Abstract
In everyday argumentation, conclusions are often left implicit, relying on the audience to infer the intended point from the provided premises. For computational systems, reconstructing these missing links is a complex task. This paper introduces a framework for argument conclusion generation that prioritizes the inference of the conclusion's "target"—the central concept or entity. By employing a hybrid approach that combines premise target ranking with a triplet neural network for embedding learning, the proposed model achieves a target adequacy rate of 89% in human evaluations. This research demonstrates that explicitly modeling the argumentative target significantly enhances the coherence and relevance of generated conclusions compared to standard sequence-to-sequence models.
Introduction
Argumentation is a cornerstone of human communication, yet it frequently relies on enthymemes—arguments where the conclusion is unstated. While humans naturally bridge the gap between premises and the intended point, AI systems struggle with this "missing link." Most existing research in computational argumentation focuses on identifying existing conclusions within a text. This paper, presented at ACL 2020, shifts the paradigm toward synthesis: generating a conclusion when it is implicit or missing.
The Importance of the Argumentative Target
A conclusion is fundamentally composed of two elements: a target (the subject of discussion) and a stance (the position taken, e.g., pro or con). The authors posit that identifying the correct target is the most challenging aspect of conclusion generation. For instance, if premises discuss "online courses" and "virtual classrooms," the appropriate conclusion target might be "distance learning"—a concept that encapsulates the premises without necessarily being explicitly stated in every one. Standard summarization techniques often fail here because they do not account for the specific argumentative "point" being made.
Methodology
The researchers decompose the task of conclusion generation into three distinct stages: (1) conclusion target inference, (2) stance inference, and (3) conclusion text generation. The primary focus of this work is the first stage, tackled through two complementary methods.
Premise Target Ranking
This approach assumes the conclusion target is already present within one of the premises. Potential targets are first identified using a BiLSTM-CRF tagger. A LambdaMART ranking model then selects the most representative target based on features such as position, word length, and sentiment word count.
Target Embedding Learning
To handle cases where the conclusion target is a related concept not explicitly found in the premises, the authors developed a "creative" approach using a Triplet Neural Network. This model maps the average embeddings of premise targets into a learned embedding space. It is trained to minimize the distance between these premise embeddings and the ground-truth conclusion target while maximizing the distance to unrelated "distractor" targets. This allows the system to select a target from a predefined lexicon (knowledge base) that best fits the semantic context of the argument.
The Hybrid Model
The final system uses a hybrid strategy: if the target identified by the embedding model overlaps with the premises, it is selected; otherwise, the system defaults to the top-ranked target from the ranking model.
Datasets and Evaluation
The model was tested across three diverse datasets:
- iDebate Dataset: Over 2,000 "pro" and "con" points from online debate portals.
- Argument Annotated Essays: 402 persuasive student essays, representing more formal, structured writing.
- Claim Stance Dataset: Used to train the initial target identification component.
Key Results and Insights
The proposed hybrid model significantly outperformed standard sequence-to-sequence (Seq2Seq) and pointer-generator baselines. On the iDebate dataset, the hybrid model achieved a BLEU score of 10.0, a stark contrast to the 0.7 score of the Seq2Seq baseline.
Human evaluation further validated the approach. Annotators found the generated targets to be "at least somewhat adequate" in 89% of cases, with 55% rated as "fully adequate." A key insight from the study is that semantic adequacy is more critical than exact lexical matching in argumentation; standard metrics like BLEU often underrepresent the quality of the generated conclusions as perceived by humans.
Conclusion and Future Directions
This research establishes that explicitly modeling the argumentative target is essential for generating coherent conclusions. While the current work focuses on target inference, it lays the groundwork for full-sentence generation. Future improvements could involve expanding the target knowledge base using external graphs like Wikidata or ConceptNet to enable the inference of entirely novel, out-of-vocabulary targets.
Summary updated