is lacking.

Datasets

Not discussed in the paper. The paper uses illustrative examples from existing literature (e.g., Cohen 1987) and a hypothetical dialogue for demonstration.

Key Insights and Ideas

  • Linguistic Structure Carries Argumentative Weight: The paper strongly emphasizes that linguistic cues (connectives, specific phrasing) are not merely decorative but actively shape how arguments are understood, often independently of their propositional content. For example, "but" signals opposition, and "even A" signals A as a strong argument.
  • Beyond Conceptual Knowledge: A purely conceptual approach to argument understanding is limited by the assumption of complete prior knowledge. Linguistic and discourse structures provide crucial constraints that can compensate for missing conceptual information.
  • Argumentation is Relational, Not Just Logical: Argumentative relations are distinct from logical implication. The relative strength of arguments towards a conclusion, rather than strict truth-conditional logic, is key to understanding discourse.
  • Integrated Analysis is Necessary: A robust model must simultaneously consider linguistic cues, conceptual knowledge, and discourse structure to achieve coherent argument understanding.
  • Learning from Constraints: The model incorporates a learning mechanism that uses linguistic and discourse constraints to infer new conceptual knowledge when direct conceptual information is absent.

Key Results

The paper presents a conceptual model and an implementation in Prolog. It demonstrates how the model can analyze dialogues and compute argumentative orientations. The key "result" is the demonstration of how linguistic constraints, particularly those associated with connectives, can disambiguate utterances and infer argumentative intent, as shown through examples like the "sister coming to the reunion" dialogue. The paper also illustrates how the model can hypothesize new argumentative relations when faced with incomplete knowledge.

Limitations and Future Work

  • Syntactic/Semantic Analysis Not Implemented: The current implementation relies on pre-analyzed feature structures and does not include a full syntactic or semantic parser.
  • Integration with Broader Discourse Processing: The authors suggest future work should integrate this model into a larger framework that includes speech act analysis and plan recognition.
  • Formalization of Linguistic Operators: While the paper proposes a formalization for some operators, further work is needed to cover the full range of subtle linguistic phenomena.
  • Handling Ambiguity: While the model aims to resolve ambiguity, the paper acknowledges that misunderstanding is common in human communication and that intentional ambiguity can be used by speakers.

The authors believe their work is a useful contribution to the automatic processing of natural language dialogues by providing a more complete account of argumentation.

Summary updated

BibTeX
@inproceedings{guez1990computational,
  author = {Stephane Guez},
  title = {{A Computational Model for Arguments Understanding}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1990},
}
Figure