Graph Based Network with Contextualized Representations of Turns in Dialogue

Abstract

Dialogue-based relation extraction (RE) aimsto extract relation(s) between two arguments that appear in a dialogue. Because dialogues have the characteristics of high personal pronoun occurrences and low information density, and since most relational facts in dialogues are not supported by any single sentence, dialogue-based relation extraction requires a comprehensive understanding of dialogue. In this paper, we propose the TUrn COntext awaRE Graph Convolutional Network (TUCORE-GCN) modeled by paying attention to the way people understand dialogues. In addition, we propose a novel approach which treats the task of emotion recognition in conversations (ERC) as a dialoguebased RE. Experiments on a dialogue-based RE dataset and three ERC datasets demonstrate that our model is very effective in various dialogue-based natural language understanding tasks.In these experiments, TUCORE-GCN outperforms the state-of-theart models on most of the benchmark datasets. Our code is available at https://github. com/BlackNoodle/TUCORE-GCN.

D-REX: Dialogue Relation Extraction with Explanations

“…Next, we see a five point absolute improvement in F1 from the baseline model when using RoBERTa. The trend from BERT to RoBERTa is similar to results found by Lee and Choi (2021), where changing from a BERT base model to RoBERTa Large (not shown here) improved their model performance significantly. Additionally, we see a 3 point improvement from R to D-REX when using RoBERTa (compared to 0.7 for BERT), which we believe is due to the better per- forming ranking model, which allows for D-REX to rely more on the input explanations.…”

“…There are no restrictions on R, it can be any algorithm which ranks relations (e.g., deep neural network, rule-based, etc.) such as (Yu et al, 2020;Lee and Choi, 2021). However, if R needs to be trained, it must be done prior to D-REX training; D-REX will not make any updates to R.…”

“…In Table 2, we compare the baseline RE model R BERT with the methods presented in this paper. We also compare with three other methods which use similarly sized language models, but additionally utilize graph neural networks (GNN): GDP-Net (Xue et al, 2021), TUCORE-GCN BERT (Lee and Choi, 2021), and SocAoG (Qiu et al, 2021).…”

“…Finally, there have been multiple recent works in dialogue RE which perform quite well by using graph neural networks (Xue et al, 2021;Qiu et al, 2021;Lee and Choi, 2021). However, they focus only on RE and not on explaining the relations.…”

“…Traditional relation extraction (RE) approaches discover relations that exist between entities within a single sentence. Recently, several approaches have been proposed which focus on cross-sentence RE, the task of extracting relations between entities that appear in separate sentences (Peng et al, 2017;Han and Wang, 2020;Yao et al, 2019) as well as cross-sentence RE in dialogues (Yu et al, 2020;Chen et al, 2020;Xue et al, 2021;Qiu et al, 2021;Lee and Choi, 2021).…”