Frame- and Entity-Based Knowledgefor Common-Sense ArgumentativeReasoning

Teresa Botschen, Daniil Sorokin, Iryna GurevychUbiquitous Knowledge Processing Lab (UKP) and Research Training Group AIPHESTechnische Universität Darmstadthttps://www.informatik.tu-darmstadt.de/ukp/

In Proceedings of the 5th Workshop on Argument Mining at the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP)

SummaryCommon-sense Argumentative Reasoning

Given a debate title, a claim and a reason choosea correct warrant to complete the argumentativereasoning chain (SemEval-2018 Task 12)

Motivation• External world knowledge is essential for reason-

ing comprehension•No shared task system focused on enriching mod-

els with external knowledge•Related work shows improvements for natural

language inference and question answering withsymbolic knowledge

Contributions• Explore external knowledge beyond lexical→WIKIDATA (fact knowledge about entities)→ FRAMENET (prototypical situations)!Both resources lead to a small improvement!More work needed on integration of multiple an-

notation types! Logic-based modeling of reasoning is missing

Code and Models

ERead the paperEDownload the code

Built withPython, PyTorch

https://github.com/UKPLab/emnlp2018-argmin-commonsense-knowledge

WIKIDATA

Corporation

Company

Intel

Santa Clara

SUBCLASS OF

INSTANCE OF HEADQUARTERS IN

FRAMENET

Companies make profit.

Intentionallycreate

Created entityCreator

Can companies be trusted?

Corporations have only one goal : to make profit .Reason

They do not have to satisfy customers tomake a profit.

Warrant 1They have to satisfy customers to make aprofit.

Warrant 2

Companies can’t be trusted .Claim

Do police use deadly force too often?

Police use the excuse of fear for life to abuse use of force .Reason

The excuse is rarely warranted.Warrant 1

The excuse is sometimes warranted.Warrant 2

Police is too willing to use force.Claim

Results•We use the intra-warrant attention model from the dataset paper (Habernal et al., 2018)+Word embeddings are concatenated with FRAMENET and WIKIDATA embeddings

Table 1: Mean and max accuracy over 10 runs on the ARC dev. and test sets

mean maxApproach Dev. Test Test

Habernal et al. (2018) 0.6712 0.5570 0.5878+WD 0.6623 0.5680 0.6036+FN 0.6741 0.5676 0.6104+FN/WD 0.6630 0.5592 0.5946

Figure 1: Performance for the ‘+WD’ approach bythe number of WD entities in an instance.

0 1 2 3 ≥ 40.45

0.5

0.55

0.6

Acc

urac

y

WD Habernal et al. (2018)

0

50

100

#in

stan

ces

Figure 2: Performance for the ‘+FN’ approach by thenumber of frames in an instance.

12 14 16 18 20 22 24 26 28 30≥ 320.3

0.4

0.5

0.6

0.7 1.0

Acc

urac

y

FN Habernal et al. (2018)

0255075

#in

stan

ces

Integrated componentsFrameNet event parsing•Multimodal Frame Identification with Multilingual Evaluation. Teresa Botschen et al. NAACL

2018

Identifying entities and linking them to Wikidata•Mixing Context Granularities for Improved Entity Linking on Question Answering Data across

Entity Categories. Daniil Sorokin and Iryna Gurevych. *SEM 2018

This work has been supported by the German Research Foundation as part of the Research Training Group AIPHES under grant No. GRK 1994/1, and via the QA-EduInf project (grant GU 798/18-1 and grant RI 803/12-1).