Despite the widespread adoption, there is a lack of research into how various critical aspects of pretrained language models (PLMs) affect their performance in hate speech detection. Through five research questions, our findings and recommendations lay the groundwork for empirically investigating different aspects of PLMs’ use in hate speech detection. We deep dive into comparing different pretrained models, evaluating their seed robustness, finetuning settings, and the impact of pretraining data collection time. Our analysis reveals early peaks for downstream tasks during pretraining, the limited benefit of employing a more recent pretraining corpus, and the significance of specific layers during finetuning. We further call into question the use of domain-specific models and highlight the need for dynamic datasets for benchmarking hate speech detection.

Multi-document Summarization (MDS) characterizes compressing information from multiple source documents to its succinct summary. An ideal summary should encompass all topics and accurately model cross-document relations expounded upon in the source documents. However, existing systems either impose constraints on the length of tokens during the encoding or falter in capturing the intricate cross-document relationships. These limitations impel the systems to produce summaries that are non-factual and unfaithful, thereby imparting an unfair comprehension of the topic to the readers. To counter these limitations and promote the information equivalence between the source document and generated summary, we propose FIBER, a novel encoder-decoder model that uses pre-trained BART to comprehensively analyze linguistic nuances, simplicial complex layer to apprehend inherent properties that transcend pairwise associations and sheaf graph attention to effectively capture the heterophilic properties. We benchmark FIBER with eleven baselines over four widely-used MDS datasets – Multinews, CQASumm, DUC and Opinosis, and show that FIBER achieves consistent performance improvement across all the evaluation metrics (syntactical, semantical and faithfulness). We corroborate these improvements further through qualitative human evaluation.

The recent transition to the online educational domain has increased the need for Automatic Short Answer Grading (ASAG). ASAG automatically evaluates a student’s response against a (given) correct response and thus has been a prevalent semantic matching task. Most existing methods utilize sequential context to compare two sentences and ignore the structural context of the sentence; therefore, these methods may not result in the desired performance. In this paper, we overcome this problem by proposing a Multi-Relational Graph Transformer, MitiGaTe, to prepare token representations considering the structural context. Abstract Meaning Representation (AMR) graph is created by parsing the text response and then segregated into multiple subgraphs, each corresponding to a particular relationship in AMR. A Graph Transformer is used to prepare relation-specific token embeddings within each subgraph, then aggregated to obtain a subgraph representation. Finally, we compare the correct answer and the student response subgraph representations to yield a final score. Experimental results on Mohler’s dataset show that our system outperforms the existing state-of-the-art methods. We have released our implementation https://github.com/kvarun07/asag-gt, as we believe that our model can be useful for many future applications.

We investigate the problem of extracting Indian-locations from a given crowd-sourced textual dataset. The problem of extracting fine-grained Indian-locations has many challenges. One challenge in the task is to collect relevant dataset from the crowd-sourced platforms that contain locations. The second challenge lies in extracting the location entities from the collected data. We provide an in-depth review of the information collection process and our annotation guidelines such that a reliable dataset annotation is guaranteed. We evaluate many recent algorithms and models, including Conditional Random fields (CRF), Bi-LSTM-CNN and BERT (Bidirectional Encoder Representations from Transformers), on our developed dataset named . The study shows the best F1-score of 72.49% for BERT, followed by Bi-LSTM-CNN and CRF. As a result of our work, we prepare a publicly-available annotated dataset of Indian geolocations that can be used by the research community. Code and dataset are available at https://github.com/vkartik2k/STHAL.