Explaining a Deep Learning Model for Aspect-Based Sentiment Classification Using Post-hoc Local Classifiers

Vlad Miron; Flavius Frasincar; Maria Mihaela Truşcǎ

doi:10.1007/978-3-031-35320-8_6

Explaining a Deep Learning Model for Aspect-Based Sentiment Classification Using Post-hoc Local Classifiers

Vlad Miron, Flavius Frasincar^*, Maria Mihaela Truşcǎ

^*Corresponding author for this work

Econometrics

Research output: Chapter/Conference proceeding › Chapter › Academic

3 Citations (Scopus)

Abstract

Aspect-Based Sentiment Classification (ABSC) models are increasingly utilised given the surge in opinionated text displayed on the Web. This paper aims to explain the outcome of a black box state-of-the-art deep learning model used for ABSC, LCR-Rot-hop++. We compare two sampling methods that feed an interpretability algorithm which is based on local linear approximations (LIME). One of the sampling methods, SS, swaps out different words from the original sentence with other similar words to create neighbours to the original sentence. The second method, SSb, uses SS and then filters its neighbourhood to better balance the sentiment proportions in the localities created. We use a 2016 restaurant reviews dataset for ternary classification and we judge the interpretability algorithms based on their hit rate and fidelity. We find that SSb can improve neighbourhood sentiment balance compared to SS, reducing bias for the majority class, while simultaneously increasing the performance of LIME.

Original language	English
Title of host publication	Natural Language Processing and Information Systems - 28th International Conference on Applications of Natural Language to Information Systems, NLDB 2023, Proceedings
Editors	Elisabeth Métais, Farid Meziane, Warren Manning, Stephan Reiff-Marganiec, Vijayan Sugumaran
Publisher	Springer Science+Business Media
Pages	79-93
Number of pages	15
ISBN (Print)	9783031353192
DOIs	https://doi.org/10.1007/978-3-031-35320-8_6
Publication status	Published - 2023
Event	28th International Conference on Applications of Natural Language to Information Systems, NLDB 2023 - Derby, United Kingdom Duration: 21 Jun 2023 → 23 Jun 2023

Publication series

Series	Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume	13913 LNCS
ISSN	0302-9743

Conference

Conference	28th International Conference on Applications of Natural Language to Information Systems, NLDB 2023
Country/Territory	United Kingdom
City	Derby
Period	21/06/23 → 23/06/23

Bibliographical note

Publisher Copyright:
© 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.

Access to Document

10.1007/978-3-031-35320-8_6

Cite this

Miron, V., Frasincar, F., & Truşcǎ, M. M. (2023). Explaining a Deep Learning Model for Aspect-Based Sentiment Classification Using Post-hoc Local Classifiers. In E. Métais, F. Meziane, W. Manning, S. Reiff-Marganiec, & V. Sugumaran (Eds.), Natural Language Processing and Information Systems - 28th International Conference on Applications of Natural Language to Information Systems, NLDB 2023, Proceedings (pp. 79-93). Springer Science+Business Media. https://doi.org/10.1007/978-3-031-35320-8_6

Miron, Vlad ; Frasincar, Flavius ; Truşcǎ, Maria Mihaela. / Explaining a Deep Learning Model for Aspect-Based Sentiment Classification Using Post-hoc Local Classifiers. Natural Language Processing and Information Systems - 28th International Conference on Applications of Natural Language to Information Systems, NLDB 2023, Proceedings. editor / Elisabeth Métais ; Farid Meziane ; Warren Manning ; Stephan Reiff-Marganiec ; Vijayan Sugumaran. Springer Science+Business Media, 2023. pp. 79-93 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Vol. 13913 LNCS).

@inbook{067afb7945b64b2fbcace1a04eca5bb9,

title = "Explaining a Deep Learning Model for Aspect-Based Sentiment Classification Using Post-hoc Local Classifiers",

abstract = "Aspect-Based Sentiment Classification (ABSC) models are increasingly utilised given the surge in opinionated text displayed on the Web. This paper aims to explain the outcome of a black box state-of-the-art deep learning model used for ABSC, LCR-Rot-hop++. We compare two sampling methods that feed an interpretability algorithm which is based on local linear approximations (LIME). One of the sampling methods, SS, swaps out different words from the original sentence with other similar words to create neighbours to the original sentence. The second method, SSb, uses SS and then filters its neighbourhood to better balance the sentiment proportions in the localities created. We use a 2016 restaurant reviews dataset for ternary classification and we judge the interpretability algorithms based on their hit rate and fidelity. We find that SSb can improve neighbourhood sentiment balance compared to SS, reducing bias for the majority class, while simultaneously increasing the performance of LIME.",

author = "Vlad Miron and Flavius Frasincar and Tru{\c s}cǎ, \{Maria Mihaela\}",

note = "Publisher Copyright: {\textcopyright} 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.; 28th International Conference on Applications of Natural Language to Information Systems, NLDB 2023 ; Conference date: 21-06-2023 Through 23-06-2023",

year = "2023",

doi = "10.1007/978-3-031-35320-8\_6",

language = "English",

isbn = "9783031353192",

series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",

publisher = "Springer Science+Business Media",

pages = "79--93",

editor = "Elisabeth M{\'e}tais and Farid Meziane and Warren Manning and Stephan Reiff-Marganiec and Vijayan Sugumaran",

booktitle = "Natural Language Processing and Information Systems - 28th International Conference on Applications of Natural Language to Information Systems, NLDB 2023, Proceedings",

address = "Germany",

}

Miron, V, Frasincar, F & Truşcǎ, MM 2023, Explaining a Deep Learning Model for Aspect-Based Sentiment Classification Using Post-hoc Local Classifiers. in E Métais, F Meziane, W Manning, S Reiff-Marganiec & V Sugumaran (eds), Natural Language Processing and Information Systems - 28th International Conference on Applications of Natural Language to Information Systems, NLDB 2023, Proceedings. Springer Science+Business Media, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 13913 LNCS, pp. 79-93, 28th International Conference on Applications of Natural Language to Information Systems, NLDB 2023, Derby, United Kingdom, 21/06/23. https://doi.org/10.1007/978-3-031-35320-8_6

Explaining a Deep Learning Model for Aspect-Based Sentiment Classification Using Post-hoc Local Classifiers. / Miron, Vlad; Frasincar, Flavius; Truşcǎ, Maria Mihaela.
Natural Language Processing and Information Systems - 28th International Conference on Applications of Natural Language to Information Systems, NLDB 2023, Proceedings. ed. / Elisabeth Métais; Farid Meziane; Warren Manning; Stephan Reiff-Marganiec; Vijayan Sugumaran. Springer Science+Business Media, 2023. p. 79-93 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Vol. 13913 LNCS).

Research output: Chapter/Conference proceeding › Chapter › Academic

TY - CHAP

T1 - Explaining a Deep Learning Model for Aspect-Based Sentiment Classification Using Post-hoc Local Classifiers

AU - Miron, Vlad

AU - Frasincar, Flavius

AU - Truşcǎ, Maria Mihaela

PY - 2023

Y1 - 2023

N2 - Aspect-Based Sentiment Classification (ABSC) models are increasingly utilised given the surge in opinionated text displayed on the Web. This paper aims to explain the outcome of a black box state-of-the-art deep learning model used for ABSC, LCR-Rot-hop++. We compare two sampling methods that feed an interpretability algorithm which is based on local linear approximations (LIME). One of the sampling methods, SS, swaps out different words from the original sentence with other similar words to create neighbours to the original sentence. The second method, SSb, uses SS and then filters its neighbourhood to better balance the sentiment proportions in the localities created. We use a 2016 restaurant reviews dataset for ternary classification and we judge the interpretability algorithms based on their hit rate and fidelity. We find that SSb can improve neighbourhood sentiment balance compared to SS, reducing bias for the majority class, while simultaneously increasing the performance of LIME.

AB - Aspect-Based Sentiment Classification (ABSC) models are increasingly utilised given the surge in opinionated text displayed on the Web. This paper aims to explain the outcome of a black box state-of-the-art deep learning model used for ABSC, LCR-Rot-hop++. We compare two sampling methods that feed an interpretability algorithm which is based on local linear approximations (LIME). One of the sampling methods, SS, swaps out different words from the original sentence with other similar words to create neighbours to the original sentence. The second method, SSb, uses SS and then filters its neighbourhood to better balance the sentiment proportions in the localities created. We use a 2016 restaurant reviews dataset for ternary classification and we judge the interpretability algorithms based on their hit rate and fidelity. We find that SSb can improve neighbourhood sentiment balance compared to SS, reducing bias for the majority class, while simultaneously increasing the performance of LIME.

UR - https://www.scopus.com/pages/publications/85164706662

U2 - 10.1007/978-3-031-35320-8_6

DO - 10.1007/978-3-031-35320-8_6

M3 - Chapter

AN - SCOPUS:85164706662

SN - 9783031353192

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 79

EP - 93

BT - Natural Language Processing and Information Systems - 28th International Conference on Applications of Natural Language to Information Systems, NLDB 2023, Proceedings

A2 - Métais, Elisabeth

A2 - Meziane, Farid

A2 - Manning, Warren

A2 - Reiff-Marganiec, Stephan

A2 - Sugumaran, Vijayan

PB - Springer Science+Business Media

T2 - 28th International Conference on Applications of Natural Language to Information Systems, NLDB 2023

Y2 - 21 June 2023 through 23 June 2023

ER -

Miron V, Frasincar F, Truşcǎ MM. Explaining a Deep Learning Model for Aspect-Based Sentiment Classification Using Post-hoc Local Classifiers. In Métais E, Meziane F, Manning W, Reiff-Marganiec S, Sugumaran V, editors, Natural Language Processing and Information Systems - 28th International Conference on Applications of Natural Language to Information Systems, NLDB 2023, Proceedings. Springer Science+Business Media. 2023. p. 79-93. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Vol. 13913 LNCS). doi: 10.1007/978-3-031-35320-8_6