Quick Links

Overview

The objective of this project is to represent factual claims in a structural way, where various aspects of such claims, including the entities involved and their relationships, quantities, points and intervals in time, comparisons, and aggregate structures, can be captured. With such a modeling capability in place, fact-check assisting tools can exploit the idiosyncrasies of different forms of factual claims. For instance, in translating claims into verification queries over knowledge bases, query templates can be carefully crafted beforehand for different types of claims, and methods can be designed to replace the variables in the query templates by entities and elements from the structured representations.

Our approach is to extend the Berkeley FrameNet project, a lexical resource for English built on a theory of meaning called frame semantics. In frame semantics, lexical units (LUs, i.e., words, phrases, and linguistic patterns) evoke frames. A frame describes a type of event, action, situation, or relation, together with frame elements (FEs). Frame elements are frame-specific semantic roles that provide additional information to the semantic structure of a sentence. We created 20 factual-claim specific frames, including 11 new frames and nine existing ones from FrameNet to represent claims in a structured format. Table 1 shows a new frame---Vote created for characterizing claims about someone's voting decision towards an issue. Agent and Issue are two of the frame elements. Agent, a conscious entity, holds a positive or negative opinion about an Issue and votes on it. The lexical units of the Vote frame are vote and (a/the) deciding vote.n in the verb and noun forms, respectively.








We created the following 11 frames to model factual claims.

• Taking sides consistency: This frame is about the consistency of an "Agent's" "Stance" towards an "Issue'". The "Agent" either alters or maintains his/her "Stance".
• Recurring action: The Recurring action frame describes a repetitive "Action" that is performed by an "Agent" at the interval of a "Time span".
• Recurring action with frequency: This frame is about a repetitive "Action" that is performed by an "Agent" at a given "Frequency".
• Correlation: It shows the connection or relationship between the occurrences of "Event_1" and "Event_2".
• Comparing two entities: This frame is about comparing two entities using a "Comparison_criterion" while qualifying with a "Degree".
• Comparing at two different points in time: This frame is about comparing an "Entity" with itself at two different points in time using a "Comparison_criterion" while qualifying with a "Degree".
• Occupy rank via ordinal numbers: This frame is about "Items" in the state of occupying a certain "Rank" specified by an ordinal number within a hierarchy.
• Occupy rank via superlatives: This frame is about "Items" in the state of occupying a certain "Rank" specified by a superlative within a hierarchy.
• Ratio: In this frame, a "Criterion" determines a "Ratio" that quantifies the size of the subset of a larger "Group".
• Uniqueness of trait: This frame distinguishes a "Unique entity" from a "Generic entity" based on a specific "Trait" where a "Trait" is some property, quality, point-of-view, or an arbitrary construct which is generally understood to be an attribute of an entity.
• Vote: This frame distinguishes a "Unique entity" from a "Generic entity" based on a specific "Trait" where a "Trait" is some property, quality, point-of-view, or an arbitrary construct which is generally understood to be an attribute of an entity.



We adopted the following nine existing FrameNet frames to represent some factual claim categories.

• Taking sides: A Cognizer has a relatively fixed positive or negative point of view towards an issue.
• Statement: A Speaker is communicating about a message.
• Causation: A cause leads to an effect.
• Cause change of position on a scale: An item’s position moves along a scale due to some agent or cause.
• Capability: An entity is able or unable to do something.
• Change position on a scale: An item moves along a scale.
• Creating: A cause leads to an entity being created.
• Occupy rank:The location of an entity in a hierarchy with other entities.
• Conditional occurrence: Something can occur if [a/some] condition(s) [is/are] met.



We also built a publicly available and web-based frame annotation tool FrameAnnotator, to aid annotating sentences. FrameAnnotator supports full-text annotation and encodes annotated sentences in the same XML format used in FrameNet.

People

• IDIR Faculty
• Chengkai Li
• IDIR Students
• Fatma Arslan
• Josue Caraballo
• Damian Jimenez
• IDIR Alumni
• Gensheng Zhang
• Sarbajit Roy

Publications

• Fatma Arslan, Josue Caraballo, Damian Jimenez, and Chengkai Li. Modeling Factual Claims with Semantic Frames. In Proceedings of the 12th International Conference on Language Resources and Evaluation (LREC), pages 2511-2520, Marseille, France, May 11-16, 2020. PDF Frame Annotator dataset video (soon)

Disclaimer

This material is based upon work partially supported by the National Science Foundation Grants 1937143, 1719054, and sub-awards from Duke University as part of a grant to the Duke Tech & Check Cooperative from the Knight Foundation and Facebook. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the funding agencies.