Domain Adaptation with Reconstruction for Disaster Tweet Classification

Li, Xukun; Caragea, Doina

doi:10.1145/3397271.3401242

Cited by 14 publications

(22 citation statements)

References 10 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The main empirical focus of our work is to measure the effectiveness of our approach when no labeled data is available from target domain. To this effect, considering an UDA setting, we pair up public datasets from disaster management as source (S) and target (T) for our experiments and supervise our model (refer to Section III-C) on only labeled instances from S and test on T. While previous works are restricted to datasets involving only binary classification [5], [6], we also report results on public datasets involving more realistic and useful multi-label multi-class classification setting. Additionally, to verify the effectiveness of our approach in limited labeled data scenarios we reduce the source domain training data.…”

Section: Methodsmentioning

confidence: 99%

“…• Feature projection [5], [8], [9] • Instance re-weighting [10]- [12] • Pivot feature centric [13], [14] • Domain Adversarial / Gradient Reversal based [6], [15]- [19] Feature projection signifies bringing the features of source and target domain to a joint latent space. [8] used stacked autoencoders to learn domain adaptive feature representations for sentiment analysis.…”

Section: Related Work a Unsupervised Domain Adaptation (Uda)mentioning

confidence: 99%

“…However, document level embedding might fail to capture the token-level distribution shift between source and target domain. SOTA performance in disaster scenario was achieved recently in [5] where the authors use a domain reconstruction classification network (DRCN). DRCN reduces the covariate shift by reconstructing the target domain data with an LSTM based autoencoder.…”

Section: B Uda For Disaster Responsementioning

confidence: 99%

“…We compare our approach against SOTA methods in disaster domain. For binary classification, we compare against the methods named DRCN and DAAT presented in [5] and [6] respectively. These methods, to the best of our knowledge, are the most recent methods with SOTA performance on NEQ and QFL datasets.…”

Section: B Baselines and Modelsmentioning

confidence: 99%

“…In this work, we focus on the problem of identification and categorization of actionable social media posts when no labeled data is available from the current disaster. Unsupervised Domain Adaptation (UDA) may be utilized in such a scenario if labeled data from a prior disaster is available [5]- [7]. However, even when labeled data from a past disaster is available, the amount of labeled data is generally minimal as annotating posts based on the actionable information is expensive.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Unsupervised Domain Adaptation with Global and Local Graph Neural Networks in Limited Labeled Data Scenario: Application to Disaster Management

Ghosh,

Maji,

Desarkar

2021

Preprint

View full text Add to dashboard Cite

Identification and categorization of social media posts generated during disasters are crucial to reduce the sufferings of the affected people. However, lack of labeled data is a significant bottleneck in learning an effective categorization system for a disaster. This motivates us to study the problem as unsupervised domain adaptation (UDA) between a previous disaster with labeled data (source) and a current disaster (target). However, if the amount of labeled data available is limited, it restricts the learning capabilities of the model. To handle this challenge, we utilize limited labeled data along with abundantly available unlabeled data, generated during a source disaster to propose a novel two-part graph neural network. The first-part extracts domain-agnostic global information by constructing a token level graph across domains and the second-part preserves local instance-level semantics. In our experiments, we show that the proposed method outperforms state-of-the-art techniques by 2.74% weighted F1 score on average on two standard public dataset in the area of disaster management. We also report experimental results for granular actionable multi-label classification datasets in disaster domain for the first time, on which we outperform BERT by 3.00% on average w.r.t weighted F1. Additionally, we show that our approach can retain performance when very limited labeled data is available. We plan to release our code publicly.

show abstract

Section: Methodsmentioning

confidence: 99%

Section: Related Work a Unsupervised Domain Adaptation (Uda)mentioning

confidence: 99%

Section: B Uda For Disaster Responsementioning

confidence: 99%

Section: B Baselines and Modelsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Unsupervised Domain Adaptation with Global and Local Graph Neural Networks in Limited Labeled Data Scenario: Application to Disaster Management

Ghosh,

Maji,

Desarkar

2021

Preprint

View full text Add to dashboard Cite

show abstract

Using Pseudo-Labelled Data for Zero-Shot Text Classification

Wang

Nulty

Lillis

2022

Natural Language Processing and Information Systems

View full text Add to dashboard Cite

Existing Zero-Shot Learning (ZSL) techniques for text classification typically assign a label to a piece of text by building a matching model to capture the semantic similarity between the text and the label descriptor. This is expensive at inference time as it requires the text paired with every label to be passed forward through the matching model. The existing approaches to alleviate this issue are based on exact-word matching between the label surface names and an unlabelled target-domain corpus to get pseudo-labelled data for model training, making them difficult to generalise to ZS classification in multiple domains, In this paper, we propose an approach called P-ZSC to leverage pseudo-labelled data for zero-shot text classification. Our approach generates the pseudo-labelled data through a matching algorithm between the unlabelled target-domain corpus and the label vocabularies that consist of in-domain relevant phrases via expansion from label names. By evaluating our approach on several benchmarking datasets from a variety of domains, the results show that our system substantially outperforms the baseline systems especially in datasets whose classes are imbalanced.

show abstract