Sentiment and topic analysis on social media

A multi-task multi-label classification approach

Shu Huang, Wei Peng, Jingxuan Li, Dongwon Lee

Research output: Chapter in Book/Report/Conference proceedingConference contribution

21 Citations (Scopus)

Abstract

Both sentiment analysis and topic classification are frequently used in customer care and marketing. They can help people understand the brand perception and customer opinions from social media, such as online posts, tweets, forums, and blogs. As such, in recent years, many solutions have been proposed for both tasks. However, we believe that the following two problems have not been addressed adequately: (1) Conventional solutions usually treat the two tasks in isolation. When the two tasks are closely related (e.g., posts about "customer care" often have a "negative" tone), exploring their correlation may yield a better accuracy; (2) Each post is usually assigned with only one sentiment label and one topic label. Since social media is, compared to traditional document corpus, more noisy, ambiguous, and sparser, single label classification may not be able to capture the post classes accurately. To address these two problems, in this paper, we propose a multi-task multi-label (MTML) classification model that performs classification of both sentiments and topics concurrently. It incorporates results of each task from prior steps to promote and reinforce the other iteratively. For each task, the model is trained with multiple labels so that they can help address class ambiguity. In the empirical validation, we compare the accuracy of MTML model against four competing methods in two different settings. Results show that MTML produces a much higher accuracy of both sentiment and topic classifications.

Original languageEnglish (US)
Title of host publicationProceedings of the 5th Annual ACM Web Science Conference, WebSci'13
PublisherAssociation for Computing Machinery
Pages172-181
Number of pages10
Volumevolume
ISBN (Print)9781450318891
DOIs
StatePublished - Jan 1 2013
Event3rd Annual ACM Web Science Conference, WebSci 2013 - Paris, France
Duration: May 2 2013May 4 2013

Other

Other3rd Annual ACM Web Science Conference, WebSci 2013
CountryFrance
CityParis
Period5/2/135/4/13

Fingerprint

Labels
Blogs
Marketing

All Science Journal Classification (ASJC) codes

  • Computer Networks and Communications

Cite this

Huang, S., Peng, W., Li, J., & Lee, D. (2013). Sentiment and topic analysis on social media: A multi-task multi-label classification approach. In Proceedings of the 5th Annual ACM Web Science Conference, WebSci'13 (Vol. volume, pp. 172-181). Association for Computing Machinery. https://doi.org/10.1145/2464464.2464512
Huang, Shu ; Peng, Wei ; Li, Jingxuan ; Lee, Dongwon. / Sentiment and topic analysis on social media : A multi-task multi-label classification approach. Proceedings of the 5th Annual ACM Web Science Conference, WebSci'13. Vol. volume Association for Computing Machinery, 2013. pp. 172-181
@inproceedings{e2b7671280b248f19656b567d378b02f,
title = "Sentiment and topic analysis on social media: A multi-task multi-label classification approach",
abstract = "Both sentiment analysis and topic classification are frequently used in customer care and marketing. They can help people understand the brand perception and customer opinions from social media, such as online posts, tweets, forums, and blogs. As such, in recent years, many solutions have been proposed for both tasks. However, we believe that the following two problems have not been addressed adequately: (1) Conventional solutions usually treat the two tasks in isolation. When the two tasks are closely related (e.g., posts about {"}customer care{"} often have a {"}negative{"} tone), exploring their correlation may yield a better accuracy; (2) Each post is usually assigned with only one sentiment label and one topic label. Since social media is, compared to traditional document corpus, more noisy, ambiguous, and sparser, single label classification may not be able to capture the post classes accurately. To address these two problems, in this paper, we propose a multi-task multi-label (MTML) classification model that performs classification of both sentiments and topics concurrently. It incorporates results of each task from prior steps to promote and reinforce the other iteratively. For each task, the model is trained with multiple labels so that they can help address class ambiguity. In the empirical validation, we compare the accuracy of MTML model against four competing methods in two different settings. Results show that MTML produces a much higher accuracy of both sentiment and topic classifications.",
author = "Shu Huang and Wei Peng and Jingxuan Li and Dongwon Lee",
year = "2013",
month = "1",
day = "1",
doi = "10.1145/2464464.2464512",
language = "English (US)",
isbn = "9781450318891",
volume = "volume",
pages = "172--181",
booktitle = "Proceedings of the 5th Annual ACM Web Science Conference, WebSci'13",
publisher = "Association for Computing Machinery",

}

Huang, S, Peng, W, Li, J & Lee, D 2013, Sentiment and topic analysis on social media: A multi-task multi-label classification approach. in Proceedings of the 5th Annual ACM Web Science Conference, WebSci'13. vol. volume, Association for Computing Machinery, pp. 172-181, 3rd Annual ACM Web Science Conference, WebSci 2013, Paris, France, 5/2/13. https://doi.org/10.1145/2464464.2464512

Sentiment and topic analysis on social media : A multi-task multi-label classification approach. / Huang, Shu; Peng, Wei; Li, Jingxuan; Lee, Dongwon.

Proceedings of the 5th Annual ACM Web Science Conference, WebSci'13. Vol. volume Association for Computing Machinery, 2013. p. 172-181.

Research output: Chapter in Book/Report/Conference proceedingConference contribution

TY - GEN

T1 - Sentiment and topic analysis on social media

T2 - A multi-task multi-label classification approach

AU - Huang, Shu

AU - Peng, Wei

AU - Li, Jingxuan

AU - Lee, Dongwon

PY - 2013/1/1

Y1 - 2013/1/1

N2 - Both sentiment analysis and topic classification are frequently used in customer care and marketing. They can help people understand the brand perception and customer opinions from social media, such as online posts, tweets, forums, and blogs. As such, in recent years, many solutions have been proposed for both tasks. However, we believe that the following two problems have not been addressed adequately: (1) Conventional solutions usually treat the two tasks in isolation. When the two tasks are closely related (e.g., posts about "customer care" often have a "negative" tone), exploring their correlation may yield a better accuracy; (2) Each post is usually assigned with only one sentiment label and one topic label. Since social media is, compared to traditional document corpus, more noisy, ambiguous, and sparser, single label classification may not be able to capture the post classes accurately. To address these two problems, in this paper, we propose a multi-task multi-label (MTML) classification model that performs classification of both sentiments and topics concurrently. It incorporates results of each task from prior steps to promote and reinforce the other iteratively. For each task, the model is trained with multiple labels so that they can help address class ambiguity. In the empirical validation, we compare the accuracy of MTML model against four competing methods in two different settings. Results show that MTML produces a much higher accuracy of both sentiment and topic classifications.

AB - Both sentiment analysis and topic classification are frequently used in customer care and marketing. They can help people understand the brand perception and customer opinions from social media, such as online posts, tweets, forums, and blogs. As such, in recent years, many solutions have been proposed for both tasks. However, we believe that the following two problems have not been addressed adequately: (1) Conventional solutions usually treat the two tasks in isolation. When the two tasks are closely related (e.g., posts about "customer care" often have a "negative" tone), exploring their correlation may yield a better accuracy; (2) Each post is usually assigned with only one sentiment label and one topic label. Since social media is, compared to traditional document corpus, more noisy, ambiguous, and sparser, single label classification may not be able to capture the post classes accurately. To address these two problems, in this paper, we propose a multi-task multi-label (MTML) classification model that performs classification of both sentiments and topics concurrently. It incorporates results of each task from prior steps to promote and reinforce the other iteratively. For each task, the model is trained with multiple labels so that they can help address class ambiguity. In the empirical validation, we compare the accuracy of MTML model against four competing methods in two different settings. Results show that MTML produces a much higher accuracy of both sentiment and topic classifications.

UR - http://www.scopus.com/inward/record.url?scp=84883117557&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84883117557&partnerID=8YFLogxK

U2 - 10.1145/2464464.2464512

DO - 10.1145/2464464.2464512

M3 - Conference contribution

SN - 9781450318891

VL - volume

SP - 172

EP - 181

BT - Proceedings of the 5th Annual ACM Web Science Conference, WebSci'13

PB - Association for Computing Machinery

ER -

Huang S, Peng W, Li J, Lee D. Sentiment and topic analysis on social media: A multi-task multi-label classification approach. In Proceedings of the 5th Annual ACM Web Science Conference, WebSci'13. Vol. volume. Association for Computing Machinery. 2013. p. 172-181 https://doi.org/10.1145/2464464.2464512