Add Batch 10cae29d-b70b-420e-8671-60bd0d97be39

enablingmultimodalgenerationonclipviavisionlanguageknowledgedistillation/774f5cc8-5cfe-41fb-905e-9679f75d52b8_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be5f9863eaf30ea1929efcbb6f9b3a911f6dec5a1114c91033651bdc34d3958d
+size 103315

enablingmultimodalgenerationonclipviavisionlanguageknowledgedistillation/774f5cc8-5cfe-41fb-905e-9679f75d52b8_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:38774f37258e4d29773a5b004c11a3ce5bfa33ea42fa57287704bd96a528d2a6
+size 124405

enablingmultimodalgenerationonclipviavisionlanguageknowledgedistillation/774f5cc8-5cfe-41fb-905e-9679f75d52b8_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0b793fbce001f05f9c92994c8b0ceb87ef82aac1b1e5d9b12878a9c55c04025
+size 1401865

enablingmultimodalgenerationonclipviavisionlanguageknowledgedistillation/full.md

@@ -0,0 +1,500 @@
+# Enabling Multimodal Generation on CLIP via Vision-Language Knowledge Distillation
+
+Wenliang Dai $^{1}$ , Lu Hou $^{2}$ , Lifeng Shang $^{2}$ , Xin Jiang $^{2}$ , Qun Liu $^{2}$ , Pascale Fung $^{1}$
+
+$^{1}$ Hong Kong University of Science and Technology,  $^{2}$ Huawei Noah's Ark Lab
+
+wdaiai@connect.ust.hk, pascale@ece.ust.hk,
+
+{houlu3, shang.lifeng, jiang.xin, qun.liu} @huawei.com
+
+# Abstract
+
+The recent large-scale vision-language pretraining (VLP) of dual-stream architectures (e.g., CLIP) with a tremendous amount of image-text pair data, has shown its superiority on various multimodal alignment tasks. Despite its success, the resulting models are not capable of multimodal generative tasks due to the weak text encoder. To tackle this problem, we propose to augment the dual-stream VLP model with a textual pre-trained language model (PLM) via vision-language knowledge distillation (VLKD), enabling the capability for multimodal generation. VLKD is pretty data- and computation-efficient compared to the pre-training from scratch. Experimental results show that the resulting model has strong zero-shot performance on multimodal generation tasks, such as open-ended visual question answering and image captioning. For example, it achieves  $44.5\%$  zero-shot accuracy on the VQAv2 dataset, surpassing the previous state-of-the-art zero-shot model with  $7\times$  fewer parameters. Furthermore, the original textual language understanding and generation ability of the PLM is maintained after VLKD, which makes our model versatile for both multimodal and unimodal tasks.
+
+# 1 Introduction
+
+Recent large-scale dual-stream Vision-Language Pre-training (VLP) models like CLIP (Radford et al., 2021) and ALIGN (Jia et al., 2021), have shown remarkable performance on various downstream multimodal alignment tasks, e.g., image-text retrieval and image classification. These models are pre-trained using cross-modal contrastive learning on tremendous image-text pairs and learn strong multimodal representations. Despite their success, as mentioned by Radford et al. (2021), their text encoder is relatively weak by only having a discriminative multimodal pre-training objective,
+
+![](images/baf86575c8eeb3a556f985782d7913f3de4f33670f3fd5be3018906ec4d6b103.jpg)  
+Figure 1: Intuition of our proposed approach. After VLKD, the model can fill in the masked locations with meaningful words to describe the image without further finetuning. Moreover, it can answer questions with proper reasoning over the given images and pre-trained knowledge inside PLMs, e.g., a napkin is for wiping the face at meals.
+
+which makes them incompetent on generative multimodal tasks such as image captioning and open-ended visual question answering (VQA).
+
+Meanwhile, the Transformer-based (Vaswani et al., 2017) auto-regressive large-scale pre-trained language models (PLMs), such as GPT (Radford and Narasimhan, 2018; Brown et al., 2020), have been dominating in the natural language generation (NLG) tasks. These models are usually trained with causal self-attention, which only allows the model to attend to past outputs (unidirectional) to satisfy their generative nature. More recently, BART (Lewis et al., 2020) and T5 (Raffel et al., 2020) propose to augment the auto-regressive decoder with a bidirectional Transformer encoder to
+
+further capture bidirectional information of the input. These encoder-decoder architectures excel on not only NLG but also understanding (NU) tasks.
+
+To tackle the aforementioned limitations of dual-stream VLP models and fully utilize PLMs, in this paper, we present Vision-Language Knowledge Distillation (VLKD), a simple yet effective approach to enable CLIP to perform generative multimodal tasks through knowledge distillation. Specifically, we align the BART encoder to CLIP's joint multimodal embedding space to gain the understanding of multimodal knowledge, along with an image-conditioned language modeling loss to consort BART encoder and decoder. During training, we freeze CLIP's weights to keep its learned multimodal space. For the finetuning and inference of downstream tasks, the original CLIP text encoder is discarded, which can be interpreted as being replaced by the distilled BART. Therefore, we leverage the strengths from both sides, the expressive multimodal representation space of CLIP and the strong text generation capability of BART.
+
+Compared to VLP from scratch, VLKD uses several magnitudes fewer image-text pairs and computational resources. As depicted in Figure 1, after VLKD pre-training, the model exhibits strong zero-shot performance on generative multimodal tasks, including open-ended VQA and image captioning. Without finetuning, it has the ability to generate answers by reasoning over the question, the visual information, and the textual knowledge embedded in the pre-trained BART. Furthermore, it can also directly generate a plausible caption given an image. Empirical results show that our model achieves  $44.5\%$  accuracy on the VQAv2 dataset and 84.6 CIDEr on COCO image caption dataset in a zero-shot manner. Moreover, the original NLU and NLG ability of BART is maintained, which makes the model versatile for both multimodal and unimodal tasks.
+
+To summarize, our contributions are: 1) We introduce an efficient approach to distill knowledge from the dual-stream VLP model CLIP to BART. The resulting model shows strong zero-shot performance on generative multimodal tasks, as well as pure NLP tasks; 2) We exhaustively quantify these capabilities on six benchmarks under various settings; and 3) We conduct comprehensive analysis and ablation study to provide insights and grease future work on this direction.
+
+# 2 Related Work
+
+# 2.1 Vision-language Pre-training
+
+Based on how the two modalities interact, recent VLP models mainly fall into two categories: single-stream and dual-stream models. Single-stream models (Chen et al., 2020; Li et al., 2019; Ramesh et al., 2021; Lin et al., 2021; Kim et al., 2021a; Shen et al., 2022) concatenate the patch-wise or regional visual features and textual embeddings and feed them into a single model. Dual-stream models (Lu et al., 2019; Radford et al., 2021; Jia et al., 2021; Zhai et al., 2021; Yao et al., 2022) use separate encoders for images and texts, allowing efficient inference for downstream multimodal alignment tasks like image-text retrieval, by pre-computing image/text features offline. However, these models can not be directly used for multimodal generation tasks. In this paper, we propose an efficient method to align the dual-stream VLP model CLIP's multimodal embedding space with a powerful PLM BART to gain multimodal generation ability.
+
+There are also VLP models that can perform multimodal generation tasks, by expensive pretraining with objective of image-conditioned autoregressive language modeling (Lin et al., 2021; Wang et al., 2021; Hu et al., 2021; Li et al., 2022). However, the pre-training of these models requires a large number of image-text pairs and numerous computation resources. Other models like (Agrawal et al., 2019; Li et al., 2019, 2020; Cho et al., 2021; Li et al., 2021) rely on an extra pretrained object detector such as Faster-RCNN with labeled bounding-box data to extract image regional features offline and are less scalable.
+
+# 2.2 Knowledge Distillation
+
+Knowledge distillation (KD) in deep learning is first proposed by Hinton et al. (2015), which transfers knowledge embedded in the logits learned in a cumbersome teacher model to a smaller student model without sacrificing too much performance. Besides logits, other forms of knowledge like the intermediate representations and attentions (Jiao et al., 2019; Hou et al., 2020) have also been used in transferring the knowledge embedded in Transformer-based models. Recently, contrastive representation distillation (Tian et al., 2019) distills the knowledge from the teacher network to the student network by maximizing the mutual information between the two networks, and is recently extended to transfer the knowledge from the pre
+
+trained multimodal model CLIP for zero-shot detection (Gu et al., 2021) and multilingual setting (Jain et al., 2021). In this paper, we apply the conventional KD as well as the contrastive KD to transfer the knowledge from the pre-trained CLIP to BART. Besides, we also propose to transfer the knowledge in CLIP image encoder to BART decoder through the cross-attention.
+
+# 3 Proposed Method
+
+We propose to distill multimodal knowledge from CLIP to BART for generative multimodal tasks, which takes the strengths from both sides (powerful multimodal representations of CLIP and text generation ability of BART). To this end, we propose three objectives (Section 3.2). The overall architecture is illustrated in Figure 2.
+
+# 3.1 Model Architecture
+
+CLIP. CLIP (Radford et al., 2021) is a dual-stream VLP model pre-trained with a contrastive loss on 400 million image-text pairs. It consists of a text encoder which is a GPT (Radford et al., 2019) style Transformer model, and an image encoder which can be either a Vision Transformer (ViT) (Dosovitskiy et al., 2020) or Residual Convolutional Neural Network (ResNet) (He et al., 2016). CLIP learns a joint multimodal embedding space with its text encoder and image encoder aligned. Given an input image-text pair, the image encoder first reshapes the image into a sequence of 2D patches and then maps them into 1D embeddings with a prepended [CLS] token using a trainable linear projection. These embeddings are fed into the CLIP image encoder together with positional encodings. The output embedding of the [CLS] token can represent the whole image. For the text sentence, it is bracketed with [SOS] and [EOS] tokens, and the output embedding of the latter is used as the sentence-level representation. In this paper, we explore four CLIP variants, including ViT-B/16, ViT-L/14, RN50×16, and RN50×64.
+
+BART. BART is a Transformer-based (Vaswani et al., 2017) sequence-to-sequence model that has a bi-directional encoder and a uni-directional (left-to-right) decoder, which can be seen as a generalization of the BERT (Devlin et al., 2019) and GPT (Radford and Narasimhan, 2018). It is pretrained on 160GB text data in a self-supervised way by performing the text span infilling task with the input sentences corrupted and shuffled. Similar to
+
+the CLIP text encoder, BART also tokenizes and converts the input text into a sequence of embeddings, which are then fed into the BART encoder. BART excels at both NLG (e.g., abstractive summarization) and NLU tasks.
+
+# 3.2 Training Objectives
+
+To distill multimodal knowledge from CLIP to BART, we propose three objective functions: 1) Text-Text Distance Minimization (TTDM); 2) Image-Text Contrastive Learning (ITCL); and 3) Image-Conditioned Text Infilling (ICTI). During training, the model parameters of CLIP are frozen constantly, i.e. no gradients will be backpropagated through them (marked as SG in Figure 2), to ensure its two encoders are still aligned and the multimodal knowledge is not forgotten.
+
+For each training batch with  $B$  image-text pairs, denote the  $k$ -th image-text pair as  $\mathbf{x}^k = \{\mathbf{x}_I^k,\mathbf{x}_T^k\}$ , and the output of multimodal encoders of CLIP and BART encoder as
+
+$$
+\operatorname {C L I P} _ {I} \left(\mathbf {x} _ {I} ^ {k}\right)\rightarrow \mathbf {V} ^ {k} = \left[ \mathbf {v} _ {c l s} ^ {k}, \mathbf {v} _ {1} ^ {k}, \dots , \mathbf {v} _ {n _ {1}} ^ {k} \right],
+$$
+
+$$
+\operatorname {C L I P} _ {T} \left(\mathbf {x} _ {T} ^ {k}\right)\rightarrow \mathbf {T} ^ {k} = \left[ \mathbf {t} _ {s o s} ^ {k}, \mathbf {t} _ {1} ^ {k}, \dots , \mathbf {t} _ {n _ {2}} ^ {k}, \mathbf {t} _ {e o s} ^ {k} \right],
+$$
+
+$$
+\mathrm {B A R T} _ {e n c} (\mathbf {x} _ {T} ^ {k}) \rightarrow \mathbf {E} ^ {k} = [ \mathbf {e} _ {b o s} ^ {k}, \mathbf {e} _ {1} ^ {k}, \dots , \mathbf {e} _ {n _ {3}} ^ {k}, \mathbf {e} _ {e o s} ^ {k} ].
+$$
+
+Here,  $n_1$  is the number of image patches,  $n_2$  and  $n_3$  denote the sequence lengths of the text encoder of CLIP and BART, respectively.  $\mathbf{v}_{*}^{k}, \mathbf{t}_{*}^{k} \in \mathbb{R}^{d_{1}}$  represents the  $\ell_2$ -normalized output embedding from the CLIP image and text encoder at a certain position.  $\mathbf{e}_{*}^{k}$  is the unnormalized raw output embedding from the BART encoder. In the following, we elaborate on the three distillation objectives.
+
+# 3.2.1 Text-Text Distance Minimization
+
+To align the CLIP text encoder and BART encoder, i.e. making their output representations close given the same input text, we propose to minimize the  $\ell_2$  distance between their sequence-level output representations. Specifically, for the  $k$ -th input text, it can be formulated as
+
+$$
+\bar {\mathbf {e}} _ {\text {n o r m}} ^ {k} = \mathbf {W} _ {e} \bar {\mathbf {e}} ^ {k} / \| \mathbf {W} _ {e} \bar {\mathbf {e}} ^ {k} \| _ {2},
+$$
+
+$$
+\mathcal {L} _ {T T D M} = \frac {1}{B} \sum_ {k = 1} ^ {B} \| \mathbf {t} _ {e o s} ^ {k} - \bar {\mathbf {e}} _ {\mathrm {n o r m}} ^ {k} \| ^ {2},
+$$
+
+where  $\bar{\mathbf{e}}^k\in \mathbb{R}^{d_2}$  is the average of all output embeddings from the BART encoder, and  $\mathbf{W}_e\in \mathbb{R}^{d_1\times d_2}$  is a weight matrix to linearly project the output of BART encoder to CLIP's multimodal space.
+
+![](images/7875b006513f23347547c68a38d7760044d1cc2493d914963d6a332b4e35ee58.jpg)  
+(a) The TTDM and ITCL losses.
+
+![](images/cb4a89b2fae35db9ea68842d5e327a23cca81816844cbb639747cc15f0431bfe.jpg)  
+(b) The ICTI loss.  
+Figure 2: Architecture of the proposed VLKD method to distill multimodal knowledge from CLIP to BART. (a) shows the  $TTDM$  and  $ITCL$  losses between the dual-stream CLIP encoders and BART encoder. (b) illustrates the ICTI loss for image-conditioned language modeling. SG denotes the stop gradient operation, indicating that no gradients will be back-propagated through that part of model parameters.
+
+# 3.2.2 Image-Text Contrastive Learning
+
+Contrastive training has been shown to be very effective in cross-modal representation learning (Tian et al., 2020; Sigurdsson et al., 2020; Zhang et al., 2020; Radford et al., 2021). To further adapt the BART encoder to CLIP's multimodal space, we optimize a symmetric InfoNCE loss between the output representations of the BART encoder and CLIP image encoder. The image-to-text contrastive loss  $\mathcal{L}_{i2t}$  is formulated as
+
+$$
+\mathcal {L} _ {i 2 t} = - \frac {1}{B} \sum_ {k = 1} ^ {B} \log \frac {\exp \big (\mathbf {v} _ {c l s} ^ {k \top} \bar {\mathbf {e}} _ {\mathrm {n o r m}} ^ {k} / \tau \big)}{\sum_ {j} \exp \big (\mathbf {v} _ {c l s} ^ {k \top} \bar {\mathbf {e}} _ {\mathrm {n o r m}} ^ {j} / \tau \big)},
+$$
+
+where  $\tau$  is a learnable temperature parameter. Different from Radford et al. (2021), we find that not clamping the  $\tau$  shows a slight improvement. Similarly, the text-to-image contrastive loss  $\mathcal{L}_{t2i}$  is
+
+$$
+\mathcal {L} _ {t 2 i} = - \frac {1}{B} \sum_ {k = 1} ^ {B} \log \frac {\exp \left(\mathbf {v} _ {c l s} ^ {k \top} \bar {\mathbf {e}} _ {\mathrm {n o r m}} ^ {k} / \tau\right)}{\sum_ {j} \exp \left(\mathbf {v} _ {c l s} ^ {j \top} \bar {\mathbf {e}} _ {\mathrm {n o r m}} ^ {k} / \tau\right)}.
+$$
+
+Then, the ITCL loss can be calculated as
+
+$$
+\mathcal {L} _ {I T C L} = \frac {1}{2} \big (\mathcal {L} _ {i 2 t} + \mathcal {L} _ {t 2 i} \big).
+$$
+
+Note that when computing the ITCL and TTDM losses, we do not introduce any new linear projections to the CLIP output features to avoid destroying the pre-trained alignment between its image and text encoders. Instead, we add one linear layer (parameterized by  $\mathbf{W}_e$ ) to project the BART encoder to CLIP's representation space and match their feature dimension.
+
+# 3.2.3 Image-Conditioned Text Infilling
+
+With only TTDM and ITCL, the BART decoder is not updated at all. To consort BART encoder and decoder, we propose to perform the text span infilling task conditioned on the corresponding image features. As depicted in Figure 2b, for the  $k$ -th image-text pair, following Lewis et al. (2020), we corrupt the input text by masking  $15\%$  of whole-word tokens with span lengths drawn from a Poisson Distribution with  $\lambda = 3$ .
+
+Considering that  $\mathbf{V}^k$  and  $\mathbf{W}_e\mathbf{E}^k$  are already aligned in the CLIP's multimodal space through TTDM and ITCL, and having a different feature dimension with the BART decoder, we further project them to the BART decoder dimension with  $\mathbf{W}_i$  and  $\mathbf{W}_e^\prime$ . Then, we concatenate them together as  $\mathbf{C}^k$  before feeding into the BART decoder as shown in Eq.(1). As mentioned in Section 3.1, we explore two variants of CLIP. With a slight abuse of notation, for ResNet-based CLIP,  $\mathbf{V}^k$  is composed of representations of all image patches  $\{\mathbf{v}_i^k\}_{i=1}^{n_1}$  while for ViT-based CLIP,  $\mathbf{V}^k$  consists of the representation of the [CLS] token  $\mathbf{v}_{cls}^k$  only.
+
+Note that the weight matrix  $\mathbf{W}_e^{\prime}$  is initialized to be the pseudo-inverse of  $\mathbf{W}_e$ , such that text representations after the two projections  $\mathbf{W}_e^{\prime}\mathbf{W}_e\mathbf{E}^k$  are the closest to the original pre-trained BART encoder space at initialization<sup>1</sup>. The BART decoder then interacts with  $\mathbf{C}^k$  through standard Transformer cross-attention layers. We optimize a lan
+
+guage modeling loss  $\mathcal{L}_{ICTI}$  by minimizing the negative log-likelihood in Eq.(2), in which  $\mathbf{w}_j$  denotes the token to be predicted at each decoding step.
+
+$$
+\mathbf {C} ^ {k} = \operatorname {c o n c a t} \left(\mathbf {W} _ {i} \mathbf {V} ^ {k}, \mathbf {W} _ {e} ^ {\prime} \mathbf {W} _ {e} \mathbf {E} ^ {k}\right), \tag {1}
+$$
+
+$$
+\mathcal {L} _ {I C T I} = - \frac {1}{B} \sum_ {k = 1} ^ {B} \sum_ {j} \log P \left(\mathbf {w} _ {j} ^ {k} \mid \mathbf {w} _ {<   j} ^ {k}, \mathbf {C} ^ {k}\right). (2)
+$$
+
+The ICTI loss is crucial for our methodology to work, as it not only coordinates the BART encoder and decoder, but also enables the BART decoder to understand the multimodal information by recovering texts with visual clues.
+
+Finally, we simultaneously optimize the summation of three losses  $\mathcal{L}$  as
+
+$$
+\mathcal {L} = \gamma \mathcal {L} _ {T T D M} + \mathcal {L} _ {I T C L} + \mathcal {L} _ {I C T I},
+$$
+
+where  $\gamma$  is set to  $10^{3}$  by default, as  $\mathcal{L}_{ITCL},\mathcal{L}_{ICTI}$  are about three magnitudes larger than  $\mathcal{L}_{TTDM}$ .
+
+# 3.3 Datasets for VLKD
+
+Our model is trained on the Conceptual Captions (CC3M) (Sharma et al., 2018) dataset, which contains 3 million image-text pairs crawled from the Internet. For larger model variants (ViT-L/14 and RN50x64), we further include the Visual Genome Caption data which contains  $\sim 700\mathrm{K}$  image-text pairs. No images for pre-training appear in the downstream datasets. Compared to previous VLP work (Radford et al., 2021; Jia et al., 2021; Wang et al., 2021), VLKD is much cheaper by leveraging several magnitudes less data. Furthermore, we experiment with even smaller data (1M, 100K) by uniformly sampling a subset of CC3M to test the limit of dataset size of VLKD, with results discussed in Section 5.
+
+# 4 Experiments
+
+To demonstrate the effectiveness of VLKD, we evaluate it on generative multimodal tasks for both zero-shot and finetuning. Specifically, we test the image captioning task, and also the VQA task under the open-ended scenario. Furthermore, we also run the model on NLU and NLG tasks to investigate the influence of VLKD on the text processing ability of the original pre-trained BART.
+
+# 4.1 Finetuning Datasets
+
+Image Captioning. Image captioning requires the model to generate a relevant description given an image. We use the COCO image caption
+
+dataset (Lin et al., 2014) with the Karpathy split (Karpathy and Fei-Fei, 2017). Additionally, we use the NoCaps (Agrawal et al., 2019) dataset to test the model performance when there are out-of-domain objects.
+
+Open-Ended VQA. Unlike previous works (Anderson et al., 2018; Chen et al., 2020; Li et al., 2020; Yu et al., 2021a; Zhang et al., 2021; Kim et al., 2021b) that treat the VQA task as a discriminative problem, we let the model generate answers freely, which is more aligned with the real-world scenario of this task. We use the standard VQAv2 (Goyal et al., 2017), and also OK-VQA (Marino et al., 2019) which requires knowledge to answer questions correctly.
+
+NLU and NLG. For NLU, we test our model on the GLUE benchmark (Wang et al., 2019), which consists of nine text classification tasks. We exclude the WNLI task as it is problematic<sup>2</sup>. For NLG, we test the abstractive summarization task on XSUM (Narayan et al., 2018) dataset, which requires the model to comprehend long texts and generate short summaries with key information.
+
+# 4.2 Implementation Details
+
+We use BART-large as the pre-trained backbone NLP model, which has 12 layers in both encoder and decoder with a hidden size of 1024 and 16 heads in each multi-head attention (MHA) layer. In total, it contains 406M parameters. For the pre-trained CLIP (Radford et al., 2021) model, we report four variants with different visual backbones, including ViT-B/16, ViT-L/14,  $\mathrm{RN}50\times 16$  and  $\mathrm{RN}50\times 64$ .
+
+We use 64 Nvidia V100 GPUs for VLKD and 8 for the finetuning of downstream tasks. In total, we pre-train the model for 10 epochs, which takes about 5 hours. We use a batch size of 4608 for ViT-B/16 and ViT-L/14, 4096 for RN50x16 and 3840 for RN50x64. All of the models are optimized by the AdamW (Loshchilov and Hutter, 2019) optimizer. The learning rate is warmed up to  $2.4e^{-4}$  within the first  $2\%$  steps and then linearly decay to 0. More information of VLKD pre-training and the finetuning of each downstream task can be found in Appendix A.
+
+![](images/9edbc86c6d239dfa02a0cf0e4237a748151d7e51b41b9a13f21a708854b5f5fd.jpg)
+
+On what holiday do people traditionally eat this bird? Answer: [MASK].
+
+Generated answer: Thanksgiving.
+
+![](images/88cf79a10f52f98f65b4d9b44fe63888a39314f7b59c1a75b39895cd907a5d91.jpg)
+
+What retractable appendage could this animal use to destroy the chair? Answer: [MASK].
+
+Generated answer: Claw.
+
+![](images/4c0b5e0c2a7fefce51054188ac5aef5395508772ea11ad98dfa6253d225cf0b7.jpg)
+
+What area of a school might this be? Answer: [MASK].
+
+Generated answer: Library
+
+![](images/1a667f53916fd562dd2f5bee3aad1d69d06e427d8151767b6d92912712063d5c.jpg)  
+(a) Zero-shot VQA.
+
+What's reflecting from the mirror?
+
+Candidate answer(s): Light; Wall; Shower.
+
+Generated answer: Light.
+
+![](images/e0334fada3ea7e348ae0006bb26978e15db68244cb0d310c562933887f665cf0.jpg)
+
+Reference caption:
+
+Two people sit on the beach with surfboards at their sides.
+
+Generated caption:
+
+A couple sitting on the beach with their surfboards in the background.
+
+![](images/6611ba1106338d324c3ad1cce6c55ee531de69f3fee489c88435ef5cb331f6c0.jpg)
+
+Reference caption:
+
+A cat is laying next to a blue book.
+
+Generated caption:
+
+A cat reading a book on a couch in the living room.
+
+![](images/b8f99d9e193917a96e127bf51a8e39099c07551d37a9f9cdbdec11fbda5e4d93.jpg)
+
+Reference caption:
+
+A woman sitting on a bench with a dog.
+
+Generated caption:
+
+A young woman sitting on a bench with her dog in the background.
+
+![](images/2cbc9f00030ddd6cd2ffc25703a498e67eb715c41e8d40a052cc58bbe9c3a479.jpg)  
+(b) Zero-shot image captioning.  
+Figure 3: Examples of (a) zero-shot VQA and (b) image captioning. Our model shows the ability to recognize visual objects and generate appropriate sentences based on their properties and relationship. Furthermore, the model can bind visual objects to text conceptual knowledge that is learned in the PLMs when generating answers given questions.
+
+Reference caption:
+
+A man holds a stick during a hockey game.
+
+Generated caption:
+
+A young man in the middle of a hockey game.
+
+# 4.3 Multimodal Zero-Shot Evaluation
+
+Benefit from the knowledge distillation, especially the ICTI loss, our model can perform various downstream multimodal tasks in a zero-shot manner.
+
+# 4.3.1 Zero-Shot Image Captioning
+
+During knowledge distillation, the ICTI loss can be seen as a simple version of the image captioning task, which asks the model to fill in the corrupted locations of image descriptions. If the masking ratio increases to  $100\%$ , it reduces to the image captioning task. Therefore, it is intuitive to test the zero-shot performance of our model.
+
+Following Radford et al. (2021) and Wang et al. (2021), we compose the input with a text prompt and also  $m$  mask tokens, i.e., "A picture of  $[\mathrm{MASK}] \times m$ ," for the model to generate the caption for the image. The zero-shot results are included in Table 1. Our zero-shot model achieves comparable overall performance to the finetuned UpDown (Agrawal et al., 2019) model on NoCaps dataset. As shown in Figure 3b, the zero-shot generated captions are plausible with correct objects, relationships, and actions. However, sometimes details like colors could be omitted.
+
+In our experiments, we use  $m = 6$  for COCO and  $m = 8$  for NoCaps. Although it could poten-
+
+tially limit the length of generation, we find that it has negligible influence to the performance, as for each [MASK] token, the model is learned to fill one to three tokens depending on the context. Furthermore, this could be used to control the length of generated texts for different scenarios. See Section 5 for a more detailed discussion about the effects of number of the masks.
+
+# 4.3.2 Zero-Shot VQA
+
+Zero-shot VQA is much more challenging than image captioning, as it requires reasoning over both the image and question, which is very different from the ICTI loss during the knowledge distillation. As illustrated in Figure 1, we construct the input by appending a text prompt "Answer:  $[\mathrm{MASK}]\times n.$ " to the question Given the context (image+question+prompt), the model is required to predict the answer by recovering the textual token in the [MASK] positions. In our experiments, we use  $n = 2$  for the VQAv2, which is found performing best among  $n\in \{1,2,3\}$ .
+
+In Table 2, compared to the strong baseline Frozen (Tsimpoukelli et al., 2021), our model improves the zero-shot accuracy by  $13.1\%$  on the VQAv2 validation set and  $7.4\%$  on the OK-VQA test set with  $7\times$  fewer parameters, indicating the
+
+<table><tr><td>Methods</td><td>#Pretrain
+Image-text
+Pairs</td><td>OD</td><td>OT</td><td colspan="4">COCO Caption
+Karpathy Test</td><td>In
+C</td><td>S</td><td colspan="4">NoCaps Validation
+Near</td><td>Out
+C</td><td>S</td><td>Overall
+C</td><td>S</td></tr><tr><td>BUTD†</td><td>1.5M</td><td>✓</td><td>✓</td><td>36.3</td><td>120.1</td><td>27.7</td><td>21.4</td><td>80.0</td><td>12.0</td><td>73.6</td><td>11.3</td><td>66.4</td><td>9.7</td><td>73.1</td><td>11.1</td><td></td><td></td></tr><tr><td>OSCAR†Large</td><td>6.5M</td><td>✓</td><td>✓</td><td>41.7</td><td>140.0</td><td>30.6</td><td>24.5</td><td>85.4</td><td>11.9</td><td>84.0</td><td>11.7</td><td>80.3</td><td>10.0</td><td>83.4</td><td>11.4</td><td></td><td></td></tr><tr><td>VinVLLarge</td><td>6.5M</td><td>✓</td><td>✓</td><td>41.0</td><td>140.9</td><td>31.1</td><td>25.2</td><td>103.7</td><td>13.7</td><td>95.6</td><td>13.4</td><td>83.8</td><td>11.9</td><td>94.3</td><td>13.1</td><td></td><td></td></tr><tr><td>VL-T5</td><td>9.2M</td><td>✓</td><td>X</td><td>34.6</td><td>116.1</td><td>28.8</td><td>21.9</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td>VL-BART</td><td>9.2M</td><td>✓</td><td>X</td><td>34.2</td><td>114.1</td><td>28.4</td><td>21.3</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td>LEMONHuge</td><td>203M</td><td>✓</td><td>✓</td><td>42.6</td><td>145.5</td><td>31.4</td><td>25.5</td><td>118.0</td><td>15.4</td><td>116.3</td><td>15.1</td><td>120.2</td><td>14.5</td><td>117.3</td><td>15.0</td><td></td><td></td></tr><tr><td>SIMVLMHuge</td><td>1.8B</td><td>X</td><td>X</td><td>40.6</td><td>143.3</td><td>33.7</td><td>25.4</td><td>113.7</td><td>-</td><td>110.9</td><td>-</td><td>115.2</td><td>-</td><td>112.2</td><td>-</td><td></td><td></td></tr><tr><td>VLKD (Zero-shot)</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>ViT-B/16</td><td>3M</td><td>X</td><td>X</td><td>16.7</td><td>58.3</td><td>19.7</td><td>13.4</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td>RN50×16</td><td>3M</td><td>X</td><td>X</td><td>18.2</td><td>61.1</td><td>20.8</td><td>14.5</td><td>52.6</td><td>9.7</td><td>52.9</td><td>9.6</td><td>58.6</td><td>9.3</td><td>54.0</td><td>9.6</td><td></td><td></td></tr><tr><td>RN50×64</td><td>3.7M</td><td>X</td><td>X</td><td>25.8</td><td>85.1</td><td>23.1</td><td>16.9</td><td>64.8</td><td>13.6</td><td>62.3</td><td>13.6</td><td>66.9</td><td>9.9</td><td>63.6</td><td>12.8</td><td></td><td></td></tr><tr><td>VLKD (Finetuned)</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>ViT-B/16</td><td>3M</td><td>X</td><td>X</td><td>37.2</td><td>128.0</td><td>28.8</td><td>22.4</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td><td></td></tr><tr><td>RN50×16</td><td>3M</td><td>X</td><td>X</td><td>38.9</td><td>131.1</td><td>29.6</td><td>23.9</td><td>92.3</td><td>12.6</td><td>82.0</td><td>11.8</td><td>70.3</td><td>10.4</td><td>81.1</td><td>11.7</td><td></td><td></td></tr><tr><td>RN50×64</td><td>3.7M</td><td>X</td><td>X</td><td>40.3</td><td>135.7</td><td>30.5</td><td>24.3</td><td>105.1</td><td>14.5</td><td>99.7</td><td>13.8</td><td>90.2</td><td>12.1</td><td>97.6</td><td>13.6</td><td></td><td></td></tr></table>
+
+Table 1: Results on the COCO caption (Karpathy test set) and NoCaps (validation set). B@4, C, M, and S denote BLEU-4, CIDEr, METEOR, and SPICE, respectively. OD and OT indicate whether object detectors and object tags are used or not. Numbers of previous models are taken from (Anderson et al., 2018; Li et al., 2020; Zhang et al., 2021; Cho et al., 2021; Hu et al., 2021; Wang et al., 2021). Models marked by  $\dagger$  additionally use the constrained beam search (CBS) (Anderson et al., 2017) for the NoCaps dataset. Note that LEMON and SIMVLM use significantly more pre-training data and have more trainable model parameters than the others.  
+
+<table><tr><td>Methods</td><td>#Params</td><td>VQAv2 val / test-dev</td><td>OK-VQA test</td></tr><tr><td colspan="4">Generative (Open-ended)</td></tr><tr><td>Frozen (Zero-shot)</td><td rowspan="2">7B</td><td>29.5 / -</td><td>5.9</td></tr><tr><td>Frozen (Finetuned)</td><td>48.4 / -</td><td>19.6</td></tr><tr><td>VLKD (Zero-shot)</td><td></td><td></td><td></td></tr><tr><td>RN50×16</td><td></td><td>37.4 / 38.2</td><td>9.9</td></tr><tr><td>ViT-B/16</td><td></td><td>38.6 / 39.7</td><td>10.5</td></tr><tr><td>ViT-L/14</td><td>&lt; 1B</td><td>42.6 / 44.5</td><td>13.3</td></tr><tr><td>VLKD (Finetuned)</td><td></td><td></td><td></td></tr><tr><td>RN50×16</td><td></td><td>67.4 / 68.8</td><td>36.2</td></tr><tr><td>ViT-B/16</td><td></td><td>69.3 / 69.8</td><td>36.3</td></tr><tr><td>ViT-L/14</td><td></td><td>73.9 / 74.5</td><td>39.0</td></tr><tr><td colspan="4">Discriminative</td></tr><tr><td>UNITERLarge</td><td>-</td><td>- / 73.8</td><td>-</td></tr><tr><td>OSCARLarge</td><td>-</td><td>- / 73.6</td><td>-</td></tr><tr><td>VinVLLarge</td><td>-</td><td>- / 76.5</td><td>-</td></tr><tr><td>SIMVLMBase</td><td>-</td><td>- / 77.9</td><td>-</td></tr></table>
+
+efficiency and effectiveness of VLKD. Our model achieves  $44.5\%$  zero-shot accuracy on the VQAv2 test-dev set, which to the best of our knowledge is the new state-of-the-art. Furthermore, as shown in Figure 3a, our model can bind visual objects to conceptual knowledge stored in the PLM to answer
+
+Table 2: Accuracies(%) on the VQAv2 and OK-VQA datasets. We categorize models into two parts: answer questions in a generative or discriminative way.  
+
+<table><tr><td>Model</td><td>In-domain</td><td>Out-of-domain</td></tr><tr><td>UNITER</td><td>74.4</td><td>10.0</td></tr><tr><td>VL-T5</td><td>71.4</td><td>13.1</td></tr><tr><td>VL-BART</td><td>72.1</td><td>13.2</td></tr><tr><td>VLKD (ViT-L/14)</td><td>74.9</td><td>23.4</td></tr></table>
+
+Table 3: Accuracies(%) on VQAv2 Karpathy test-split.
+
+questions. For example, it connects the visual object Turkey with the traditional food people usually eat at the Thanksgiving festival.
+
+# 4.4 Multimodal Finetuning Evaluation
+
+When finetuning VLKD on downstream multimodal tasks, we keep the same input format as zero-shot to obtain outputs in a generative way. The CLIP model parameters are still frozen during finetuning.
+
+# 4.4.1 Finetuning Image Captioning
+
+In Table 1, we demonstrate that our model can achieve decent performance when finetuned on the COCO dataset. The SCST CIDEr optimization method (Rennie et al., 2017) is used to further improve the performance. Our model outperforms VL-T5/BART (Cho et al., 2021) without using an extra object detector, which is fairly time-consuming as explained by Kim et al. (2021b). Compared to state-of-the-art models, however,
+
+<table><tr><td>Model</td><td>CoLA</td><td>SST-2</td><td>RTE</td><td>MRPC</td><td>QQP</td><td>MNLI</td><td>QNLI</td><td>Avg.</td></tr><tr><td>BERTLARGE (Devlin et al., 2019)</td><td>60.6</td><td>93.2</td><td>70.4</td><td>82.9/88.0</td><td>91.3/87.9</td><td>86.4</td><td>92.3</td><td>82.6</td></tr><tr><td>BARTLARGE (Lewis et al., 2020)</td><td>62.8</td><td>96.6</td><td>87.0</td><td>86.7/90.4</td><td>92.5/89.3</td><td>90.0</td><td>94.9</td><td>87.2</td></tr><tr><td>VisualBERT† (Li et al., 2019)</td><td>38.6</td><td>89.4</td><td>56.6</td><td>71.9/82.1</td><td>89.4/86.0</td><td>81.6</td><td>87.0</td><td>74.0</td></tr><tr><td>UNITER† (Chen et al., 2020)</td><td>37.4</td><td>89.7</td><td>55.6</td><td>69.3/80.3</td><td>89.2/85.7</td><td>80.9</td><td>86.0</td><td>73.1</td></tr><tr><td>VL-BERT† (Su et al., 2020)</td><td>38.7</td><td>89.8</td><td>55.7</td><td>70.6/81.8</td><td>89.0/85.4</td><td>81.2</td><td>86.3</td><td>73.6</td></tr><tr><td>VilBERT† (Lu et al., 2019)</td><td>36.1</td><td>90.4</td><td>53.7</td><td>69.0/79.4</td><td>88.6/85.0</td><td>79.9</td><td>83.8</td><td>72.1</td></tr><tr><td>LXMERT† (Tan and Bansal, 2019)</td><td>39.0</td><td>90.2</td><td>57.2</td><td>69.8/80.4</td><td>75.3/75.3</td><td>80.4</td><td>84.2</td><td>71.6</td></tr><tr><td>SIMVLM‡ (Wang et al., 2021)</td><td>46.7</td><td>90.9</td><td>63.9</td><td>75.2/84.4</td><td>90.4/87.2</td><td>83.4</td><td>88.6</td><td>77.4</td></tr><tr><td>VLKD (RN50×16)</td><td>59.1</td><td>95.5</td><td>81.2</td><td>87.5/91.1</td><td>92.1/89.2</td><td>89.6</td><td>94.3</td><td>85.7</td></tr></table>
+
+there is still a small performance gap, which we conjecture is mainly due to their usage of object detector-tags and much more pre-training image-text pairs. We also evaluate our VLKD models with ResNet visual backbones on the NoCaps dataset (Table 1). For zero-shot image caption, the CIDEr score on the out-of-domain set is even higher than the in- and near-domain sets, which shows the generalization of our knowledge distillation method to common visual objects. After finetuned on the COCO training set, the performance on NoCaps of our model with the  $\mathrm{RN}50\times 64$  backbone is comparable to the state-of-the-art models.
+
+# 4.4.2 Finetuning VQA
+
+From Table 2, the best performance of VQAv2 is achieved by VLP models that tackle this task in a discriminative way with a set of pre-defined answers. However, this approach does not generalize to real-world scenarios and cannot be directly applied to more diverse datasets (e.g., OK-VQA). Differently, Frozen (Tsimpoukelli et al., 2021) and our proposed VLKD formulate VQA as a generative problem to generate answers conditioned on the questions and images in an open-ended manner, which also enables zero-shot VQA. Specifically, for each question-answer pair in the VQAv2 dataset, we optimize the model to generate the answer with the cross-entropy loss and a label-smoothing of 0.1. The loss is weighted by the weight of each answer candidate. In addition, we augment the training data with VG-QA (Krishna et al., 2016).
+
+Furthermore, following (Cho et al., 2021), we test the performance on out-of-domain questions with rare answers using the Karpathy test-split. As
+
+Table 4: Results on the GLUE development set (single task single models). We report the Matthews correlation for CoLA, accuracy/F1 for MRPC and QQP, and accuracy for the rest of the tasks. The performance of models that are marked by  $\diamond$  are taken from (Lewis et al., 2020),  $\dagger$  are from (Iki and Aizawa, 2021), and  $\ddagger$  are from (Wang et al., 2021). Compared to other VLP models, our VLKD model has a great advantage in text-only NLP tasks.  
+
+<table><tr><td>Model</td><td>ROUGE-1</td><td>ROUGE-2</td><td>ROUGE-L</td></tr><tr><td>BARTLarge</td><td>45.14</td><td>22.27</td><td>37.25</td></tr><tr><td>VLKD</td><td>44.86</td><td>22.06</td><td>36.95</td></tr></table>
+
+Table 5: Results of abstractive summarization on XSUM. We use the best performing checkpoint of the  $\mathrm{RN}50\times 16$  variant.
+
+shown in Table 3, our method shows a salient advantage on out-of-domain questions due to the benefit from VLKD and its generative nature without defining the answer list.
+
+# 4.5 Evaluation of NLU and NLG
+
+Table 4 shows results on the GLUE benchmark. Although prior VLP models are either initialized from the pre-trained BERT model, or trained by a text-only language modeling loss together with the vision-language (VL) losses, they generally suffer from the weakened performance of NLU. For example, SIMVLM performs significantly worse than BART, though trained with five times more textual data. We speculate that the weakened NLU ability of these models is caused by the catastrophic forgetting of the pre-trained BERT weights during the multimodal pre-training. Moreover, simultaneous optimization of multimodal and text-only objectives potentially shifts the latter to be an auxiliary loss, making the NLP ability not as effective.
+
+On the other hand, the resulting model of VLKD performs only slightly worse than the original BART and significantly outperforms BERT, as the original knowledge embedded in BART is well maintained.
+
+Additionally, as presented in Table 5, we also run VLKD on the abstractive summarization task to evaluate its NLG performance, since BART-based methods excel on the summarization (Lewis et al., 2020; Dou et al., 2021; Yu et al., 2021b). The gap between VLKD and its backbone BART is negligible. Overall, we empirically demonstrate that VLKD enables the backbone PLM to perform multimodal tasks without hurting its original NLP ability.
+
+# 5 Ablation Study
+
+Knowledge Distillation Objectives. Table 6 shows the ablation on the knowledge distillation objectives, except the ICTI loss which is necessary for our method to work. Without TTDM or ITCL, we observe a clear degradation of zero-shot performance on both VQAv2 and COCO image caption datasets. It is worth noting that ITCL contributes more to the image captioning task, which requires a deeper perception of visual features to generate captions. Oppositely, TTDM helps more for the VQA task, which involves reasoning over the question and image features. Removing both of them incurs a large performance drop, which demonstrates the importance of aligning the embedding space between CLIP and BART.
+
+<table><tr><td>Model</td><td>VQAv2 (val)</td><td>COCO Caption (test)</td></tr><tr><td>VLKDViT-B/16ZERO-SHOT</td><td>38.6</td><td>58.3</td></tr><tr><td>w/o TTDM</td><td>35.5</td><td>55.7</td></tr><tr><td>w/o ITCL</td><td>36.3</td><td>54.1</td></tr><tr><td>w/o Both</td><td>30.1</td><td>48.6</td></tr></table>
+
+Number of Masks. Furthermore, we also test the influence of the number of masks for zero-shot image captioning in Table 7. As discussed in Section 4.3.1, it has a trivial influence as the model learns to fill a variable length of tokens for each masked position. We achieve the best performance on the COCO caption dataset when  $m = 6$  and NoCaps when  $m = 8$ .
+
+Table 6: Ablation study on three distillation objectives.  
+
+<table><tr><td>#masks</td><td>5</td><td>6</td><td>7</td><td>8</td></tr><tr><td>CIDEr</td><td>59.7</td><td>61.1</td><td>60.6</td><td>59.6</td></tr></table>
+
+Dataset Size of Distillation. In Table 8, we vary the size of dataset used for knowledge distillation. VLKD only has a slight performance drop when the size is reduced from 3M to 1M, and a sharp drop when further reduced to 100K.
+
+Table 7: Zero-shot image captioning on COCO test set using VLKD (RN50×16), with varying number of masks.  
+
+<table><tr><td></td><td>VQAv2 (val)</td><td>COCO Caption (test)</td></tr><tr><td>VLKD3M</td><td>38.6</td><td>58.3</td></tr><tr><td>VLKD1M</td><td>38.3</td><td>56.2</td></tr><tr><td>VLKD100K</td><td>33.8</td><td>45.1</td></tr></table>
+
+Table 8: Zero-shot performance of VLKD (ViT-B/16) on two datasets, with varying dataset size for distillation.
+
+Unfreeze CLIP Weights. To quantitatively measure the importance of freezing the model weights of CLIP during the VLKD pre-training, we tried unfreezing CLIP's weights and conduct the VLKD pre-training using the ViT-B/16 variant on CC3M without modifying other settings. It achieves 31.7 zero-shot accuracy on the VQAv2 validation set and 44.8 CIDEr on the COCO Caption test set. We speculate that unfreezing CLIP harms its pretrained multimodal space, which further downgrades the performance of VLKD.
+
+# 6 Conclusion
+
+Recent dual-stream VLP models (e.g., CLIP) are powerful in various multimodal classification and retrieval tasks. However, their ability of multimodal generation or pure NLP tasks is highly restricted. In this paper, we propose a novel knowledge distillation method to efficiently align CLIP's multimodal encoders and BART's textual encoder to the same multimodal space, as well as a cross-modal LM loss to consort BART encoder and decoder. This enables multimodal generation under zero-shot and also fully-finetuned settings without losing the original BART's NLP ability. Empirical results show that our model achieves new state-of-the-art zero-shot performance on VQA and excellent performance on both NLP and multimodal tasks when finetuned, demonstrating the effectiveness of our proposed method.
+
+# References
+
+Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. 2019. nocaps: novel object captioning at scale. In International Conference on Computer Vision, pages 8947-8956.  
+Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. 2017. Guided open vocabulary image captioning with constrained beam search. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing.  
+Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. 2018. Bottom-up and top-down attention for image captioning and visual question answering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6077-6086.  
+Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems.  
+Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. 2020. Uniter: Universal image-text representation learning. In European Conference on Computer Vision.  
+Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. 2021. Unifying vision-and-language tasks via text generation. Preprint arXiv:2102.02779.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics.  
+Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations.  
+Zi-Yi Dou, Pengfei Liu, Hiroaki Hayashi, Zhengbao Jiang, and Graham Neubig. 2021. GSum: A general framework for guided neural abstractive summarization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4830-4842, Online. Association for Computational Linguistics.  
+Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2017. Making the V in VQA matter: Elevating the role of image understanding in Visual Question Answering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition.
+
+Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. 2021. Zero-shot detection via vision and language knowledge distillation. Preprint arXiv:2104.13921.  
+Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778.  
+Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. Preprint arXiv:1503.02531.  
+Lu Hou, Zhiqi Huang, Lifeng Shang, Xin Jiang, Xiao Chen, and Qun Liu. 2020. Dynabert: Dynamic bert with adaptive width and depth. In Advances in Neural Information Processing Systems, volume 33.  
+Xiaowei Hu, Zhe Gan, Jianfeng Wang, Zhengyuan Yang, Zicheng Liu, Yumao Lu, and Lijuan Wang. 2021. Scaling up vision-language pre-training for image captioning. ArXiv, abs/2111.12233.  
+Taichi Iki and Akiko Aizawa. 2021. Effect of vision-and-language extensions on natural language understanding in vision-and-language models. Preprint arXiv:2104.08066.  
+Aashi Jain, Mandy Guo, Krishna Srinivasan, Ting Chen, Sneha Kudugunta, Chao Jia, Yinfei Yang, and Jason Baldridge. 2021. Mural: Multimodal, multitask retrieval across languages. Preprint arXiv:2109.05125.  
+Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. 2021. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning.  
+Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. 2019. Tinybert: Distilling bert for natural language understanding. Preprint arXiv:1909.10351.  
+Andrej Karpathy and Li Fei-Fei. 2017. Deep visualsemantic alignments for generating image descriptions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39:664-676.  
+Wonjae Kim, Bokyung Son, and Ildoo Kim. 2021a. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning.  
+Wonjae Kim, Bokyung Son, and Ildoo Kim. 2021b. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning.  
+Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. 2016. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International Journal of Computer Vision, 123:32-73.
+
+Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Annual Meeting of the Association for Computational Linguistics.  
+Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation.  
+Lianian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. 2019. Visualbert: A simple and performant baseline for vision and language. Preprint arXiv:1908.03557.  
+Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Liu, Hua Wu, and Haifeng Wang. 2021. Unimo: Towards unified-modal understanding and generation via cross-modal contrastive learning. In Annual Meeting of the Association for Computational Linguistics.  
+Xiujun Li, Xi Yin, Chunyuan Li, Xiaowei Hu, Pengchuan Zhang, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. 2020. Oscar: Object-semantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision.  
+Junyang Lin, Rui Men, An Yang, Chang Zhou, Ming Ding, Yichang Zhang, Peng Wang, Ang Wang, Le Jiang, Xianyan Jia, et al. 2021. M6: A chinese multimodal pretrainer. Preprint arXiv:2103.00823.  
+Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In European Conference on Computer Vision.  
+Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations.  
+Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. 2019. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc.  
+Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. 2019. Ok-vqa: A visual question answering benchmark requiring external knowledge. In IEEE/CVF Conference on Computer Vision and Pattern Recognition.  
+Shashi Narayan, Shay B. Cohen, and Mirella Lapata. 2018. Don't give me the details, just the summary! Topic-aware convolutional neural networks for extreme summarization. In Conference on Empirical Methods in Natural Language Processing.
+
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning.  
+Alec Radford and Karthik Narasimhan. 2018. Improving language understanding by generative pretraining.  
+Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21:1-67.  
+Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. 2021. Zero-shot text-to-image generation. Preprint arXiv:2102.12092.  
+Steven J. Rennie, Etienne Marcheret, Youssef Mroueh, Jerret Ross, and Vaibhava Goel. 2017. Self-critical sequence training for image captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1179-1195.  
+Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. 2018. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Annual Meeting of the Association for Computational Linguistics, pages 2556-2565.  
+Sheng Shen, Liunian Harold Li, Hao Tan, Mohit Bansal, Anna Rohrbach, Kai-Wei Chang, Zhewei Yao, and Kurt Keutzer. 2022. How much can CLIP benefit vision-and-language tasks? In International Conference on Learning Representations.  
+Gunnar A. Sigurdsson, Jean-Baptiste Alayrac, Aida Nematzadeh, Lucas Smaira, Mateusz Malinowski, João Carreira, Phil Blunsom, and Andrew Zisserman. 2020. Visual grounding in video for unsupervised word translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition.  
+Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. 2020. Vl-bert: Pre-training of generic visual-linguistic representations. In International Conference on Learning Representations.  
+Hao Tan and Mohit Bansal. 2019. Lxmert: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing.  
+Yonglong Tian, Dilip Krishnan, and Phillip Isola. 2019. Contrastive representation distillation. In International Conference on Learning Representations.
+
+Yonglong Tian, Dilip Krishnan, and Phillip Isola. 2020. Contrastive multiview coding. In European Conference on Computer Vision, pages 776-794.  
+Maria Tsimpoukelli, Jacob Menick, Serkan Cabi, S. M. Ali Eslami, Oriol Vinyals, and Felix Hill. 2021. Multimodal few-shot learning with frozen language models. Preprint arXiv:abs/2106.13884.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems.  
+Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In International Conference on Learning Representations.  
+Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. 2021. Simvlm: Simple visual language model pretraining with weak supervision. Preprint arXiv:2108.10904.  
+Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. 2022. FILIP: Finegrained interactive language-image pre-training. In International Conference on Learning Representations.  
+Yang You, Jing Li, Sashank Reddi, Jonathan Hseu, Sanjiv Kumar, Srinadh Bhojanapalli, Xiaodan Song, James Demmel, Kurt Keutzer, and Cho-Jui Hsieh. 2020. Large batch optimization for deep learning: Training bert in 76 minutes. In International Conference on Learning Representations.  
+Fei Yu, Jiji Tang, Weichong Yin, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. 2021a. Ernie-vil: Knowledge enhanced vision-language representations through scene graph. In AAAI Conference on Artificial Intelligence.  
+Tiezheng Yu, Wenliang Dai, Zihan Liu, and Pascale Fung. 2021b. Vision guided generative pre-trained language models for multimodal abstractive summarization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3995-4007, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. 2021. Lit: Zero-shot transfer with locked-image text tuning. CoRR, abs/2111.07991.  
+Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. 2021. Vinvl: Revisiting visual representations in vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5575-5584.
+
+Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D Manning, and Curtis P Langlotz. 2020. Contrastive learning of medical visual representations from paired images and text. Preprint arXiv:2010.00747.
+
+<table><tr><td>Hyper-params</td><td>Values</td></tr><tr><td>Batch size</td><td>4608 (ViT-B/16 and ViT-L/14), 4096 (RN50x16), 3840 (RN50x64)</td></tr><tr><td>Optimizer</td><td>AdamW, β = (0.99, 0.999)</td></tr><tr><td>Learning rate</td><td>2.4e-4</td></tr><tr><td>Weight decay</td><td>0.01</td></tr><tr><td>Eps</td><td>1e-6</td></tr><tr><td>Temperature τ</td><td>Initialized to 0.07</td></tr><tr><td>Warmup steps</td><td>2%</td></tr><tr><td>#Epochs</td><td>10</td></tr><tr><td>Gradient clipping</td><td>3.0</td></tr></table>
+
+Table 9: Hyper-parameters of VLKD pre-training.  
+
+<table><tr><td>Hyper-params</td><td>VQA</td><td>Image captioning</td></tr><tr><td>Batch size</td><td>72</td><td>64</td></tr><tr><td>Total epochs</td><td>10</td><td>10</td></tr><tr><td>#Masks</td><td>2</td><td>6 (COCO), 8 (NoCaps)</td></tr><tr><td>Beam search size</td><td>1 (greedy)</td><td>6</td></tr><tr><td>Optimizer</td><td colspan="2">AdamW, β = (0.99, 0.999)</td></tr><tr><td>Learning rate</td><td colspan="2">1e-4</td></tr><tr><td>Weight decay</td><td colspan="2">0.01</td></tr><tr><td>Eps</td><td colspan="2">1e-8</td></tr><tr><td>LR warmup</td><td colspan="2">First epoch</td></tr><tr><td>Gradient clipping</td><td colspan="2">5.0</td></tr></table>
+
+Table 10: Hyper-parameters for two multimodal tasks.
+
+# A Hyper-parameters
+
+In this section, we show the hyper-parameters of vision-language knowledge distillation (VLKD), as well as downstream task finetuning.
+
+For VLKD, the hyper-parameters are shown in Table 9, for both two CLIP variants we explored. For finetuning multimodal downstream tasks, we use the hyper-parameters shown in Table 10. Within each task, we use the same setting for multiple datasets.
+
+For the GLUE benchmark, we use the LAMB optimizer (You et al., 2020) to train for 10 epochs. We conduct a hyper-parameter grid search with batch size  $= \{16, 32, 64\}$ , lr  $= \{1e-4, 5e-4, 1e-3\}$ , weight decay  $= \{1e-4, 1e-3\}$ . We warm up the learning rate in the first epoch, then linearly decay it to zero.
+
+For XSUM, we directly follow the hyperparameters used in Lewis et al. (2020).
+
+# B More Examples of Zero-shot Inference
+
+In Figure 4, we show more examples of zero-shot image captioning. In Figure 5, we depict more cases of the results of zero-shot open-ended VQA.
+
+![](images/ea92d6f765a042f0624dcc89a24c6ad9d6fa2cca9799f2f9401600ea2ccd0728.jpg)
+
+Reference caption: A big cat laying down in a chair on a porch.
+
+Generated caption: A cat lounging on a chair in a hammock.
+
+![](images/5523612a712a487b205c4acc75e24b2840ba4ae12f6a036893ac0ee054f8c0b4.jpg)
+
+Reference caption: A little girl holding up a pink umbrella.
+
+Generated caption: A girl holding a pink umbrella in the rain.
+
+![](images/2143964338965bb0bfb852538c836ee7a80520b5e7194793d381257d0f4211b1.jpg)
+
+Reference caption: A white boat out in the middle of the ocean.  
+Generated caption: A small fishing boat in the middle of the ocean.
+
+![](images/fe81b9bec31e0f38be1d3e675b563482ae5b24924a4bf108c9bcff3cecd36887.jpg)  
+Figure 4: More examples of zero-shot image captioning.
+
+Reference caption:  
+A small herd of elephants standing in the grass.  
+Generated caption:  
+A herd of elephants in a field of grasses.
+
+![](images/c50bf24349ced4cd1a9ce5ec4b63eb55cec46a0cdaa1bf564e71322b738f0edb.jpg)
+
+What fruit is present on 3 items? Candidate answer(s): Apple.   
+Generated answer: Apple.
+
+![](images/2c65f66d2b7fa83a082ba8ad2de306d5299d741d8500d2c2e6ccfa55f6f48485.jpg)
+
+Where is the cell phone?
+
+Candidate answer(s):
+
+On table; In bowl; Yes.
+
+Generated answer: On table.
+
+![](images/357b77ddbc31415089bb07644006b4bea406fd701b022d75c0fdc42b2b945c9a.jpg)
+
+What are the people doing?
+
+Candidate answer(s):
+
+Standing; Playing; Talking.
+
+Generated answer: Playing.
+
+![](images/5378f3290f114a41820d370f09506c5199336ab74699a5e2379e9bee140d5f06.jpg)
+
+What type of fabric is the hat made of? Candidate answer(s): Cotton; Wool; Denim.
+
+Generated answer: Cotton.
+
+![](images/ccb59504bc266630f7de337b46d04321f612f12a2e308ed47c7e8c3bc1e17082.jpg)
+
+What is the animal on top of? Candidate answer(s): Laptop; Cat; Computer.
+
+Generated answer: Computer:
+
+![](images/f54a612b6cd74ba4c429f087203415f87417889965153aa9915bc6131e763dc0.jpg)  
+Figure 5: More examples of zero-shot VQA.
+
+Why is there a line? Candidate answer(s): No parking; Parking; Caution; Curb. Generated answer: Parking.

enablingmultimodalgenerationonclipviavisionlanguageknowledgedistillation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6914c3e9224b3275f8e499fcd9c99315b6df0711e10cd3098758e59d14cfb5a6
+size 683773

enablingmultimodalgenerationonclipviavisionlanguageknowledgedistillation/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e611f54b64b1d35ac4ba04fb8ab9abe77d0c22a720b982f7b5dc3567b70a6eec
+size 498709

encbpanewbenchmarkdatasetforfinergrainedculturalbackgroundpredictioninenglish/3adbe831-41a8-4970-9083-e485cc560e35_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:acfd7d7feca161acb8ff9c425a094bf87ebd1e87fbfe800dcfadf428af0d9736
+size 82791

encbpanewbenchmarkdatasetforfinergrainedculturalbackgroundpredictioninenglish/3adbe831-41a8-4970-9083-e485cc560e35_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:106f42923a5c0a24003aa3b055c2ee6466c11f4d5fc856984f349477296de09d
+size 95585

encbpanewbenchmarkdatasetforfinergrainedculturalbackgroundpredictioninenglish/3adbe831-41a8-4970-9083-e485cc560e35_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:217a6f5ade5cdb23c5e3a035efa254eea34acb02bcb69eae39a71a253a27a2c8
+size 422320

encbpanewbenchmarkdatasetforfinergrainedculturalbackgroundpredictioninenglish/full.md

@@ -0,0 +1,261 @@
+# EnCBP: A New Benchmark Dataset for Finer-Grained Cultural Background Prediction in English
+
+Weicheng Ma, Samiha Datta, Lili Wang, and Soroush Vosoughi
+
+Department of Computer Science, Dartmouth College
+
+{weicheng.ma.gr, samiha.datta.23, lili.wang.gr, soroush.vosoughi}  $@$  dartmouth.edu
+
+# Abstract
+
+While cultural backgrounds have been shown to affect linguistic expressions, existing natural language processing (NLP) research on culture modeling is overly coarse-grained and does not examine cultural differences among speakers of the same language. To address this problem and augment NLP models with cultural background features, we collect, annotate, manually validate, and benchmark EnCBP, a finer-grained news-based cultural background prediction dataset in English. Through language modeling (LM) evaluations and manual analyses, we confirm that there are noticeable differences in linguistic expressions among five English-speaking countries and across four states in the US. Additionally, our evaluations on nine syntactic (CoNLL-2003), semantic (PAWS-Wiki, QNLI, STS-B, and RTE), and psycholinguistic tasks (SST-5, SST-2, Emotion, and Go-Emotions) show that, while introducing cultural background information does not benefit the Go-Emotions task due to text domain conflicts, it noticeably improves deep learning (DL) model performance on other tasks. Our findings strongly support the importance of cultural background modeling to a wide variety of NLP tasks and demonstrate the applicability of EnCBP in culture-related research.
+
+# 1 Introduction
+
+Psychological research has revealed that people from different cultural background behave differently in the ways they think (Nisbett et al., 2001), talk (Kim, 2002), write (Krampetz, 2005; Almuhailib, 2019; Kitano, 1990), and express emotions (Hareli et al., 2015; Sun et al., 2021; Acheampong et al., 2020). NLP researchers have applied cultural background information to model differences in linguistic expressions across culture groups especially for psycholinguistic tasks<sup>1</sup>, e.g.,
+
+distributional perspective identification (Tian et al., 2021) and sentiment analysis (Sun et al., 2021). In prior research, culture groups are usually defined by official language (Tian et al., 2021) (e.g., US, UK, and India are considered part of the same culture group) or, even more coarse-grained, by ideology (Imran et al., 2020) (e.g., "Western" countries and "Eastern" countries). These settings typically overlook the nuanced cultural differences across or within countries, and they do not provide useful information for modeling different language use behavior in mono-lingual contexts.
+
+To study culture-specific linguistic expressions in the same language and to apply culture-related knowledge to other NLP tasks, we build EnCBP, a cultural background prediction dataset in English. Following (Tambassi, 2018), we assume that language use patterns are more consistent inside each country or each district in a large country, e.g., states in the US. As such, we first construct news corpora by sampling news articles covering five frequently discussed and controversial topics from major news outlets in five English-speaking countries and four geographically dispersed states in the US. We then break the articles down to paragraphs and annotate them with the country and state codes of the news outlets to construct the country- and district-level subsets of EnCBP. We refer to the two subsets as EnCBP-country and EnCBP-district. To ensure annotation quality, we randomly sample 20 instances from each culture group and have them validated manually by local residents using Amazon Mechanical Turk (MTurk). The annotation accuracies and inter-validator agreement rates are both high for all the validation sets, supporting the correctness of the labels and demonstrating the differences in writing style across culture groups. In addition, we benchmark EnCBP for cultural background prediction with three widely-used NLP model architectures, namely BiLSTM, BERT (Vaswani et al., 2017), and RoBERTa (Liu et al.,
+
+2019). Among the three models, the roberta-base model achieves the best overall performance, scoring 82.96 in F1-macro on the EnCBP-country and 73.96 on EnCBP-district. The better performance of BERT and RoBERTa over BiLSTM implies the importance of deep neural network architectures and large-scale pre-training for the challenging text-based cultural background prediction task.
+
+We conduct both quantitative and qualitative analyses on EnCBP to show the differences in linguistic expressions across culture groups. For the quantitative analysis, we fine-tune a BERT model on the corpus with each cultural background label and evaluate it on the corpora of all the culture groups. Results show that all the fine-tuned models are more compatible with the cultural domains of their training corpora and less compatible with the those of other corpora, with perplexity differences ranging from 0.43 to 14.90. For the qualitative analysis, we manually analyze sentence structures and the choices of words or phrases in instances randomly sampled from EnCBP to illustrate culture-specific English expressions.
+
+Furthermore, we evaluate a BERT model on nine psycholinguistic (sentiment analysis and emotion recognition), syntactic (named entity recognition), and semantic (paraphrase identification, natural language inference, semantic textual similarity, and text entailment) tasks to examine how modeling culture-specific English writing styles benefits the performance of NLP models. The models that incorporate cultural background information perform noticeably better on the named entity recognition (NER) task, most semantic tasks, and the sentiment analysis (SA) tasks. In our emotion recognition (ER) evaluation on the Go-Emotions dataset, however, the performance is slightly harmed by incorporating cultural background information. This is likely due to the imbalanced cultural background distribution in the dataset, as the evaluation performance of BERT clearly improves on Emotion, another ER dataset. On the paraphrase identification (PI) task, while the model performs better with cultural information incorporated, the improvement is lower than those on SA and NER tasks. This result suggests that differentiating linguistic expressions with the same semantic meaning may introduce additional noise to semantic tasks.
+
+Our analyses and evaluations support the importance of cultural background modeling for a wide range of NLP tasks and show that EnCBP can con
+
+tribute to future culture-related NLP research.
+
+The contributions of this paper are three-fold:
+
+- we construct, manually validate, and benchmark EnCBP, a mono-lingual news-based cultural background prediction dataset;  
+- we qualitatively and quantitatively examine the distinctions in writing style from different culture groups; and  
+- we show the effect of introducing cultural background information to nine downstream NLP tasks to showcase the importance of cultural information in natural language understanding.
+
+# 2 Dataset Construction
+
+This section introduces the construction, validation, and benchmarking of the EnCBP dataset. The EnCBP dataset adopts a multi-class classification objective. The labels are country codes of news outlets for the coarse-grained subset (EnCBP-country) and US state codes for the finer-grained subset (EnCBP-district).
+
+# 2.1 Data Collection and Annotation
+
+Our work relies on the hypothesis that news articles from mainstream news outlets of a country or district reflect the local language use patterns. Thus, we construct 5 text corpora with news articles posted by New York Times, Fox News, and the Wall Street Journal in the US, BBC in UK, Big News Network - Canada in Canada (CAN), Sydney Morning Herald in Australia (AUS), and Times of India in India (IND) for EnCBP-country. For EnCBP-district, we construct 4 corpora from Coosa Valley News, WJCL, and Macon Daily in Georgia (GA), Times Union, Gotham Gazette, and Newsday in New York (NY), NBC Los Angeles, LA Times, and San Diego Union Tribune in California (CA), and Hardin County News, Jasper Newsboy, and El Paso Times in Texas (TX). We stream news articles from Media Cloud  $^{2}$ , a platform that collects articles from a large number of media outlets, using its official API.
+
+To maintain consistent mentions of events and named entities (NEs) in the corpora, we limit the articles to those under five frequently discussed topics, namely "global warming", "abortion", "immigration", "social safety net", and "mandatory vaccination". 1,000 news articles published between Jan. 1, 2020 and Jun. 30, 2021 are sampled from each news outlet to form our corpora.
+
+<table><tr><td rowspan="2" colspan="2"></td><td colspan="5">Topics</td><td colspan="4">Splits</td></tr><tr><td>Global Warming</td><td>Abortion</td><td>Immigration</td><td>Social Safety Net</td><td>Mandatory Vaccination</td><td>Total</td><td>Train</td><td>Dev</td><td>Test</td></tr><tr><td rowspan="9">Labels</td><td>US</td><td>332</td><td>455</td><td>253</td><td>336</td><td>624</td><td>2,000</td><td>1,600</td><td>200</td><td>200</td></tr><tr><td>UK</td><td>648</td><td>129</td><td>383</td><td>456</td><td>384</td><td>2,000</td><td>1,600</td><td>200</td><td>200</td></tr><tr><td>AUS</td><td>532</td><td>188</td><td>439</td><td>402</td><td>439</td><td>2,000</td><td>1,600</td><td>200</td><td>200</td></tr><tr><td>CAN</td><td>418</td><td>379</td><td>430</td><td>315</td><td>458</td><td>2,000</td><td>1,600</td><td>200</td><td>200</td></tr><tr><td>IND</td><td>478</td><td>171</td><td>540</td><td>371</td><td>440</td><td>2,000</td><td>1,600</td><td>200</td><td>200</td></tr><tr><td>NY</td><td>206</td><td>134</td><td>443</td><td>704</td><td>513</td><td>2,000</td><td>1,600</td><td>200</td><td>200</td></tr><tr><td>CA</td><td>274</td><td>242</td><td>473</td><td>556</td><td>455</td><td>2,000</td><td>1,600</td><td>200</td><td>200</td></tr><tr><td>GA</td><td>245</td><td>384</td><td>214</td><td>389</td><td>768</td><td>2,000</td><td>1,600</td><td>200</td><td>200</td></tr><tr><td>TX</td><td>365</td><td>328</td><td>468</td><td>585</td><td>254</td><td>2,000</td><td>1,600</td><td>200</td><td>200</td></tr></table>
+
+After data collection, we remove duplicates and overly short documents (less than 100 words) to ensure data quality. We also replace the mentions of countries and districts with the "[country]" and "[district]" special tokens. Then, we chunk the remaining news articles into paragraphs and label the documents with the country or district codes of the news outlets by which they are posted. We adopt paragraph-level annotations since asking the annotators to read an overly-long document may cause them to lose track of culture-specific information when they are making judgments. Most state-of-the-art DL models also have input length limits that are not capable of encoding full-length news article. To avoid overly simplifying the task, we remove paragraphs containing NE mentions that are mainly used by news media in specific countries or districts. We quantify the specificity of NEs using inverse document frequency (IDF) scores.
+
+From the filtered news paragraphs, we sample 2,000 paragraphs from the corpus of each culture group to form the annotated dataset. Table 1 provides the statistics of the label and topic distribution of the instances in EnCBP.
+
+# 2.2 Manual Validation
+
+To ensure that the cultural background labels in EnCBP correlate with writing styles, we randomly sample 50 instances from each class and manually validate them on MTurk. In each questionnaire, we pair the sampled instance with another random news paragraph from EnCBP and ask three annotators whether the first, second, or both paragraphs are posted by media outlets in a specific country
+
+Table 1: Number of documents associated with each label and under each topic in EnCBP. For each country or district label, the documents under each topic are randomly sampled into the training, development, and test sets with a  $80\% / 10\% / 10\%$  split.  
+
+<table><tr><td>Culture Groups</td><td>ACC (%)</td><td>IAA</td></tr><tr><td>US</td><td>64.00</td><td>0.61</td></tr><tr><td>UK</td><td>76.67</td><td>0.73</td></tr><tr><td>AUS</td><td>74.00</td><td>0.71</td></tr><tr><td>CAN</td><td>58.67</td><td>0.57</td></tr><tr><td>IND</td><td>61.43</td><td>0.61</td></tr><tr><td>NY</td><td>81.33</td><td>0.78</td></tr><tr><td>CA</td><td>64.67</td><td>0.59</td></tr><tr><td>GA</td><td>70.00</td><td>0.66</td></tr><tr><td>TX</td><td>72.00</td><td>0.68</td></tr></table>
+
+Table 2: Validation results of the EnCBP dataset. ACC and IAA refer to validation accuracy and inter-annotator agreement rate in Fleiss'  $\kappa$ , respectively.
+
+or district. We manually check the instances to ensure there are no country- or district-specific mentions remaining to avoid potential information leakage. For quality control purposes, we only hire crowdsourcing workers from the country or US state matching the label of the instances sampled for validation.
+
+To ensure the quality of annotations in EnCBP, we hire crowdsourcing workers from MTurk to validate randomly sampled data points. Since all the news articles are written by native English speakers and the culture groups are not strictly separated from each other, it is difficult for a validator to identify whether a news paragraph is written by a journalist from the same cultural background as them. Instead, we provide each validator with a news paragraph posted by an international or do
+
+![](images/116f638c707495fe1f083ddc83673ffb90c526d75b27aa0ee52779b8d9d49172.jpg)  
+Figure 1: An example of the questionnaire used for validating the annotations in EnCBP.
+
+mestic news outlet in the country or district they live in (MTurk allows for filtering based on location) and a randomly selected news paragraph from our dataset. The validators are asked to compare the two news paragraphs and decide which of the two paragraphs (or both) were written by their local news outlets through analyzing the use of words, phrases, and sentence structures. To avoid information leak and bias in the validation process, the mentions of countries and districts are replaced with "[country]" and "[district]" special tokens at the pre-processing stage of the dataset. An example questionnaire is shown in Figure 1.
+
+We display the validation accuracy (ACC), i.e., the proportion of the annotators' answers that match the labels of those instances in EnCBP, and inter-annotator agreement rate (IAA) in Table 2. Since we have three options in each of the questionnaires, the ACC of random guess is around  $33\%$  for each culture group. We quantify IAA with Fleiss'  $\kappa$  (Fleiss, 1971), a widely used metric for evaluating IAA. The Fleiss'  $\kappa$  in Table 2 range from moderate ( $>0.40$ ) to substantial agreement ( $>0.60$ ). We infer from the relatively high ACC and IAA that: 1) news writing styles are affected by the cultural backgrounds of journalists and 2) writing styles in each culture group are identifiable by local residents. Since we removed country- or state-specific NEs and mentions of countries or states from the paragraphs, and as the distributions of topics and sentiments are balanced across corpora, the chance that the annotators make their judgments based on these external information is low.
+
+# 2.3 Dataset Benchmarking
+
+After data validation, we divide both EnCBP-country and EnCBP-district into training, development, and test sets with a  $80\% / 10\% / 10\%$  split and a random state of 42. To show the predictability of cultural background labels with NLP models,
+
+<table><tr><td>Model</td><td>EnCBP-country</td><td>EnCBP-district</td></tr><tr><td>BiLSTM</td><td>50.89 (0.98)</td><td>44.53 (1.39)</td></tr><tr><td>BERT</td><td>78.13 (0.67)</td><td>72.09 (1.84)</td></tr><tr><td>RoBERTa</td><td>82.96 (0.89)</td><td>73.96 (1.01)</td></tr></table>
+
+Table 3: Benchmark performance of BiLSTM, bert-base-cased (BERT), and robert-base (RoBERTa) models on EnCBP-country and EnCBP-district. Average F1-macro scores over five runs with different random seeds are reported and standard deviations are shown in parentheses.
+
+we benchmark the EnCBP-country and EnCBP-district separately with BiLSTM, bert-base-cased, and roberta-base models. We train the BiLSTM model for 20 epochs with a learning rate of 0.25 and fine-tune the other models for five epochs with a learning rate of 1e-4 on both subsets.
+
+Table 3 displays the average F1-macro scores across five runs with different random seeds for model initialization. For all the models, the standard deviations of the five runs are at most 0.98 on EnCBP-country and 1.84 on EnCBP-district, indicating that randomness does not severely affect the predictions of models, and that the culture-specific writing styles can be modeled by DL models. Both the BERT and RoBERTa models outperform the BiLSTM model with large margins, which suggests the importance of deep neural network architectures and large-scale pre-training for the task. We also note that all the three models perform worse on EnCBP-district, which may be caused by both the more difficult task setting and the higher level of noise in EnCBP-district, since local news outlets target audiences from all over the country. In the rest of this paper, we use the bert-base-cased model for the analyses and discussions since it is less resource-consuming than the roberta-base model, while the findings potentially apply to other model architectures as well.
+
+# 3 Cultural Domain Compatibility
+
+This section examines whether linguistic expressions are clearly separable across culture groups in EnCBP through LM evaluations. We also manually examine representative linguistic expressions associated with each label to illustrate the differences in linguistic expression across cultures.
+
+# 3.1 Language Modeling Analysis
+
+Since all the documents in EnCBP come from news articles, we assume they are well-written and grammatically correct. In addition, LMs trained on a grammatical corpus should produce similar perplexities on the corpus with each label if the writing styles are consistent across corpora. Thus, to examine culture-specific differences in writing styles, we fine-tune a bert-base-cased model on the training corpus of each class in EnCBP with the MLM objective and evaluate perplexity of the fine-tuned models on all the test corpora.
+
+As Table 4 shows, BERT models usually produce the lowest perplexities on the test portions of their training corpora, and the cross-corpus perplexities are usually considerably higher. This supports our hypothesis that English writing styles are culture-dependent, and that the writing styles across cultures are different enough to be detected by LMs. Meanwhile, we find that the cultural domain compatibility differs for different pairs of corpora, e.g., the IND corpus is more compatible with the UK corpus than other countries or districts. The relations are not symmetric either, e.g., while the LM trained on the CAN corpus well adapts to the US corpus, the US LM performs the worst on the CAN corpus among the five countries. These potentially result from the effects of geographical, geo-political, and historical backgrounds on the formation of cultural backgrounds. For instance, the US could be said to have greater influence on Canadian culture than vice versa. Potentially for similar reasons, compared to TX and GA, NY has a more consistent writing style with CAN. We also note clear cultural domain compatibility gaps between liberal (NY and CA) and conservative states (GA and TX), which, agreeing with Imran et al. (2020), shows that ideologies and policies of a district potentially has an effect on its culture-specific writing styles. We provide additional topic-level LM analysis in Appendix A.
+
+# 3.2 Topic and Sentiment Distributions
+
+To verify if the different expressions across classes in the EnCBP datasets are triggered by cultural differences, we analyze the distributions of topics and sentiment scores for each class. Specifically, we model the topics of each corpus using BERTopic (Grootendorst, 2020) and analyze sentiments of text using Stanza (Qi et al., 2020).
+
+We apply two-sided Kolmogorov-Smirnov (KS) tests on the topic distributions of each pair of classes to see whether the topic distributions for each country or state are similar. For all pairwise comparisons, the null hypothesis (which is that the distributions are identical) cannot be rejected using the KS test, with all p-values being above 0.1, and most in fact being above 0.7. This potentially results from both topic control at the data collection phase and data filtering eliminating paragraphs containing NEs with high IDF scores. Additionally, the sentiment score distribution is relatively consistent across classes (28.02% to 34.97% instances with negative sentiments). Since the classes in EnCBP contain documents that are similar in topics and sentiments, it is likely that the differences in linguistic expressions across classes are caused by cultural differences.
+
+# 3.3 Manual Analysis
+
+In addition to automatic evaluations, we manually examine distinguishable English expressions for each culture group in EnCBP. Specifically, we extract phrases with high TF-IDF values for each corpus in EnCBP, retrieve news paragraphs that contain these phrases, and examine sentence structures and phrase usages in these representative instances.
+
+From our analyses, we find that the different writing styles of countries and districts in EnCBP are affected by the choice of words or phrases, the ordering of phrases, and degrees of formality. For example, the phrases "in the wake of", "in the lead up to", and "the rest of the world" are much more frequently used by AUS news outlets than the others. Also, the use of auxiliaries, especially the word "may", is more frequent in the UK corpus, in the context of politeness. The US corpus is in general more colloquial than the other corpora, as the journalists often write subjective comments in the news articles. Additionally, the ways of referencing speeches differ across corpora, e.g., the quoted text usually appears prior to the "[name] said" in the UK corpus but reversely in the US corpus. In the
+
+<table><tr><td rowspan="2" colspan="2"></td><td colspan="9">Evaluation Corpus</td></tr><tr><td>US</td><td>UK</td><td>AUS</td><td>CAN</td><td>IND</td><td>NY</td><td>CA</td><td>GA</td><td>TX</td></tr><tr><td rowspan="9">Training Corpus</td><td>US</td><td>22.80</td><td>24.13</td><td>25.08</td><td>27.67</td><td>26.54</td><td>28.08</td><td>24.54</td><td>27.54</td><td>24.41</td></tr><tr><td>UK</td><td>24.77</td><td>14.09</td><td>28.76</td><td>28.99</td><td>27.30</td><td>25.50</td><td>22.37</td><td>26.30</td><td>24.14</td></tr><tr><td>AUS</td><td>22.49</td><td>27.56</td><td>21.82</td><td>26.53</td><td>27.26</td><td>25.31</td><td>24.18</td><td>23.69</td><td>25.61</td></tr><tr><td>CAN</td><td>26.13</td><td>37.45</td><td>30.60</td><td>23.30</td><td>28.41</td><td>24.32</td><td>31.04</td><td>26.30</td><td>25.56</td></tr><tr><td>IND</td><td>27.87</td><td>24.63</td><td>29.36</td><td>30.19</td><td>23.91</td><td>29.69</td><td>26.46</td><td>34.42</td><td>26.40</td></tr><tr><td>NY</td><td>22.65</td><td>22.98</td><td>25.68</td><td>21.82</td><td>25.66</td><td>20.53</td><td>21.22</td><td>22.98</td><td>25.88</td></tr><tr><td>CA</td><td>24.23</td><td>29.50</td><td>25.53</td><td>24.41</td><td>24.45</td><td>24.77</td><td>23.80</td><td>28.27</td><td>27.92</td></tr><tr><td>GA</td><td>19.21</td><td>24.61</td><td>29.29</td><td>26.76</td><td>27.16</td><td>21.44</td><td>22.78</td><td>20.25</td><td>20.97</td></tr><tr><td>TX</td><td>24.99</td><td>26.96</td><td>30.91</td><td>29.97</td><td>30.09</td><td>30.31</td><td>27.46</td><td>26.64</td><td>23.83</td></tr></table>
+
+Table 4: Perplexity of LMs fine-tuned on the training corpora of EnCBP with the MLM objective and evaluated on the test corpora. The lowest perplexity for each fine-tuned LM is in bold and the highest perplexity is underlined.
+
+EnCBP-district subset, the sentence structures are more consistent across corpora, while the mentions of NEs and wordings differ more. For example, the word "border" appears frequently in the TX corpus but less in the other corpora when discussing the "immigration" topic. Though the observations summarized from EnCBP may not be universally applicable to other datasets or text domains, they are validated by native speakers of English to be accounting for the high ACC in manual validations.
+
+# 4 Experiments and Analyses
+
+Since cultural background labels are expensive to annotate, most NLP models forego the use of this information to opt for larger training data amount. For example, BERT is trained on Wikipedia text written in styles from mixed cultural backgrounds without access to cultural background information of the writers. Using the EnCBP dataset we constructed, this section examines the relatedness between the cultural background prediction task and multiple other NLP tasks via model probing. We also examine the effectiveness of cultural feature augmentation, i.e., augmenting DL models on downstream NLP tasks with culture-specific writing style information. Specifically, we evaluate a bert-base-cased model with two common information injection methods, namely two-stage training and MTL, on nine syntactic, semantic, and psycholinguistic tasks.
+
+# 4.1 Tasks and Datasets
+
+The datasets used in our evaluations are:
+
+PAWS-Wiki (Zhang et al., 2019) is a PI dataset containing English Wikipedia articles. Each instance in PAWS-Wiki consists of a pair of sentences and
+
+a label indicating whether the two sentences are paraphrase (1) or not (0). There are 49,401 training instances, 8,000 development instances, and 8,000 test instances in this dataset.
+
+CoNLL-2003 English NER dataset (Tjong Kim Sang and De Meulder, 2003) contains news articles from Reuters news only, so the dataset has a more consistent UK writing style, compared to the other datasets we utilize. Each word in the documents is annotated with persons (PER), organizations (ORG), locations (LOC), or miscellaneous names (MISC) NE label in the IOB-2 format. We adopt the official data split of the CoNLL-2003 dataset in the experiments, where there are 7,140, 1,837, and 1,668 NEs in the training, development, and test sets, respectively.
+
+Go-Emotions (Demszky et al., 2020) is an ER dataset containing 58,009 English Reddit comments. Instances in this dataset are labeled with 28 emotion types including neutral, in the multi-label classification form. We split the dataset into training, development, and test sets with a  $80\% /10\% /10\%$  split using 42 as the random seed. To be consistent with other evaluations, we switch the annotations to the multi-class classification form by duplicating the data points associated with multiple labels and assigning one emotion label to each copy. This results in an ER dataset containing 199,461 training instances, 35,057 development instances, and 34,939 test instances after removing instances with no labels.
+
+Stanford Sentiment Treebank (SST-5) (Socher et al., 2013) is a document-level SA dataset containing sentences from movie reviews. The documents are annotated with sentiment scores, which are turned to fine-grained (5-class) sentiment labels
+
+after pre-processing. Using the official data split, we divide the dataset into training, development, and test splits containing 156,817, 1,102, and 2,211 instances, respectively. Note that the training set of SST-5 contains a mixture of phrases and sentences, while the development and test sets contain only complete sentences.
+
+SST-2 is the coarse-grained SST-5 dataset, in which each document is labeled with positive (1) or negative (0) sentiments. There are 67,349 training instances, 872 development instances, and 1,821 test instances in this dataset.
+
+QNLI (Wang et al., 2019) is a natural language inference (NLI) dataset with a question answering background. Each instance in QNLI contains a question, a statement, and a label indicating whether the statement contains the answer to the question (1) or not (0). There are 104,743 training instances, 5,463 development instances, and 5,463 test instances in this dataset.
+
+STS-B (Cer et al., 2017) is a benchmarked semantic textual similarity (STS) dataset. Each instance in STS-B is a pair of sentences manually annotated with a semantic similarity score from 0 to 5. The dataset contains 5,749 training instances, 1,500 development instances, and 1,379 test instances.
+
+RTE is a textual entailment (TE) dataset. Each instance in RTE contains a pair of sentences and a label indicating whether the second sentence is an entailment (1) or not (0) of the first sentence. The RTE dataset we use is a combination of RTE1 (Dagan et al., 2005), RTE2 (Bar-Haim et al., 2006), RTE3 (Giampiccolo et al., 2007), and RTE5 (Bentivogli et al., 2009) datasets, which contains 2,490 training instances, 277 development instances, and 3,000 test instances.
+
+Emotion (Saravia et al., 2018) is a Twitter-based ER dataset labeled with six emotion types, i.e., sadness (0), joy (1), love (2), anger (3), fear (4), and surprise (5). There are 16,000 training instances, 2,000 development instances, and 2,000 test instances in this dataset.
+
+# 4.2 Feature Augmentation
+
+# 4.2.1 Experimental Settings
+
+We use the Huggingface (Wolf et al., 2020) implementation of BERT in all our evaluations. On each task, we fine-tune a bert-base-cased model for five epochs with different random seeds, and we report the average evaluation score on the test sets of downstream tasks over the five runs to avoid the
+
+influence of randomness. Each experiment is run on a single RTX-6000 GPU with a learning rate of 1e-4 and a batch size of 32.
+
+# 4.2.2 Two-Stage Training
+
+We first explore the two-stage training method which successively fine-tunes the pre-trained BERT model on a cultural background prediction dataset and the target task. We use EnCBP-country here to examine the efficacy of coarse-grained cultural feature augmentation, and we study the effect of using EnCBP-district in Section 4.2.4.
+
+As Table 5 shows, the two-stage training strategy brings noticeable performance improvements to the SA models. This agrees with prior psychological research (Sun et al., 2021), since the expressions of sentiments and attitudes differ across culture groups. Similarly, since NEs are usually mentioned differently across cultures, training the model to distinguish culture-specific writing styles helps resolve the conflict between the training domain of BERT and that of the CoNLL-2003 dataset and improves the performance of the NER model. On the PI task, while two-stage training has a positive effect on the performance of the model, the score improvement is not as significant as those on SA and NER tasks. The same trend holds for two other semantic tasks (QNLI and STS-B), where two-stage training brings only marginal performance improvements. We attribute this to the additional noise introduced by the cultural background labels for a semantic task, since expressions with the same semantic meaning can be associated with different cultural background labels in EnCBP. To verify this assumption, we conduct an additional experiment by applying the MLM objective instead of the classification objective in the first training stage. The model performance on PI is raised to 94.11 in F1-macro score, outperforming the previous two-stage training model by 2.44. The two-stage training performance also improves by 0.81 and  $0.49 / 0.53$  for QNLI and STS-B when using the MLM objective at the first fine-tuning stage. These results imply that while the cultural background labels are noisy for semantic tasks, enhancing the LM with English expressions from multiple cultural backgrounds is beneficial. Quite differently, however, two-stage training brings noticeable performance improvements to the RTE model. One possible explanation is, as is supported by the large standard deviations of evaluation scores in five runs, that the RTE dataset is too small and the performance tend
+
+<table><tr><td></td><td>PAWS-Wiki (PI)</td><td colspan="2">CoNLL-2003 (NER)</td><td>Go-Emotions (ER)</td><td>SST-5 (SA)</td></tr><tr><td>BERTorig</td><td>90.01 (0.35)</td><td colspan="2">91.73 (0.39)</td><td>31.67 (0.59)</td><td>52.41 (1.20)</td></tr><tr><td>+ two-stage training</td><td>91.67 (0.20)</td><td colspan="2">94.41 (0.10)</td><td>30.72 (0.16)</td><td>54.54 (0.45)</td></tr><tr><td>+ multi-task learning</td><td>91.58 (0.19)</td><td colspan="2">92.92 (0.18)</td><td>30.71 (0.24)</td><td>54.47 (0.70)</td></tr><tr><td></td><td>QNLI (NLI)</td><td>STS-B (STS)</td><td>RTE (TE)</td><td>SST-2 (SA)</td><td>Emotion (ER)</td></tr><tr><td>BERTorig</td><td>90.89 (0.06)</td><td>89.22/88.83 (0.05/0.02)</td><td>64.69 (1.13)</td><td>91.86 (0.46)</td><td>88.25 (0.49)</td></tr><tr><td>+ two-stage training</td><td>91.77 (0.09)</td><td>89.47/89.08 (0.11/0.13)</td><td>68.45 (1.71)</td><td>93.09 (0.33)</td><td>91.94 (0.50)</td></tr><tr><td>+ multi-task learning</td><td>91.20 (0.22)</td><td>89.32/88.94 (0.10/0.11)</td><td>70.76 (0.93)</td><td>92.34 (0.42)</td><td>91.70 (0.35)</td></tr></table>
+
+Table 5: The performance of BERT model without cultural feature augmentation (BERT-orig), and models with cultural feature augmentation via two-stage training and multi-task learning. EnCBP-country is used as the auxiliary dataset. We report accuracy for QNLI, RTE, and SST-2, Pearson's and Spearman's correlations for STS-B, and F1-macro for the other tasks. The average score and standard deviation (in parentheses) in five runs with different random seeds are reported for each experiment  
+
+<table><tr><td></td><td>PAWS-Wiki (PI)</td><td colspan="2">CoNLL-2003 (NER)</td><td>Go-Emotions (ER)</td><td>SST-5 (SA)</td></tr><tr><td>BERTorig</td><td>90.01 (0.35)</td><td colspan="2">91.73 (0.39)</td><td>31.67 (0.59)</td><td>52.41 (1.20)</td></tr><tr><td>+ two-stage training</td><td>91.40 (0.20)</td><td colspan="2">94.25 (0.11)</td><td>30.21 (0.37)</td><td>53.82 (0.45)</td></tr><tr><td>+ multi-task learning</td><td>91.70 (0.23)</td><td colspan="2">93.64 (0.14)</td><td>30.47 (0.14)</td><td>53.52 (0.54)</td></tr><tr><td></td><td>QNLI (NLI)</td><td>STS-B (STS)</td><td>RTE (TE)</td><td>SST-2 (SA)</td><td>Emotion (ER)</td></tr><tr><td>BERTorig</td><td>90.89 (0.06)</td><td>89.22/88.83 (0.05/0.02)</td><td>64.69 (1.13)</td><td>91.86 (0.46)</td><td>88.25 (0.49)</td></tr><tr><td>+ two-stage training</td><td>91.77 (0.08)</td><td>89.45/89.01 (0.12/0.13)</td><td>67.87 (1.09)</td><td>92.52 (0.32)</td><td>91.65 (0.24)</td></tr><tr><td>+ multi-task learning</td><td>91.21 (0.24)</td><td>89.34/89.14 (0.11/0.10)</td><td>69.68 (1.04)</td><td>92.89 (0.36)</td><td>92.07 (0.52)</td></tr></table>
+
+Table 6: The performance of BERT model without cultural feature augmentation (BERT-orig), and models with cultural feature augmentation via two-stage training and multi-task learning. The EnCBP-district is used as the auxiliary dataset. We report accuracy for QNLI, RTE, and SST-2, Pearson's and Spearman's correlations for STS-B, and F1-macro for the other tasks. The average score and standard deviation (in parentheses) in five runs with different random seeds are reported for each experiment
+
+to be affected more greatly by other issues such as model initialization. Unlike the other tasks, the performance of BERT drops on Go-Emotions in our evaluations, which is counter-intuitive since expressions of emotion are culture-specific (Hareli et al., 2015). We hypothesize that the negative effect of cultural feature augmentation is mainly caused by the imbalanced distribution of users' cultural backgrounds in the Go-Emotions dataset, as the dataset is constructed over a Reddit<sup>3</sup> corpus and nearly  $50\%$  Reddit users are from the US<sup>4</sup>. Supporting our hypothesis, cultural feature augmentation on the Emotion dataset notably improves the perfor
+
+mance of BERT, despite the domain differences between the EnCBP-country (news domain) and Emotion (social media domain) datasets.
+
+# 4.2.3 Multi-Task Learning
+
+We further explore MTL methods for cultural feature augmentation when training the BERT model on downstream tasks. Specifically, we use EnCBP-country as the auxiliary task and train the model alternatively on the primary and auxiliary tasks.
+
+According to Table 5, introducing cultural background information via MTL improves the performance of BERT on all the datasets except for GoEmotions, similar to the two-stage training method. However, the performance on NER is noticeably lower with MTL than with two-stage training. This
+
+potentially results from the mono-cultural nature of the CoNLL-2003 dataset, which is constructed on Reuters news, a UK news outlet. While the information and expressions in countries other than UK fade gradually during the second training stage, the MTL method strengthens the irrelevant information in the entire training process and harms the evaluation performance of the model more severely. To validate our hypothesis, we generate a binary cultural background prediction dataset by treating the UK documents as positive instances and the others as negative instances, and we re-run the MTL evaluation on the CoNLL-2003 dataset. The performance of BERT under this setting is raised to 93.97 in F1-macro score, which implies the importance of careful text domain selection for cultural feature augmentation on DL models.
+
+# 4.2.4 Finer-Grained Feature Augmentation
+
+We repeat the two-stage training and MTL evaluations on the nine downstream tasks using EnCBP-district to examine the effects of cultural feature augmentation with cultural background information with different granularity levels. The evaluation results are shown in Table 6. While the scores are very consistent with those in Table 5, we observe better MTL performance on CoNLL-2003 and Emotion and worse performance with both two-stage training and MTL on SST-5. Based on our analysis of EnCBP-country and EnCBP-district, the larger gaps in writing style among countries than those across states are likely the cause of the lower NER evaluation performance. In EnCBP-district, the linguistic expressions are more consistent since they all come from news outlets in the US, which relieves the problem and improves the MTL performance on CoNLL-2003. On the contrary, the lower diversity in expressions potentially negatively affects the performance of the SST-5 model since the SA task benefits from identifying culture-specific linguistic expressions, and since the corpus of SST-5 contains writings from all over the world. In addition, using EnCBP-district does not relieve the problem on the Go-Emotions dataset either, which suggests the limitation of cultural feature augmentation: trying to distinct expressions in different cultural backgrounds may introduce unexpected noise into models especially when the cultural background of a dataset is mostly the same. The performance of BERT on the Emotion dataset which consists of writings from more diverse cultural backgrounds, for example, is subject
+
+to comparable or even greater improvements when the model is augmented using the finer-grained EnCBP-district dataset.
+
+To summarize, while cultural feature augmentation using EnCBP is beneficial for a wide range of NLP tasks, the necessity of conducting cultural feature augmentation has to be carefully evaluated. We also examine the effect of feature augmentation with less auxiliary data in Appendix B, showing that the size of the auxiliary data has an affect on the performance of DL models.
+
+# 5 Conclusion and Future Work
+
+This paper presents EnCBP, a mono-lingual news-based cultural background prediction dataset containing country-level (coarse-grained) and district-level (finer-grained) cultural background labels. Through manual validation on MTurk and cultural domain compatibility evaluations, we find that writing style clearly differs across countries and districts, confirming that cultural background has a substantial effect on writing style even in the same language. We also benchmark the dataset with state-of-the-art NLP models to show that, though challenging, different English expressions across cultural backgrounds can be identified and classified into culture categories by DL models. Additionally, our evaluations on downstream NLP tasks of various types show that cultural feature augmentation is able to improve the performance of DL models on various semantic, syntactic, and psycholinguistic tasks. While the performance of the BERT model is negatively affected by introducing cultural background information on an ER dataset, the imbalanced distribution of cultural backgrounds in its corpus may account for the performance drop. Our results demonstrate that cultural feature augmentation with EnCBP is a practical way of improving the performance of DL models on various NLP tasks, as long as the text domains of EnCBP and the downstream tasks are not too divergent.
+
+Future work can extend our research to examine cultural differences in social media writings, which reflect even finer-grained cultural distinctions and are much noisier and difficult to annotate or validate than news articles.
+
+# 6 Acknowledgement
+
+We thank Mr. Aadil Islam for conducting additional evaluations on QNLI, STS-B, RTE, SST-2, and Emotion to respond to the reviewers' comments.
+
+# 7 Ethics Statement
+
+This paper presents and releases a news-based cultural background prediction dataset. The dataset is constructed on publicly available news outlets using the public API of Media Cloud and the labels are generated based on the country and district codes of the media outlets. Thus, there is no sensitive or private information in the dataset. Additionally, since we use mainstream news outlets for our data collection we believe there is less risk of overtly unethical information (though we cannot be sure given the current sociopolitical climate). Given the relatively large size of our dataset, we cannot manually examine all articles, however, the publicly released dataset will warn users of the possibility of the dataset containing unethical information and will allows users to flag unethical articles in our dataset. We also hired annotators from MTurk to validate the quality of annotations for a sample instances from our dataset. To ensure the quality of dataset validation, we require the annotators to be native English speakers from the same country or district as the label of each instance to be validated. The annotators were given clear instructions to choose the news paragraph(s) written by journalists in their countries or districts from a pair of paragraphs. We paid \(0.14 (USD)\) for validating each instance, which translates to over \)25 per hour since each data point takes no more than 1 minute to validate. This hourly rate is considerably higher than the federal minimum wage in the US. The entire annotation process was anonymized and the annotators were not asked for their personally identifiable information, so there was not any risk of harm associated with their participation.
+
+This paper presents one of the first attempts at tailoring NLP models to the writing styles of specific regions, thus reducing the out-sized influence of the linguistic style of larger countries in these models.
+
+# References
+
+Francisca Adoma Acheampong, Chen Wenyu, and Henry Nunoo-Mensah. 2020. Text-based emotion detection: Advances, challenges, and opportunities. Engineering Reports, 2(7):e12189.  
+Badar Almuailib. 2019. Analyzing cross-cultural writing differences using contrastive rhetoric: A critical review. Advances in Language and Literary Studies, 10(2):102-106.
+
+Roy Bar-Haim, Ido Dagan, Bill Dolan, Lisa Ferro, Danilo Giampiccolo, Bernardo Magnini, and Idan Szpektor. 2006. The second pascal recognising textual entailment challenge. In Proceedings of the second PASCAL challenges workshop on recognising textual entailment, volume 6, pages 6-4. Venice.  
+Luisa Bentivogli, Peter Clark, Ido Dagan, and Danilo Giampiccolo. 2009. The fifth pascal recognizing textual entailment challenge. In TAC.  
+Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. 2017. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. arXiv preprint arXiv:1708.00055.  
+Ido Dagan, Oren Glickman, and Bernardo Magnini. 2005. The pascal recognising textual entailment challenge. In *Machine Learning Challenges Workshop*, pages 177-190. Springer.  
+Dorottya Demszky, Dana Movshovitz-Attias, Jeongwoo Ko, Alan Cowen, Gaurav Nemade, and Sujith Ravi. 2020. GoEmotions: A dataset of fine-grained emotions. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4040–4054, Online. Association for Computational Linguistics.  
+Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378.  
+Danilo Giampiccolo, Bernardo Magnini, Ido Dagan, and Bill Dolan. 2007. The third pascal recognizing textual entailment challenge. In Proceedings of the ACL-PASCAL workshop on textual entailment and paraphrasing, pages 1-9. Association for Computational Linguistics.  
+Maarten Grootendorst. 2020. Bertopic: Leveraging bert and c-tfidf to create easily interpretable topics.  
+Shlomo Hareli, Konstantinos Kafetsios, and Ursula Hess. 2015. A cross-cultural study on emotion expression and the learning of social norms. Frontiers in psychology, 6:1501.  
+Ali Shariq Imran, Sher Mohammad Doudpota, Zenun Kastrati, and Rakhi Bhatra. 2020. Cross-cultural polarity and emotion detection using sentiment analysis and deep learning—a case study on COVID-19. arXiv preprint arXiv:2008.10031.  
+Heejung S Kim. 2002. We talk, therefore we think? a cultural analysis of the effect of talking on thinking. Journal of personality and social psychology, 83(4):828.  
+Hiroko Kitano. 1990. Cross-cultural differences in written discourse patterns: a study of acceptability of japanese expository compositions in american universities.  
+Erin McClure Krampetz. 2005. International educational administration and policy analysis.
+
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.  
+Richard E Nisbett, Kaiping Peng, Incheol Choi, and Ara Norenzayan. 2001. Culture and systems of thought: holistic versus analytic cognition. *Psychological review*, 108(2):291.  
+Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. Stanza: A Python natural language processing toolkit for many human languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations.  
+Elvis Saravia, Hsien-Chi Toby Liu, Yen-Hao Huang, Junlin Wu, and Yi-Shin Chen. 2018. CARER: Contextualized affect representations for emotion recognition. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3687-3697, Brussels, Belgium. Association for Computational Linguistics.  
+Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, Seattle, Washington, USA. Association for Computational Linguistics.  
+Jimin Sun, Hwijeen Ahn, Chan Young Park, Yulia Tsvetkov, and David R. Mortensen. 2021. Cross-cultural similarity features for cross-lingual transfer learning of pragmatically motivated tasks. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 2403-2414, Online. Association for Computational Linguistics.  
+Timothy Tambassi. 2018. From geographical lines to cultural boundaries: Mapping the ontological debate. Rivista di estetica, 67:150-164.  
+Yufei Tian, Tuhin Chakrabarty, Fred Morstatter, and Nanyun Peng. 2021. Identifying distributional perspectives from collingual groups. In Proceedings of the Ninth International Workshop on Natural Language Processing for Social Media, pages 178-190, Online. Association for Computational Linguistics.  
+Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142-147.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all
+
+you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008.  
+Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In the Proceedings of ICLR.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.  
+Yuan Zhang, Jason Baldridge, and Luheng He. 2019. PAWS: Paraphrase Adversaries from Word Scrambling. In Proc. of NAACL.
+
+# A Language Modeling Analysis Based on Topic
+
+We study the cultural domain compatibility across news topics in EnCBP by repeating the LM evaluations with the news paragraphs grouped by their topics. As Table A1 shows, for the topics "Immigration" and "Social Safety Net", the LMs do not achieve the lowest perplexities on their training topics. We speculate that this reflects the more controversial nature of these two topics, since linguistic expressions are heavily affected by attitudes and stances. In addition, since each country or state news outlet has a relatively stable attitude towards each topic, the discrepancy between each trained LM and the test set in the cultural domain of its training set implies that the EnCBP dataset is constructed over diverse culture groups. The diverse writing styles in EnCBP make it appropriate for improving DL models on downstream tasks via cultural feature augmentation, since EnCBP does not bias extremely towards the writing styles of a single culture group.
+
+# B Feature Augmentation with Less Data
+
+We repeat the joint modeling and two-stage training experiments on PAWS-Wiki, CoNLL-2003, Go-Emotions, and SST-5 datasets with randomly downsampled EnCBP-country and EnCBP-district training datasets to examine the effect of auxiliary data size. Specifically, we randomly reduce  $20\%$ ,  $40\%$ , and  $80\%$  of training instances from EnCBP-country and EnCBP-district with a random seed of 42 and use the reduced datasets in the evaluations. The experimental results are shown in Table B1 (EnCBP-country) and Table B2 (EnCBP-district).
+
+While removing  $20\%$  of the training instances from EnCBP-country and EnCBP-district generally does not greatly affect the feature augmentation evaluation results, there is noticeable performance gap on all the tasks when over  $40\%$  of the training instances are eliminated. This may be due to the poorer predictability of cultural background labels from the much smaller training datasets, as the BERT performance drops greatly from 78.13 to 60.92 (on EnCBP-country) and from 72.09 to 60.03 (on EnCBP-district) when  $40\%$  of the training data is removed (see Table 3 for the original BERT performance results). On the other hand, though using more training data from EnCBP has positive overall effects on the performance of feature-augmented models, the improvements
+
+become gradually smaller when the training data amount increases.
+
+In brief, through these experiments we hypothesize that a cultural background prediction dataset of a moderate size such as EnCBP is sufficient for cultural feature augmentation. Even if datasets larger in size could potentially lead to better performance improvements, the gains are likely to be small compared to the effort required for constructing a larger dataset.
+
+<table><tr><td rowspan="2" colspan="2"></td><td colspan="5">Evaluation Corpus</td></tr><tr><td>Global Warming</td><td>Abortion</td><td>Immigration</td><td>Social Safety Net</td><td>Mandatory Vaccines</td></tr><tr><td rowspan="5">Training Corpus</td><td>Global Warming</td><td>21.42</td><td>25.79</td><td>25.29</td><td>26.36</td><td>24.18</td></tr><tr><td>Abortion</td><td>26.40</td><td>20.79</td><td>30.66</td><td>24.38</td><td>25.80</td></tr><tr><td>Immigration</td><td>30.00</td><td>25.00</td><td>28.70</td><td>25.50</td><td>24.88</td></tr><tr><td>Social Safety Net</td><td>25.54</td><td>26.80</td><td>27.78</td><td>29.01</td><td>27.88</td></tr><tr><td>Mandatory Vaccines</td><td>25.48</td><td>25.13</td><td>29.53</td><td>28.18</td><td>23.22</td></tr></table>
+
+Table A1: Perplexity of each BERT model fine-tuned on a training topic with the MLM objective and evaluated on an evaluation topic. The lowest perplexity for each fine-tuned LM is in bold and the highest perplexity is underlined.  
+
+<table><tr><td rowspan="2">DR</td><td colspan="2">PAWS-Wiki (PI)</td><td>CoNLL-2003 (NER)</td><td>Go-Emotions (ER)</td><td>SST-5 (SA)</td></tr><tr><td>BERT-orig</td><td>90.01</td><td>91.73</td><td>31.67</td><td>52.41</td></tr><tr><td rowspan="2">80%</td><td>+ two-stage training</td><td>91.24</td><td>94.07</td><td>29.76</td><td>53.86</td></tr><tr><td>+ multi-task learning</td><td>91.50</td><td>93.88</td><td>29.42</td><td>54.37</td></tr><tr><td rowspan="2">60%</td><td>+ two-stage training</td><td>90.60</td><td>92.50</td><td>28.98</td><td>50.54</td></tr><tr><td>+ multi-task learning</td><td>90.84</td><td>92.00</td><td>28.97</td><td>51.24</td></tr><tr><td rowspan="2">20%</td><td>+ two-stage training</td><td>90.15</td><td>91.75</td><td>28.84</td><td>50.04</td></tr><tr><td>+ multi-task learning</td><td>90.23</td><td>91.53</td><td>28.81</td><td>50.71</td></tr></table>
+
+Table B1: The performance of BERT without cultural feature augmentation (BERT-orig), and models with cultural feature augmentation via two-stage training (+two-stage training) and multi-task learning (+multi-task learning). The downsampled EnCBP-country datasets are used as auxiliary datasets. DR represents the percentile of remaining data.  
+
+<table><tr><td rowspan="2">DR</td><td colspan="2">PAWS-Wiki (PI)</td><td>CoNLL-2003 (NER)</td><td>Go-Emotions (ER)</td><td>SST-5 (SA)</td></tr><tr><td>BERT-orig</td><td>90.01</td><td>91.73</td><td>31.67</td><td>52.41</td></tr><tr><td rowspan="2">80%</td><td>+ two-stage training</td><td>91.18</td><td>93.48</td><td>29.57</td><td>53.34</td></tr><tr><td>+ multi-task learning</td><td>90.91</td><td>93.29</td><td>29.96</td><td>53.38</td></tr><tr><td rowspan="2">60%</td><td>+ two-stage training</td><td>90.23</td><td>93.34</td><td>28.43</td><td>51.86</td></tr><tr><td>+ multi-task learning</td><td>90.46</td><td>92.85</td><td>28.54</td><td>51.06</td></tr><tr><td rowspan="2">20%</td><td>+ two-stage training</td><td>89.98</td><td>92.00</td><td>28.81</td><td>50.71</td></tr><tr><td>+ multi-task learning</td><td>90.00</td><td>91.65</td><td>28.39</td><td>50.02</td></tr></table>
+
+Table B2: The performance of BERT without cultural feature augmentation (BERT-orig), and models with cultural feature augmentation via two-stage training (+two-stage training) and multi-task learning (+multi-task learning). The downsampled EnCBP-district datasets are used as auxiliary datasets. DR represents the percentile of remaining data.

encbpanewbenchmarkdatasetforfinergrainedculturalbackgroundpredictioninenglish/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e541ee9570b158c4969b63a999cf05c1bf565cb02e8ccc30f3b1386cdfa660f
+size 593190

encbpanewbenchmarkdatasetforfinergrainedculturalbackgroundpredictioninenglish/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f7672e29a1edf36952ba28abcf05f463bce8e0f8041b92ae621a1f379af027f
+size 280821

encodingandfusingsemanticconnectionandlinguisticevidenceforimplicitdiscourserelationrecognition/ee9fd883-524a-4da0-8fc0-dae76b155cec_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd739551591c39f1f4490377aa8a6cb9b913635a5dd24dd01de93a627778e5ab
+size 74782

encodingandfusingsemanticconnectionandlinguisticevidenceforimplicitdiscourserelationrecognition/ee9fd883-524a-4da0-8fc0-dae76b155cec_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f5970c09eef59c8a1e17c9f3bf81e928a224631a494b9a485529347340404770
+size 92376

encodingandfusingsemanticconnectionandlinguisticevidenceforimplicitdiscourserelationrecognition/ee9fd883-524a-4da0-8fc0-dae76b155cec_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70c16046643a2e052997697c109efa80065f806d4c9f7773dc746359907a83c8
+size 2896056

encodingandfusingsemanticconnectionandlinguisticevidenceforimplicitdiscourserelationrecognition/full.md

@@ -0,0 +1,344 @@
+# Encoding and Fusing Semantic Connection and Linguistic Evidence for Implicit Discourse Relation Recognition
+
+Wei Xiang $^{1}$ , Bang Wang $^{1}$ , Lu Dai $^{1}$ , Yijun Mo $^{2*}$
+
+$^{1}$ School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan, China
+
+$^{2}$ School of Computer Science and Technology,
+
+Huazhong University of Science and Technology, Wuhan, China  
+xiangwei, wangbang, dailu18, moyj}@hust.edu.
+
+# Abstract
+
+Prior studies use one attention mechanism to improve contextual semantic representation learning for implicit discourse relation recognition (IDRR). However, diverse relation senses may benefit from different attention mechanisms. We also argue that some linguistic relation in between two words can be further exploited for IDRR. This paper proposes a Multi-Attentive Neural Fusion (MANF) model to encode and fuse both semantic connection and linguistic evidence for IDRR. In MANF, we design a Dual Attention Network (DAN) to learn and fuse two kinds of attentive representation for arguments as its semantic connection. We also propose an Offset Matrix Network (OMN) to encode the linguistic relations of word-pairs as linguistic evidence. Our MANF model achieves the state-of-the-art results on the PDTB 3.0 corpus.
+
+# 1 Introduction
+
+Implicit Discourse Relation Recognition (IDRR) is to detect and classify some latent relation in between a pair of text segments (called arguments) without an explicit connective word. It is of great importance for many downstream Natural Language Processing (NLP) applications, such as question answering (Liakata et al., 2013), machine translation (Guzmán et al., 2014), information extraction (Xiang and Wang, 2019), sentiment analysis (Wang and Wang, 2020), and etc. However, due to the absence of an explicit connective word, inferring discourse relations from the contextual semantics of arguments is still a challenging task.
+
+Conventional machine learning based methods usually train a relation classifier by using many handmade features to capture lexical, syntactic regularity and contextual information of arguments, which is time-consuming and labor-intensive (Pitler et al., 2009, 2008). Deep learning based methods
+
+design diverse neural networks to automatic learn the contextual semantic representation of each argument, such as the Shallow Conventional Neural Network (SCNN) (Zhang et al., 2015), Tree-like Long Short-Term Memory (Tree-LSTM) (Rutherford et al., 2017), and BiLSTM-CNN framework (Guo et al., 2019). Although these neural networks can autonomously learn a kind of deeper contextual semantics of arguments, they do not differentiate arguments' words in the representation learning.
+
+Recently, some attention mechanisms have been employed in neural networks to unequally treat words in representation learning. For example, the self-attention computes the local contextual importance of each word in one argument, which generally prioritizes content words for better learning substantive meaning of an argument (Zhou et al., 2016). The interactive attention weights each word in one argument according to its interaction with the representation of another argument, which usually focuses on the rhetorical device of two arguments, like prioritizing some function words with little substantive meaning but potentially indicating the connection of two arguments (Liu and Li, 2016; Guo et al., 2018).
+
+Both kinds of attention mechanisms have been proven effective for IDRR, as each can well exploit either content semantics or rhetorical devices of an argument pair. We regard these contextual semantic information derived from argument content as a kind of semantic connection for relation recognition. However, the IDRR task normally needs to recognize diverse senses of relations, while different senses may benefit from different attentions. To enjoy both advantages, we propose to learn two kinds of argument representation, each based on one attention mechanism. They are next fused to encode an argument pair as the semantic connection for relation recognition in this paper.
+
+Besides semantic connection, we argue that a
+
+kind of linguistic evidence can be obtained from word distributed representation for relation recognition. Indeed, many pre-trained language models, like the word2vec (Mikolov et al., 2013) and BERT (Devlin et al., 2019), are learned from a large amount of unlabeled text by encoding the linguistic regularities and patterns in an unsupervised way. As a word embedding contains inherent meaning of the word, it can be used to infer some linguistic relation in between two words by linear translation. This motivates us to encode such linguistic relations of word-pairs as linguistic evidence.
+
+In this paper, we propose a Multi-Attentive Neural Fusion (MANF) model to encode and fuse both semantic connection and linguistic evidence for the IDRR task. The MANF model contains two modules. One is a Dual Attention Network (DAN): It builds upon a BiLSTM to first encode a self-attentive representation and an inter-attentive representation for each argument. To adapt to different relation senses, we next use a fusion gate to integrate the two representations into the semantic connection representation. Another is an Offset Matrix Network (OMN): It first computes the offset between word embeddings of a word-pair that contains one word from the first argument and another word from the second argument. Upon the offset matrix, we next design an offset attention layer and a multilayer perceptron to encode the linguistic evidence representation. Finally, we design another fusion gate to integrate both semantic connection and linguistic evidence representation for relation recognition. Our MANF Model achieve the state-of-the-art results on the PDTB 3.0 corpus. Our main contributions are as follows:
+
+- Propose a MANF model to encode and fuse semantic connection and linguistic evidence for the IDRR task.  
+- Propose a DAN to enjoy both self-attention and interactive attention for semantic connection encoding.  
+- Propose an OMN to encode word-pairs' offsets as linguistic evidence.  
+- Provide a new baseline result for the IDRR task on the PDTB 3.0 corpus.
+
+# 2 The Multi-Attentive Neural Fusion Model
+
+Fig. 1 illustrates our MANF model, including the DAN, the OMN, and a hierarchical fusion mechanism.
+
+# 2.1 Dual Attention Network
+
+Our DAN is built upon a BiLSTM or BERT to encode a self-attentive representation and an inter-attentive representation for each argument, which are next fused to output the semantic connection representation for an argument pair. We note that a BiLSTM has been widely used to capture word contextual semantics for its good sequential encoding capability. In our experiments, we also replace the word2ve by a fine-tuned BERT for comparison.
+
+The DAN model is illustrated in the left part of Fig. 1, which consists of a BiLSTM layer, a dual attention layer, and a fusion gate layer. We use pretrained word2vec word embeddings  $\mathbf{x} \in \mathbb{R}^{d_w}$  to input the BiLSTM. An argument pair  $(\text{Arg}_1; \text{Arg}_2)$  can be denoted by:
+
+$$
+\operatorname {A r g} _ {1}: \left[ \mathbf {x} _ {1} ^ {1}, \mathbf {x} _ {2} ^ {1}, \dots , \mathbf {x} _ {L _ {1}} ^ {1} \right]; \tag {1}
+$$
+
+$$
+A r g _ {2}: [ \mathbf {x} _ {1} ^ {2}, \mathbf {x} _ {2} ^ {2}, \dots , \mathbf {x} _ {L _ {2}} ^ {2} ], \tag {2}
+$$
+
+where  $\mathbf{x}_i^1$  and  $\mathbf{x}_j^2$  represents the  $i$ -th word embedding in the 1st argument and the  $j$ -th word embedding in the 2nd argument respectively, and  $d_w$  the word embedding dimension.
+
+BiLSTM layer: After the BiLSTM, we obtain two hidden states  $\vec{\mathbf{h}}_i$  and  $\vec{\mathbf{h}}_i$  for each word in one argument from the forward and backward sequence respectively, which are concatenated to obtain an intermediate state  $\mathbf{h}_i = [\vec{\mathbf{h}}_i, \vec{\mathbf{h}}_i]$ . We use a matrix  $\mathbf{H} = [\mathbf{h}_1, \mathbf{h}_2, \dots, \mathbf{h}_L]$  to denote an argument encoding after the BiLSTM, where  $\mathbf{h}_i \in \mathbb{R}^{2d_h}$ ,  $\mathbf{H} \in \mathbb{R}^{L \times 2d_h}$ ,  $d_h$  is the dimension of hidden state.
+
+Dual attention layer: To enjoy both advantages of self-attention and interactive attention, we propose a dual attention mechanism to encode an argument pair. For self-attention, the representation of each argument  $\mathbf{r_s}$  is formed by weighted sum of intermediate state vectors produced by BiLSTM (Zhou et al., 2016):
+
+$$
+\boldsymbol {\alpha} _ {s} = \operatorname {s o f t m a x} \left(\mathbf {w} _ {s} ^ {\top} \mathbf {H}\right), \tag {3}
+$$
+
+$$
+\mathbf {r} _ {s} = \mathbf {H} \boldsymbol {\alpha} _ {s} ^ {\mathsf {T}}, \tag {4}
+$$
+
+where  $\alpha_{s} \in \mathbb{R}^{L}$  is the self-attention weight vector of an argument computed by local contextual importance of each word,  $\mathbf{w}_{s}$  a learnable parameter vector.
+
+For interactive attention, we use the representation of one argument to weight each word in another argument (Ma et al., 2017; Meng et al., 2016). We sum up the intermediate states  $\mathbf{h}_{\mathrm{i}}$  to obtain an intermediate argument representation, i.e.,
+
+![](images/8401663af6876aec31b0bc640ea6bcd225afb2cb86318902c0826df5e70d507e.jpg)  
+Figure 1: Illustration of our multi-attentive neural fusion model.
+
+$\mathbf{S} = \sum_{i=1}^{L} \mathbf{h}_{i}$ . The weight vector  $\alpha_{i} \in \mathbb{R}^{L}$  is computed by taking inner product between  $\mathbf{S}$  and  $\mathbf{H}$  cross two arguments, and followed by a softmax function as follows:
+
+$$
+\boldsymbol {\alpha} _ {i} ^ {1} = \operatorname {s o f t m a x} \left(\mathbf {H} _ {1} \mathbf {S} _ {2} ^ {\top}\right) \tag {5}
+$$
+
+$$
+\boldsymbol {\alpha} _ {i} ^ {2} = \operatorname {s o f t m a x} \left(\mathbf {H} _ {2} \mathbf {S} _ {1} ^ {\top}\right) \tag {6}
+$$
+
+Finally, we weighted sum the intermediate state vectors with corresponding weight vector to form the interactive attention representation  $\mathbf{r}_i$  for each argument:
+
+$$
+\mathbf {r} _ {i} ^ {1} = \mathbf {H} _ {1} \left(\boldsymbol {\alpha} _ {i} ^ {1}\right) ^ {\top}, \quad \mathbf {r} _ {i} ^ {2} = \mathbf {H} _ {2} \left(\boldsymbol {\alpha} _ {i} ^ {2}\right) ^ {\top}. \tag {7}
+$$
+
+Fusion gate layer: Considering the importance of the two attentions not always the same for different relation sense classification, we use a fusion gate to integrate their representations. First, we concatenate the representation of  $Arg_{1}$  and  $Arg_{2}$  to model their discourse relation as  $\mathbf{v}_s = [\mathbf{r}_s^1,\mathbf{r}_s^2 ]$  and  $\mathbf{v}_i = [\mathbf{r}_i^1,\mathbf{r}_i^2 ]$ , where  $\mathbf{v}_s,\mathbf{v}_i\in \mathbb{R}^{4d_h}$ . The transition functions of fusion gate layer are computed as follows:
+
+$$
+\mathbf {g} _ {d} = \operatorname {s i g m o i d} \left(\mathbf {W} _ {d} \mathbf {v} _ {s} + \mathbf {U} _ {d} \mathbf {v} _ {i} + \mathbf {b} _ {d}\right), \tag {8}
+$$
+
+$$
+\mathbf {v} _ {d} = \mathbf {g} _ {d} \odot \mathbf {v} _ {s} + (1 - \mathbf {g} _ {d}) \odot \mathbf {v} _ {i}, \tag {9}
+$$
+
+where  $\mathbf{W}_d\in \mathbb{R}^{4d_h\times 4d_h}$ $\mathbf{U}_d\in \mathbb{R}^{4d_h\times 4d_h}$  and  $\mathbf{b}_d\in$ $\mathbb{R}^{4d_h}$  are learnable parameters and  $\odot$  donates the element-wise product of vectors.
+
+With the fusion gate, our DAN adaptively assigns different importance to self-attention and interactive attention, and outputs  $\mathbf{v}_d\in \mathbb{R}^{4d_h}$  as the semantic connection vector for an argument pair.
+
+# 2.2 Offset Matrix Network
+
+We propose an OMN to encode the linguistic evidence representation based on the offsets of pretrained word embeddings, as shown in the right part of Fig. 1. First, we compute the offset between word embeddings of a word-pair that contains one word from the first argument and another word from the second argument. Then all the word-pair offsets of an argument pair compose an offset matrix  $\mathbf{M} \in \mathbb{R}^{L_1 \times L_2 \times d_h}$ , where  $\mathbf{e}_{ij} \in \mathbb{R}^{d_h}$  is the offset vector between the  $i$ -th word in the 1st argument and the  $j$ -th word in the 2nd argument.
+
+Considering that each word-pair in the offset matrix may have different contribution to the relation classification, we assign a weight score  $\alpha_{ij}$  to every offset vectors, and the weight scores are compute as follows:
+
+$$
+\mathbf {A} = \operatorname {s o f t m a x} \left(\mathbf {w} _ {o} ^ {\top} \mathbf {M}\right), \tag {10}
+$$
+
+where  $\mathbf{A} \in \mathbb{R}^{L_1 \times L_2}$  is the weight matrix,  $\mathbf{w}_o$  is a learnable parameter vector. We compute a word-pair interaction vector  $\mathbf{m} \in \mathbb{R}^{d_h}$  as the weighted
+
+sum of all word-pair offset vectors:
+
+$$
+\mathbf {m} = \sum_ {i = 1} ^ {L _ {1}} \sum_ {j = 1} ^ {L _ {2}} \mathbf {e} _ {i j} \boldsymbol {\alpha} _ {i j} \tag {11}
+$$
+
+Next, we input  $\mathbf{m}$  into a multilayer perceptron (MLP) followed by a tanh activation function to output the linguistic evidence vector  $\mathbf{v}_o\in \mathbb{R}^{4d_h}$  for an argument pair:
+
+$$
+\mathbf {v} _ {o} = \tanh (\mathbf {W} _ {o} \mathbf {m} + \mathbf {b} _ {o}), \tag {12}
+$$
+
+where  $\mathbf{W}_o\in \mathbb{R}^{d_h\times 4d_h}$  and  $\mathbf{b}_o\in \mathbb{R}^{4d_h}$  are learnable parameters.
+
+# 2.3 Implicit Discourse Relation Classification
+
+After obtaining the semantic connection vector  $\mathbf{v}_d$  and the linguistic evidence vector  $\mathbf{v}_o$ , we also argue that they may have different importance for diverse relation sense classification. So we use another fusion gate to integrate the two kinds of representation vectors and obtain the final representation  $\mathbf{v} \in \mathbb{R}^{4d_h}$  of an argument pair for relation classification. The transition functions are:
+
+$$
+\mathbf {g} _ {o} = \operatorname {s i g m o i d} \left(\mathbf {W} _ {f} \mathbf {v} _ {d} + \mathbf {U} _ {f} \mathbf {v} _ {o} + \mathbf {b} _ {f}\right), \tag {13}
+$$
+
+$$
+\mathbf {v} = \mathbf {g} _ {o} \odot \mathbf {v} _ {d} + (1 - \mathbf {g} _ {o}) \odot \mathbf {v} _ {o}, \tag {14}
+$$
+
+where  $\mathbf{W}_f, \mathbf{U}_f \in \mathbb{R}^{4d_h \times 4d_h}$  and  $\mathbf{b}_f \in \mathbb{R}^{4d_h}$  are learnable parameters.
+
+The classifier is a fully connected layer with softmax to transform the final argument pair representation  $\mathbf{v}$  to a probability distribution  $\hat{\mathbf{y}}\in \mathbb{R}^n$  for predicting the discourse relation sense:
+
+$$
+\hat {\mathbf {y}} = \operatorname {s o f t m a x} \left(\mathbf {W} _ {c} \mathbf {v} + \mathbf {b} _ {c}\right),
+$$
+
+where  $\mathbf{W}_c\in \mathbb{R}^{4d_h\times n}$ ,  $\mathbf{b}_c\in \mathbb{R}^n$  are learnable parameters.
+
+For model training, we adopt the cross entropy loss as the cost function:
+
+$$
+J (\theta) = - \frac {1}{K} \sum_ {k = 1} ^ {K} \mathbf {y} ^ {(k)} \log \left(\hat {\mathbf {y}} ^ {(k)}\right) + \lambda \| \theta \| ^ {2}, \tag {15}
+$$
+
+where  $\mathbf{y}^{(k)}$  and  $\hat{\mathbf{y}}^{(k)}$  are the gold label and predicted label of the  $k$ -th training instance respectively.  $\lambda$  and  $\theta$  are the regularization hyper-parameters. We use the Adam optimizer and combine dropout with  $L2$  regularization for model training.
+
+<table><tr><td>Relation</td><td>Train</td><td>Dev.</td><td>Test</td></tr><tr><td>Expansion</td><td>8645</td><td>748</td><td>643</td></tr><tr><td>Comparison</td><td>1937</td><td>190</td><td>154</td></tr><tr><td>Contingency</td><td>5916</td><td>579</td><td>529</td></tr><tr><td>Temporal</td><td>1447</td><td>136</td><td>148</td></tr><tr><td>Total</td><td>17945</td><td>1653</td><td>1474</td></tr></table>
+
+Table 1: Statistics of implicit discourse relation instances in PDTB 3.0 with four top-level relation senses.
+
+# 3 Experiment Setting
+
+# 3.1 The PDTB 3.0 Dataset
+
+We conduct experiments on the latest version 3.0 of Penn Discourse TreeBank (PDTB) corpus, which was released on March 2019 and updated on February 2020. Following the conventional data splitting in PDTB 2.0, we use sections 2-20 as the training set, sections 21-22 as the testing set and 0-1 as the development set (Ji and Eisenstein, 2015). Our experiments are conducted on the four top-level classes of relation sense as the existing studies, including Comparison, Contingency, Expansion and Temporal. The statistics of implicit discourse relation instances in the PDTB 3.0 corpus are summarized in Table 1. More details about PDTB 3.0 are provided in the supplementary material.
+
+# 3.2 Competitors
+
+- NNMA (Liu and Li, 2016) combines two arguments' representation for stacked interactive attentions.  
+- ANN (Lan et al., 2017) applies interactive attention into a multi-task learning framework.  
+IPAL (Ruan et al., 2020) propagates selfattention into interactive attention by a cross-coupled network.  
+- DAGRN (Chen et al., 2016b) encodes word-pair interactions by a neural tensor network.
+
+# 3.3 Parameter Setting
+
+We obtain the pre-trained word embeddings from the 300-dimensional English word2vec model  $(d_w = 300)$  provided by Google  $^{1}$  and the 768-dimensional English BERT model  $(d_w = 768)$  provided by HuggingFace  $^{2}$ . From our statistics,  $99.46\%$  of arguments do not exceed 50 words in PDTB3.0. So we set the maximum length of argument to 50 ( $L = 50$ ). For the word2vec model, we set the mini-batch size to 32 and the
+
+initial learning rate to 5e-4; while for the BERT model, the mini-batch size and initial learning rate is 16 and 1e-5. The trainable parameters are randomly initialized from normal distributions, and the dropout rate is set to 0.2 in the fusion gates and 0.5 in the MLP. We release the code at: https://github.com/HustMinsLab/MANF.
+
+# 4 Result and Analysis
+
+# 4.1 Overall Result
+
+We implement four-way classification and binary classification (i.e. one-versus-others) on the PDTB 3.0, in which macro  $F_{1}$  score and accuracy (Acc) are used for four-way classification and  $F_{1}$  score is used for binary classification.
+
+Table 2 compares the overall performance between our MANF and the competitors. In four-way classification, our MANF achieves significant improvements over competitors in terms of both macro  $F_{1}$  and Acc. In binary-classification, ours also achieves the best performance in three relation sense classification, while the second with a small  $F_{1}$  gap to that of NNMA in the Temporal sense classification.
+
+We note that the first three competitors are neural models mainly for learning argument representation from contextual semantic connections; While the DAGRN is a neural model for learning representation from word linguistic evidences. The first observation is that the DARGN cannot outperform the first three competitors, though the performance gaps are not obvious. This might suggest that latent semantic connections learned from sequential contexts play the main role in relation recognition. This, however, is not unexpected. A relation is usually used for linking the meanings of two arguments, i.e., semantic connections, no matter with or without an explicit connective.
+
+The second observation is that in the first three competitors, the ANN cannot outperform either NNMA or IPAL, not even once in all the performance metrics. We note that although they all employ attention mechanisms in learning semantic connection, the ANN applies a straightforward interactive attention to learn argument representation; While the NNMA designs a sophisticated mechanism for stacking multiple levels of attention, and the IPAL employs a kind of sequential attention mechanism, i.e., interactive attention after self-attention.
+
+Finally, we attribute the outstanding perfor
+
+mance of our MANF model to its fusion of two attentions for learning semantic connection, as well as its exploitation of word linguistic evidence. This will be further analyzed in our ablation studies.
+
+# 4.2 Ablation Study
+
+Linguistic Evidence: We have argued that the inherent meaning of a word, other than its contextual semantics, can be exploited as a kind of linguistic evidence between two arguments for relation classification. To this end, we have designed the OMN module with the pre-trained word embeddings as its input. This input choice is from such considerations: A pre-trained word embedding is normally learned from a huge corpus containing materials from diverse backgrounds<sup>3</sup>, which not only could capture some polysemous property for one word, but also could encode some linguistic regularity and pattern in between words from different contexts. While such properties might be compromised, if we input the OMN with the contextual semantic encodings.
+
+To verify our arguments, we design two variants for the input of the OMN module. (1) Shared: It replaces the input of pre-trained  $\mathbf{x}_i^1 (\mathbf{x}_j^2)$  by the hidden state  $\mathbf{h}_i^1 (\mathbf{h}_j^2)$  of the respective BiLSTM in the DAN module. That is, two modules share the same BiLSTM for encoding word contextual semantics. (2) Parallel: We adopt additional BiLSM networks with their hidden states to replace pre-trained word embeddings. That is, two modules adopt parallel BiLSTM networks.
+
+Table 3 presents the results of the three input choices for the OMN module. The better performance of using pre-trained word embedding can support our arguments. Although a BiLSTM network is well capable of encoding a word contextual semantics for its sequential processing mechanism, our design objective is to exploit the inherent meaning of a word to capture linguistic evidence for an argument pair. This is particular evident in the binary classification of Comparison and Temporal relation sense for its larger improvements. So using the pre-trained word embedding is a wise choice.
+
+Module ablation study: To examine the effectiveness of different modules, we design the following ablation study.
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">Four-way Classification</td><td colspan="4">Binary Classification (F1)</td></tr><tr><td>F1</td><td>Acc</td><td>Expa.</td><td>Comp.</td><td>Cont.</td><td>Temp.</td></tr><tr><td>NNMA (EMNLP, 2016)</td><td>46.13%</td><td>57.67%</td><td>65.10%</td><td>29.15%</td><td>63.33%</td><td>41.03%</td></tr><tr><td>ANN (EMNLP, 2017)</td><td>47.29%</td><td>57.06%</td><td>64.03%</td><td>30.10%</td><td>60.91%</td><td>33.71%</td></tr><tr><td>IPAL (COLING, 2020)</td><td>49.45%</td><td>58.01%</td><td>64.28%</td><td>30.37%</td><td>61.95%</td><td>34.74%</td></tr><tr><td>DAGRN (ACL, 2016)</td><td>45.11%</td><td>57.33%</td><td>64.71%</td><td>27.34%</td><td>62.56%</td><td>38.91%</td></tr><tr><td>Our MANF</td><td>53.14%</td><td>60.45%</td><td>67.82%</td><td>34.16%</td><td>65.48%</td><td>40.22%</td></tr></table>
+
+Table 2: Overall result of comparison models for implicit discourse relation classification.  
+Four-way Classification  
+
+<table><tr><td>Method</td><td>Pre-trained</td><td>Shared</td><td>Parallel</td></tr><tr><td>F1</td><td>53.14 %</td><td>50.41%</td><td>51.54%</td></tr><tr><td>Acc</td><td>60.45%</td><td>58.82%</td><td>60.85%</td></tr></table>
+
+Binary Classification (F1)  
+
+<table><tr><td>Method</td><td>Pre-trained</td><td>Shared</td><td>Parallel</td></tr><tr><td>Expa.</td><td>67.82%</td><td>67.13%</td><td>67.47%</td></tr><tr><td>Comp.</td><td>34.16%</td><td>31.43%</td><td>30.48%</td></tr><tr><td>Cont.</td><td>65.48%</td><td>63.06%</td><td>64.93%</td></tr><tr><td>Temp.</td><td>40.22%</td><td>38.83%</td><td>38.36%</td></tr></table>
+
+Table 3: Ablation study for linguistic evidence by using different word encodings as the OMN input.
+
+- BiLSTM (B) is the building block of DAN, without two attentions and word-pair offsets.  
+- B+SelfAtt is a subpart of DAN, with only self-attention, but without interactive attention and word-pair offsets.  
+- B+InterAtt is a subpart of DAN, with only interactive attention, but without self-attention and word-pair offsets.  
+- B+DualAtt (DAN) is only the DAN module, with two attentions, but without word-pair offsets.  
+- WordPair (OMN) is only the OMN module, without argument representation for semantic connection.  
+- B+WordPair combines the OMN with a BiLSTM for encoding semantic connection but without any attention.  
+- B+DualAtt+WordPair is our MANF model.
+
+Table 4 presents the results of our module ablation study. Among the first four models without using word-pair offsets, we first observe that the bare BiLSTM cannot outperform those employing attention(s) to differentiate words in argument representation learning. On the other hand, the B+DualAtt achieves better performance compared with the B+SelfAtt and B+InterAtt each using only one kind of attention, except a slight gap of Acc in the four-way classification. This indicates that our
+
+fusion of both attention mechanisms is an effective approach to augment semantic connection learning for an argument pair.
+
+We also observe that the WordPair(OMN) only exploiting word-pair offsets performs the worst among all models. This, however, is not unexpected, as it totally ignores an argument semantics as well as latent semantic connection between arguments. On the other hand, the B+WordPair model, fusing linguistic evidence with semantic connection even learned by a bare BiLSTM without any attention, can greatly improve the performance of WordPair(OMN). The B+WordPair model can even achieve the best or the second best in some cases. This again validates our arguments of encoding and fusing both semantic connection and linguistic evidence to improve relation recognition.
+
+Table 5 presents experiments using fine-tuned BERT to replace the word2vec based BiLSTM for semantic connection encoding. In contrast, the OMN module uses the BERT without fine-tuning to exploit linguistic evidence. The first three ablation modules correspond to the BiLSTM (B), B+DualAtt (DAN) and B+WordPair, respectively. We can observe that the BERT+DualAtt and BERT+WordPair models achieve better performance than the baseline BERT model. This further confirms the necessity of fusing both attention mechanisms and exploiting linguistic evidence. Finally, our MANF model yields substantial improvements overall ablation modules, and the outstanding performance approves our arguments and design objectives.
+
+# 4.3 Case Study
+
+We use case study to visualize and compare different attention mechanisms. Fig. 2 visualizes the word weight obtained by self-attention and interactive attention for four cases of different relation senses. We observe that the two attentions assign different weights to different words. In particular, the interactive attention seems to mainly focus on
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">Four-way Classification</td><td colspan="4">Binary Classification (F1)</td></tr><tr><td>F1</td><td>Acc</td><td>Expa.</td><td>Comp.</td><td>Cont.</td><td>Temp.</td></tr><tr><td>BiLSTM (B)</td><td>47.80%</td><td>57.67%</td><td>63.07%</td><td>28.05%</td><td>61.79%</td><td>36.40%</td></tr><tr><td>B+SelfAtt</td><td>49.39%</td><td>59.16%</td><td>66.79%</td><td>30.80%</td><td>64.72%</td><td>36.57%</td></tr><tr><td>B+InterAtt</td><td>50.70%</td><td>59.63%</td><td>67.30%</td><td>30.15%</td><td>62.36%</td><td>36.33%</td></tr><tr><td>B+DualAtt (DAN)</td><td>51.64%</td><td>59.50%</td><td>67.50%</td><td>32.18%</td><td>65.42%</td><td>38.53%</td></tr><tr><td>WordPair (OMN)</td><td>39.62%</td><td>51.22%</td><td>60.81%</td><td>25.95%</td><td>57.37%</td><td>26.87%</td></tr><tr><td>B+WordPair</td><td>50.95%</td><td>60.31%</td><td>67.01%</td><td>34.30%</td><td>63.15%</td><td>36.81%</td></tr><tr><td>B+DualAtt+WordPair (MANF)</td><td>53.14%</td><td>60.45%</td><td>67.82%</td><td>34.16%</td><td>65.48%</td><td>40.22%</td></tr></table>
+
+Table 4: Experiment results of module ablation study.  
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">Four-way Classification</td><td colspan="4">Binary Classification (F1)</td></tr><tr><td>F1</td><td>Acc</td><td>Expa.</td><td>Comp.</td><td>Cont.</td><td>Temp.</td></tr><tr><td>BERT</td><td>54.74%</td><td>62.69%</td><td>68.01%</td><td>34.75%</td><td>64.45%</td><td>40.25%</td></tr><tr><td>BERT+DualAtt (DAN)</td><td>55.23%</td><td>62.21%</td><td>68.18%</td><td>35.70%</td><td>65.07%</td><td>40.37%</td></tr><tr><td>BERT+WordPair</td><td>55.02%</td><td>61.67%</td><td>68.49%</td><td>36.12%</td><td>65.45%</td><td>42.65%</td></tr><tr><td>BERT+DualAtt+WordPair (MANF)</td><td>56.63%</td><td>64.04%</td><td>70.00%</td><td>35.83%</td><td>66.77%</td><td>42.13%</td></tr></table>
+
+Table 5: Experiment results with the fine-tuned BERT language model.
+
+one word with a very high weight in each argument, which is generally a kind of function word, such as the "in", "back" in the Temporal case, "His", "and" in Contingency case, and "I" in Comparison case. Such function words may be regarded as serving a kind of rhetorical devices for some common linguistic regularities and patterns.
+
+On the other hand, the self-attention tends to assign several words in one argument with similar yet non-ignorable weights, which are often kinds of content words, such as "slithered", "and", "slipped" in the Expansion case. Such a few of content words might be more important to capture the contextual semantics of an argument, which can be next exploited for encoding semantic connection between two arguments. Such functionality differences of the two attentions indeed have motivated us to try a fusion mechanism, so as for each to excel in relation recognition of different senses.
+
+Fig. 3 visualizes the weight matrix of word-pair offsets in the OMN module but with different input. It can be observed that using pre-trained word embeddings can help emphasizing the word-pair "don't-did" probably for their generally containing fewer contextual information. On the other hand, the other two using word contextual encoding pay attentions to word-pairs much similar to those words in the self-attention and interactive attention, such as "I-I", "I-think", "I-don't". As word contextual encoding has already been exploited in the DAN module, we argue that using
+
+pre-trained word embeddings for word-pair offsets could complete argument representation learning from another view of common linguistic evidence.
+
+# 5 Related Work
+
+The IDRR task is usually approached as a classification problem, and the key is the argument representation.
+
+Machine learning approaches, like using a Naive Bayes, Support Vector Machine (SVM) classifier, have designed various features to capture lexical, syntactic regularity and contextual information as argument representation (Pitler et al., 2008; Lin et al., 2009; Pitler et al., 2009; Louis et al., 2010). However, manually crafting features is not only time-consuming and labor-intensive, but also suffers from data sparsity problem due to the use of one-hot feature encoding.
+
+Deep learning models have prevailed for their capabilities of automatic learning argument representation (Zhang et al., 2015; Rutherford et al., 2017). For example, the SCNN model (Zhang et al., 2015) obtains each argument representation via a single convolution layer, and the concatenation of two arguments' representations is used for relation classification. Rutherford et al. (Rutherford et al., 2017) employ a LSTM network to capture word contextual semantics for argument representation. Some hybrid models have attempted to combine CNN, LSTM, graph convolutional networks and etc. for more sophisticated argument representa
+
+![](images/70a4d964f1eeb74f60527b51bd3e32db4ff339f1a1647ad0b86d5270badd068b.jpg)  
+Figure 2: Visualization of attention weights for four cases of relations senses.
+
+![](images/70578b81fa6752b3c86995aeca02ca6f2b04df45d85c4f63de9affac790483b6.jpg)  
+(a) Pre-trained word embedding
+
+![](images/2b8b4c0ceaa5890e259cc95e44ff75471738105658932a68fb7fc62b920687ed.jpg)  
+(b) Shared BiLSTM  
+Figure 3: Visualization of the weight matrix of word-pair offsets in the OMN module with the input of (a) pretrained word embeddings, (b) hidden states of shared BiLSTM, and (c) hidden states of parallel BiLSTM.
+
+![](images/46f8ab4182efbbc022b982f70511965a83303b81ed7fa6086cdaa3aa7cde1c70.jpg)  
+(c) Parallel BiLSTM
+
+tion (Guo et al., 2019; Xu et al., 2019; Zhang et al., 2021). These approaches, however, have ignored the fact that different words may contribute differently in argument representation learning.
+
+Attention mechanisms can guide a neural model to unequally encode each word according to its contextual importance for argument representation (Zhou et al., 2016; Bai and Zhao, 2018; Liu and Li, 2016; Guo et al., 2018, 2020). For example, Zhou et al. (Zhou et al., 2016) apply self-attention to weight a word according to its similarity to its belonging argument. Guo et al. (Guo et al., 2018, 2020) adopt an interactive attention to differentiae words in one argument, where a word is weighted according to the similarity between its encoding and another argument representation. Liu and Li (Liu and Li, 2016) design a multi-level attention to repeatedly compute word importance in a hierarchical way. Ruan et al. (Ruan et al., 2020) propose a pipeline workflow to apply interactive attention after self-attention.
+
+Word pair features have been exploited in machine learning and deep learning approaches for argument representation (Blair-Goldensohn et al., 2007; Biran and McKeown, 2013; Zhou et al., 2013; Chen et al., 2016a,b). For example, Biran
+
+and McKeown (Biran and McKeown, 2013) compute the appearance probabilities of aggregated word pairs to train a logistic regression classifier. Chen et al. (Chen et al., 2016b) construct a relevance score word-pair interaction matrix based on a bilinear model (Jenatton et al., 2012) and a single layer neural model (Collobert and Weston, 2008).
+
+The proposed MANF model is a deep neural model, employing a hierarchical fusion mechanism to fuse two kinds of attentive word encodings as well as word pair offset encodings in argument representation learning.
+
+# 6 Concluding Remarks
+
+In this paper, we argue that implicit relation recognition can benefit from both semantic connection and linguistic evidence between arguments. Motivated from such considerations, we have designed the MANF model to encode and fuse them for the IDRR task. The MANF model consists a DAN module to fuse both self-attentive and interattentive contextual semantics for learning representation of semantic connection, and a OMN module to attentively encode word-pair offsets for learning representation of linguistic evidence. Both kinds of representations are finally fused for rela
+
+tion recognition. Experiments on the latest PDTB 3.0 corpus have validated our design objectives for the new benchmark performance established by our MANF model.
+
+This paper has employed the pre-trained word embeddings trained by the word2vec and BERT; While other pre-training models shall also be adopted and tested in our future work. The performance differences of recognizing different relation senses also motivate to further investigate other advanced fusion mechanisms.
+
+# Acknowledgements
+
+This work is supported in part by National Natural Science Foundation of China (Grant No: 62172167). We also want to use our MANF model on MindSpore<sup>4</sup>, which is a new deep learning computing framework. These problems are left for future work.
+
+# References
+
+Hongxiao Bai and Hai Zhao. 2018. Deep Enhanced Representation for Implicit Discourse Relation Recognition. In Proceedings of the 27th International Conference on Computational Linguistics, pages 571-583, Stroudsburg, PA. The Association for Computational Linguistics.  
+Or Biran and Kathleen McKeown. 2013. Aggregated Word Pair Features for Implicit Discourse Relation Disambiguation. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 69-73, Stroudsburg, PA. The Association for Computational Linguistics.  
+Sasha Blair-Goldensohn, Kathleen McKeown, and Owen Rambow. 2007. Building and Refining Rhetorical-Semantic Relation Models. In Human Language Technology Conference of the North American Chapter of the Association of Computational Linguistics, pages 428-435, Stroudsburg, PA. The Association for Computational Linguistics.  
+Jifan Chen, Qi Zhang, Pengfei Liu, and Xuanjing Huang. 2016a. Discourse relations detection via a mixed generative-discriminative framework. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, pages 2921-2927, Menlo Park, Calif. AAAI Press.  
+Jifan Chen, Qi Zhang, Pengfei Liu, Xipeng Qiu, and Xuan-Jing Huang. 2016b. Implicit discourse relation detection via a deep architecture with gated relevance network. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1726-1735, Stroudsburg, PA. The Association for Computational Linguistics.
+
+Ronan Collobert and Jason Weston. 2008. A unified architecture for natural language processing: Deep neural networks with multitask learning. In Proceedings of the 25th international conference on Machine learning, pages 160-167, New York, NY. ACMPress.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association Computational Linguistics, pages 4171-4186, Stroudsburg, PA. The Association for Computational Linguistics.  
+Fengyu Guo, Ruifang He, and Jianwu Dang. 2019. Implicit discourse relation recognition via a bilstm-cnn architecture with dynamic chunk-based max pooling. IEEE Access, 7:169281-169292.  
+Fengyu Guo, Ruifang He, Jianwu Dang, and Jian Wang. 2020. Working Memory-Driven Neural Networks with a Novel Knowledge Enhancement Paradigm for Implicit Discourse Relation Recognition. In Proceedings of the Thirty-Fourth AAAI Conference on Artificial Intelligence, pages 7822-7829, Menlo Park, Calif. AAAI Press.  
+Fengyu Guo, Ruifang He, Di Jin, Jianwu Dang, Longbiao Wang, and Xiangang Li. 2018. Implicit discourse relation recognition using neural tensor network with interactive attention and sparse learning. In Proceedings of the 27th International Conference on Computational Linguistics, pages 547-558, Stroudsburg, PA. The Association for Computational Linguistics.  
+Francisco Guzmán, Shafiq Joty, Lluis Márquez, and Preslav Nakov. 2014. Using discourse structure improves machine translation evaluation. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 687-698, Stroudsburg, PA. The Association for Computational Linguistics.  
+Rodolphe Jenatton, Nicolas L Roux, Antoine Bordes, and Guillaume R Obozinski. 2012. A latent factor model for highly multi-relational data. In Advances in neural information processing systems, pages 3167-3175, Cambridge, MA. MIT Press.  
+Yangfeng Ji and Jacob Eisenstein. 2015. One vector is not enough: Entity-augmented distributed semantics for discourse relations. Transactions of the Association for Computational Linguistics, 3:329-344.  
+Man Lan, Jianxiang Wang, Yuanbin Wu, Zheng-Yu Niu, and Haifeng Wang. 2017. Multi-task attention-based neural networks for implicit discourse relationship representation and identification. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1299-1308, Stroudsburg, PA. The Association for Computational Linguistics.
+
+Maria Liakata, Simon Dobnik, Shyamasree Saha, Colin Batchelor, and Dietrich Rebholz Schuhmann. 2013. A discourse-driven content model for summarising scientific articles evaluated in a complex question answering task. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 747-757, Stroudsburg, PA. The Association for Computational Linguistics.  
+Ziheng Lin, Min-Yen Kan, and Hwee Tou Ng. 2009. Recognizing Implicit Discourse Relations in the Penn Discourse Treebank. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 343-351, Stroudsburg, PA. The Association for Computer Linguistics.  
+Yang Liu and Sujian Li. 2016. Recognizing Implicit Discourse Relations via Repeated Reading: Neural Networks with Multi-Level Attention. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1224-1233, Stroudsburg, PA. The Association for Computational Linguistics.  
+Annie Louis, Aravind Joshi, Rashmi Prasad, and Ani Nenkova. 2010. Using entity features to classify implicit discourse relations. In Proceedings of the SIG-DIAL 2010 Conference, pages 59-62, Stroudsburg, PA. The Association for Computer Linguistics.  
+Dehong Ma, Sujian Li, Xiaodong Zhang, and Houfeng Wang. 2017. Interactive Attention Networks for Aspect-Level Sentiment Classification. In Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, pages 4068-4074, Amsterdam, Netherlands. Elsevier.  
+Fandong Meng, Zhengdong Lu, Hang Li, and Qun Liu. 2016. Interactive Attention for Neural Machine Translation. In Proceedings of the 26th International Conference on Computational Linguistics: Technical Papers, pages 2174-2185, Stroudsburg, PA. The Association for Computational Linguistics.  
+Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems, pages 3111-3119, Cambridge, MA. MIT Press.  
+Emily Pittler, Annie Louis, and Ani Nenkova. 2009. Automatic sense prediction for implicit discourse relations in text. In Proceedings of the 47th Annual Meeting of the Association for Computational Linguistics and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 683-691, Stroudsburg, PA. The Association for Computer Linguistics.  
+Emily Pitler, Mridhula Raghupathy, Hena Mehta, Ani Nenkova, Alan Lee, and Aravind K. Joshi. 2008.
+
+Easily Identifiable Discourse Relations. In Proceedings of the 22th International Conference on Computational Linguistics, pages 87-90, Stroudsburg, PA. The Association for Computer Linguistics.  
+Huibin Ruan, Yu Hong, Yang Xu, Zhen Huang, Guodong Zhou, and Min Zhang. 2020. Interactively-Propagative Attention Learning for Implicit Discourse Relation Recognition. In Proceedings of the 28th International Conference on Computational Linguistics, pages 3168-3178, Stroudsburg, PA. The Association for Computational Linguistics.  
+Attapol Rutherford, Vera Demberg, and Nianwen Xue. 2017. A systematic study of neural discourse models for implicit discourse relation. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics, pages 281-291, Stroudsburg, PA. The Association for Computational Linguistics.  
+Chang Wang and Bang Wang. 2020. An End-to-end Topic-Enhanced Self-Attention Network for Social Emotion Classification. In Proceedings of The Web Conference 2020, pages 2210–2219, New York, NY. ACMPress.  
+Wei Xiang and Bang Wang. 2019. A survey of event extraction from text. IEEE Access, 7:173111-173137.  
+Sheng Xu, Peifeng Li, Fang Kong, Qiaoming Zhu, and Guodong Zhou. 2019. Topic Tensor Network for Implicit Discourse Relation Recognition in Chinese. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 608-618, Stroudsburg, PA. The Association for Computational Linguistics.  
+Biao Zhang, Jinsong Su, Deyi Xiong, Yaojie Lu, Hong Duan, and Junfeng Yao. 2015. Shallow convolutional neural network for implicit discourse relation recognition. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2230-2235, Stroudsburg, PA. The Association for Computational Linguistics.  
+Yingxue Zhang, Fandong Meng, Li Peng, Jian Ping, and Jie Zhou. 2021. Context tracking network: Graph-based context modeling for implicit discourse relation recognition. In Proceedings of the 2021 Conference of the North American Chapter of the Association Computational Linguistics, pages 1592-1599, Stroudsburg, PA. The Association for Computational Linguistics.  
+Peng Zhou, Wei Shi, Jun Tian, Zhenyu Qi, Bingchen Li, Hongwei Hao, and Bo Xu. 2016. Attention-Based Bidirectional Long Short-Term Memory Networks for Relation Classification. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, Stroudsburg, PA. The Association for Computational Linguistics.  
+Xiaopei Zhou, Yu Hong, Tingting Che, Jianmin Yao, and Qiaoming Zhu. 2013. An unsupervised ap
+
+proach to inferring implicit discourse relation. Journal of Chinese Information Processing, 27(2):17-25.

encodingandfusingsemanticconnectionandlinguisticevidenceforimplicitdiscourserelationrecognition/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e5658467dbd100651e101b80499ea5a0253adcf8672eff04224a2a110d2f2b2
+size 553722

encodingandfusingsemanticconnectionandlinguisticevidenceforimplicitdiscourserelationrecognition/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b5db36b28337f76dd2c4620468118faad198f525a1b46fcfefae73bd30dd684
+size 366171

endtoendsegmentationbasednewssummarization/c3927bc1-bfb7-48fd-bc81-3ec43ca46ad8_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7ccbe6f74e884b721577b40bf05a24c889f1c63f8f4aa4c6536d3c74ce0f774
+size 82002

endtoendsegmentationbasednewssummarization/c3927bc1-bfb7-48fd-bc81-3ec43ca46ad8_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2a807d772439f0c99fc4a86f1a81a961557bde4a31289cb3ad47a581e3dd1cc2
+size 98908

endtoendsegmentationbasednewssummarization/c3927bc1-bfb7-48fd-bc81-3ec43ca46ad8_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:adcec14f90325f7377a3a341cc43ccde1f87b82fa5bb4639ad1b0b63b7a4e3ea
+size 1033441

endtoendsegmentationbasednewssummarization/full.md

@@ -0,0 +1,336 @@
+# End-to-End Segmentation-based News Summarization
+
+Yang Liu, Chenguang Zhu and Michael Zeng  
+Microsoft Cognitive Services Research
+
+{yaliu10, chezhu, nzeng}@microsoft.com
+
+# Abstract
+
+In this paper, we bring a new way of digesting news content by introducing the task of segmenting a news article into multiple sections and generating the corresponding summary to each section. We make two contributions towards this new task. First, we create and make available a dataset, SEGNEWS, consisting of 27k news articles with sections and aligned heading-style section summaries. Second, we propose a novel segmentation-based language generation model adapted from pretrained language models that can jointly segment a document and produce the summary for each section. Experimental results on SEG NEWS demonstrate that our model can outperform several state-of-the-art sequence-to-sequence generation models for this new task.
+
+# 1 Introduction
+
+In recent years, automatic summarization has received extensive attention in the natural language processing community, due to its potential for processing redundant information. The evolution of neural network models and availability of large-scale datasets have driven the rapid development of summarization systems.
+
+Despite promising results, there are specific characteristics of the traditional summarization task that impedes it to provide more beneficial ways of digesting long news articles. For instance, current news summarization system only provides one genetic summary of the whole article, and when users want to read in more details, the generated summary is not capable of helping navigate the reading. For example, given a news report, current system will output several highlight summaries (Nallapati et al., 2017; Liu and Lapata, 2019; Zhang et al., 2020). Under this circumstance, if a user expect to read more details about one highlight, he will still need to browse the whole article to locate related paragraphs. Meanwhile, when processing a long
+
+news article, current systems usually truncate the text and only generate a summary based on the partial article (Cheng and Lapata, 2016a; Zhang et al., 2020). Although this is reasonable since most important content usually lies in the initial portion, it also makes it difficult for users to quickly access information beyond the truncated portion.
+
+In this paper, we propose a new task of Segmentation-based News Summarization. Given a news article, we aim to identify its potential sections and at the same time, to generate the corresponding summary for each section. This new task provides a novel alternative to summarizing a news article. We argue that it can lead to a more organized way of understanding long articles and facilitates a more effective style of reading documents.
+
+First, segmenting a news article can provide a structural organisation of the content, which is not only helpful to reading but also benefit many important NLP tasks. For example, Brown et al. (1983) states that this kind of multi-paragraph division is one of the most fundamental tasks in discourse. However, many expository texts, like news articles, instruction manuals, or textbooks consist of long sequences of paragraphs with very little structural demarcation (Hearst, 1994), and for these documents a subtopical segmentation can be useful. Second, generating concise text descriptions of each sections further reduces the cognitive burden of reading the article (Florax and Ploetzner, 2010). Previous studies (Paice, 1990; Hearst, 1997) present that subtopic segments with their headings is an effective alternative to traditional summarization tasks.
+
+In this paper, we make two main contributions towards the development of Segmentation-based News Summarization systems.
+
+First, we create and publicize a large-scale
+
+![](images/e9817e63bda356ad0b3e7f06b5d6e953dd19f7f4ba1576048d568f6c58084d52.jpg)  
+Figure 1: One example from the segmentation-based summarization task SEGNEWS. The news article is taken from a CNN news article and we truncate the article for display. CNN editors have divided this article into several sections and written a heading to section. The goal of this task is to automatically identify sub-topic segments of multiple paragraphs, and generate the heading-style summary for each segment. Dotted lines in the figure indicate segment boundaries. In this article, paragraphs 1,2 are annotated as the first segment, paragraphs 3,4 are annotated as the second segment, paragraphs 5,6 are annotated as the third segment, and paragraphs 7,8 are annotated as the forth segment. To the right of the article are the heading-style summaries for segments. Since the first segment is usually an overview of the news, we do not assign a summary to it.
+
+benchmark, SEGNEWS, for Segmentation-based News Summarization task. Figure 4 shows one example article and its aligned segmentation and summaries from SEGNEWS.
+
+Second, we propose a novel end-to-end approach for this task, which can jointly segment an article while generating the corresponding summaries. These two sub-tasks can learn from each other via a shared encoder. The model is equipped with a segmentation-aware attention mechanism, allowing it to capture segmentation information during summary generation. One important advantage of our framework is that it is a non-invasive adaptation of the Transformer (Vaswani et al., 2017) model, i.e. it does not alter the inner structure of Transformers. And our framework can integrate many pretrained language generation models, including BART (Lewis et al., 2020), GPT (Radford et al., 2019) and UNILM (Bao et al., 2020). This enables our framework to enjoy a high degree of flexibility and better performance.
+
+We compare the proposed framework with several state-of-the-art methods on the SEGNEWS benchmark. Both automatic evaluation and human evaluation demonstrate the superiority of our model.
+
+# 2 Related Work
+
+# 2.1 Document Summarization
+
+Document summarization is the task of automatically generating a shorter version text of one or multiple documents while retaining its most important information (Radev et al., 2002). The task has received much attention in the natural language processing community due to its potential for various information access applications. Most large-scale summarization datasets are built on news articles. Popular single-document summarization benchmarks include CNN/DM (Hermann et al., 2015; Nallapati et al., 2016; Cheng and Lapata, 2016a), NYT (Durrett et al., 2016) and XSum (Narayan et al., 2018).
+
+Document summarization can be classified into different paradigms by different factors (Nenkova and McKeown, 2011). And among them, two have consistently attracted attention. extractive approaches form summaries by copying and concatenating the most important spans in a document; while in abstractive summarization, various text rewriting operations generate summaries using words or phrases that are not in the original text.
+
+Recent approaches to extractive summarization frame the task as a sequence labeling problem by taking advantage of the success of neural network architectures (Bahdanau et al., 2015). The idea is to predict a label for each sentence specify
+
+ing whether it should be included in the summary. Existing systems mostly rely on recurrent neural networks (Hochreiter and Schmidhuber, 1997) or Transformer model (Vaswani et al., 2017) to encode the document and obtain a vector representation for each sentence (Nallapati et al., 2017; Cheng and Lapata, 2016b; Liu et al., 2019).
+
+In recent years, neural sequence-to-sequence approaches dominate abstractive summarization methods. Rush et al. (2015) and Nallapati et al. (2016) are among the first to apply the neural encoder-decoder architecture to text summarization. See et al. (2017) enhance this model with a pointer-generator network and a coverage mechanism. Pretrained language models have recently emerged as a key technology for improving abstractive summarization systems. These models first pretrain a language model with self-supervised objectives on large corpora and then fine-tune it on summarization datasets. Liu and Lapata (2019) combine a pretrained encoder based on BERT (Devlin et al.) with a randomly initialized decoder, demonstrating substantial gains on summarization performance. MASS (Song et al., 2019) is an encoder-decoder neural model pretrained with the objective of reconstructing a masked text and can be fine-tuned on summarization tasks. BART (Lewis et al., 2020) is an encoder-decoder Transformer (Vaswani et al., 2017) pretrained by reconstructing a text corrupted with several arbitrary noising functions. Bao et al. (2020) design UNILMv2, a Transformer-based neural network pretrained as a pseudo-masked language model.
+
+# 2.2 Text Segmentation and Outline Generation
+
+Text segmentation has been widely used in the fields of natural language processing and information extraction. Existing methods for text segmentation fall into two categories: unsupervised and supervised. TextTiling (Hearst, 1997) is one of the first unsupervised topic segmentation algorithms. It segments texts in linear time by calculating the similarity between two blocks of words based on the cosine similarity. Choi (2000) introduce a statistical model which can calculate the maximum-probability segmentation of a given text. The TopicTiling (Riedl and Biemann, 2012) algorithm is based on TextTiling, which uses the Latent Dirichlet Allocation to find topical changes within documents. LCSeg (Galley et al., 2003) computes
+
+lexical chains of documents and segments texts by a score which captures the sharpness of the change in lexical cohesion.
+
+Supervised methods have also been proposed for text segmentation. Hsueh et al. (2006) integrate lexical and conversation-based features for topic and sub-topic segmentation. Hernault et al. (2010) use CRF to train a discourse segmenter with a set of lexical and syntactic features. Li et al. (2018) propose SEGBOT which uses a neural network model with a bidirectional recurrent neural network together with a pointer network to select text boundaries in the input sequence.
+
+Recently, Zhang et al. (2019) propose Outline Generation task, aiming to identify potential sections of a multi-paragraph document and generate the corresponding section headings as outlines. This task is in form similar to segmentation-based summarization. However, there are two main differences. First, outline generation focused on academic or encyclopaedic documents, where the section headings are extremely short (on average less than two words) and cannot be considered as a summarization task. Second, since outlines care more about briefly describing their corresponding sections, headings in outlines are independently from each other. In segmentation-based summarization, despite describing the sections, heading-style summaries also devote to navigating the reading, and they are usually related and coherent in content.
+
+# 3 The SEGNEWS Benchmark
+
+# 3.1 Data Collection
+
+In order to study and evaluate the Segmentation-based News Summarization task, we build a new benchmark dataset SEGNEWS. We take CNN website as our article source. As shown in Figure 1, there are a large part of CNN articles which are divided by editors into several sub-topic sections (see Appendix for details). And each section is assigned a heading-style summary also written by these editors. We collect articles published from 2017 to 2021, covering multiple CNN news channels, including US Politics, Business, Health, Entertainment, Travel and Sports. We filter articles with no sub-topic structures or editor written heading-style summaries. Since the first segment is usually an overview of the news, editors do not assign a summary to it. The resulting dataset contains 26,876 news articles. For each article, it has human annotated segmentation structures and each segment
+
+<table><tr><td>#news articles</td><td>26,876</td></tr><tr><td>#paragraphs</td><td>40.31</td></tr><tr><td>#sections per article</td><td>3.17</td></tr><tr><td>#tokens per article</td><td>1362.24</td></tr><tr><td>#tokens per section summary</td><td>4.70</td></tr></table>
+
+Table 1: Data statistics of the SEGNEWS dataset.
+
+![](images/5203a8170e267f74f76bc82223cc2dc662ecd4cdd31d0ea28986dcdd510307db.jpg)  
+Figure 2: The frequency of the non-stop words in summary appearing at different positions of the source article. The positions range from [0, 1024].
+
+has a human-written heading-style summary.
+
+# 3.2 Data Statistics
+
+Table 1 shows the overall statistics of our SEG-NEWS benchmark dataset. We can see that the news articles in SEG-NEWS contain rich structural information and are much longer (1,362 tokens per article) than traditional news summarization datasets: articles in CNN/DM (Cheng and Lapata, 2016b) dataset has an average length of 686.63 tokens and articles in NYT (Sandhaus, 2008) dataset has an average length of 800.04 tokens. This is in line with our motivation that segmentation-based summarization can help readers better understand longer articles.
+
+It has been found that in many news articles, the most important information is often shown at the beginning (Kedzie et al., 2018). We compare SEGNEWS with CNN summarization dataset (Cheng and Lapata, 2016b) to investigate the difference of their positional bias. In Figure 2, we record the position of each non-stop word in the summary that also appears in the article. For both datasets, he beginning of article contains more summary words. However, different from conventional summarization dataset, SEGNEWS dataset has a much smoother position distribution and information in the middle of the article still contributes a lot to the summary.
+
+# 4 Task Formulation
+
+Given a multi-paragraph article, the segmentation-based summarization task aims to: i) identify sections of the article to unveil its inherent sub-topic structure, where each section consists of neighboring paragraphs with a coherent topic, and ii) generate the heading-style summary for each section to concisely summarize the section. Particularly, in one article, summaries of different sections should be coherent in content and consistent in style.
+
+Formally, let  $d$  indicate a document consisting of paragraphs  $[p_1,p_2,\dots ,p_M]$ . The segmentation-based summarization task aims to recognize a sequence of section boundaries  $[b_{1},b_{2},\dots ,b_{N - 1}]$ . These boundaries divide the document into  $N$  sections  $s_1 = [p_1,\ldots ,p_{b_1}]$ ,  $s_2 = [p_{b_1 + 1},\ldots ,p_{b_2}],\dots ,s_N = [p_{b_{N - 1} + 1},\ldots ,p_M]$ . Meanwhile, summarization systems will generate the corresponding section summaries  $[y_1,y_2,\dots ,y_N]$ .
+
+# 5 Systems for Segmentation-based News Summarization
+
+In this section, we present two different frameworks to tackle the segmentation-based summarization task. In Pipeline approach, we first apply a segmentation model to identify potential sections, and then apply a generation model to produce the headings. In Joint approach, one neural model is able to jointly segment an article and produce the summaries. To achieve this, we design a novel segmentation-aware attention mechanism, which allows the model to capture segmentation information when generating summaries. This new attention mechanism can also be considered as a non-invasive adaption for conventional Transformer models. Thus, to take the most advantage of existing pre-trained models, we propose SEGUNILM and SEGBART which are respectively based on pre-trained UNILM model and BART model. They can be initialized completely from pre-trained models and achieve substantial improvement on segmentation-based summarization.
+
+# 5.1 Pipeline Approach
+
+Segmentation model We formulate the section identification process as a sequence labeling task. We insert a special symbol [X_SEP] at the boundary of paragraph  $p_i$  and  $p_{i+1}$ , and then concatenate all paragraphs into a single text input. A neural encoder is then applied to encode this input. Define  $u_i$
+
+![](images/a0b118109a40c987e7a48142455bf4cb1fd16b73009751cc6305ff4fa0042fc3.jpg)  
+Figure 3: The overall framework of SEGTRANS model. The blue circles indicate input source text, where dark blue circles indicate paragraph boundaries. The yellow circles indicate output target text, where orange circles indicate heading boundaries. Dotted red lines indicate attention heads with segmentation-aware attention mechanism and dotted blue lines indicate attention heads with original full attention mechanism.
+
+as the output vector of [X_SEP] after paragraph  $p_i$ . We then apply a binary classifier over  $u_i$  to obtain  $y_i \in \{0,1\}$ .  $y_i = 0$  indicates paragraph  $p_i$  and  $p_{i + 1}$  are in one segmentation, and  $y_{i} = 1$  indicates  $p_{i + 1}$  should be the start of a new segment.
+
+Generation model We then generate an aligned heading-style summary for each identified section  $s_j$ . The generation of each heading is independent. Here, we can choose existing extractive or abstractive summarization methods.
+
+- TOPICRANK (Bouguin et al., 2013) is an extractive method for keyphrase extraction which represents a document as a complete graph depending on topical representations. We use the top ranked phrase as the summary for input section;  
+- SEQ2SEQ represents the sequence-to-sequence neural model, which is usually used in abstractive summarization. It first encodes the concatenated text of all paragraphs within this section, and the decodes the heading in an auto-regressive manner. In experiments, we try both non-pretrained Transformer model and pretrained UNILM and BART models as SEQ2SEQ models.
+
+# 5.2 Joint Approach
+
+Instead of relying on a pipeline framework, we can also tackle the segmentation-based summarization task with a single encoder-decoder neural model. This brings two main advantages. First, the encoders for segmentation and generation can be
+
+shared, benefiting both tasks as a multi-task learner. Second, we can decode all summaries in an auto-regressive manner. In this way, when the decoder generates the  $l$ -th heading, it will be exposed to the 1st to  $(l - 1)$ -th generated headings. This is considerably helpful since in a news article, many headings are highly related and coherent in their content.
+
+We use Transformer (Vaswani et al., 2017) as base model for the encoder and decoder. Formally, the encoder maps a sequence of tokens in the source document  $\pmb{x} = [x_{1},\dots,x_{n}]$  into a sequence of continuous representations  $\pmb{t} = [t_1,\dots,t_n]$ . Then a segment classifier is applied over output vectors of paragraph boundaries to identify correct segments  $B = [b_{1},b_{2},\dots,b_{N - 1}]$  for the input article. The decoder then generates the tokens of target text  $y = (y_{1},\dots,y_{m})$  auto-regressively based on the conditional probability:  $p(y_{1},\dots,y_{m}|x_{1},\dots,x_{n},B)$ . As the decoder produces summaries for all sections in one pass, we add a special symbol [Y_SEP] between summaries from neighboring sections to indicate their boundaries. However, in this vanilla sequence-to-sequence model, during inference, the decoder is not aware of the segmentation results and can only implicitly use this information when decoding the summaries. Thus, to better jointly learn segmentation and generation tasks, we propose SEGTRANS model, which is equipped with Segmentation-aware Attention mechanism.
+
+Segmentation-aware attention The multi-head decoder-to-encoder attention in a Transformer decoder defines that for a head  $z \in \{1, \dots, n_{head}\}$  at each layer, the model calculates attention probabilities  $a_{ij}^{z}$  against each source token  $x_{j}$  when generating the  $i$ -th token  $y_{i}$ .
+
+$$
+q _ {i} ^ {z} = W _ {q} ^ {z} Y _ {i}; k _ {j} ^ {z} = W _ {k} ^ {z} X _ {j}, \tag {1}
+$$
+
+$$
+a _ {i j} ^ {z} = \frac {\exp \left(q _ {i} ^ {z T} k _ {j} ^ {z}\right)}{\sum_ {o = 1} ^ {n} \exp \left(q _ {i} ^ {z T} k _ {o} ^ {z}\right)}, \tag {2}
+$$
+
+where  $Y_{i},X_{j}\in \mathbb{R}^{d}$  are the layer's input vectors corresponding to the token  $y_{i}$  and  $x_{j}$ , respectively.  $W_{q}^{z},W_{k}^{z}\in \mathbb{R}^{d_{head}*d}$  are learnable weights.  $n$  is the number of tokens in source input.
+
+However, in segmentation-based summarization, when generating the heading for the  $i$ -th section, the decoder should focus more on the input tokens belonging to that section. Thus, we propose the segmentation-aware attention as follows.
+
+We select a subset  $\hat{z}$  of decoder heads to apply a segmentation mask to enforce that these heads only
+
+attemto the corresponding section.For a head in  $\hat{z}$  ,Eq.2is modified to:
+
+$$
+a _ {i j} ^ {z} = \frac {\exp \left(q _ {i} ^ {z T} k _ {j} ^ {z}\right) \operatorname {s e g} \left(y _ {i} , x _ {j}\right)}{\sum_ {o = 1} ^ {n} \exp \left(q _ {i} ^ {z T} k _ {o} ^ {z}\right) \operatorname {s e g} \left(y _ {i} , x _ {j}\right)} \tag {3}
+$$
+
+where  $\text{seg}(y_i, x_j)$  is a indicator function. It equals 1 if and only if  $y_i$  and  $x_j$  both belong to the same section, and 0 otherwise. In this manner, parts of the heads in multi-head attention are able to dynamically capture segmentation information, while the other heads still model global features of the entire input article.
+
+We illustrate a detailed example of our framework with segmentation-aware attention in Figure 3. We first encode the source text, and apply a segmentation classification layer over output vectors of paragraph boundaries. For this example input, the model classifies the first and the third paragraph boundaries to be segmentation points. Then the decoder will apply a segmentation-aware multi-head attention over the source outputs. It generates the summary for the first identified section with parts of the attention heads over only the first and the second paragraphs. After generating the first heading ending symbol [Y_SEP], the decoder changes the segmentation-aware attention to the third paragraph for generating the summary for the second section.
+
+The final loss for training SEGTRANS is the summation of the segmentation loss (binary classification loss)  $\mathcal{L}_{\mathrm{seg}}$  and generation loss (negative likelihood loss)  $\mathcal{L}_{\mathrm{gen}}$ .
+
+One advantage of our framework is that it is a non-invasive adaptation of the Transformer model, i.e. it does not alter the inner structure of Transformers. This is important since this adaptation can be applied to many popular pretrained language generation models (e.g. MASS, BART and UNILM), offering our framework a high degree of flexibility and better performance. In this paper, we also augment pre-trained UNILM and BART with this mechanism and propose SEGUNILM and SEGBART to further boost their performance.
+
+# 6 Experiments
+
+In this section, we conduct experiments on SEG-News dataset by comparing our proposed model with several strong baselines.
+
+# 6.1 Experimental Settings
+
+In pre-processing, all the words in news articles and headings are transformed to lower case and to-
+
+kenized with wordpiece tokenizer from BERT (Devlin et al.). In data splitting, we guarantee the headings of articles in the test set have low bigram overlap with articles in the training set. We obtain a splitting of 21,748 articles in training set, 2,688 in validation set and 2,444 in test set.
+
+We experiment under both non-pretrained and pretrained settings. In non-pretrained setting, we use a 6-layer Transformer encoder-decoder model (SEGTRANS) with 512 hidden size and 2,048 feedforward size. In pretrained setting, we propose SEGUNILM and SEGBART which adopts the base version of UNILMv2 (Bao et al., 2020) and the large version of BART (Lewis et al., 2020) as the pretrained model. UNILMv2 is a Transformer-based neural network with 12 Transformer layers and 12 attention heads, pretrained as a pseudo-masked language model. BART is a Transformer-based neural encode-decoder model with 12 layers and 16 attention heads, pretrained via a denoising auto-encoder loss. Label smoothing is used with smoothing factor 0.1. For segmentation-aware attention, we choose the best  $c$  (number of segmentation-aware heads) by experiments on the validation set, and  $c = 9$  for SEGUNILM and  $c = 13$  for SEGBART provide the best performance.
+
+During all decoding we use beam search (size 5), and tune  $\alpha$  for the length penalty (Wu et al., 2016) between 0.6 and 1 on the validation set. To guarantee the number of generated headings can match the number of predicted source segments, we take a trick of only generating the end-of-generation token (EOS) when these two numbers match.
+
+We compare the proposed joint models with two sets of strong baselines. The first set of baselines are vanilla sequence-to-sequence models. These models take complete raw articles as input and output the concatenated headings. The second set are pipeline models. As described, these systems first use a segmentor to divide the article into several sections, and then apply a generator to produce summary for each section.
+
+In segmentation-based summarization, summarization systems require segmentation results. We set two settings of segmentation. For the first setting, we provide golden segments to the models to evaluate their performance of generating the summaries when given the correct segments. For the second setting, we require the models to first segment the article and then generate summaries for the predicted segments.
+
+<table><tr><td colspan="2">Vanilla Seq2Seq</td><td colspan="2">R1</td><td colspan="2">R2</td><td colspan="2">RL</td></tr><tr><td colspan="2">TRANS</td><td colspan="2">8.66</td><td colspan="2">1.51</td><td colspan="2">8.16</td></tr><tr><td colspan="2">UNILM</td><td colspan="2">19.22</td><td colspan="2">7.18</td><td colspan="2">16.99</td></tr><tr><td colspan="2">Pipeline</td><td colspan="3">With Gold Segments</td><td colspan="3">With Predicted Segments</td></tr><tr><td>Segmentor</td><td>Generator</td><td>R1</td><td>R2</td><td>RL</td><td>R1</td><td>R2</td><td>RL</td></tr><tr><td>Transformer</td><td>Transformer</td><td>8.69</td><td>1.83</td><td>9.09</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Transformer</td><td>TopicRank</td><td>5.09</td><td>1.14</td><td>6.28</td><td>-</td><td>-</td><td>-</td></tr><tr><td>BART</td><td>BART</td><td>21.42</td><td>7.76</td><td>19.28</td><td>16.01</td><td>5.27</td><td>14.37</td></tr><tr><td>UNILM</td><td>UNILM</td><td>21.76</td><td>8.22</td><td>19.75</td><td>16.27</td><td>5.45</td><td>14.65</td></tr><tr><td colspan="2">Joint</td><td>R1</td><td>R2</td><td>RL</td><td>R1</td><td>R2</td><td>RL</td></tr><tr><td colspan="2">SEGTRANS</td><td>8.94</td><td>1.85</td><td>9.35</td><td>-</td><td>-</td><td>-</td></tr><tr><td colspan="2">SEGBART</td><td>21.49</td><td>8.29</td><td>19.52</td><td>16.36</td><td>5.14</td><td>14.96</td></tr><tr><td colspan="2">SEGUNILM</td><td>22.17</td><td>8.86</td><td>20.17</td><td>17.59</td><td>6.20</td><td>15.90</td></tr></table>
+
+Table 2: ROUGE F1 results on SEGNEWS test set. R1 and R2 are shorthands for ROUGE scores of unigram and bigram overlap; RL is the ROUGE score of longest common subsequence. In pipeline approach, we try combinations of different segmentators and generators. Due to their failure on segmentation, non-pretraind models have very low ROUGE scores with predicted segments, and we do not compare them in the table.  
+
+<table><tr><td>Models</td><td>R1</td><td>R2</td><td>RL</td></tr><tr><td>SEGUNILM</td><td>22.17</td><td>8.86</td><td>20.17</td></tr><tr><td>(c=12)</td><td>22.14</td><td>8.81</td><td>20.09</td></tr><tr><td>(c=8)</td><td>22.13</td><td>8.84</td><td>20.10</td></tr><tr><td>(c=4)</td><td>21.39</td><td>7.99</td><td>19.23</td></tr><tr><td>(c=0)</td><td>19.85</td><td>7.74</td><td>17.62</td></tr><tr><td>(w/o seg loss)</td><td>22.06</td><td>8.66</td><td>20.02</td></tr></table>
+
+Table 3: Ablation study results on SEGNEWS. We compare multiple variants of SEGUNILM.  $c$  indicates the number of decoder heads modified into segmentation-aware attention. Be default, SEGUNILM uses  $c = 9$  to achieve the best performance. We also present a SEGUNILM model without (w/o) segmentation classification loss, and it is trained solely by generation loss.
+
+# 6.2 Evaluation Metrics
+
+Evaluation metrics for summarization performance are ROUGE (Lin, 2004) F1 scores of the generated headings against the gold headings. We report unigram and bigram overlap (ROUGE-1 and ROUGE-2) as a means of assessing informativeness and the longest common subsequence (ROUGE-L) as a means of assessing fluency.
+
+We use standard metrics  $Pk$  (Beeferman et al., 1999) and WinDiff (Pevzner and Hearst, 2002) to evaluate segmentation results. Lower scores of these two metrics indicate that the predicted segmentation is closer to the ground truth. A EVEN baseline is included for comparison where it segments the whole article evenly.
+
+# 6.3 Results
+
+Table 2 describes our summarization results on the SEGNEWS dataset. The first vertical block includes the results of vanilla sequence-to-sequence models. TRANS is the non-pretrained Transformer encoder-decoder model. UNILM and BART are two
+
+pretrained baseline models. The second vertical block contains the results of pipeline models. We present the combinations of different segmentation models and generation models. For segmentor, we experiment non-pretrained Transformer model and pretrained BART and UNILM models. For generator, we also include TOPICRANK, which is a classical extractive summarization method.
+
+The last vertical block includes the results of our joint models: SEGTRANS, SEGBART and SEGUNILM. They respectively rely on non-pretrained Transformer and pretrained BART and UNILM as backbone models. Segmentation-aware attention mechanism is used to augment these jointly trained systems.
+
+We can see vanilla sequence-to-sequence models with no segmentation information input perform poorly on this task. End-to-end SEGUNILM model achieves the best performance among all systems. SEGUNILM outperforms the best pipeline system under both settings when gold segments or predicted segments are provided. This indicates SEGUNILM has better overall performance and will be more useful when applied as practical applications. It also shows higher summarization results than vanilla UNILM model, confirming the effectiveness of segmentation-aware attention mechanism. SEGBART and SEGTRANS also show similar superiority over their pipeline versions. Examples of system output are shown in Table 4.
+
+Table 3 summarizes ablation studies aiming to assess the contribution of individual components of SEGUNILM. We first modify SEGUNILM by varying  $c$ , the number of heads of segmentation-aware attention. We can see the best results of ROUGE
+
+<table><tr><td></td><td>Title: One JFK conspiracy theory that could be true</td></tr><tr><td>GOLD</td><td>1. LBJ had it done; 2. The military industrial complex did it; 3. The mob did it; 4. Oswald acted alone as part of an unknown conspiracy; 5. The CIA did it</td></tr><tr><td>Pipeline UNILM</td><td>Those Kennedys will never embarrass me again; Did Kennedy want to withdraw us troops from Vietnam?; 3. Different mobs; other conspirators ?; Would America be OK with that ?</td></tr><tr><td>SEGBART</td><td>1. They thought he was a crook; 2. He was going to pull American troops out of Vietnam; 3. The mob did this; 4. There were others, but who were they?; 5. The CIA ordered the killing</td></tr><tr><td>SEGUNILM</td><td>1. Those Kennedy&#x27;s will never embarrass me again; 2. He said he&#x27;d pull troops out of Vietnam; 3. Mob members claim they were witnesses to the alleged shootings; 4. there were more people who knew where Oswald was; 5. The CIA didn&#x27;t release any of the good stuff</td></tr><tr><td></td><td>Title: This man is tasked with finding out who failed Larry Nassar&#x27;s victims</td></tr><tr><td>GOLD</td><td>Seeking justice; A very youthful 68-year-old; A model independent prosecutor</td></tr><tr><td>Pipeline UNILM</td><td>Searching for truth; He couldn&#x27;t stay retired; He didn&#x27;t have an agenda</td></tr><tr><td>SEGBART</td><td>Searching for the truth; Working with juveniles; No stone unturned</td></tr><tr><td>SEGUNILM</td><td>Searching for the truth; He&#x27;s has to do something; He doesn&#x27;t have an agenda</td></tr></table>
+
+Table 4: GOLD reference summaries and automatic summaries produced by pipeline UNILM, SEGBART and SEGUNILM on the SEGNEWS datasets. Semicolons indicate the boundaries of headings.  
+
+<table><tr><td>Model</td><td>WD</td><td>PK</td></tr><tr><td>EVEN</td><td>0.469</td><td>0.450</td></tr><tr><td>Transformer</td><td>0.563</td><td>0.462</td></tr><tr><td>BART</td><td>0.484</td><td>0.411</td></tr><tr><td>UNILM</td><td>0.479</td><td>0.391</td></tr><tr><td>SEGBART</td><td>0.471</td><td>0.405</td></tr><tr><td>SEGUNILM</td><td>0.462</td><td>0.380</td></tr></table>
+
+Table 5: Experimental results on document segmentation task. WD indicates WinDiff metric.  
+
+<table><tr><td>Model</td><td>Quality</td><td>Fluency</td></tr><tr><td>Pipeline UNILM</td><td>1.93</td><td>2.62</td></tr><tr><td>SEGUNILM</td><td>2.17</td><td>2.59</td></tr><tr><td>Gold</td><td>2.44</td><td>2.79</td></tr></table>
+
+Table 6: Human evaluation results based on summary quality and fluency.
+
+are achieved when  $c = 9$ . With more or less heads modified as segmentation-aware attention heads, the summarization performance show a clear trend of decreasing. Also, as shown in the last column, when segmentation layer and segmentation loss are removed, we observe a sharp decrease on ROUGE scores. The results prove that both segmentation-aware attention and joint training provide improvement to the summarization results.
+
+Table 5 describes the results on news segmentation task. SEGUNILM achieves the lowest WD and PK scores, revealing its ability to identify the structure of a news article. Compared with UNILM model without the segmentation-aware attention, SEGUNILM shows clear superiority on both metrics. The same trend is also observed in BART related models.
+
+# 6.4 Human Evaluation
+
+In addition to automatic evaluation, we also assess system performance by eliciting human judgments on 20 randomly selected test instances. The evaluation study assess the overall quality and fluency of the summaries by asking participants to rate them. We present the news article to evaluators along with system generated heading-style summaries, and we ask evaluators to read the complete article, and give scores based on summary quality and fluency respectively. Participants can have three scores (1-low quality/fluency, 2-median quality/fluency, 3-high quality/fluency).
+
+Gold summaries, outputs from pipeline UNILM and SEGUNILM models are compared in evaluation. We invite three evaluators with linguist background to conduct the human evaluation. The averaged results are shown in Table 4. Overall, we observe pipeline UNILM and SEGUNILM perform similarly on fluency, but SEGUNILM shows its superiority on summary quality. Gold summaries are marginally better than automatic generated summaries.
+
+# 7 Conclusion
+
+In this work, we proposed a new task, segmentation-based news summarization. It aims to segment a news article into multiple sections and generate the corresponding summary to each section. This new task provides a novel alternative to digesting a news article. We built a new benchmark dataset SEGNEWS to study and evaluate the task. Furthermore, we designed a segmentation-aware attention mechanism, which allows neural decoder to capture segmentation information in the source
+
+texts. We jointly train the model for generating summaries and recognizing news segments. Experimental results on SEGNEWS demonstrate that our framework produces better segmentation-based summaries than competitive systems.
+
+# 8 Ethical Statement
+
+We honor and support the ACL Code of Ethics. We have used only the publicly available news articles from the CNN website and adhere to their only-for-research-purpose guideline. Meanwhile, to make sure the downstream usage of the data will not break the permission of CNN website, we only release the URLs of these articles along with a script to download and process them.
+
+The content of the news and summaries only reflect the views of the media, and should be viewed with discretion.
+
+# References
+
+Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In *In Proceedings of the 3rd International Conference on Learning Representations*, San Diego, California.  
+Hangbo Bao, Li Dong, Furu Wei, Wenhui Wang, Nan Yang, Xiaodong Liu, Yu Wang, Songhao Piao, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. 2020. UniLMv2: Pseudo-masked language models for unified language model pre-training. arXiv preprint arXiv:2002.12804.  
+Doug Beeferman, Adam Berger, and John Lafferty. 1999. Statistical models for text segmentation. Machine Learning, 1(34):177-210.  
+Adrien Bougouin, Florian Boudin, and Beatrice Daille. 2013. Topicrank: Graph-based topic ranking for keyphrase extraction. In International joint conference on natural language processing (IJCNLP), pages 543-551.  
+Gillian Brown, Gillian D Brown, Gillian R Brown, Brown Gillian, and George Yule. 1983. Discourse analysis. Cambridge university press.  
+Jianpeng Cheng and Mirella Lapata. 2016a. Neural summarization by extracting sentences and words. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 484-494, Berlin, Germany.  
+Jianpeng Cheng and Mirella Lapata. 2016b. Neural summarization by extracting sentences and words. In Proceedings of the ACL Conference.  
+Freddy Y. Y. Choi. 2000. Advances in domain independent linear text segmentation. In 1st Meeting of
+
+the North American Chapter of the Association for Computational Linguistics.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the NAACL Conference.  
+Greg Durrett, Taylor Berg-Kirkpatrick, and Dan Klein. 2016. Learning-based single-document summarization with compression and anaphoricity constraints. In Proceedings of the ACL Conference.  
+Mareike Florax and Rolf Ploetzner. 2010. What contributes to the split-attention effect? the role of text segmentation, picture labelling, and spatial proximity. Learning and instruction, 20(3):216-224.  
+Michel Galley, Kathleen McKeown, Eric Fosler-Lussier, and Hongyan Jing. 2003. Discourse segmentation of multi-party conversation. In Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics, pages 562-569.  
+Marti A. Hearst. 1994. Multi-paragraph segmentation expository text. In 32nd Annual Meeting of the Association for Computational Linguistics, pages 9–16, Las Cruces, New Mexico, USA. Association for Computational Linguistics.  
+Marti A Hearst. 1997. Text tiling: Segmenting text into multi-paragraph subtopic passages. Computational linguistics, 23(1):33-64.  
+Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. Advances in neural information processing systems, 28.  
+Hugo Hernault, Danushka Bollegala, and Mitsuru Ishizuka. 2010. A sequential model for discourse segmentation. In International Conference on Intelligent Text Processing and Computational Linguistics, pages 315-326. Springer.  
+Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735-1780.  
+Pei-Yun Hsueh, Johanna D Moore, and Steve Renals. 2006. Automatic segmentation of multiparty dialogue. In 11th Conference of the European Chapter of the Association for Computational Linguistics.  
+Chris Kedzie, Kathleen McKeown, and Hal Daumé III. 2018. Content selection in deep learning models of summarization. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1818-1828, Brussels, Belgium.  
+Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training
+
+for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.  
+Jing Li, Aixin Sun, and Shafiq R Joty. 2018. Segbot: A generic neural text segmentation model with pointer network. In IJCAI, pages 4166-4172.  
+Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out: Proceedings of the ACL-04 Workshop, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.  
+Yang Liu and Mirella Lapata. 2019. Text summarization with pretrained encoders. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3730-3740, Hong Kong, China.  
+Yang Liu, Ivan Titov, and Mirella Lapata. 2019. Single document summarization as tree induction. In Proceedings of the NAACL Conference.  
+Ramesh Nallapati, Feifei Zhai, and Bowen Zhou. 2017. Summarrunner: A recurrent neural network based sequence model for extractive summarization of documents. In Proceedings of the AAAI Conference.  
+Ramesh Nallapati, Bowen Zhou, Cicero dos Santos, Caglar Gulcehre, and Bing Xiang. 2016. Abstractive text summarization using sequence-to-sequence RNNs and beyond. In Proceedings of The 20th SIGNLL Conference on Computational Natural Language Learning, pages 280-290, Berlin, Germany. Association for Computational Linguistics.  
+Shashi Narayan, Shay B. Cohen, and Mirella Lapata. 2018. Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1797-1807, Brussels, Belgium.  
+Ani Nenkova and Kathleen McKeown. 2011. Automatic summarization. Foundations and Trends in Information Retrieval, 5(2-3):103-233.  
+Chris D Paice. 1990. Constructing literature abstracts by computer: techniques and prospects. Information Processing & Management, 26(1):171-186.  
+Lev Pevzner and Marti A Hearst. 2002. A critique and improvement of an evaluation metric for text segmentation. Computational Linguistics, 28(1):19-36.  
+Dragomir R. Radev, Eduard Hovy, and Kathleen McKeown. 2002. Introduction to the special issue on summarization. Computational Linguistics, 28(4):399-408.
+
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.  
+Martin Riedl and Chris Biemann. 2012. Topictiling: a text segmentation algorithm based on lda. In Proceedings of ACL 2012 Student Research Workshop, pages 37-42.  
+Alexander M. Rush, Sumit Chopra, and Jason Weston. 2015. A neural attention model for abstractive sentence summarization. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 379-389, Lisbon, Portugal.  
+Evan Sandhaus. 2008. The New York Times Annotated Corpus. Linguistic Data Consortium, Philadelphia, 6(12).  
+Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointer-generator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1073-1083, Vancouver, Canada.  
+Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and TieYan Liu. 2019. MASS: Masked sequence to sequence pre-training for language generation. In Proceedings of International Conference on Machine Learning, pages 5926-5936, Long Beach, California.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems 30, pages 5998-6008.  
+Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. 2016. Google's neural machine translation system: Bridging the gap between human and machine translation. In arXiv preprint arXiv:1609.08144.  
+Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter Liu. 2020. Pegasus: Pre-training with extracted gap-sentences for abstractive summarization. In International Conference on Machine Learning, pages 11328-11339. PMLR.  
+Ruqing Zhang, Jiafeng Guo, Yixing Fan, Yanyan Lan, and Xueqi Cheng. 2019. Outline generation: Understanding the inherent content structure of documents. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 745-754.
+
+# A Build SEGNEWS from CNN website
+
+The SEGNEWS dataset is built from news articles on CNN website. For many news reports on CNN, news editors manually divide them into several sections and write a heading-style summary for each section. As illustrated in Figure 1, in a display of this news article<sup>2</sup>, it has a general title "Global businesses must address climate change before it's too late". Below the title, there are several paragraphs of news content. This news article is divided into 5 sections. Despite the first section, the other 4 sections are assigned with their heading-style summaries: "Reduce their own emissions", "Disclose risks and adopt new reporting standards", "Educate employees" and "Advocate for climate policies".
+
+We crawl news articles like this from CNN website. Articles without segmentation information or headings are filtered. The resulting SEGNEWS dataset contains 26,876 articles. Each instance in SEGNEWS consists of a news article, its segmentation structure and heading-style summaries for each segments.
+
+![](images/58c8296abfba324b3b53f6565ad8e5b8f156203bf9c160e928a6aabe86aa2c01.jpg)
+
+# PERSPECTIVES
+
+# Global businesses must address climate change before it's too late
+
+Updated by: Panit Ramani & Henry Business Media Ltd.  
+Updated: 1221 GMT (2021 HKT) August, 9, 2021
+
+Editor's Note: Punit Renjen is the CEO of Deloitte Global. The opinions expressed in this commentary are his own.
+
+Climate change poses one of the greatest threats humanity has ever faced. In the past few weeks alone, wildlife have emerged across the globe, brutal heatwaves have devastated American cities and floodings has claimed the lives of hundreds in Europe and Asia. The global warming crisis is also a major concern for many countries.
+
+Making meaningful, measurable progress is a monumental task. For the sake of the planet and future generations, it is vital that global businesses, and the professionals who run them, step up in the fight against climate change and take urgent action.
+
+Here's how:
+
+# Reduce their own emissions
+
+At the recent GT Summit, world leaders doubled down on their climate pledges, focusing on the opportunity for a just energy transition that will likely create clean energy jobs around the world. Among other things, they committed to reach net-zero carbon emissions by 2030 and to achieve net-zero carbon emissions by 2050.
+
+The business community must match the ambition of world governments by cutting emissions across their own operations.
+
+Hundreds of companies have now taken the first step in doing so: committing to  $100\%$  renewable energy. The tech sector in particular is leading this shift, with a number of companies already reaching their  $100\%$  goal, and many more setting science-based plans to get there.
+
+# Disclose risks and adopt new reporting standards
+
+Companies must analyze the financial risks that climate change poses to them — and publicly disclose that information.
+
+Global capital markets need high-quality, consistent and comparable data to understand drivers of risk and return, allocate capital efficiently and finance the transition to a more resilient, low-carbon economy.
+
+![](images/6c4c0d12775a261a796ce3eb49b0cdbad7c12ea8262e2065b739e353eb4ea58b.jpg)  
+Related Article: How the business community can help protect voting rights.  
+Figure 4: One example news article on CNN website. It contains human-annotated segments and heading-style summaries.
+
+By providing financial markets with the right information, we can build confidence that money flows where it needs to go to boost resiliency and curb emissions across the globe. This means companies exposed to long-term climate-related risks may see higher costs of doing business in the long term. The current climate solutions could have access to cheaper capital. We are seeing this play out already.
+
+Allianz, for instance, will no longer offer property or casualty insurance coverage to mining companies. The company's "risk-free" stock market backrock is setting limits on its investments in companies that are exposed to climate risks.
+
+I am encouraged that businesses are increasingly reporting on the impact of climate change, and taking steps to increase transparency and accountability through initiatives like the World Economic Forum's Stakeholder Capitalism Metrics, a set of environmental, social and economic indicators that measure the impacts of climate change. I also encourage the G3T to adopt climate reporting standards reinforces the direction that the regulatory environment is
+
+heading in.
+
+# Educate employees
+
+As business leaders, we need to recognize that our people are our greatest asset - our "superpower." With organization-wide climate education programs, businesses can develop a culture of sustainability and climate-conscious thinking at the very core of their work. That is why, starting this month, Deloitte has begun to roll out a new climate learning program for all 330,000 of its professionals worldwide.
+
+Developed in collaboration with World Wildlife Fund (WWF), the program is designed to engage our people on the impacts of climate change and its consequences on biodiversity. The WWF's goal is to develop a framework that supports the implementation of sustainable business models, culture of climate action, we will create a network of support for a transition to sustainable business models with far-reaching influence.
+
+# Advocate for climate policies
+
+Businesses must leverage their collective power to advocate for climate policies that meet the ambition of reaching net-zero by 2050 or earlier.
+
+Alongside fellow CEOs from across the world, I recently signed on to a call for government leaders to place a price on carbon and invest in the infrastructure needed to accelerate a transition away from fossil fuels. Companies ranging from shipping and logistics giants to major retailers have taken similar steps to drive systemic change in recent years. And even the oil and gas sector has lent its support for carbon pricing.
+
+I call on other leaders to rise to this challenge. They must make climate knowledge a core competency across their businesses — a springboard for greater action — and take concrete steps to address the climate crisis head-on.
+
+Watching the world come together to fight Covid-19 has been nothing short of inspiring, but as communities begin to recover from the pandemic, we are more aware that our efforts will be built on this global cooperation that we know it possible and step up to the next greatest challenge humanity has ever confronted.
+
+For companies to build long-term sustainable value for all stakeholders, they must do their part to build an equitable and sustainable future. Our future depends on it.

endtoendsegmentationbasednewssummarization/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:740dbba2ecf542ae09a9a55b019fa63d92d47ee020c72a7788beaf508ffc0228
+size 437816

endtoendsegmentationbasednewssummarization/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:405eecc59e72ae73a2fe4a1cbb56ecfa2a7d52da29431f742bb07f71a56dc7d0
+size 353798

endtoendspeechtranslationforcodeswitchedspeech/0d727901-bfd5-475f-af31-7c070c33cbaa_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d653e9b8302d95f5039a739439d89e2666dc198fab87f308c3a7901a6876e02
+size 92645

endtoendspeechtranslationforcodeswitchedspeech/0d727901-bfd5-475f-af31-7c070c33cbaa_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:15485dfc7877b7cedb826cae09c85268e43ffe12f69de6bf9ae8fb77d94b4418
+size 111998

endtoendspeechtranslationforcodeswitchedspeech/0d727901-bfd5-475f-af31-7c070c33cbaa_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bbc50fc058ad14f88d64642e7b741cba206664ff155f7b77d3444cef73200b4f
+size 840584

endtoendspeechtranslationforcodeswitchedspeech/full.md

@@ -0,0 +1,300 @@
+# End-to-End Speech Translation for Code Switched Speech
+
+Orion Weller $^{1*}$ , Matthias Sperber $^{2}$ , Telmo Pires $^{2}$ , Hendra Setiawan $^{2}$ , Christian Gollan $^{2}$ , Dominic Telaar $^{2}$ , Matthias Paulik $^{2}$ $^{1}$ Johns Hopkins University  
+ $^{2}$ Apple oweller@cs.jhu.edu, sperber@apple.com
+
+# Abstract
+
+Code switching (CS) refers to the phenomenon of interchangeably using words and phrases from different languages. CS can pose significant accuracy challenges to NLP, due to the often monolingual nature of the underlying systems. In this work, we focus on CS in the context of English/Spanish conversations for the task of speech translation (ST), generating and evaluating both transcript and translation. To evaluate model performance on this task, we create a novel ST corpus derived from existing public data sets. We explore various ST architectures across two dimensions: cascaded (transcribe then translate) vs end-to-end (jointly transcribe and translate) and unidirectional (source  $\rightarrow$  target) vs bidirectional (source  $\leftrightarrow$  target). We show that our ST architectures, and especially our bidirectional end-to-end architecture, perform well on CS speech, even when no CS training data is used.
+
+# 1 Introduction
+
+Over half of the world's population is estimated to be bilingual.2 Those that know multiple languages are prone to code switch, i.e., to interchangeably use words and phrases from two (or more) languages in situations such as casual dialog, while traveling abroad, or simply to use a word they find more fitting (Myers-Scotton and Ury, 1977; Heredia and Altarriba, 2001). In CS, the base language is referred to as the matrix language while the contributing language is called the embedded language (Myers-Scotton, 1995), where speakers often use the matrix language the majority of the time.
+
+Code switched language is challenging to both automatic speech recognition (ASR) and machine translation (MT) - and therefore also to the composite task of speech translation (ST). While a rich
+
+![](images/4623a9e72e02742d32899754a7a0c4d97894814db5ed56c29a00c845999456ae.jpg)  
+Figure 1: An example instance of the joint speech recognition and translation task for code-switching (CS). Red indicates English words in the transcript and their corresponding words in the translation, whereas blue indicates Spanish words in the transcript and their corresponding translation.
+
+amount of prior works exist on CS in the context of ASR (Lyu et al., 2006; Ahmed and Tan, 2012; Vu et al., 2012; Johnson et al., 2017; Yue et al., 2019) and MT (Sinha and Thakur, 2005; Winata et al., 2021; Zhang et al., 2021; Yang et al., 2020), there is little prior work in the context of ST.
+
+The aforementioned challenges to ASR, MT and ST arise largely due to the lack of CS data as well as the often monolingual nature of ASR systems, and of encoders of MT and ST systems. The lack of CS data is often addressed via synthetic data, e.g. as seen in Xu and Yvon (2021); Nakayama et al. (2019). Instead, in this work we derive two novel natural CS datasets from existing public corpora. CS is also difficult for modeling due to its mixed multilingual nature. In order to support multiple languages on the utterance level, automatic language identification (LID) is often performed before applying monolingual systems on a per utterance basis. However, this does not address within-utterance CS, where embedded foreign words and phrases result in recognition errors for monolingual ASR systems, making multilingual models an attractive alternative. Furthermore, CS increases speech recognition errors, significantly increasing the problem of error propagation (Ruiz and Fed
+
+erico, 2014) in cascaded ST systems, where MT is then performed on the erroneous ASR output. Thus, multilingual end-to-end (E2E) ST systems may be especially appropriate to tackle CS speech.
+
+As both the transcript and translation are important in many CS ST use cases, we focus on the joint transcription and translation ST setting (Anastasopoulos and Chiang, 2018; Weller et al., 2021), extending it to CS data. We follow the methodology of these previous works and focus on the triangle E2E ST model to jointly generate both a transcript of the CS utterance and a translation of that utterance into text containing only one language (c.f. Figure 1 for an illustration). We perform a comparison along two axes: (1) comparing this E2E model to the standard cascaded ST systems, and (2) exploring the difference between bilingual systems and primarily monolingual systems gated by utterance-level LID. Following recent work that has shown the effectiveness of pre-trained models for ST (Li et al., 2020; Gallego et al., 2021), we use Wav2Vec 2.0 (Baevski et al., 2020) as our encoder model and the multilingual mBART 50-50 (Tang et al., 2020) as our decoder model.
+
+We also make several modeling contributions in order to use these pre-trained models for joint transcription and translation. For the E2E ST model, we extend Li et al. (2020) to adapt the mBART decoder to jointly produce both transcription and translation. Furthermore, we introduce a triangle E2E ST model with a shared bilingual decoder and show that this improves transcription and translation accuracy. Our model analysis shows a surprising amount of robustness to CS speech, with the amount (or proportion) of CS words in a sentence not affecting model accuracy. Overall, we observe strong accuracy scores (WER, BLEU) on the CS task, both without CS training data and in the low-resource setting. We believe this opens the door to new and exciting progress in this area.
+
+# 2 Related Work
+
+Code-switching in NLP has seen a rise of interest in recent years, including a dedicated workshop starting in 2014 (Diab et al., 2014) and still ongoing (Solorio et al., 2021). CS in machine translation also has a long history (Le Féal, 1990; Climent et al., 2003; Sinha and Thakur, 2005; Johnson et al., 2017; Elmadany et al., 2021; Xu and Yvon, 2021),
+
+but has seen a rise of interest with the advent of large multilingual models such as mBART (Liu et al., 2020) or mT5 (Xue et al., 2020; Gautam et al., 2021; Jawahar et al., 2021). Due to the lack of available CS data and the ease of single-word translation, most of these recent related MT works have synthetically created CS data for either training or testing by translating one or more of the words in a sentence (Song et al., 2019; Nakayama et al., 2019; Xu and Yvon, 2021; Yang et al., 2020). We differ from those works by using naturally occurring CS data (Section 3) which models the real-world CS distribution rather than arbitrary language mixing.
+
+For spoken input, as present in ASR and ST, synthetically creating realistic CS data is more challenging than it is for MT. However, dedicated ASR corpora that contain natural CS exist, including the Bangor Miami (Deuchar et al., 2014), SEAME (Zeng et al., 2018), and the recent large-scale ASRU 2019 task (Shi et al., 2020). These corpora generally do not contain translations of the ASR annotations, since they were designed for the ASR task only. However, there exist two exceptions, which we leverage to derive our ST CS data set, described in Section 3.
+
+There also exists a wide range of prior modeling work on CS in ASR models, for a variety of strategies (Lyu et al., 2006; Ahmed and Tan, 2012; Seki et al., 2018; Luo et al., 2018; Lu et al., 2020; Du et al., 2021; Zhang et al., 2021). However, the recently introduced large multilingual models for speech, such as Wav2Vec, Wav2Vec 2.0, Schneider et al. (2019); Baevski et al. (2020) and HuBERT (Hsu et al., 2021), are still underexplored with regards to their CS performance.
+
+Handling mixed languages also requires understanding what languages are being spoken. Systems that support mixed language input therefore require some form of automatic LID – either as an explicit component on the utterance (Mabokela et al., 2014; Xu and Yvon, 2021) or word-level (Lyu and Lyu, 2008a; Nakayama et al., 2019), or implicitly learned by the underlying model(s) via a multi-task learning setup (Lyu and Lyu, 2008b; Watanabe et al., 2017; Hou et al., 2020). In our work, we leverage both, exploring utterance-level LID components as well as implicit learning of utterance and word level LID.
+
+In both MT and ASR, prior publications have also included the study of intra-word mixing of languages (Yilmaz et al., 2018; Mager et al., 2019),
+
+a phenomenon we do not explore in our work.
+
+Finally, our work builds off of advances made by Gallego et al. (2021); Li et al. (2020) that show that combining large multilingual speech and text models provide consistent improvements. We differ however, by exploring ST in the novel CS setting.
+
+# 3 Task Description & Data Used
+
+# 3.1 Task Description
+
+We investigate systems suitable for bilingual English/Spanish conversational scenarios where some of the English and Spanish utterances may include some amount of words and phrases of the respective other language. That is, we are focusing on ST systems that can automatically and seamlessly handle utterances that are either purely English, purely Spanish, English with some Spanish words/phrases embedded or Spanish with some English words/phrases embedded. For transcription, we aim for models to generate the exact mixed-language transcript with each word written in its original spoken language. For translation, we aim to generate purely monolingual translations. See Figure 1 for an example. The experiments and results presented in this paper focus on translating into monolingual English only due to data availability, although we expect similar results for Spanish translations, due the bidirectional model training on standard ST data (Appendix D). We will leave it to future work to more closely examine translation into Spanish – or even a third language not present in the original utterance.
+
+It must be noted that word-level language categorization is sometimes ambiguous. A word in one language may also be considered part of a different language. That is for example true for loan words (Baugh, 1935), e.g.,  $e$ -mail in many non-English languages such as German. This issue can be further complicated by attempting to categorize what language named entities fall under: is a Spanish speaker saying Joe Biden or New York code-switching? Although we acknowledge the complexity of separating words between languages, our work, following previous work (Modipa et al., 2013; Nakayama et al., 2018), uses data annotated by crowd-sourced workers, counting any sentence annotated as having a least one foreign word as being CS. This approach also makes intuitive sense for speech, as the CS words (classified as foreign) will have phonemes that will align more with the embedded language, while the non-CS phonemes
+
+will align more with the matrix language.
+
+# 3.2 Code-Switched Speech Datasets
+
+We use the Fisher (Cieri et al., 2004) and Bangor Miami<sup>3</sup> (Deuchar et al., 2014) corpora for CS data, as they are the only publicly available corpora we are aware of that contains both annotated CS ASR transcripts, as well as translations of those transcripts (Table 1). Although these corpora contain the translations, to our knowledge they have not been used to study CS translation before.
+
+The Miami corpus was collected for linguistic code-switching analysis and gathered from recorded conversations between bilingual English/Spanish speakers in casual settings, primarily in Miami, Florida. These conversations include a high proportion of naturally occurring CS speech. However, in order to collect these naturally occurring conversations, the participants were recorded throughout their day using a small digital recorder worn on belts and lapels. Due to this, the Miami audio contains lower audio quality and much noiser background conditions than standard ASR datasets.
+
+The Fisher dataset was collected for ASR and was gathered by pairing sets of Spanish speakers, located in the U.S. and Canada, to each other through phone calls. Although the Fisher dataset is not a CS focused dataset, we found that it contains a large amount of (annotated) CS utterances, due to the speakers being situated in English-speaking contexts. The recording method (phone recordings in 2004) makes this a noisy ASR dataset, although significantly less so than Miami.
+
+To prepare the data for the joint ST CS task, we separate the data with CS utterances (utterances that contain at least one word annotated as CS) from those with none, creating a CS set and a monolingual set for each dataset. We note that for the Miami dataset the monolingual split contains both English-only and Spanish-only monolingual audio. As the Miami corpus was also annotated with both ambiguous and unambiguous code-switching, we only include utterances in the CS set if the annotations were tagged as unambiguously code-switched (i.e. excluding words such as ok, aha, and named entities). The Fisher CS dataset consists of majority (matrix<sup>4</sup>) Spanish  $77\%$  of the time, English-majority  $17\%$ , and  $6\%$  evenly
+
+<table><tr><td>Dataset</td><td>Raw Transcript</td><td>Clean Transcript</td></tr><tr><td>Fisher</td><td>un &lt;foreign lang=&quot;English&quot;&gt; show &lt;\foreign&gt;, a mi me gusta ver是多么這些 &lt;foreign lang=&quot;English&quot;&gt; shows &lt;\foreign&gt; de la medicina forense</td><td>un show, a mi me gusta ver是多么這些 show de la medicina forense</td></tr><tr><td>Miami</td><td>hay una [/] una que dice (.) it&#x27;s@s:eng five@s:eng o&#x27;clock@s:eng somewhere@s:eng</td><td>hay una una que dice it&#x27;s five o&#x27;clock somewhere</td></tr></table>
+
+![](images/886e928acbce02c24d728f1b30efc11d3f5cddc7d195d352e343843d66553a53.jpg)  
+Figure 2: Histogram of the proportions of code-switched words in a sentence for the CS test sets (Fisher on the left, Miami on the right). For example, 0.2 means that  $20\%$  of the words in the sentence are CS.
+
+![](images/5cb6fffa3bd145370f48e4178904b50e72d6cb58601716c832597033f4a428fe.jpg)
+
+Table 1: Examples of the raw and clean data for Miami and Fisher. Text in red indicates English text while blue text indicates Spanish. The Miami dataset uses the CHAT annotation format (MacWhinney and Snow, 1990).  
+
+<table><tr><td>Dataset</td><td>Split</td><td>Type</td><td>Hours</td><td>Instances</td></tr><tr><td rowspan="3">Miami</td><td>Train</td><td>Mono</td><td>3.60</td><td>6,489</td></tr><tr><td rowspan="2">Test</td><td>CS</td><td>2.82</td><td>3,296</td></tr><tr><td>Mono</td><td>3.61</td><td>6,490</td></tr><tr><td rowspan="5">Fisher</td><td rowspan="2">Train</td><td>CS</td><td>13.28</td><td>7,398</td></tr><tr><td>Mono</td><td>157.3</td><td>130,600</td></tr><tr><td>Dev</td><td>CS</td><td>1.45</td><td>821</td></tr><tr><td rowspan="2">Test</td><td>CS</td><td>1.63</td><td>986</td></tr><tr><td>Mono</td><td>12.15</td><td>10,595</td></tr></table>
+
+Table 2: Dataset Statistics. CS stands for Code-Switched and Mono for Monolingual.
+
+split between English/Spanish. For the Miami CS dataset the languages are more evenly distributed, with  $51\%$  majority-Spanish,  $35\%$  majority-English, and  $9\%$  evenly split. $^{5}$
+
+The Fisher data consists of three evaluation sets (Dev/Dev2/Test) that together contain approximately a thousand instances of CS with corresponding translations in monolingual English. We com
+
+bine them into a Fisher CS Test set. The Fisher dataset also contains a large amount of CS utterances in the training set (appx. 8k or 15 hrs) which we use as fine-tuning  $(90\%)$  and validation data  $(10\%)$ . As the Miami dataset contains no splits, we use all CS data for the test set and split the monolingual data into even train/test sets. We include basic summary statistics in Table 2. Note that when compared to standard ST datasets, these CS ST datasets would be considered low-resource settings.
+
+In Figure 2, we see the proportion of CS words in a sentence for the CS test sets. We note that there are no sentences with more than  $50\%$  of the words CS since the minority language cannot be more than  $50\%$  by definition. For instances that are exactly  $50\%$  code switched their language identification was chosen by randomly selecting either English or Spanish. We see that for the Fisher dataset there are more sentences with less than  $15\%$  CS with a small uptick around  $50\%$ . For Miami it is more uniform, with a large amount of sentences being approximately  $25\%$  CS.
+
+To prepare our models for Spanish-English CS, we use the CoVoST (Wang et al., 2020a,b) and MuST-C (Cattoni et al., 2019) datasets for standard
+
+![](images/bba6a5f125044be8349dae8558174cc15b123f2c11cba23f20e872e2ee1434cb.jpg)  
+Figure 3: Illustration of model architectures, with cascaded architectures on the top and E2E architectures on the bottom. Left to right shows the progression of models with the least and the most amount of shared parameters respectively. Subscripts are present to indicate shared modules within each model. Dotted lines indicate a decision where only one path is chosen using the LID. Note that there is no cascade equivalent to the BIDIRECTIONAL E2E SHARED model, as the cascaded model by definition generates transcript then translation separately. The numbers in parentheses stand for the number of model parameters in billions.
+
+ST training, as CoVoST contains only  $\mathrm{Es} \rightarrow \mathrm{En}$  and MuST-C contains only  $\mathrm{En} \rightarrow \mathrm{Es}$ . Although high scores on these datasets are not our primary target, we note that our scores come close to or improve the state of the art (SoTA) on these tasks (see Appendix A, Table 9) albeit with different data used in training, showing that our base ST models are representative of current SoTA techniques.
+
+# 4 Experimental Settings
+
+# 4.1 Models
+
+Joint Transcript/Translation Models Many different types of E2E models exist for joint transcript/translation ST (Sperber and Paulik, 2020). Here, we focus on the triangle E2E architecture due to its strong performance in previous work (Anastasopoulos and Chiang, 2018; Sperber et al., 2020). Following recent work (Gallego et al., 2021; Li et al., 2020) we use pre-trained modules as a starting place for our ST model, using a Wav2Vec 2.0 (Baevski et al., 2020) encoder and a mBART 50-50 (Liu et al., 2020; Tang et al., 2020) decoder.
+
+Because our task involves joint ASR and ST, we need to adapt the pre-trained decoder to work with the E2E triangle architecture. Specifically, the triangle model's second decoder computes cross attention separately over both the first decoder and the encoder states. We place an additional cross-attention layer after each encoder-attention layer in mBARTs decoder blocks, initializing them with the pre-trained encoder-attention weights. To make sure these weights converge properly, we freeze the entire model for approximately the first epoch while training only the bridge and additional cross attention layers (c.f. Appendix A).
+
+As described in Section 3, our task involves modeling intra-sentence CS. This means that any model used for this task must either explicitly or implicitly learn to model the language of each word in the sentence. Furthermore, as more than one language is being modeled, each sub-component of the model can either be unidirectional or bidirectional. We can thus categorize potential models by how much information is shared within the parameters: the
+
+least shared models would be unidirectional and joined together by explicit LID, whereas the most shared would be bidirectional models that learn the LID implicitly. Models and their categorization along this scale are shown in Figure 3.
+
+For cascade models, the most basic would be separate unidirectional cascaded models joined by an LID model. The LID model will explicitly decide what the matrix language is and send the utterance to the model that is best equipped to handle that language (Figure 3A). Note that this approach may suffer from error propagation issues due to incorrect LID. A more parameter-shared version of this model is to make the cascaded model encoder shared between both unidirectional models (Figure 3B). Finally, we can examine a bidirectional cascade model that shares each component across both languages. This architecture implicitly learns to model the language of the input, removing the need for an explicit LID model (Figure 3C).
+
+We also examine similar analogues for the E2E triangle model: unidirectional models joined by LID (Figure 3D) and a bidirectional model with LID and a shared encoder (Figure 3E). We can also use the standard triangle model (see Anastasopoulos and Chiang (2018) for implementation details) that includes one encoder and two decoders (one for each sub-task) (Figure 3F). Furthermore, we propose to alter the standard triangle model and share both decoder parameters for both languages with a joint bidirectional decoder (Figure 3G, note that the cascade model cannot do this due to the definition of the cascade). By doing so, we hope to provide an inductive bias for the model to more easily handle code-switched data, as the weights of that decoder will already be used to handling multiple languages for both tasks (compared to the bidirectional cascade model, which only shares multilingual parameters for each task of transcript and translation).
+
+Language Identification Model We train the language identification (LID) model to identify the matrix language. For consistency with our other models (and similar to concurrent work, e.g. Tjandra et al. (2021)), we use a pre-trained Wav2Vec2 along with a classifier layer to predict whether the utterance is majority Spanish or majority English. We train the model in the same fashion as the joint transcription and translation models (Section 4.1 and Appendix A) but train on the LID data instead.
+
+The data for the LID model was gathered by tak
+
+ing the CS data from the training set of the Fisher corpus and combining it with randomly sampled data from several different datasets in order to help the model learn despite the domain of the audio. We use MuST-C English audio, CoVoST English audio, CoVoST Spanish audio, and the monolingual Spanish audio from the training sets of Fisher and Miami. We found that upsampling the CS training set by 2 and using the same amount of data (2x the number of the CS set) for CoVoST and MuST-C provided the best results:  $98\%+$  accuracy on CoVoST and MuST-C,  $89\%$  on the Fisher CS validation and test sets, and  $72\%$  on the Miami CS test set (due to the noisy data). As a large proportion of the CS data is close to  $50\%$  code-switched (see Figure 2), it becomes more difficult for the model to predict the matrix language correctly.
+
+# 4.2 Training Process and Evaluation
+
+For all dataset evaluations, we use word error rate (WER) and character error rate (CER) for the transcript and Charcut (CCT) (Lardilleux and Lepage, 2017) and sacreBLEU (Post, 2018) for the translation. However, we found that there was no difference in conclusions between each of the two metrics (WER vs CER and BLEU vs Charcut) and thus we only report BLEU/WER in the main text (see Appendix A for implementation details). For tables showing all metrics, see Appendix E.
+
+We evaluate our models on the Fisher and Miami test sets (with both CS-only and monolingual-only test sets) in two different settings: (1) without finetuning them on CS data (No-FT) and (2) after finetuning the already trained ST models on the Fisher CS Training set (FT). For models consisting of two monolingual sub-models we fine-tune both on the CS data. During fine-tuning we employ the same hyperparameters as in the original experiment, but perform early stopping on the Fisher CS Dev set. We use significance tests to verify the reliability of our results (Koehn, 2004). We run bootstrap resampling tests against the best performing model, using  $\alpha = 0.05$ . More training parameters such as learning rates, etc. can be found in Appendix A.
+
+# 5 Results
+
+# 5.1 Scores on Test Sets
+
+In this section, we explore the results of doing ST for CS data along the two axes of unidirectional vs
+
+<table><tr><td rowspan="3">Models</td><td colspan="4">Not Fine-Tuned</td><td colspan="4">Fine-Tuned</td></tr><tr><td colspan="2">CS</td><td colspan="2">Mono.</td><td colspan="2">CS</td><td colspan="2">Mono.</td></tr><tr><td>↓WER</td><td>↑BLEU</td><td>↓WER</td><td>↑BLEU</td><td>↓WER</td><td>↑BLEU</td><td>↓WER</td><td>↑BLEU</td></tr><tr><td rowspan="2">CASCADE UNIDIRECT</td><td>37.1</td><td>22.5</td><td>26.6</td><td>24.7</td><td>33.5</td><td>24.6</td><td>24.8</td><td>25.5</td></tr><tr><td>(-0.8)</td><td>(-0.4)</td><td>(-3.1)</td><td>(+0.9)</td><td>(-0.4)</td><td>(0.0)</td><td>(-1.0)</td><td>(+0.2)</td></tr><tr><td rowspan="2">CASCADE UNI SHARED ENC</td><td>36.0</td><td>21.6</td><td>25.6</td><td>24.3</td><td>31.2</td><td>25.4</td><td>25.6</td><td>24.8</td></tr><tr><td>(0.0)</td><td>(+0.6)</td><td>(0.0)</td><td>(+0.5)</td><td>(+0.1)</td><td>(+0.2)</td><td>(-0.3)</td><td>(+0.1)</td></tr><tr><td rowspan="2">E2E UNIDIRECT</td><td>36.6</td><td>22.3</td><td>26.7</td><td>25.0</td><td>33.4</td><td>24.4</td><td>25.3</td><td>25.5</td></tr><tr><td>(-0.9)</td><td>(-0.1)</td><td>(-3.5)</td><td>(+1.0)</td><td>(-0.2)</td><td>(+0.1)</td><td>(-1.4)</td><td>(+0.4)</td></tr><tr><td rowspan="2">E2E BIDIRECT BY LANG</td><td>37.0</td><td>23.4</td><td>27.2</td><td>25.0</td><td>36.7</td><td>22.8</td><td>27.3</td><td>25.0</td></tr><tr><td>(-0.9)</td><td>(-0.1)</td><td>(-1.9)</td><td>(+0.5)</td><td>(-0.8)</td><td>(+0.2)</td><td>(-2.0)</td><td>(+0.4)</td></tr></table>
+
+Table 3: Comparison of Oracle vs Predicted LID results on the Fisher dataset. Numbers in parenthesis are the difference to the corresponding model with oracle LID. Note that the Oracle LID improves upon the Predicted LID in most cases. Conclusions are similar for the Miami corpus (see Appendix B Table 7)  
+
+<table><tr><td rowspan="3"></td><td rowspan="3">Model</td><td colspan="4">Not Fine-Tuned</td><td colspan="4">Fine-Tuned</td></tr><tr><td colspan="2">CS</td><td colspan="2">Mono.</td><td colspan="2">CS</td><td colspan="2">Mono.</td></tr><tr><td>↓WER</td><td>↑BLEU</td><td>↓WER</td><td>↑BLEU</td><td>↓WER</td><td>↑BLEU</td><td>↓WER</td><td>↑BLEU</td></tr><tr><td rowspan="7">Fisher</td><td>CASCADE UNIDIRECT</td><td>37.1</td><td>22.5</td><td>26.6</td><td>24.7</td><td>33.5</td><td>24.6</td><td>24.8</td><td>25.5</td></tr><tr><td>CASCADE UNI SHARED ENC</td><td>36.0</td><td>21.6</td><td>25.6</td><td>24.3</td><td>31.2</td><td>*25.4</td><td>25.6</td><td>24.8</td></tr><tr><td>CASCADE BIDIRECT</td><td>37.2</td><td>21.8</td><td>26.5</td><td>24.1</td><td>33.2</td><td>23.2</td><td>28.1</td><td>23.2</td></tr><tr><td>E2E UNIDIRECT</td><td>36.6</td><td>22.3</td><td>26.7</td><td>25.0</td><td>33.4</td><td>24.4</td><td>25.3</td><td>25.5</td></tr><tr><td>E2E BIDIRECT BY LANG</td><td>37.0</td><td>23.4</td><td>27.2</td><td>25.0</td><td>36.7</td><td>22.8</td><td>27.3</td><td>25.0</td></tr><tr><td>E2E BIDIRECT BY TASK</td><td>*34.1</td><td>*23.0</td><td>23.6</td><td>26.0</td><td>*30.1</td><td>25.6</td><td>*24.3</td><td>25.6</td></tr><tr><td>E2E BIDIRECT SHARED</td><td>33.8</td><td>*23.3</td><td>23.2</td><td>26.2</td><td>30.0</td><td>*25.4</td><td>24.1</td><td>26.1</td></tr><tr><td rowspan="7">Miami</td><td>CASCADE UNIDIRECT</td><td>65.2</td><td>8.8</td><td>52.3</td><td>16.8</td><td>64.8</td><td>10.8</td><td>51.5</td><td>16.8</td></tr><tr><td>CASCADE UNI SHARED ENC</td><td>60.2</td><td>9.7</td><td>53.8</td><td>15.7</td><td>55.0</td><td>14.7</td><td>55.6</td><td>15.3</td></tr><tr><td>CASCADE BIDIRECT</td><td>61.4</td><td>9.3</td><td>54.0</td><td>14.8</td><td>57.4</td><td>10.6</td><td>58.2</td><td>14.0</td></tr><tr><td>E2E UNIDIRECT</td><td>65.6</td><td>10.1</td><td>53.0</td><td>17.2</td><td>65.1</td><td>11.7</td><td>*51.4</td><td>17.6</td></tr><tr><td>E2E BIDIRECT BY LANG</td><td>69.5</td><td>12.4</td><td>55.2</td><td>16.5</td><td>69.3</td><td>11.5</td><td>54.5</td><td>16.6</td></tr><tr><td>E2E BIDIRECT BY TASK</td><td>59.9</td><td>11.0</td><td>*50.0</td><td>*18.1</td><td>*53.6</td><td>*13.8</td><td>52.6</td><td>*17.5</td></tr><tr><td>E2E BIDIRECT SHARED</td><td>58.9</td><td>*11.8</td><td>49.9</td><td>18.3</td><td>53.0</td><td>*14.1</td><td>52.1</td><td>*17.4</td></tr></table>
+
+Table 4: Test set scores, with results from the Fisher corpus on the top half and the Miami corpus on the bottom half. Bold scores indicate the best score in the column, while asterisks indicate results that are statistically similar to the best score in the column group using a bootstrap resampling test with  $\alpha = 0.05$ .
+
+bidirectional and end-to-end vs cascade.
+
+We see results for models using explicit LID prediction in Table 3, showing that models that use the predicted LID perform worse than those that use Oracle LID (e.g. 36.6 vs 35.7 WER for the E2E UNIDIRECT). This provides a slight advantage for the bidirectional models that learn LID implicitly. However, the predicted LID case is the realistic setting, and thus we use it for the remainder of our experiments.
+
+When we examine the models along the scale of unidirectional to bidirectional, we see that higher amounts of shared parameters are correlated with higher scores, e.g. bidirectional is better. We see that on all datasets and evaluation settings (Table 4) that the E2E BIDIRECT SHARED model is either statistically similar or outperforms all other models, except for the Miami Monolingual FT case, where it comes in 3rd. Thus, the inductive bias of
+
+![](images/fcb723395ccdba822f96a17c962081c6a2428ea64e8844647ee52a7cf12a4b75.jpg)  
+Figure 4: Accuracy of the models in generating the CS spans. Note that this excludes all non-exact matches and is a lower bound on performance.
+
+sharing the multilingual task parameters provides a gain of approximately 3.5 WER points (33.8 vs 37.3) and 1.5 BLEU points (23.3 vs 21.9) for the E2E BIDIRECT SHARED model over the E2E UNIDIRECT model on the Fisher dataset, with similar
+
+<table><tr><td>Model</td><td>Transcript</td><td>Translation</td></tr><tr><td>Reference</td><td>si entonces volví ahora a la casa si el fall break</td><td>yes so I returned here to the house yes the fall break</td></tr><tr><td>Cascade</td><td>si entonces volví ahora a la casa si esolvereak</td><td>yes then I returned here at home yes itsolvereak</td></tr><tr><td>E2E</td><td>si entonces volví ahora a la casa si es fallbreak</td><td>yes so I came back to the house yes its fallbreak</td></tr></table>
+
+Table 5: Example generated output from the CASCADE BIDIRECT and E2E BIDIRECT SHARED models. Note the error propagation in the cascade model.
+
+performance on the Miami dataset.
+
+We can also examine Table 4 to see how the cascade models compare to the E2E models. The results show that the cascaded models perform the same or worse than the E2E models they compare to w.r.t. parameter sharing, with the best overall model being the E2E BIDIRECT SHARED, beating the CASCADE BIDIRECT (e.g. 33.8 vs 37.2 WER or 23.3 vs 21.8 BLEU on Fisher No-FT).
+
+Table 4 also illustrates that fine-tuning models on CS data improves scores on CS test sets (33.8 vs 30.0 WER for the E2E BIDIRECT SHARED on Fisher, 58.9 vs 53.0 for Miami). These gains are consistent for the Fisher dataset, which is the domain of the CS training set, however there are still gains for the out-of-domain Miami CS data. These results suggest that additional pre-training on natural or synthetic data (in both audio/text modalities) would likely be fruitful future work. When we examine how fine-tuning on CS data changes the model's monolingual scores, we find that they generally improve the monolingual results for the unidirectional models, but tend to make bidirectional models slightly worse, perhaps due to interference between the languages and tasks in the same weights. However, overall we find that fine-tuning provides large gains for CS with only minor decreases in monolingual performance.
+
+# 5.2 Model Analysis
+
+We also provide further analysis of the CS output of the best model and its cascaded counterpart (BIDIRECT CASCADE and E2E BIDIRECT SHARED). We perform three analyses: (1) comparing utterance level scores vs the proportion of CS words in the utterance, (2) computing the exact match accuracy of the CS spans in the model's output, and (3) qualitatively examining model output.
+
+We check the correlation between the proportion of CS words in a sentence and the model's score, using a linear model to find the  $R^2$  values. We found that surprisingly, there was no correlation between the proportion of CS words and the models score for any of the different models
+
+or metrics  $(R^2 < 0.025$  for all models and metrics). A graphical depiction of the model's scores over CS proportions is in the Appendix, Figure 5. We note that this finding was the same for comparing the number of CS words instead of the proportion. This finding implies that the models are surprisingly robust to the amount of CS in a sentence.
+
+Although BLEU andWER scores show how well the models do on the CS data, we can further isolate the performance of these models on only the code-switched parts of the utterances. To do so, we isolate all CS spans in the sentences and check to see if the model's output contains the exact-match of those spans. We note that this metric does not take into account synonyms or different tenses of the same word, making it a stricter metric serving as a lower bound of absolute performance. We see in Figure 4 that the E2E model still outperforms the cascade on CS spans, with Fisher No-FT scores around  $20 - 30\%$  and Fisher FT scores around  $45\%$ .
+
+Finally, we can also examine the model's outputs. We inspected 200 output sentences for the monolingual subsets and found that both models generated the correct language in every case, indicating that they correctly learned the implicit LID. However, we can see that the cascade model does struggle with error propagation (especially so in the CS setting, Table 5), likely causing part of the difference between the E2E and cascade models.
+
+Although the CS WER and BLEU scores are not as high as they are on cleaner monolingual datasets such as CoVoST (Appendix A), their performance is competitive with their respective monolingual performance on Miami and Fisher, even in the NoFT setting. We believe that with additional data and improvements ST models will be well-equipped to handle CS in practical situations and that overall, models show strong CS performance.
+
+# 6 Conclusion
+
+In this work, we expand the ST literature to explore code-switching, contributing a new task framework for ST that extends the joint transcription and translation setup. To further progress, we built and open
+
+sourced a new ST corpus for CS from existing public datasets. We evaluated a range of models, showing that using bilingual joint decoders provides gains over using separate task decoders. We also showed that E2E systems provide better performance than their cascading counterparts on the CS task. Overall, our work shows that ST models can perform well on CS applications with both no fine-tuning and in low-resource settings, opening the door to new and exciting areas of future work.
+
+# References
+
+Basem HA Ahmed and Tien-Ping Tan. 2012. Automatic speech recognition of code switching speech using 1-best rescoring. In 2012 International Conference on Asian Language Processing, pages 137-140. IEEE.  
+Antonios Anastasopoulos and David Chiang. 2018. Tied multitask learning for neural speech translation. arXiv preprint arXiv:1802.06655.  
+Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. arXiv preprint arXiv:2006.11477.  
+Albert C Baugh. 1935. The chronology of french loan-words in english. Modern Language Notes, 50(2):90-93.  
+Roldano Cattoni, Mattia Antonino Di Gangi, Luisa Bentivogli, Matteo Negri, and Marco Turchi. 2019. Must-c: A multilingual corpus for end-to-end speech translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2012-2017, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Christopher Cieri, David Miller, and Kevin Walker. 2004. The fisher corpus: A resource for the next generations of speech-to-text. In LREC, volume 4, pages 69-71.  
+Salvador Climent, Joaquim Moré, Antoni Oliver, Miriam Salvatierra, Imma Sánchez, Mariona Taule, and Lluisa Vallmanya. 2003. Bilingual newsgroups in catalonia: A challenge for machine translation. Journal of Computer-Mediated Communication, 9(1):JCMC919.  
+Margaret Deuchar, Peredur Davies, Jon Russell Herring, M Carmen Parafita Couto, and Diana Carter. 2014. 5. building bilingual corpora. In Advances in the Study of Bilingualism, pages 93-110. Multilingual Matters.
+
+Mona Diab, Julia Hirschberg, Pascale Fung, and Thamar Solorio. 2014. Proceedings of the first workshop on computational approaches to code switching. In Proceedings of the First Workshop on Computational Approaches to Code Switching.  
+Chenpeng Du, Hao Li, Yizhou Lu, Lan Wang, and Yanmin Qian. 2021. Data augmentation for end-to-end code-switching speech recognition. In 2021 IEEE Spoken Language Technology Workshop (SLT), pages 194-200. IEEE.  
+AbdelRahim Elmadany, Muhammad Abdul-Mageed, et al. 2021. Investigating code-mixed modern standard arabic-egyptian to english machine translation. In Proceedings of the Fifth Workshop on Computational Approaches to Linguistic Code-Switching, pages 56-64.  
+Gerard I Gallego, Ioannis Tsiamas, Carlos Escolano, José AR Fonollosa, and Marta R Costa-jussà. 2021. Upc's speech translation system for iwslt 2021. arXiv preprint arXiv:2105.04512.  
+Devansh Gautam, Prashant Kodali, Kshitij Gupta, Anmol Goel, Manish Shrivastava, and Ponnurangam Kumaraguru. 2021. Comet: Towards code-mixed translation using parallel monolingual sentences. In Proceedings of the Fifth Workshop on Computational Approaches to Linguistic Code-Switching, pages 47-55.  
+Roberto R Heredia and Jeanette Altarriba. 2001. Bilingual language mixing: Why do bilinguals code-switch? Current Directions in Psychological Science, 10(5):164-168.  
+Wenxin Hou, Yue Dong, Bairong Zhuang, Longfei Yang, Jiatong Shi, and Takahiro Shinozaki. 2020. Large-scale end-to-end multilingual speech recognition and language identification with multi-task learning. In INTERSPEECH, pages 1037-1041.  
+Wei-Ning Hsu, Yao-Hung Hubert Tsai, Benjamin Bolte, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. Hubert: How much can a bad teacher benefit asr pre-training? In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6533-6537. IEEE.  
+Ganesh Jawahar, El Moatez Billah Nagoudi, Muhammad Abdul-Mageed, and Laks VS Lakshmanan. 2021. Exploring text-to-text transformers for english to henglish machine translation with synthetic code-mixing. arXiv preprint arXiv:2105.08807.  
+Melvin Johnson, Mike Schuster, Quoc V. Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, Fernanda Viégas, Martin Wattenberg, Greg Corrado, Macduff Hughes, and Jeffrey Dean. 2017. Google's Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation. Transactions of the Association for Computational Linguistics, 5:339-351.
+
+Philipp Koehn. 2004. Statistical Significance Tests for Machine Translation Evaluation. In *Conference on Empirical Methods in Natural Language Processing (EMNLP)*, pages 388-395.  
+Adrien Lardilleux and Yves Lepage. 2017. Charcut: Human-targeted character-based mt evaluation with loose differences. In Proceedings of IWSLT 2017.  
+Karla Dejean Le Féal. 1990. A different approach to machine translation. *Meta: journal des traducteurs*, 35(4):710-719.  
+Xian Li, Changhan Wang, Yun Tang, Chau Tran, Yuqing Tang, Juan Pino, Alexei Baevski, Alexis Conneau, and Michael Auli. 2020. Multilingual speech translation with efficient finetuning of pretrained models. arXiv preprint arXiv:2010.12829.  
+Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742.  
+Yizhou Lu, Mingkun Huang, Hao Li, Jiaqi Guo, and Yanmin Qian. 2020. Bi-encoder transformer network for mandarin-english code-switching speech recognition using mixture of experts. In INTER-SPEECH, pages 4766-4770.  
+Ne Luo, Dongwei Jiang, Shuaijiang Zhao, Caixia Gong, Wei Zou, and Xiangang Li. 2018. Towards end-to-end code-switching speech recognition. arXiv preprint arXiv:1810.13091.  
+D. Lyu and Ren-Yuan Lyu. 2008a. Language identification on code-switching utterances using multiple cues. In INTERSPEECH.  
+Dau-Cheng Lyu and Ren-Yuan Lyu. 2008b. Language identification on code-switching utterances using multiple cues. In Ninth Annual Conference of the International Speech Communication Association.  
+Dau-Cheng Lyu, Ren-Yuan Lyu, Yang-chin Chiang, and Chun-Nan Hsu. 2006. Speech recognition on code-switching among the chinese dialects. In 2006 IEEE International Conference on Acoustics, Speech and Signal Processing Proceedings, volume 1, pages I-I. IEEE.  
+Koena Ronny Mabokela, Madimetja Jonas Manamela, and Mabu Manaileng. 2014. Modeling codeswitching speech on under-resourced languages for language identification. In Spoken Language Technologies for Under-Resourced Languages.  
+Brian MacWhinney and Catherine Snow. 1990. The child language data exchange system: An update. Journal of child language, 17(2):457-472.  
+Manuel Mager, Özlem Çetinoğlu, and Katharina Kann. 2019. Subword-level language identification for intra-word code-switching. In arXiv preprint arXiv:1904.01989.
+
+Thipe Modipa, Marelie Hattingh Davel, and F. D. Wet. 2013. Implications of sepedi/english code switching for asr systems. In researchgate.  
+Carol Myers-Scotton. 1995. Social motivations for codeswitching: Evidence from Africa. Oxford University Press.  
+Carol Myers-Scotton and William Ury. 1977. Bilingual strategies: The social functions of code-switching. International journal of the sociology of language, pages 5-20.  
+Sahoko Nakayama, Takatomo Kano, Andros Tjandra, S. Sakti, and Satoshi Nakamura. 2019. Recognition and translation of code-switching speech utterances. 2019 22nd Conference of the Oriental COCOSDA International Committee for the Co-ordination and Standardisation of Speech Databases and Assessment Techniques (O-COCOSDA), pages 1-6.  
+Sahoko Nakayama, Andros Tjandra, S. Sakti, and Satoshi Nakamura. 2018. Speech chain for semi-supervised learning of japanese-english code-switching asr and tts. 2018 IEEE Spoken Language Technology Workshop (SLT), pages 182-189.  
+Matt Post. 2018. A call for clarity in reporting bleu scores. arXiv preprint arXiv:1804.08771.  
+Nicholas Ruiz and Marcello Federico. 2014. Assessing the impact of speech recognition errors on machine translation quality. In 11th Conference of the Association for Machine Translation in the Americas (AMTA), Vancouver, BC, Canada.  
+Steffen Schneider, Alexei Baevski, Ronan Collobert, and Michael Auli. 2019. wav2vec: Unsupervised pre-training for speech recognition. arXiv preprint arXiv:1904.05862.  
+Hiroshi Seki, Shinji Watanabe, Takaaki Hori, Jonathan Le Roux, and J. Hershey. 2018. An end-to-end language-tracking speech recognizer for mixed-language speech. 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4919-4923.  
+Xian Shi, Qiangze Feng, and Lei Xie. 2020. The asru 2019 mandarin-english code-switching speech recognition challenge: Open datasets, tracks, methods and results. arXiv preprint arXiv:2007.05916.  
+R Mahesh K Sinha and Anil Thakur. 2005. Machine translation of bi-lingual Hindi-english (hinglish) text. 10th Machine Translation summit (MT Summit X), Phuket, Thailand, pages 149-156.  
+Thamar Solorio, Shuguang Chen, Alan W. Black, Mona Diab, Sunayana Sitaram, Victor Soto, Emre Yilmaz, and Anirudh Srinivasan, editors. 2021. Proceedings of the Fifth Workshop on Computational Approaches to Linguistic Code-Switching. Association for Computational Linguistics, Online.
+
+Kai Song, Yue Zhang, Heng Yu, Weihua Luo, Kun Wang, and M. Zhang. 2019. Code-switching for enhancing nmt with pre-specified translation. In NAACL.  
+Matthias Sperber and Matthias Paulik. 2020. Speech Translation and the End-to-End Promise: Taking Stock of Where We Are. In Association for Computational Linguistic (ACL), Seattle, USA.  
+Matthias Sperber, Hendra Setiawan, Christian Gollan, Udhyakumar Nallasamy, and Matthias Paulik. 2020. Consistent Transcription and Translation of Speech. Transactions of the Association for Computational Linguistics (TACL).  
+Yuqing Tang, Chau Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, and Angela Fan. 2020. Multilingual translation with extensible multilingual pretraining and finetuning. arXiv preprint arXiv:2008.00401.  
+Andros Tjandra, Diptanu Gon Choudhury, Frank Zhang, Kritika Singh, Alexei Baevski, Assaf Sela, Yatharth Saraf, and Michael Auli. 2021. Improved language identification through crosslingual self-supervised learning. arXiv preprint arXiv:2107.04082.  
+Ngoc Thang Vu, Dau-Cheng Lyu, Jochen Weiner, Dominic Telaar, Tim Schilpe, Fabian Blaicher, Eng-Siong Chng, Tanja Schultz, and Haizhou Li. 2012. A first speech recognition system for mandarin- english code-switch conversational speech. In 2012 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4889-4892. IEEE.  
+Changhan Wang, Juan Pino, Anne Wu, and Jiatao Gu. 2020a. Covost: A diverse multilingual speech-to-text translation corpus. arXiv preprint arXiv:2002.01320.  
+Changhan Wang, Anne Wu, and Juan Pino. 2020b. Covost 2 and massively multilingual speech-to-text translation. arXiv preprint arXiv:2007.10310.  
+Shinji Watanabe, Takaaki Hori, and John R Hershey. 2017. Language independent end-to-end architecture for joint language identification and speech recognition. In 2017 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 265-271. IEEE.  
+Orion Weller, Matthias Sperber, Christian Gollan, and Joris Kluivers. 2021. Streaming models for joint speech recognition and translation. arXiv preprint arXiv:2101.09149.  
+Genta Indra Winata, Samuel Cahyawijaya, Zihan Liu, Zhaojiang Lin, Andrea Madotto, and Pascale Fung. 2021. Are multilingual models effective in code-switching? ArXiv, abs/2103.13309.  
+Jitao Xu and François Yvon. 2021. Can you traducir this? machine translation for code-switched input. arXiv preprint arXiv:2105.04846.
+
+Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2020. mt5: A massively multilingual pre-trained text-to-text transformer. arXiv preprint arXiv:2010.11934.  
+Zhen Yang, Bojie Hu, Ambyera Han, Shen Huang, and Qi Ju. 2020. Csp: Code-switching pre-training for neural machine translation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2624-2636.  
+Emre Yilmaz, Astik Biswas, Ewald van der Westhuizen, Febe de Wet, and Thomas Niesler. 2018. Building a unified code-switching asr system for south african languages. arXiv preprint arXiv:1807.10949.  
+Xianghu Yue, Grandee Lee, Emre Yilmaz, Fang Deng, and Haizhou Li. 2019. End-to-end code-switching asr for low-resourced language pairs. In 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 972-979. IEEE.  
+Zhiping Zeng, Yerbolat Khassanov, Van Tung Pham, Haihua Xu, Eng Siong Chng, and Haizhou Li. 2018. On the end-to-end solution to mandarin- english code-switching speech recognition. arXiv preprint arXiv:1811.00241.  
+Shuai Zhang, Jiangyan Yi, Zhengkun Tian, Jianhua Tao, and Ye Bai. 2021. Rnn-transducer with language bias for end-to-end mandarin-english code-switching speech recognition. In 2021 12th International Symposium on Chinese Spoken Language Processing (ISCSLP), pages 1-5. IEEE.
+
+<table><tr><td rowspan="2">Models</td><td colspan="2">CS</td><td colspan="2">Mono.</td></tr><tr><td>↓WER</td><td>↑BLEU</td><td>↓WER</td><td>↑BLEU</td></tr><tr><td>Random Init</td><td>69.6</td><td>11.0</td><td>59.6</td><td>13.2</td></tr><tr><td>Pre-trained</td><td>33.8</td><td>23.3</td><td>23.2</td><td>26.2</td></tr></table>
+
+Table 6: Comparison of the E2E bidirectional shared model with pre-training vs random initialization on the Fisher code-switched test sets.
+
+# A Training and Evaluation Details
+
+We follow Gallego et al. (2021); Li et al. (2020) and use a triangular learning rate, adapting the step count to depend on the batch size (as not all models could fit the same batch size) with (64 / batch size) * 500 warm up steps, (64 / batch size) * 500 hold steps, (64 / batch size) * 3000 decay steps, a beta of 0.9, and a beta2 of 0.98. The learning rate was selected from running a search over {0.01, 0.005, 0.001, 0.0005, 0.0001, 0.0005}. We found that 0.0005 was best for all models, so we examined learning rates again between 0.0001 to 0.001 (by 0.0001) and found that they all performed similarly, thus we use 0.0005 in our experiments. For efficiency in batch size while training, we removed all instances whose audio length was longer than 20 seconds. We freeze the attention layers for the first  $500 \times (64 / \text{batch size})$  steps, which is approximately the first epoch of training.
+
+We initially trained the models on only CoVoST and MuST-C and found there was a large domain shift between these datasets and the comparatively noisier Fisher and Miami datasets. As domain shift was not the focus of this paper, we further trained the models on the Fisher and Miami monolingual training sets to reduce the effect of domain shift.
+
+As a sanity check of the effectiveness of our training, we also include scores in Table 9 for the test sets of CoVoST and MuST-C. We note that our scores are close to the SoTA scores of Li et al. (2020) on CoVoST (and they use the large Wav2Vec2 model while we use the base version) and our MuST-C scores are higher than that of Gallego et al. (2021).
+
+We evaluate using word error rate, character error rate, charcut, and BLEU. As the models learn different punctuation techniques from a variety of sources, including MuST-C, CoVoST, Miami, and Fisher, we remove all punctuation from the output before evaluating on the CS/Mono test sets, in order to only measure scores on the content. For BLEU, we use SacreBLEU with parameters
+
+![](images/6f1cfa3fe88bdb338b978e28eea7b0dc654bc5dd2de3a11ecab55f15ef343445.jpg)  
+Figure 5: Charcut performance of the E2E BIDIRECT SHARED model on sentences with various levels of CS proportions. Note that there is no clear correlation, as described in Section 5.2. Black lines indicate error bars of 2 standard deviations while the bar represents the average.
+
+case.lc+numrefs.1+smooth.4.0+tok.13a.
+
+# B More LID Comparisons
+
+We show results for all models that use LID on both datasets in Table 7. Note the conclusions remain the same as Table 3.
+
+# C Random Initialization Results
+
+We also perform an ablation of these pre-trained scores (Table 6) for the E2E BIDIRECT SHARED model, as it is the best performing model overall. We tried many different setups for training it from scratch rather than loading the pre-trained weights. We found that it was very difficult for this model to converge, and when it did, the results were sub-par.
+
+# D Training Results
+
+We include the scores of evaluating our models on the test sets of the ST training data (MuST-C and CoVoST) in Table 9. We also include the results of fine-tuning performance on the CS dev set in Table 8, which roughly mirrors the main results.
+
+# E Expanded Results
+
+For brevity, we do not include the CER and Charcut metrics in the main text. In this section we included tables with all metrics for all results (Table 10 for Miami and Table 11 for Fisher). We note however, that the WER and BLEU scores align with the CER and Charcut scores, and thus our conclusions remain the same.
+
+<table><tr><td rowspan="3" colspan="2">Model</td><td colspan="4">No Fine-Tuning</td><td colspan="4">Fine-Tuned</td></tr><tr><td colspan="2">CS</td><td colspan="2">Mono.</td><td colspan="2">CS</td><td colspan="2">Mono.</td></tr><tr><td>↓WER</td><td>↑BLEU</td><td>↓WER</td><td>↑BLEU</td><td>↓WER</td><td>↑BLEU</td><td>↓WER</td><td>↑BLEU</td></tr><tr><td rowspan="4">Fisher</td><td>CASCADE UNIDIRECT</td><td>37.1(36.3)</td><td>22.5(22.1)</td><td>26.6(23.5)</td><td>24.7(25.6)</td><td>33.5(33.1)</td><td>24.6(24.6)</td><td>24.8(23.8)</td><td>25.5(25.7)</td></tr><tr><td>E2E UNIDIRECT</td><td>36.6(35.7)</td><td>22.3(22.2)</td><td>26.7(23.2)</td><td>25.0(26.0)</td><td>33.4(33.2)</td><td>24.4(24.5)</td><td>25.3(23.9)</td><td>25.5(25.9)</td></tr><tr><td>CASCADE UNI SHARED ENC</td><td>36.0(36.0)</td><td>21.6(22.2)</td><td>25.6(25.6)</td><td>24.3(24.8)</td><td>31.2(31.3)</td><td>25.4(25.6)</td><td>25.6(25.3)</td><td>24.8(24.9)</td></tr><tr><td>E2E BIDIRECT BY LANG</td><td>37.0(36.1)</td><td>23.4(23.3)</td><td>27.2(25.3)</td><td>25.0(25.5)</td><td>36.7(35.9)</td><td>22.8(23.0)</td><td>27.3(25.3)</td><td>25.0(25.4)</td></tr><tr><td rowspan="4">Miami</td><td>CASCADE UNIDIRECT</td><td>65.2(61.4)</td><td>8.8(8.3)</td><td>52.3(50.0)</td><td>16.8(17.3)</td><td>64.8(64.4)</td><td>10.8(10.8)</td><td>51.5(50.9)</td><td>16.8(16.9)</td></tr><tr><td>E2E UNIDIRECT</td><td>65.6(63.1)</td><td>10.1(9.4)</td><td>53.0(51.2)</td><td>17.2(17.7)</td><td>65.1(65.6)</td><td>11.7(11.7)</td><td>51.4(50.7)</td><td>17.6(17.7)</td></tr><tr><td>CASCADE UNI SHARED ENC</td><td>60.2(60.2)</td><td>9.7(8.8)</td><td>53.8(53.8)</td><td>15.7(16.0)</td><td>55.0(56.0)</td><td>14.7(14.4)</td><td>55.6(55.5)</td><td>15.3(15.3)</td></tr><tr><td>E2E BIDIRECT BY LANG</td><td>69.5(69.7)</td><td>12.4(10.7)</td><td>55.2(53.4)</td><td>16.5(16.7)</td><td>69.3(69.7)</td><td>11.5(10.4)</td><td>54.5(53.2)</td><td>16.6(16.6)</td></tr></table>
+
+Table 7: Scores on the code-switched test sets for the models using LID, with results from zero CS training on the left and results after fine-tuning on the right.  
+
+<table><tr><td rowspan="2">Models</td><td colspan="4">Fisher CS Dev Set</td></tr><tr><td>↓WER</td><td>↓CER</td><td>↓CCT</td><td>↑BLEU</td></tr><tr><td>CASCADE UNIDIRECT</td><td>34.2</td><td>19.3</td><td>38.4</td><td>26.4</td></tr><tr><td>E2E UNIDIRECT</td><td>33.0</td><td>18.9</td><td>37.3</td><td>27.8</td></tr><tr><td>CASCADE UNI SHARED ENC</td><td>32.3</td><td>17.9</td><td>38.4</td><td>24.9</td></tr><tr><td>E2E BIDIRECT BY LANG</td><td>36.3</td><td>23.0</td><td>39.3</td><td>26.3</td></tr><tr><td>E2E BIDIRECT BY TASK</td><td>31.1</td><td>17.0</td><td>35.1</td><td>29.0</td></tr><tr><td>CASCADE BIDIRECT</td><td>35.1</td><td>19.2</td><td>39.7</td><td>23.8</td></tr><tr><td>E2E BIDIRECT SHARED</td><td>31.7</td><td>17.5</td><td>35.2</td><td>28.3</td></tr></table>
+
+Table 8: Scores on the Fisher CS Dev set. CCT stands for Charcut. Note that this mirrors the main results in Table 4.  
+
+<table><tr><td rowspan="2">Models</td><td colspan="4">MuST-C Test Set</td><td colspan="4">CoVoST Test Set</td></tr><tr><td>↓WER</td><td>↓CER</td><td>↓CCT</td><td>↑BLEU</td><td>↓WER</td><td>↓CER</td><td>↓CCT</td><td>↑BLEU</td></tr><tr><td>CASCADE UNIDIRECT</td><td>11.2</td><td>7.6</td><td>36.3</td><td>29.4</td><td>17.2</td><td>5.8</td><td>35.6</td><td>26.9</td></tr><tr><td>E2E UNIDIRECT</td><td>13.0</td><td>8.9</td><td>37.3</td><td>27.8</td><td>18.6</td><td>6.4</td><td>36.0</td><td>26.2</td></tr><tr><td>CASCADE UNI SHARED ENC</td><td>12.0</td><td>8.1</td><td>37.7</td><td>26.9</td><td>22.9</td><td>7.3</td><td>36.2</td><td>26.0</td></tr><tr><td>E2E BIDIRECT BY LANG</td><td>11.6</td><td>7.8</td><td>36.6</td><td>28.6</td><td>19.7</td><td>7.7</td><td>37.1</td><td>25.4</td></tr><tr><td>E2E BIDIRECT BY TASK</td><td>11.4</td><td>7.6</td><td>36.6</td><td>28.4</td><td>17.9</td><td>6.0</td><td>35.3</td><td>26.8</td></tr><tr><td>CASCADE BIDIRECT</td><td>13.6</td><td>9.5</td><td>39.7</td><td>24.5</td><td>22.9</td><td>7.3</td><td>38.8</td><td>22.8</td></tr><tr><td>E2E BIDIRECT SHARED</td><td>11.6</td><td>7.7</td><td>36.6</td><td>28.5</td><td>18.1</td><td>6.2</td><td>35.0</td><td>27.4</td></tr></table>
+
+Table 9: Scores on the MustC and CovoST datasets. CCT stands for Charcut.
+
+<table><tr><td rowspan="3" colspan="2">Models</td><td colspan="8">Mami</td></tr><tr><td colspan="4">CS Test Set</td><td colspan="4">Monolingual Test Set</td></tr><tr><td>↓WER</td><td>↓CER</td><td>↓CCT</td><td>↑BLEU</td><td>↓WER</td><td>↓CER</td><td>↓CCT</td><td>↑BLEU</td></tr><tr><td rowspan="11">Not Fine-Tuned</td><td>CASCADE UNIDIRECT</td><td>63.6</td><td>43.3</td><td>65.2</td><td>8.7</td><td>52.3</td><td>34.1</td><td>51.9</td><td>17.0</td></tr><tr><td>CASCADE UNIDIRECT ORA.</td><td>61.4</td><td>41.6</td><td>67.4</td><td>8.3</td><td>50.0</td><td>32.4</td><td>50.9</td><td>17.3</td></tr><tr><td>E2E UNIDIRECT</td><td>64.0</td><td>43.0</td><td>64.0</td><td>9.9</td><td>53.0</td><td>34.6</td><td>51.0</td><td>17.4</td></tr><tr><td>E2E UNIDIRECT ORA.</td><td>63.1</td><td>42.2</td><td>66.5</td><td>9.4</td><td>51.2</td><td>33.1</td><td>50.1</td><td>17.7</td></tr><tr><td>CASCADE UNI SHARED ENC</td><td>60.2</td><td>39.7</td><td>63.7</td><td>9.3</td><td>53.8</td><td>34.1</td><td>52.7</td><td>15.9</td></tr><tr><td>CASCADE UNI SHARED ENC ORA.</td><td>60.2</td><td>39.7</td><td>66.1</td><td>8.8</td><td>53.8</td><td>34.1</td><td>52.2</td><td>16.0</td></tr><tr><td>E2E BIDIRECT BY LANG</td><td>68.8</td><td>48.4</td><td>61.2</td><td>11.5</td><td>54.6</td><td>37.3</td><td>52.0</td><td>16.7</td></tr><tr><td>E2E BIDIRECT BY LANG ORA.</td><td>69.7</td><td>49.4</td><td>63.6</td><td>10.7</td><td>53.4</td><td>36.3</td><td>51.5</td><td>16.7</td></tr><tr><td>E2E BIDIRECT BY TASK</td><td>59.9</td><td>39.6</td><td>59.4</td><td>11.0</td><td>50.0</td><td>32.6</td><td>49.7</td><td>18.1</td></tr><tr><td>CASCADE BIDIRECT</td><td>61.4</td><td>39.8</td><td>62.2</td><td>9.3</td><td>54.0</td><td>34.1</td><td>53.1</td><td>14.8</td></tr><tr><td>E2E BIDIRECT SHARED</td><td>58.9</td><td>39.1</td><td>58.5</td><td>11.8</td><td>49.9</td><td>32.2</td><td>49.3</td><td>18.3</td></tr><tr><td rowspan="11">Fine-Tuned</td><td>CASCADE UNIDIRECT</td><td>64.8</td><td>42.0</td><td>56.5</td><td>10.8</td><td>51.5</td><td>33.4</td><td>51.1</td><td>16.8</td></tr><tr><td>CASCADE UNIDIRECT ORA.</td><td>64.4</td><td>41.8</td><td>56.4</td><td>10.8</td><td>50.9</td><td>32.9</td><td>50.6</td><td>16.9</td></tr><tr><td>E2E UNIDIRECT</td><td>65.1</td><td>43.0</td><td>56.9</td><td>11.7</td><td>51.4</td><td>33.7</td><td>50.4</td><td>17.6</td></tr><tr><td>E2E UNIDIRECT ORA.</td><td>65.6</td><td>43.1</td><td>57.0</td><td>11.7</td><td>50.7</td><td>33.2</td><td>49.9</td><td>17.7</td></tr><tr><td>CASCADE UNI SHARED ENC</td><td>55.0</td><td>35.2</td><td>51.4</td><td>14.7</td><td>55.6</td><td>35.9</td><td>52.9</td><td>15.3</td></tr><tr><td>CASCADE UNI SHARED ENC ORA.</td><td>56.0</td><td>35.7</td><td>51.7</td><td>14.4</td><td>55.5</td><td>35.9</td><td>52.7</td><td>15.3</td></tr><tr><td>E2E BIDIRECT BY LANG</td><td>69.3</td><td>48.6</td><td>61.3</td><td>11.5</td><td>54.5</td><td>37.2</td><td>52.1</td><td>16.6</td></tr><tr><td>E2E BIDIRECT BY LANG ORA.</td><td>69.7</td><td>49.5</td><td>63.8</td><td>10.4</td><td>53.2</td><td>36.1</td><td>51.5</td><td>16.6</td></tr><tr><td>E2E BIDIRECT BY TASK</td><td>53.6</td><td>35.0</td><td>53.3</td><td>13.8</td><td>52.6</td><td>34.4</td><td>50.5</td><td>17.5</td></tr><tr><td>CASCADE BIDIRECT</td><td>57.4</td><td>36.3</td><td>58.8</td><td>10.6</td><td>58.2</td><td>36.6</td><td>55.1</td><td>14.0</td></tr><tr><td>E2E BIDIRECT SHARED</td><td>53.0</td><td>35.0</td><td>54.4</td><td>14.1</td><td>52.1</td><td>33.9</td><td>50.4</td><td>17.4</td></tr></table>
+
+Table 10: Scores on the Miami dataset. CCT stands for Charcut. Results from zero CS training are on the top half and results after fine-tuning are on the bottom half. Ora stands for Oracle.  
+
+<table><tr><td rowspan="3" colspan="2">Models</td><td colspan="8">Fisher</td></tr><tr><td colspan="4">CS Test Set</td><td colspan="4">Monolingual Test Set</td></tr><tr><td>↓WER</td><td>↓CER</td><td>↓CCT</td><td>↑BLEU</td><td>↓WER</td><td>↓CER</td><td>↓CCT</td><td>↑BLEU</td></tr><tr><td rowspan="11">Not Fine-Tuned</td><td>CASCADE UNIDIRECT</td><td>37.3</td><td>22.2</td><td>45.6</td><td>21.9</td><td>28.0</td><td>15.3</td><td>40.0</td><td>24.4</td></tr><tr><td>CASCADE UNIDIRECT ORA.</td><td>36.3</td><td>21.5</td><td>45.0</td><td>22.1</td><td>23.5</td><td>12.0</td><td>38.0</td><td>25.6</td></tr><tr><td>E2E UNIDIRECT</td><td>36.9</td><td>22.0</td><td>45.1</td><td>21.8</td><td>28.2</td><td>15.6</td><td>39.7</td><td>24.7</td></tr><tr><td>E2E UNIDIRECT ORA.</td><td>35.7</td><td>21.3</td><td>44.4</td><td>22.2</td><td>23.2</td><td>12.0</td><td>37.6</td><td>26.0</td></tr><tr><td>CASCADE UNI SHARED ENC</td><td>36.0</td><td>20.5</td><td>44.7</td><td>21.9</td><td>25.6</td><td>13.0</td><td>39.8</td><td>24.3</td></tr><tr><td>CASCADE UNI SHARED ENC ORA.</td><td>36.0</td><td>20.5</td><td>44.2</td><td>22.2</td><td>25.6</td><td>13.0</td><td>38.8</td><td>24.8</td></tr><tr><td>E2E BIDIRECT BY LANG</td><td>36.9</td><td>23.6</td><td>43.0</td><td>23.2</td><td>27.2</td><td>15.5</td><td>39.2</td><td>25.1</td></tr><tr><td>E2E BIDIRECT BY LANG ORA.</td><td>36.1</td><td>22.9</td><td>42.6</td><td>23.3</td><td>25.3</td><td>14.0</td><td>38.4</td><td>25.5</td></tr><tr><td>E2E BIDIRECT BY TASK</td><td>34.1</td><td>19.4</td><td>42.3</td><td>23.0</td><td>23.6</td><td>11.9</td><td>37.4</td><td>26.0</td></tr><tr><td>CASCADE BIDIRECT</td><td>37.2</td><td>21.3</td><td>43.8</td><td>21.8</td><td>26.5</td><td>13.3</td><td>39.5</td><td>24.1</td></tr><tr><td>E2E BIDIRECT SHARED</td><td>33.8</td><td>19.3</td><td>41.5</td><td>23.3</td><td>23.2</td><td>11.8</td><td>37.1</td><td>26.2</td></tr><tr><td rowspan="11">Fine-Tuned</td><td>CASCADE UNIDIRECT</td><td>33.5</td><td>18.5</td><td>39.6</td><td>24.6</td><td>24.8</td><td>12.9</td><td>38.2</td><td>25.5</td></tr><tr><td>CASCADE UNIDIRECT ORA.</td><td>33.1</td><td>18.4</td><td>39.4</td><td>24.6</td><td>23.8</td><td>12.1</td><td>37.7</td><td>25.7</td></tr><tr><td>E2E UNIDIRECT</td><td>33.4</td><td>19.1</td><td>40.0</td><td>24.4</td><td>25.3</td><td>13.3</td><td>38.3</td><td>25.5</td></tr><tr><td>E2E UNIDIRECT ORA.</td><td>33.2</td><td>19.0</td><td>39.9</td><td>24.5</td><td>23.9</td><td>12.2</td><td>37.7</td><td>25.9</td></tr><tr><td>CASCADE UNI SHARED ENC</td><td>31.2</td><td>17.1</td><td>38.4</td><td>25.4</td><td>25.6</td><td>13.0</td><td>38.7</td><td>24.8</td></tr><tr><td>CASCADE UNI SHARED ENC ORA.</td><td>31.3</td><td>17.1</td><td>38.2</td><td>25.6</td><td>25.3</td><td>12.8</td><td>38.5</td><td>24.9</td></tr><tr><td>E2E BIDIRECT BY LANG</td><td>36.7</td><td>23.3</td><td>42.9</td><td>22.8</td><td>27.3</td><td>15.5</td><td>39.3</td><td>25.0</td></tr><tr><td>E2E BIDIRECT BY LANG ORA.</td><td>35.9</td><td>22.7</td><td>42.5</td><td>23.0</td><td>25.3</td><td>14.0</td><td>38.4</td><td>25.4</td></tr><tr><td>E2E BIDIRECT BY TASK</td><td>30.1</td><td>16.2</td><td>38.3</td><td>25.6</td><td>24.3</td><td>12.3</td><td>37.8</td><td>25.6</td></tr><tr><td>CASCADE BIDIRECT</td><td>33.2</td><td>18.3</td><td>41.0</td><td>23.2</td><td>28.1</td><td>14.3</td><td>40.1</td><td>23.2</td></tr><tr><td>E2E BIDIRECT SHARED</td><td>30.0</td><td>16.4</td><td>38.0</td><td>25.4</td><td>24.1</td><td>12.2</td><td>37.3</td><td>26.1</td></tr></table>
+
+Table 11: Scores on the Fisher dataset. CCT stands for Charcut. Results from zero CS training are on the top half and results after fine-tuning are on the bottom half. Ora stands for Oracle.

endtoendspeechtranslationforcodeswitchedspeech/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da035a6384c3928b5edf5d55020e208ff9f76a409da2f75b0b12027a16933fea
+size 998766

endtoendspeechtranslationforcodeswitchedspeech/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5cf7c247b0612bb7206355a9de240780e3582e0354a134ac05d692024d4cca1
+size 361726

enhancingnaturallanguagerepresentationwithlargescaleoutofdomaincommonsense/e46f025a-b8b2-4be0-afb2-586cc6a74893_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3308a39ac7141279f1f7f68597f0d722f9f36b44050a141593fe5f99b5d57410
+size 75693

enhancingnaturallanguagerepresentationwithlargescaleoutofdomaincommonsense/e46f025a-b8b2-4be0-afb2-586cc6a74893_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:483ecd7af2f499aa77076888830e68fc85fce1095dcb38bbc68c78140c9b91bb
+size 90139

enhancingnaturallanguagerepresentationwithlargescaleoutofdomaincommonsense/e46f025a-b8b2-4be0-afb2-586cc6a74893_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13c3f459e2be223e4a63366b49d2c4539c851f5de015c93d4b84208420b8a708
+size 633043

enhancingnaturallanguagerepresentationwithlargescaleoutofdomaincommonsense/full.md

@@ -0,0 +1,316 @@
+# Enhancing Natural Language Representation with Large-Scale Out-of-Domain Commonsense
+
+Wanyun Cui, Xingran Chen
+
+Shanghai University of Finance and Economics
+
+cui.wanyun@sufe.edu.cn, xingran.chen.sufe@gmail.com
+
+# Abstract
+
+We study how to enhance text representation via textual commonsense. We point out that commonsense has the nature of domain discrepancy. Namely, commonsense has different data formats and is domain-independent from the downstream task. This nature brings challenges to introducing commonsense in general text understanding tasks. A typical method of introducing textual knowledge is continuing pre-training over the commonsense corpus. However, it will cause catastrophic forgetting to the downstream task due to the domain discrepancy. In addition, previous methods of directly using textual descriptions as extra input information cannot apply to large-scale commonsense.
+
+In this paper, we propose to use large-scale out-of-domain commonsense to enhance text representation. In order to effectively incorporate the commonsense, we proposed OK-Transformer (Out-of-domain Knowledge enhanced Transformer). OK-Transformer effectively integrates commonsense descriptions and enhances them to the target text representation. In addition, OK-Transformer can adapt to the Transformer-based language models (e.g. BERT, RoBERTa) for free, without pre-training on large-scale unsupervised corpora. We have verified the effectiveness of OK-Transformer in multiple applications such as commonsense reasoning, general text classification, and low-resource commonsense settings.
+
+# 1 Introduction
+
+Although unsupervised language models have achieved big success on many tasks (Devlin et al., 2019), they are incapable of learning low-frequency knowledge. For example, in the masked language model task in Fig. 1, even if we replace "Kevin was" (left) with "Jim was" (right), BERT (Devlin
+
+et al., 2019) still predicts the masked word as sick, crying, dying, etc. This is because similar texts in its training corpus rarely describe the subject of "comforted". To improve the model's ability to generalize and understand low-frequency knowledge, we propose to incorporate commonsense into language models. In Fig. 1, to make correct predictions, we need to enhance the language model with the commonsense  $c_{1}$ .
+
+However, commonsense has the nature of domain discrepancy. The downstream task and the commonsense knowledge have distribution discrepancies. Taking the commonsense knowledge base we use (i.e. ATOMIC2020 (Hwang et al., 2020)) as an example, the distribution discrepancy is specifically manifested in (1) their data formats. The format of a commonsense description usually belongs to some specific patterns (e.g. "... As a result ..., "... Because ..."), while the downstream tasks can have arbitrary patterns. (2) The commonsense belongs to the domain of event causality, while the downstream tasks may belong to arbitrary domains.
+
+Here we highlight the challenges caused by the domain discrepancy. To introduce external textual knowledge to a pre-trained language model, a common practice is to continue pre-training the language model on the corpus of the external knowledge (Gururangan et al., 2020; Sun et al., 2019). However, the study (Gururangan et al., 2020) also found that continuing pre-training requires external knowledge and downstream tasks to have similar domains. Due to its domain discrepancy, introducing commonsense through continuing pre-training will cause catastrophic forgetting to downstream tasks, thereby injuring the effectiveness. We have verified this empirically in Sec 6.3. Therefore, the domain discrepancy prevents us from introducing commonsense by continuing pre-training.
+
+To enhance the representation of the target text with external commonsense, we propose to directly use its candidate commonsense as an extra input.
+
+![](images/39dd2f9d26b97517fdc8301aef66c45c00db715c21e7f5847056794f53a6e3f7.jpg)  
+Figure 1: The prediction of [MASK] by BERT. BERT cannot distinguish between Jim and Kevin in Jim comforted Kevin because.
+
+Our setup is different from a typical natural language understanding setup since the latter one only takes the target text as the input (Devlin et al., 2019). We argue that our setup – where the commonsense is introduced explicitly as input – is a more practicable setup to introduce out-of-domain commonsense that cannot be learned through pretraining. As far as we know, ExpBERT (Murty et al., 2020) is the closest setup to us. It also uses external knowledge (manually constructed templates) as the input.
+
+Another challenge is the scale of the commonsense. Although ExpBERT also allows extra textual commonsense as input, it only captures small-scale commonsense with a fixed size. In addition, when we introduce commonsense from a large-scale knowledge base for general purpose (i.e. ATOMIC2020), unrelated commonsense (e.g.  $c_{2}$  and  $c_{3}$  in Fig. 1) will certainly occur. However, ExpBERT lacks the ability to distinguish related and unrelated commonsense. Therefore, the power of large-scale commonsense knowledge was restricted in ExpBERT. We will verify this empirically in Sec 6.3.
+
+In order to incorporate the large-scale out-of-domain commonsense, we propose the OK-Transformer (Out-of-domain Knowledge enhanced Transformer) on the basis of Transformer (Vaswani et al., 2017). OK-Transformer has two modules. The knowledge enhancement module is used to encode the target text with commonsense, and the knowledge integration module is used to encode and integrate all candidate commonsense. OK-Transformer has two advantages. First, it fully represents the contextual information of the textual commonsense. Second, it can be adapted to existing pre-trained language models (e.g. BERT and RoBERTa) for free. That is, we are able to
+
+adapt OK-Transformer to the pre-trained language models, without pre-training OK-Transformer over large-scale unsupervised corpora from scratch.
+
+Some other methods are related to our work, such as introducing structured knowledge (Peters et al., 2019; Zhang et al., 2019; Guan et al., 2020; Zhou et al., 2018) and plain text knowledge (Guu et al., 2020) in language models. These methods do not represent the specific inductive bias of commonsense knowledge and therefore are not suitable to introduce commonsense. We will compare these studies with more details in Sec 2.
+
+# 2 Related work
+
+In this section, we compare different ways to introduce knowledge into language models. We divide the knowledge introduction methods into (1) continuing pre-training method (Gururangan et al., 2020; Sun et al., 2019) and (2) explicit introduction in the downstream task (Guu et al., 2020; Murty et al., 2020).
+
+Continuing pre-training the language model is effective when the external knowledge is similar to the downstream task (Gururangan et al., 2020; Sun et al., 2019). However, commonsense and downstream tasks have domain discrepancies, so continuing pre-training is unsuitable for introducing commonsense. We have empirically verified this in Sec 6.3.
+
+Introducing explicit knowledge in downstream tasks We classify the knowledge into structured knowledge, plain text, and semi-structured knowledge, depending on its form. The entries of structured knowledge are represented as individual embeddings (Peters et al., 2019; Zhang et al., 2019; Guan et al., 2020; Zhou et al., 2018), while commonsense descriptions in this paper can be represented more accurately by the contextual
+
+information of their word sequences.
+
+# 3 Problem Setup: Commonsense as the Extra Input
+
+We consider a text classification task where the text  $x$  and its label  $y$  are provided for training. Assuming that the candidate commonsense descriptions for enhancing  $x$  come from a large-scale commonsense knowledge base (i.e. ATOMIC2020), we retrieve candidate commonsense for  $x$  as the extra input. We denote the commonsense descriptions for  $x$  as  $cs(x) = \{c_1 \cdots c_n\}$ , where each  $c_i$  is a commonsense description. The retrieval process will be shown in Sec 6. The model takes both  $x$  and  $cs(x)$  as the input. Since ATOMIC2020 contains if-then knowledge for general purposes, the problem setup can be expanded to a broad range of text understanding tasks. The goal of training is to find parameter  $\theta$  that minimizes the loss of training examples given the texts and candidate commonsense descriptions:
+
+$$
+\arg \min  _ {\theta} \mathbb {E} _ {(\mathrm {x}, \mathrm {y}) \in \operatorname {t r a i n}} \mathcal {L} (\mathrm {f} (\mathrm {x}, \operatorname {c s} (\mathrm {x}) | \theta), \mathrm {y}) \tag {1}
+$$
+
+where  $\mathrm{f}(\cdot |\theta)$  is the model taking  $x$  and  $cs(x)$  as inputs,  $\mathcal{L}$  is the loss function.
+
+# 4 OK-Transformer
+
+In this section, we propose OK-Transformer based on Transformer to introduce extra commonsense descriptions. We first show OK-Transformer on an abstract level in Sec 4.1. Then we elaborate two modules within it, i.e. knowledge enhancement and knowledge integration, in Sec 4.2 and Sec 4.3, respectively.
+
+# 4.1 Framework
+
+In this subsection, we show how our OK-Transformer works at an abstract level. For the target sentence  $x$ , OK-Transformer takes both  $x$  and  $cs(x)$  as inputs. To incorporate all the information of  $x$  and  $cs(x)$ , the OK-Transformer contains three vanilla Transformers, denoted by  $\mathrm{Transformer}^{(1)(2)(3)}$ . The knowledge enhancement module uses  $\mathrm{Transformer}^{(1)}$  to encode the target text. Compared with the vanilla Transformer,  $\mathrm{Transformer}^{(1)}$  leverages a new knowledge token to represent the commonsense that interacts with other words. The knowledge integration module encodes each individual commonsense description by  $\mathrm{Transformer}^{(2)}$ , and then integrates all candidate commonsense descriptions by  $\mathrm{Transformer}^{(3)}$ . This is shown in Fig. 2.
+
+![](images/5fb7ab70d61758207e82f2b823d08155ab06be918d1d49dbf55652ec6f911474.jpg)  
+Figure 2: OK-Transformer.  $\mathrm{Transformer}^{(1)}$  encodes the target text  $x$  with enhanced commonsense  $k_{i}$ .  $\mathrm{Transformer}^{(2)}$  encodes each individual commonsense description.  $\mathrm{Transformer}^{(3)}$  integrates all candidate commonsense descriptions and transfers knowledge to  $\mathrm{Transformer}^{(1)}$ .
+
+# 4.2 Knowledge Enhancement Module
+
+The knowledge enhancement module allows commonsense knowledge to enhance the representation of the target text.
+
+Interaction between words and commonsense. We use  $\mathrm{Transformer}^{(1)}$  to represent the interaction between words of the target text  $x$ . In addition, we introduce a special token  $[k]$  to represent the commonsense knowledge. We denote it as the knowledge token.  $\mathrm{Transformer}^{(1)}$  encodes all words and the knowledge token together via multi-head attention. Formally, given word sequence  $\mathrm{x = w_1,\dots,w_n}$ ,  $\mathrm{Transformer}^{(1)}$  accepts a sequence of  $n + 1$  word-piece tokens:  $[\mathbf{k}]$ ,  $\mathrm{w}_1,\dots ,\mathrm{w}_n$ . We denote the knowledge embedding and word embeddings produced by the  $i$ -th layer of  $\mathrm{Transformer}^{(1)}$  as  $k_{i}\in \mathbb{R}^{d}$  and  $\mathbf{H_i}\in \mathbb{R}^{\mathbf{n}\times \mathbf{d}}$ , respectively. The  $\mathrm{Transformer}^{(1)}$  block first uses a multi-head self-attention layer followed by a residual connection and a layer normalization to model their interactions:
+
+$$
+\mathrm {k} _ {\mathrm {i}} ^ {\prime}, \mathbf {H} _ {\mathrm {i}} ^ {\prime} = \text {L a y e r N o r m} ([ \mathrm {k} _ {\mathrm {i} - 1}, \mathbf {H} _ {\mathrm {i} - 1} ] +
+$$
+
+$$
+\text {M u l t i H e a d A t t n} \left(\left[ k _ {i - 1}, \mathbf {H} _ {i - 1} \right], \left[ k _ {i - 1}, \mathbf {H} _ {i - 1} \right], \left[ k _ {i - 1}, \mathbf {H} _ {i - 1} \right]\right) \tag {2}
+$$
+
+where  $[\mathrm{k_{i - 1}},\mathbf{H}_{\mathrm{i - 1}}]\in \mathbb{R}^{(\mathrm{n + 1})\times \mathrm{d}}$  means appending  $k_{i - 1}$  at the front of  $\mathbf{H}_{i - 1}$ .  $[\mathrm{k_{i - 1}},\mathbf{H}_{\mathrm{i - 1}}]$  is used as the query, key, and value in the multi-head attention.
+
+Knowledge update The vanilla Transformer projects  $\mathbf{k}_{\mathrm{i}}^{\prime}, \mathbf{H}_{\mathrm{i}}^{\prime}$  in Eq. (2) to the output space with a multi-layer perceptron neural network (MLP). Compared to the vanilla Transformer, we use an extra update operation to update the knowledge token by the integrated commonsense knowledge after the MLP. As in the vanilla Transformer, the update layer is followed by a residual connection and a layer normalization. This can be formulated by:
+
+$$
+\begin{array}{l} \mathrm {k} _ {\mathrm {i}} = \text {L a y e r N o r m} \left(\mathrm {k} _ {\mathrm {i}} ^ {\prime} + \mathrm {M L P} \left(\mathrm {k} _ {\mathrm {i}} ^ {\prime}\right) + \mathrm {c s} _ {\mathrm {e m b}}\right) \\ \mathbf {H} _ {\mathrm {i}} = \operatorname {L a y e r N o r m} \left(\mathbf {H} _ {\mathrm {i}} ^ {\prime} + \operatorname {M L P} \left(\mathbf {H} _ {\mathrm {i}} ^ {\prime}\right)\right) \\ \end{array}
+$$
+
+where  $c s_{emb}$  is the embedding of the commonsense computed by the knowledge integration module in Sec 4.3.
+
+# 4.3 Knowledge Integration Module
+
+The knowledge integration module encodes all candidate commonsense descriptions and integrates them. We first use  $\mathrm{Transformer}^{(2)}$  to represent each candidate commonsense description. Then, we use  $\mathrm{Transformer}^{(3)}$  to integrate all candidate commonsense, and transfer the integrated knowledge to the knowledge enhancement module.
+
+Representing single commonsense We use a vanilla Transformer as  $\mathrm{Transformer^{(2)}}$  to model each candidate commonsense description. For all the retrieved commonsense  $cs(x) = \{c_1,\dots ,c_n\}$ , we compute the embedding  $emb_{j}$  of each commonsense description  $c_{j}$  by:
+
+$$
+\mathrm {e m b} _ {\mathrm {j}} = \operatorname {T r a n s f o r m e r} ^ {(2)} \left(\mathrm {c} _ {\mathrm {j}}\right) \tag {4}
+$$
+
+Knowledge integration We integrate all candidate commonsense by  $\mathrm{Transformer^{(3)}}$ . Since not all the candidate commonsense leads to high confidence prediction as we have discussed in Sec 1, we need to select relevant commonsense and ignore irrelevant commonsense. Transformer is adequate to conduct this selection. Specifically, in the query-key-value mechanism in Transformer, we use the embedding of the knowledge token in  $\mathrm{Transformer^{(1)}}$  as the query of  $\mathrm{Transformer^{(3)}}$ , and the commonsense embeddings by  $\mathrm{Transformer^{(2)}}$  as keys and values of  $\mathrm{Transformer^{(3)}}$ . Then, we integrate representations of all different commonsense descriptions based on their similarities with the knowledge token.
+
+Transformer $^{(3)}$  also uses multi-head attention to allow the knowledge token to interact with the candidate commonsense in multiple ways. The output of multi-head self-attention is followed by a residual connection and a layer normalization.
+
+$$
+\begin{array}{l} \mathrm {c s} _ {\text {e m b}} = \text {L a y e r N o r m} \left(\mathrm {k} _ {\mathrm {i} - 1}\right) \tag {5} \\ + \text {M u l t i H e a d A t t n} \left(\mathrm {k} _ {\mathrm {i} - 1}, \mathbf {e m b}, \mathbf {e m b}\right)) \\ \end{array}
+$$
+
+where  $\mathbf{emb} = [\mathrm{emb}_1, \dots, \mathrm{emb}_n]$  denotes the sequence of embeddings of all candidate commonsense descriptions. We then apply a residual connection and a layer normalization to it.
+
+Null Commonsense Some target texts may not have valid commonsense from ATOMIC2020 to enhance their representations. Therefore, we refer to the settings of REALM (Guu et al., 2020) to add a null commonsense into the candidate commonsense of all target texts. We denote the null commonsense as  $c_{0}$ . Matching to the null commonsense indicates that the commonsense knowledge base cannot help enhance the target text.
+
+# 5 Adaptation to Pre-trained Language Models
+
+In this section, we take BERT as an example to illustrate how we adapt OK-Transformer to existing pre-trained language models. We denote the adapted model as OK-BERT. An important manifestation of the effectiveness of the Transformer structure is its applications in large-scale pre-trained models (e.g. BERT, RoBERTa). In order to introduce external knowledge, many other studies conduct training over large-scale unsupervised corpus (Peters et al., 2019; Xiong et al., 2019). However, OK-Transformer is able to directly adapt to the existing pre-trained language models for free. In other words, when adapting OK-Transformer to OK-BERT, we directly use the parameters of each Transformer layer of BERT to initialize the OK-Transformer layers of OK-BERT. This property greatly improves the applicability of OK-BERT. In the rest of this section, we will describe how  $\mathrm{Transformer}^{(1)}$ ,  $\mathrm{Transformer}^{(2)}$ , and  $\mathrm{Transformer}^{(3)}$  are adapted respectively in Sec 5.1, and how to fine-tune OK-BERT in Sec 5.2.
+
+# 5.1 Layer-by-Layer Adaptation
+
+The OK-BERT we designed uses two original BERTs to serve as  $\mathrm{Transformer}^{(1)}$  and  $\mathrm{Transformer}^{(2)}$ , respectively. We denote them as BERT1 and BERT2. We connect the
+
+Transformer $^{(1)}$  and Transformer $^{(2)}$  in the corresponding layer of each BERT by Transformer $^{(3)}$ . Therefore, OK-BERT makes full use of the multi-layer structure of BERT, while allowing commonsense in the knowledge token to fully interact with the target text in each layer. The architecture is shown in Fig. 3.
+
+![](images/2706676cda0e054f3e7bd4dad48527154f50e4a6a03e4bfbea44dc4668b23913.jpg)  
+Figure 3: The architecture of OK-BERT. We only draw edges that connect to the  $i$ -th layer.
+
+$\mathbf{Transformer^{(1)}}$  We adapt the Transformer of BERT1 to  $\mathbf{Transformer^{(1)}}$  in the knowledge enhancement module of OK-Transformer. Note that the original BERT's tokens are [CLS]  $\mathrm{w}_1\dots \mathrm{w_L}$  [SEP] (for a single sentence) or [CLS]  $\mathrm{w}_1\dots \mathrm{w_m}$  [SEP]  $\mathrm{w}_{\mathrm{m} + 1}\dots \mathrm{w_L}$  [SEP] (for a sentence pair). We follow (Wang et al., 2020) and use a special token  $[k]$  as the knowledge token. When tokenizing sentences, we insert the  $[k]$  token after the [CLS] token for each given text. In this way, the input tokens become [CLS]  $[\mathrm{k}]$  w1 wL [SEP] or [CLS]  $[\mathrm{k}]$  w1...wm [SEP]  $\mathrm{w}_{\mathrm{m} + 1}\dots \mathrm{w}_{\mathrm{L}}$  [SEP], respectively. This simple modification allows us to use  $[k]$  as the knowledge token in the knowledge enhancement module.
+
+$\mathbf{Transformer}^{(2)}$  We adapt each Transformer layer of BERT2 to the  $\mathbf{Transformer}^{(2)}$  layer. The adaptation is straightforward since  $\mathbf{Transformer}^{(2)}$  uses the vanilla Transformer structure. We use the encoding of the [CLS] token in each corresponding layer as the commonsense representation  $emb_{j}$  to enhance the representation of the corresponding layer in BERT1.
+
+$\mathbf{Transformer}^{(3)}$  For each pair of corresponding  $\mathbf{Transformer}^{(1)}$  and  $\mathbf{Transformer}^{(2)}$  from the same layer, we use one  $\mathbf{Transformer}^{(3)}$  to connect them to transfer the information from BERT2 to BERT1.
+
+In summary, when adapting to BERT-base with 12 Transformer layers, OK-BERT con
+
+tains 12 Transformer $^{(1)}$  layers for BERT1, 12 Transformer $^{(2)}$  layers for BERT2, and 12 Transformer $^{(3)}$  layers for layer-wise knowledge integration.
+
+# 5.2 Parameter Initialization and Model Training
+
+In our implementation, BERT1 and BERT2 have independent parameters. We use the parameters of BERT to initialize both BERT1 and BERT2. The parameters of Transformer $^{(3)}$  layers are randomly initialized. For downstream tasks, we then finetune all the parameters in the fashion of end2end.
+
+# 6 Experiments
+
+We evaluate the effectiveness of our proposed models in three scenarios: cloze-style commonsense reasoning, text classification, and low-resource commonsense settings. All the experiments run over a computer with 4 Nvidia Tesla V100 GPUs.
+
+Models We consider adapting OK-Transformer to BERT and RoBERTa, which are denoted as OK-BERT and OK-RoBERTa, respectively. We use the BERT-base and RoBERTa-large from the Hugging-Face Transformer library (Wolf et al., 2020).
+
+Implementation details for candidate knowledge retrieval For a given text  $x$ , we retrieve candidate commonsense from ATOMIC2020. We use the if-then descriptions in ATOMIC2020 (e.g. Fig. 1). Since these descriptions cover 173k different verb phrases – one of the fundamental elements of language – the retrieval is applicable to a broad range of downstream text understanding tasks.
+
+We use a simple retrieval method. We simply consider word segments with window size 5 of the input text  $x$ . All the commonsense descriptions matching one of these text segments will be regarded as the candidate commonsense descriptions  $c_{i} \in cs(x)$ .
+
+# 6.1 Commonsense Reasoning
+
+# 6.1.1 Setup
+
+Datasets We consider the following commonsense reasoning benchmarks: WSC273 (Levesque et al., 2012), PDP (Morgenstern et al., 2016), Winogender (Rudinger et al., 2018), WinoGrande (Sakaguchi et al., 2019), CommonsenseQA (Talmor et al., 2019) and PhysicalQA (Bisk et al., 2020).
+
+Model details Due to the different implementations between (Kocijan et al., 2019b) and (Sakaguchi et al., 2019), in this paper, we also follow
+
+their settings to compare with them, respectively. For (Kocijan et al., 2019b), we conduct disambiguation tasks directly through masked language modeling in OK-BERT. For the latter one, we convert cloze-style problems to multiple-choice classification problems in OK-RoBERTa. In particular, we replace the target pronoun of one query sentence with each candidate reference, then put the new sentences into the language model. We use a single linear layer and a softmax layer over the encoding of its [CLS] token to compute the probability of each new sentence, and select the one with the highest probability as the pronoun disambiguation result.
+
+Hyperparameters of pre-training We follow (Kocijan et al., 2019b; Sakaguchi et al., 2019) to first pre-train models for 30 and 3 epochs over WSCR (Kocijan et al., 2019b) or WinoGrande (Sakaguchi et al., 2019), respectively. Then we fine-tune models over specific tasks. We use AdamW as the optimizer with learning rate 5e-6, which is selected from  $\{2e - 5, 1e - 5, 5e - 6\}$ . We set the batch size to 8.
+
+<table><tr><td>Model</td><td>WSC</td><td>PDP</td></tr><tr><td>KEE(Liu et al., 2016)</td><td>52.8</td><td>58.3</td></tr><tr><td>WKH (Emami et al., 2018)</td><td>57.1</td><td>-</td></tr><tr><td>MAS (Klein and Nabi, 2019)</td><td>60.3</td><td>68.3</td></tr><tr><td>DSSM (Wang et al., 2019)</td><td>63.0</td><td>75.0</td></tr><tr><td>LM(Trinh and Le, 2018)</td><td>63.8</td><td>70.0</td></tr><tr><td>CSS (Klein and Nabi, 2020)</td><td>69.6</td><td>90.0</td></tr><tr><td>GPT2 (Radford et al., 2019)</td><td>70.7</td><td>-</td></tr><tr><td>BERT-large+WSCR (Kocijan et al., 2019b)</td><td>71.4</td><td>79.2</td></tr><tr><td>HNN (He et al., 2019)</td><td>75.1</td><td>90.0</td></tr><tr><td>Human (Sakaguchi et al., 2019)</td><td>96.5</td><td>92.5</td></tr><tr><td>BERT+WSCR</td><td>66.3</td><td>85.0</td></tr><tr><td>OK-BERT+WSCR</td><td>67.4</td><td>86.7</td></tr><tr><td>RoB.+WinoGrande</td><td>90.1</td><td>87.5</td></tr><tr><td>OK-RoB.+WinoGrande</td><td>91.6</td><td>91.7</td></tr></table>
+
+Table 1: Results on WSC and PDP. RoB. denotes RoBERTa.  
+
+<table><tr><td>Model</td><td>WinoGen.</td><td>WinoGran.</td></tr><tr><td>WikiCREM (Kocijan et al., 2019a)</td><td>82.1</td><td>-</td></tr><tr><td>WinoGrande (Sakaguchi et al., 2019)</td><td>94.6</td><td>79.3</td></tr><tr><td>BERT+WSCR</td><td>68.2</td><td>51.4</td></tr><tr><td>OK-BERT+WSCR</td><td>72.4</td><td>53.4</td></tr><tr><td>RoB.+WinoGrande</td><td>94.6</td><td>79.3</td></tr><tr><td>OK-RoB.+WinoGrande</td><td>96.2</td><td>79.6</td></tr></table>
+
+# 6.1.2 Results
+
+We compare our models with state-of-the-art commonsense reasoning models in Table 1, 2, and 3.
+
+Table 2: Results on WinoGender and WinoGrande.  
+
+<table><tr><td>Model</td><td>CommonsenseQA</td><td>PhysicalQA</td></tr><tr><td>BERT</td><td>55.86</td><td>68.71</td></tr><tr><td>OK-BERT</td><td>56.27</td><td>69.09</td></tr><tr><td>RoBERTa</td><td>73.55</td><td>79.76</td></tr><tr><td>OK-RoBERTa</td><td>75.92</td><td>80.09</td></tr></table>
+
+Table 3: Results on CommonsenseQA and PhysicalQA.
+
+It can be seen that our models outperform other models in most settings. This verifies the effectiveness of our proposed models for commonsense reasoning.
+
+Ablations In Table 1, 2, and 3 we also compare OK-BERT with BERT. We found that OK-BERT with OK-Transformers effectively improved the accuracy of BERT with Transformers. Similar results can be found between OK-RoBERTa and RoBERTa. This shows that the proposed OK-Transformer improves pre-trained language models by adapting to them for free, i.e. without retraining on large-scale unsupervised corpora.
+
+# 6.2 General Text Classification
+
+We use MRPC, CoLA, RTE, STS-B, SST-2, and QNLI in the GLUE dataset (Wang et al., 2018) to verify the effectiveness of the proposed models on general text classification tasks. We did not evaluate over MNLI, because our model needs to represent the corresponding  $n$  commonsense for each sentence, which is too costly for MNLI. We believe that this efficiency problem can be solved by further applying model compression (Iandola et al., 2020), but this is beyond the scope of this paper. It can be seen from Table 4 that OK-BERT and OK-RoBERTa outperform their baselines.
+
+# 6.3 Commonsense Introduction Methods
+
+Continue pre-train In the introduction section, we mentioned that a typical method of introducing textual knowledge is continuing pre-training (Gururangan et al., 2020; Sun et al., 2019). However, due to the domain discrepancy of commonsense, this method will cause catastrophic forgetting. To verify this intuition, in this subsection we compare with the continuing pre-trained model. We first continue pre-training the language model over ATOMIC2020, then fine-tune it over the target task.
+
+ExpBERT (Murty et al., 2020) We also compare our OK-Transformer with ExpBERT, another model that is able to introduce textual knowledge. In Sec 1, we mentioned that ExpBERT is not appli
+
+<table><tr><td>GLUE Task</td><td>MRPC</td><td>CoLA</td><td>RTE</td><td>QNLI</td><td>STS-B</td><td>SST-2</td></tr><tr><td>BERT</td><td>86.27/90.21</td><td>59.50</td><td>71.43</td><td>91.20</td><td>89.35/88.93</td><td>91.97</td></tr><tr><td>OK-BERT</td><td>87.25/90.84</td><td>58.29</td><td>73.65</td><td>91.58</td><td>89.82/89.46</td><td>93.69</td></tr><tr><td>RoBERTa</td><td>90.44/93.15</td><td>66.57</td><td>84.11</td><td>94.00</td><td>91.83/91.95</td><td>95.70</td></tr><tr><td>OK-RoBERTa</td><td>91.91/94.24</td><td>66.89</td><td>86.28</td><td>94.41</td><td>92.41/92.20</td><td>96.10</td></tr></table>
+
+Table 4: Results on text classification tasks. Models are evaluated by the dev split from GLUE.
+
+cable to large-scale commonsense knowledge bases for its disability to select related commonsense and ignore unrelated commonsense. To verify this, we use the retrieved candidate commonsense descriptions from ATOMIC2020 as the additional explanations for ExpBERT. ExpBERT concatenates all the embedding of a fixed number of commonsense, which is inflexible for ATOMIC2020. For this reason, we fix the number of commonsense to 48. If there are more than 48 candidate commonsense descriptions for one sample, we will randomly select 48 of them. Otherwise, we will pad null commonsense to it. In our experiments, we also apply ExpBERT to RoBERTa (Liu et al., 2019) (i.e. ExpRoBERTa).
+
+We show the results in Table 5. We do not report the results of ExpBERT on WSC273, as ExpBERT cannot solve the cloze-style problems. It can be seen that the performance of language models was suffered when we simply continue pre-training the models on the commonsense knowledge base. This verifies that the continuing pre-training on the out-of-domain commonsense will cause catastrophic forgetting and injure the effectiveness. On the other hand, using OK-Transformer to introduce commonsense as the extra input significantly improves the accuracy. The results also suggest that ExpBERT is not applicable to large-scale commonsense knowledge bases.
+
+# 6.4 Why is OK-Transformer effective?
+
+We now analyze why OK-Transformer can effectively introduce out-of-domain commonsense without pre-training. We are inspired by an observation of language model fine-tuning LMs (Radiya-Dixit and Wang, 2020), i.e., the parameters after fine-tuning are close to those before fine-tuning. Therefore, we argue that the key to effective introduction is whether the parameters of the meta LM is good initialization for the commonsense-enhanced LM, that the parameters do not change much before and after fine-tuning.
+
+To verify this, we compare the parameter
+
+![](images/5a4ba77195b7a4ea3455504788feab1f4688abf499a57eda47f2c09cdc7cf781.jpg)  
+Figure 4:  $L_{1}$  distances in parameter space between pretrained and fine-tuned meta LMs. We show the metrics of  $W_{I}$  across the 12 Transformer layers.
+
+![](images/fc19e1f83ac81aee8dfe44d42ed89dfa548ed8c76bb5864600a28b6cdb4067df.jpg)  
+Figure 5: Losses of different knowledge integration methods in SST-2. The [CLS] token method does not converge.
+
+![](images/6be8aeaadbe27bddf16792b5b166a38542624de272b14d6b9cd8977459b14145.jpg)  
+Figure 6: Effect in lowresource commonsense settings with different  $k$ s over SST-2.
+
+changes of different knowledge integration methods. These methods include (1) OK-Transformer, (2) KnowBERT (Peters et al., 2019), (3) using the original  $[CLS]$  token instead of the proposed knowledge token, and (4) abandoning the knowledge token and instead calculating the  $cs_{emb}$  of each verb phrase of the target sentence separately, and adding them to these verb phrases' hidden states in  $\mathbf{H}_{\mathrm{i - 1}}$ . We follow (Radiya-Dixit and Wang, 2020) to use the  $L1$  as the distance metric. (Radiya-Dixit and Wang, 2020) found that the main change in parameters occurs on the  $W_{I}$  matrix of the Transformer. Our experimental results also follow this phenomenon. Therefore, for greater clarity, we only show the distances of the  $W_{I}$  matrices after fine-tune. We show the distances of different methods in Fig. 4, and their training losses in Fig. 5.
+
+<table><tr><td></td><td>MRPC</td><td>CoLA</td><td>RTE</td><td>QNLI</td><td>STS-B</td><td>SST-2</td><td>WSC273</td></tr><tr><td>BERT</td><td>86.27/90.21</td><td>59.50</td><td>71.43</td><td>91.20</td><td>89.35/88.93</td><td>91.97</td><td>66.30</td></tr><tr><td>BERT-continue</td><td>83.58/88.81</td><td>54.70</td><td>62.09</td><td>90.24</td><td>87.41/87.46</td><td>91.74</td><td>63.00</td></tr><tr><td>ExpBERT</td><td>85.78/89.79</td><td>58.29</td><td>62.82</td><td>87.06</td><td>84.78/84.67</td><td>91.51</td><td>-</td></tr><tr><td>OK-BERT</td><td>87.25/90.84</td><td>58.29</td><td>73.65</td><td>91.58</td><td>89.82/89.46</td><td>93.69</td><td>67.40</td></tr><tr><td>RoBERTa</td><td>90.44/93.15</td><td>66.57</td><td>84.11</td><td>94.00</td><td>91.83/91.95</td><td>95.70</td><td>90.10</td></tr><tr><td>RoBERTa-continue</td><td>87.01/90.38</td><td>61.74</td><td>74.01</td><td>93.61</td><td>89.57/89.66</td><td>95.99</td><td>87.91</td></tr><tr><td>ExpRoBERTa</td><td>89.46/92.22</td><td>66.90</td><td>83.39</td><td>93.78</td><td>89.81/89.94</td><td>95.99</td><td>-</td></tr><tr><td>OK-RoBERTa</td><td>91.91/94.24</td><td>66.89</td><td>86.28</td><td>94.41</td><td>92.41/92.20</td><td>96.10</td><td>91.58</td></tr></table>
+
+Table 5: Comparison of different commonsense introduction approaches. Continuing pre-training even injures the effectiveness. On the other hand, using OK-Transformers to introduce external knowledge achieves better results than using Transformer.
+
+It can be seen that the distances of OK-Transformer are much smaller than other methods, except the [CLS] token method, which does not converge as shown in Fig. 5. This fits our intuition of reducing the parameter variations to introduce external knowledge more effectively.
+
+# 6.5 Effect in Low-Resource Commonsense Settings
+
+Since there is a large number of commonsense descriptions in ATOMIC2020, a large portion of descriptions only occur a few times in the training set. In this subsection, we want to verify for these rare descriptions, can the model still benefit from it? If so, we think it means that the model uses the contextual information of the commonsense to improve the understanding of the commonsense.
+
+To do this, we proposed a low-resource commonsense setting. We evaluate the effect of the model if the training dataset only contains  $k = 8 / 16 / 32 / 64$  samples. Therefore the frequency of the appeared commonsense descriptions is low. In order to exclude the influence of other samples, we only use test samples whose candidate commonsense descriptions have already occurred in the  $k$  training samples. For example, when  $k = 8$ , we randomly select 8 samples from the training set for training, and use all samples in the test set which contains the commonsense of the 8 training samples for evaluation. We show the results over the SST-2 dataset in Fig. 6. It can be seen that our models still benefit from low-frequency commonsense.
+
+# 6.6 Does OK-Transformer Provide Interpretability?
+
+In this subsection, we try to answer if the integration of candidate commonsense descriptions by
+
+OK-Transformer is interpretable. To answer this question, we calculate the influence of different commonsense descriptions on the model's predictions. We follow (Wu et al., 2020) to quantify the influence of a commonsense description  $c_{i}$  as: If  $c_{i}$  is removed from  $cs(x)$ , how much will the prediction change? This change is measured by the Euclidean distance between the prediction by  $cs(x) - c_{i}$  and by  $cs(x)$ . The greater the change in the prediction, the greater the influence of this commonsense.
+
+John promised Bill to leave, so an hour later [John] left.  
+
+<table><tr><td>PersonX promises PersonY.</td></tr><tr><td>1.···As a result, PersonX wants to fulfill his promise.</td></tr><tr><td>2.···PersonX is seen as truthful</td></tr><tr><td>3.···PersonX is seen as trustworthy.</td></tr><tr><td>4.···Before, PersonX needed to talk to PersonY.</td></tr><tr><td>5.···Before, PersonX needed to go to PersonY&#x27;s house.</td></tr></table>
+
+Table 6: A case study of top 5 commonsense descriptions.
+
+Through the case studies of the samples in WSC273, we found that although commonsense with higher influence is somewhat interpretable for people, the interpretability is not significant. We show some examples in Table 6. We believe that this is because some commonsense for people has been learned in pre-training. Therefore, the out-of-domain commonsense that these pre-trained language models need to incorporate for downstream tasks is inconsistent with human understanding.
+
+# 7 Conclusion
+
+In this paper, we study how to use commonsense to enhance the general text representation. We first analyzed the challenges brought by the domain discrepancy of commonsense. Then, we propose OK-
+
+Transformer to allow commonsense integration and enhancement. In the experiments, we verified the effectiveness of our proposed models in a variety of scenarios, including commonsense reasoning, general text classification, and low-resource commonsense. Our models consistently outperform the baselines. We have also empirically analyzed other properties (e.g. interpretability) of the model.
+
+# Acknowledgments and Disclosure of Funding
+
+We thank Wenting Ba for her valuable plotting assistance. This paper was supported by National Natural Science Foundation of China (No. 61906116), by Shanghai Sailing Program (No. 19YF1414700).
+
+# References
+
+Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. 2020. Piqa: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 7432-7439.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In *NAACL*, pages 4171–4186.  
+Ali Emami, Adam Trischler, Kaheer Suleman, and Jackie Chi Kit Cheung. 2018. A generalized knowledge hunting framework for the winograd schema challenge. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Student Research Workshop, pages 25-31.  
+Jian Guan, Fei Huang, Zhihao Zhao, Xiaoyan Zhu, and Minlie Huang. 2020. A knowledge-enhanced pretraining model for commonsense story generation. Transactions of the Association for Computational Linguistics, 8:93-108.  
+Suchin Gururangan, Ana Marasovic, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A Smith. 2020. Don't stop pretraining: Adapt language models to domains and tasks. arXiv preprint arXiv:2004.10964.  
+Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. Realm: Retrievalaugmented language model pre-training. arXiv preprint arXiv:2002.08909.  
+Pengcheng He, Xiaodong Liu, Weizhu Chen, and Jianfeng Gao. 2019. A hybrid neural network model for commonsense reasoning. arXiv preprint arXiv:1907.11983.
+
+Jena D Hwang, Chandra Bhagavatula, Ronan Le Bras, Jeff Da, Keisuke Sakaguchi, Antoine Bosselut, and Yejin Choi. 2020. Comet-atomic 2020: On symbolic and neural commonsense knowledge graphs. arXiv preprint arXiv:2010.05953.  
+Forrest Iandola, Albert Shaw, Ravi Krishna, and Kurt Keutzer. 2020. Squeezebert: What can computer vision teach nlp about efficient neural networks? In Proceedings of SustainNLP: Workshop on Simple and Efficient Natural Language Processing, pages 124-135.  
+Tassilo Klein and Moin Nabi. 2019. Attention is (not) all you need for commonsense reasoning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4831-4836.  
+Tassilo Klein and Moin Nabi. 2020. Contrastive self-supervised learning for commonsense reasoning. arXiv preprint arXiv:2005.00669.  
+Vid Kocijan, Oana-Maria Camburu, Ana-Maria Cretu, Yordan Yordanov, Phil Blunsom, and Thomas Lukasiewicz. 2019a. Wikicrem: A large unsupervised corpus for coreference resolution. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4294-4303.  
+Vid Kocijan, Ana-Maria Cretu, Oana-Maria Camburu, Yordan Yordanov, and Thomas Lukasiewicz. 2019b. A surprisingly robust trick for winograd schema challenge. In The 57th Annual Meeting of the Association for Computational Linguistics (ACL).  
+Hector Levesque, Ernest Davis, and Leora Morgenstern. 2012. The winograd schema challenge. In Thirteenth International Conference on the Principles of Knowledge Representation and Reasoning.  
+Quan Liu, Hui Jiang, Zhen-Hua Ling, Xiaodan Zhu, Si Wei, and Yu Hu. 2016. Commonsense knowledge enhanced embeddings for solving pronoun disambiguation problems in winograd schema challenge. arXiv preprint arXiv:1611.04146.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.  
+Leora Morgenstern, Ernest Davis, and Charles L Ortiz. 2016. Planning, executing, and evaluating the winograd schema challenge. AI Magazine, 37(1):50-54.  
+Shikhar Murty, Pang Wei Koh, and Percy Liang. 2020. Expert: Representation engineering with natural language explanations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2106-2113.
+
+Matthew E Peters, Mark Neumann, Robert Logan, Roy Schwartz, Vidur Joshi, Sameer Singh, and Noah A Smith. 2019. Knowledge enhanced contextual word representations. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 43-54.  
+Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners.  
+Evani Radiya-Dixit and Xin Wang. 2020. How fine can fine-tuning be? learning efficient language models. In International Conference on Artificial Intelligence and Statistics, pages 2435-2443. PMLR.  
+Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. 2018. Gender bias in coreference resolution. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 8-14.  
+Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2019. Winogrande: An adversarial winograd schema challenge at scale. arXiv preprint arXiv:1907.10641.  
+Chi Sun, Xipeng Qiu, Yige Xu, and Xuanjing Huang. 2019. How to fine-tune bert for text classification? In Proceedings of China National Conference on Computational Linguistics, pages 194-206.  
+Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2019. *Commonsenseqa: A question answering challenge targeting commonsense knowledge.* In *Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies*, Volume 1 (Long and Short Papers), pages 4149-4158.  
+Trieu H Trinh and Quoc V Le. 2018. A simple method for commonsense reasoning. arXiv preprint arXiv:1806.02847.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 6000-6010.  
+Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. Glue: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355.
+
+Shuohang Wang, Yuwei Fang, Siqi Sun, Zhe Gan, Yu Cheng, Jingjing Liu, and Jing Jiang. 2020. Cross-thought for sentence encoder pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 412-421.  
+Shuohang Wang, Sheng Zhang, Yelong Shen, Xiaodong Liu, Jingjing Liu, Jianfeng Gao, and Jing Jiang. 2019. Unsupervised deep structured semantic models for commonsense reasoning. arXiv preprint arXiv:1904.01938.  
+Thomas Wolf, Julien Chaumont, Lysandre Debut, Victor Sanh, Clement Delangue, Anthony Moi, Pierric Cistac, Morgan Funtowicz, Joe Davison, Sam Shleifer, et al. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45.  
+Zhiyong Wu, Yun Chen, Ben Kao, and Qun Liu. 2020. Perturbed masking: Parameter-free probing for analyzing and interpreting bert. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4166-4176.  
+Wenhan Xiong, Jingfei Du, William Yang Wang, and Veselin Stoyanov. 2019. Pretrained encyclopedia: Weakly supervised knowledge-pretrained language model. In International Conference on Learning Representations.  
+Zhengyan Zhang, Xu Han, Zhiyuan Liu, Xin Jiang, Maosong Sun, and Qun Liu. 2019. Ernie: Enhanced language representation with informative entities. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1441-1451.  
+Hao Zhou, Tom Young, Minlie Huang, Haizhou Zhao, Jingfang Xu, and Xiaoyan Zhu. 2018. Commonsense knowledge aware conversation generation with graph attention. In *IJCAI*, pages 4623-4629.
+
+# A Experimentation Details
+
+When continuing pre-training BERTcontinue/RoBERTa-continue in Table 5, we follow (Kocijan et al., 2019b) and set learning rate to  $1e - 5$ , batch size to 64, and train the model for only one epoch.
+
+When fine-tuning the models in Sec 6.2 and Sec 6.3, we train the models for 10 epochs. We use grid search to select their learning rates and batch sizes from  $\{1e - 5,2e - 5,5e - 5\}$  and  $\{8,16,32,64\}$ , respectively.
+
+<table><tr><td>Dataset</td><td>WSC</td><td>PDP</td><td>WinoGender</td><td>WinoGrande</td></tr><tr><td>Dataset size</td><td>273</td><td>60</td><td>720</td><td>40938/1267</td></tr><tr><td>Matched ratio</td><td>67%</td><td>83%</td><td>65%</td><td>71%</td></tr><tr><td>Average |cs(x)|</td><td>129.71</td><td>189.68</td><td>80.63</td><td>140.56</td></tr><tr><td>Average length of c</td><td>17.88</td><td>17.91</td><td>16.83</td><td>17.91</td></tr></table>
+
+Table 7: Statistical results on commonsense reasoning datasets.  
+
+<table><tr><td>Dataset</td><td>MRPC</td><td>CoLA</td><td>RTE</td><td>QNLI</td><td>STS-B</td><td>SST-2</td></tr><tr><td>Dataset size</td><td>3668/408</td><td>8551/1043</td><td>2490/277</td><td>104743/5463</td><td>5749/1500</td><td>67349/872</td></tr><tr><td>Matched ratio</td><td>59%</td><td>40%</td><td>72%</td><td>52%</td><td>56%</td><td>25%</td></tr><tr><td>Average |cs(x)|</td><td>80.71</td><td>84.85</td><td>122.60</td><td>81.35</td><td>117.00</td><td>83.07</td></tr><tr><td>Average length of c</td><td>17.47</td><td>17.60</td><td>17.71</td><td>17.59</td><td>17.34</td><td>17.59</td></tr></table>
+
+Table 8: Statistical results on sentence classification datasets.
+
+# B Statistics of Commonsense Descriptions
+
+In Table 7 and Table 8, we report statistics about down-stream tasks and their commonsense descriptions. Our report includes the size of the train/test splits for the downstream tasks, the proportion of samples that matched to at least one commonsense description (Matched proportion) in each task, the average number of matched commonsense descriptions per sample (Average  $|cs(x)|$ ), and the average length of each matched commonsense description (Average length of  $c$ ).
+
+From the results, we found that more than half of the samples matched to at least one commonsense description in most of the datasets. This indicates that the OOD commonsense used in this paper is generalizable to different datasets. Also, the average length of the matched commonsense descriptions is short (about 17), thus encoding them via Transformer is efficient.

enhancingnaturallanguagerepresentationwithlargescaleoutofdomaincommonsense/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:66528a962832893796dd6c7e044080d1a91069322d548e5ec9c267b7294a65c2
+size 515259

enhancingnaturallanguagerepresentationwithlargescaleoutofdomaincommonsense/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bbc158012a60151747ce3c05097ef25af453c93b94ae33e941e1bc669f6f2d41
+size 392673

entropybasedattentionregularizationfreesunintendedbiasmitigationfromlists/661d542b-40b1-43e3-bac7-49df4d2ed7b8_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c6b916f055cbcd281d02d42bfb58e64e9fa04d337aaa0bb89b9f27f77b1a269
+size 98425

entropybasedattentionregularizationfreesunintendedbiasmitigationfromlists/661d542b-40b1-43e3-bac7-49df4d2ed7b8_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:15c5bd768348000efccd4d4939ee152f250cc270cc3f27792ef47f6ffd172ed3
+size 119568

entropybasedattentionregularizationfreesunintendedbiasmitigationfromlists/661d542b-40b1-43e3-bac7-49df4d2ed7b8_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd0ccd9c015357229940b5e08b40a9d4d020d3f5574d3115bea964e1aba1828a
+size 586020

entropybasedattentionregularizationfreesunintendedbiasmitigationfromlists/full.md

@@ -0,0 +1,380 @@
+# Entropy-based Attention Regularization Frees Unintended Bias Mitigation from Lists
+
+Giuseppe Attanasio $^{1,2}$ , Debora Nozza $^{1}$ , Dirk Hovy $^{1}$ , Elena Baralis $^{2}$
+
+<sup>1</sup>Bocconi University, Milan, Italy
+
+$^{2}$ Politecnico di Torino, Turin, Italy
+
+{giuseppe.attenasio3,debora.nozza,dirk.hovy}@unibocconi.it,
+
+elena.baralis@polito.it
+
+# Abstract
+
+Warning: This paper contains examples of language that some people may find offensive.
+
+Natural Language Processing (NLP) models risk overfitting to specific terms in the training data, thereby reducing their performance, fairness, and generalizability. E.g., neural hate speech detection models are strongly influenced by identity terms like gay, or women, resulting in false positives, severe unintended bias, and lower performance. Most mitigation techniques use lists of identity terms or samples from the target domain during training. However, this approach requires a-priori knowledge and introduces further bias if important terms are neglected. Instead, we propose a knowledge-free Entropy-based Attention Regularization (EAR) to discourage overfitting to training-specific terms. An additional objective function penalizes tokens with low self-attention entropy. We fine-tune BERT via EAR: the resulting model matches or exceeds state-of-the-art performance for hate speech classification and bias metrics on three benchmark corpora in English and Italian. EAR also reveals overfitting terms, i.e., terms most likely to induce bias, to help identify their effect on the model, task, and predictions.
+
+# 1 Introduction
+
+Online hate speech is growing at a rapid pace, with effects that can result in dangerous criminal acts offline. Due to its verbal nature, various Natural Language Processing approaches have been proposed (Qian et al., 2018; Indurthi et al., 2019; Atanasio and Pastor, 2020; Kennedy et al., 2020; Vidgen et al., 2021, inter alia). Recently, detection performance has significantly improved with the use of large pre-trained language models based on Transformers (Vaswani et al., 2017), such as Bidirectional Encoder Representations from Transformers (BERT) (Devlin et al., 2019). However,
+
+![](images/8a61bd4038890f12dcc45ff7b6ec5dbff8b48075783fbf2eff3f3c0f3d468802.jpg)  
+Figure 1: False positive from BERT as a hate speech detector. The darker and taller the bar, the higher the overfitting on the term.
+
+several works have shown that by fine-tuning neural language models on hate speech detection, the classifiers obtained contain severe unintended bias (Dixon et al., 2018), i.e. they perform better or worse when texts mention specific identity terms (such as gay, Muslim, or woman). As a result, a sentence like "As a Muslim woman, I agree" would be wrongly classified as hate speech, purely due to the presence of two identity terms, i.e., terms referring to specific groups based on their socio-demographic features. One cause of false positives is selection bias in the keyword-driven collection of corpora (Ousidhoum et al., 2020). Figure 1 shows a false positive example for a fine-tuned BERT model on hate speech detection. Ideally, the model should rely on the words adore and you. Instead, BERT overfitted to the word Girl and associated it with a hateful context. This unwanted effect demonstrates the issues of lexical overfitting, and how they cause unintended bias on identity terms.
+
+Various methods have been proposed to mitigate and measure (unintended) bias (Elazar and Goldberg, 2018; Park et al., 2018; Dixon et al., 2018; Nozza et al., 2019; Kennedy et al., 2020; Vaidya et al., 2020). However, all those methods rely on the availability of a set of identity terms. This is a severe limitation, which hinders the generalizability and applicability of hate detection models
+
+to real-world contexts. For example, a model designed to reduce the unintended bias on gender-related terms (such as woman, wife) will not address unintended bias for religious affiliation. So practitioners must decide a-priori "which vulnerable groups are present in our data?"
+
+We propose an Entropy-based Attention Regularization (EAR) that forces the model to build token representations by attending to a wider context, i.e., consider a larger number of tokens from the rest of the sentence. We measure the attended context as the entropy of the self-attention weight distribution over the input sequence. We use EAR as a regularization term in the loss computation to maximize each token's entropy. We apply EAR to BERT. The resulting model (BERT+EAR) significantly improves performance on unintended bias mitigation in English and Italian. In addition, it requires no a-priori knowledge (e.g., sets of identity terms), making it fairer and more general. The contextualized representations EAR induces avoid basing the classification on individual terms and, ultimately, mitigate lexical overfitting and intrinsic bias from pre-trained weights.
+
+As a training by-product, EAR lets us extract the overfitting terms, i.e., terms accounting for narrower context that most likely induce unintended bias. These terms can highlight possible weaknesses in the model: from the over-sensitivity of pre-trained weights to specific words (Sheng et al., 2019; Nangia et al., 2020; Vig et al., 2020), to overspecialization of training corpora on the keywords used for collecting data (Ousidhoum et al., 2020).
+
+Note that while we show results on BERT, EAR is applicable to any attention-based architecture.
+
+Contributions. EAR is a novel entropy-based attention regularization method to mitigate unintended bias by reducing lexical overfitting. It is applied to all terms, so it does not need a-priori domain knowledge (e.g., predefined term lists). Independent of domain-specific information, EAR generalizes better to different languages and contexts compared to similar approaches. Attention entropy is used to extract a list of the most likely biased terms. EAR code is available at https://github.com/g8a9/ear.
+
+# 2 Entropy-based Attention Regularization
+
+Attention was originally designed for aligning target and source sequences in machine translation
+
+![](images/a1fa958aadb8f0f836afe004acb20966c8c8bbb2620e22ff923508be7d0cf139.jpg)  
+Figure 2: Self-attention distribution on tokens Girl (solid orange) and you (shaded blue). Attention for Girl is concentrated on its representation: its entropy is low. Attention for you is spread: its entropy is high.
+
+(Graves, 2013; Bahdanau et al., 2015). However, in the Transformer architecture (Vaswani et al., 2017), it has become a means to account for lexical influence and long-range dependencies. It also provides useful information about the importance of a term for the output (Wiegreffe and Pinter, 2019; Brunner et al., 2020; Sun and Marasovic, 2021). Here, we use the notion of attention entropy, and EAR's use of it in BERT. Note, though, that EAR can be used with any attention-based architecture.
+
+Self-attention in Transformers. The Transformer model consists of two connected units, an encoder and a decoder, designed for sequence-to-sequence tasks.
+
+A transformer encoder applies scaled-dot product self-attention over the input tokens to compute  $N$  independent attention heads. Let  $E = [e_0, \dots, e_{d_s}]$  be the sequence of input embeddings, with  $e_i \in \mathbb{R}^{d_m}$ . For the  $h$ -th attention head and  $i$ -th position, each embedding  $e_i$  is projected into a query  $q_{h,i}$ , a key  $k_{h,i}$  and value  $v_{h,i}$ . So each token expresses an attention distribution over all input embeddings as
+
+$$
+a _ {h, i} = \operatorname {s o f t m a x} \left(\frac {\mathbf {q} _ {h , i} ^ {T} K _ {h}}{\sqrt {d _ {k}}}\right) \tag {1}
+$$
+
+where  $K_{h}$  is the matrix of keys and  $d_{k}$  their dimension.
+
+Attention weights  $a_{h,i} = [a_{h,i,0},\dots,a_{h,i,d_s}]$  where  $a_{h,i,j}\in [0,1]$  and  $\sum_{j}a_{h,i,j} = 1$ , can be seen as a soft-indexing over the values. Since the values are projections of the tokens themselves, each weight in self-attention measures the contribution of its token to the attention head and, in turn, to the new token representation. We provide additional details to the self-attention mechanism in Appendix A.
+
+Attention entropy. Information entropy was first introduced in Shannon (1948), and measures the average information content of a random variable  $X$  with the set  $[x_0, \dots, x_n]$  of possible outcomes. It is defined as
+
+$$
+H (X) = - \sum_ {i} P \left(x _ {i}\right) \log P \left(x _ {i}\right) \tag {2}
+$$
+
+Following Ghader and Monz (2017), we compute the entropy in the self-attention heads by interpreting each token's attention distribution as a probability mass function of a discrete random variable. The input embeddings are the possible outcomes, and the attention weights their probability.
+
+For the sake of simplicity, we now discuss the computation of attention entropy of a single token in a standard transformer encoder. Attention weights are first averaged over heads by defining  $a_{i,j}^{\prime} = \frac{1}{h}\sum_{h}a_{h,i,j}$  as the mean attention that the token at position  $i$  pays to the token at position  $j$ . Then, we define a probability mass function by applying a softmax operator:
+
+$$
+a _ {i, j} = \frac {e ^ {a _ {i , j} ^ {\prime}}}{\sum_ {j} e ^ {a _ {i , j} ^ {\prime}}} \tag {3}
+$$
+
+We define the attention entropy as follows
+
+$$
+H _ {i} = - \sum_ {j = 0} ^ {d _ {s}} a _ {i, j} \log a _ {i, j} \tag {4}
+$$
+
+Intuitively, attention entropy measures the degree of contextualization while constructing the model's upper level's embedding. A large entropy suggests that a wider context contributes to the new embedding, while a small entropy tells the opposite: only a few tokens are deemed relevant. From a broader viewpoint, contextualized tokens improve the information passage between continuous layers by re-distributing the information content for every unit involved.
+
+Figure 2 shows a toy example of self-attention distributions for two arbitrary tokens. Solid orange bars correspond to  $a_{\mathrm{Girl},j}$ , while shaded blue bars correspond to  $a_{\mathrm{you},j}$ . The toy example illustrates the correlation between attention distributions and entropy. The representation of you uses a wider context and, thus, it has a higher attention entropy. Note that, if present, we discard padding tokens from the attention entropy computation. Conversely, we include special tokens when required by the downstream task.
+
+![](images/7fd7db2e754567e0302acb893a5e7e60420c67e3095282cf09498a7881e39f98.jpg)  
+Figure 3: Overview of BERT+EAR. Grey boxes are Transformer layers. Each builds a token with attention entropy  $H_{i}^{\ell}$ . Right green box pools layer-wise contextualization contributions and outputs regularization loss. First layer self-attention distribution (bottom) shown for you (shaded blue) and Girl (solid orange).
+
+EAR in BERT. We introduced attention entropy as a proxy for the degree of contextualization of token representations above. Following this intuition, we propose BERT with EAR mitigation (BERT+EAR), a novel model trained to learn tokens with maximal self-attention entropy over the input sequence. We fine-tune BERT+EAR in the downstream task of hate speech detection. Note, though, that the approach is feasible for any classification task. In classification models, having more contextualized tokens avoids individual terms driving the classification outcome because they got over-attentioned.
+
+Although EAR is applicable to any Transformer-based model, we base our approach here on the BERT (Devlin et al., 2019) base architecture. BERT provides an informative case study, given the number of architectures it has spawned and the recent interest in its attention patterns (Clark et al., 2019b; Kovaleva et al., 2019; Serrano and Smith, 2019). BERT consists of twelve stacked transformer encoders, each running self-attention on the output of the previous encoder. In BERT+EAR, we build new tokens with the maximal information content coming from the previous layer for every transformer layer in the architecture. Using Equation 4, we first compute the attention entropy of each token in the input sentence. We then take their
+
+mean and define the average contextualization for the  $\ell$ -th layer as
+
+$$
+H ^ {\ell} = \frac {1}{d _ {s}} \sum_ {i = 0} ^ {d _ {s}} H _ {i} ^ {\ell} \tag {5}
+$$
+
+where  $H_{i}^{\ell}$  is the attention entropy of the token at position  $i$ , and  $d_{s}$  is the length of the input sequence (excluding the padding tokens but including the [CLS] and [SEP] special tokens). Finally, we introduce a new regularization term to the model loss to maximize the entropy at each layer:
+
+$$
+\mathcal {L} = \mathcal {L} _ {C} + \mathcal {L} _ {R}, \quad \mathcal {L} _ {R} = - \alpha \sum_ {l} H ^ {\ell} \tag {6}
+$$
+
+$\mathcal{L}$  is the total loss,  $\mathcal{L}_C$  and  $\mathcal{L}_R$  are the classification and regularization loss, respectively, and  $\alpha \in \mathbb{R}$  is the regularization strength. As in previous work,  $\mathcal{L}_C$  is the Cross Entropy loss obtained with a linear layer on top of the last encoder as a classification head. It receives the [CLS] embedding and outputs the probability of the positive class (Hate).
+
+The new regularization term  $\mathcal{L}_R$  frames the task of maximal contextualization learning in the network. This framing has several advantages over existing approaches. First, it is a sum of differentiable terms and is hence differentiable. We can thus optimize BERT+EAR with classical back-propagation updates. Second, the regularization is agnostic to specific identity terms. It instead induces the network to learn contextualized tokens globally. This induction is crucial to regularize biased terms that might not be known in advance. Finally, note that the  $\mathcal{L}_R$  pools each layer's entropy-based contributions  $H^{\ell}$ . Each term  $H^{\ell}$  is in turn dependent on the sole attention entropy defined in Equation 4. This makes the setup a general framework not limited to BERT.  $\mathcal{L}_R$  can be used to evaluate and maximize the token contextualization in any attention-based architecture.
+
+Figure 3 shows a graphical overview of BERT+EAR. Each layer provides a contextualization contributing to the loss independently, where layers with a low average contextualization increase the loss the most. Note also that, similarly to He et al. (2016),  $\mathcal{L}_R$  introduces skip connections between layers and the classification head, so shorter paths for the contextualization information to flow.
+
+Insights from attention entropy. On the one hand, we use attention entropy maximization to
+
+train BERT+EAR and test its classification and bias mitigation performance. On the other hand, we can leverage attention entropy to automatically extract the tokens with the lowest contextualization, which are the most likely to induce unintended bias. When a sentence is fed through a model like BERT, we can inspect the attention distribution of its terms<sup>2</sup>.
+
+We propose to exploit entropy, and hence contextualization, to gain insights into any attention-based model. Given a corpus and a model we want to inspect, we repeatedly query the model with sentences from the corpus and collect each token's attention entropy. Finally, we take each token's mean to measure the impact it has on bias, where lower is worse. Note that the same term can impact bias differently depending on the sentence.
+
+While our approach works for any attention-based model and data set, we test it on fine-tuned classifiers to extract the biased terms learned on the training data set. We discuss this functionality in Section 5.
+
+# 3 Experimental settings
+
+In this work, we consider the problem of unintended bias (Dixon et al., 2018): "a model contains unintended bias if it performs better for comments containing some particular identity terms than for comments containing others".
+
+Datasets. Unintended bias is measured on synthetic test sets, artificially generated by filling manually defined contexts with identity terms (e.g., I hate all _, I love all ). By construction, each identity term appears  $50\%$  of the time in hateful contexts and  $50\%$  in non-hateful ones. If a model then classifies the instances related to one identity term differently than the others, it means that the model contains unintended bias towards that term, e.g., if every instance containing the term women is labelled hateful, independently of the context. Synthetic test sets simulate new data, so a model that has low performance on this set demonstrates low generalization abilities and incapacity to be used in real-world contexts and applications.
+
+We test BERT+EAR on hate speech datasets with associated synthetic test sets to measure unintended bias.
+
+MISOGNY (EN) (Fersini et al., 2018) is a state-of-the-art corpus for misogyny detection in English.
+
+<table><tr><td></td><td>MISOGINY (EN)</td><td>MISOGINY (IT)</td><td>MLMA</td></tr><tr><td># Train</td><td>4,000</td><td>5,000</td><td>5082</td></tr><tr><td># Test</td><td>1,000</td><td>1,000</td><td>565</td></tr><tr><td>% Validation</td><td>10</td><td>10</td><td>10</td></tr><tr><td>% Hate (train, test)</td><td>45, 46</td><td>47, 53</td><td>88, 88</td></tr><tr><td>\(B_{2}\)</td><td>0.858</td><td>0.852</td><td>0.881</td></tr><tr><td># Synthetic</td><td>1,464</td><td>1,908</td><td>77,000</td></tr><tr><td># Identity terms</td><td>12</td><td>18</td><td>50</td></tr><tr><td>% Hate (Synthetic)</td><td>50</td><td>50</td><td>50</td></tr></table>
+
+Table 1: Statistics of the data sets.
+
+The related synthetic test set (Nozza et al., 2019) was created via several manually defined templates and synonyms for "woman" as identity terms.
+
+MISOGNY (ITA) (Fersini et al., 2020) is the benchmark corpus for misogyny detection in Italian. The synthetic test set has been generated similarly to the English one. This dataset allows us to study EAR's impact on cross-lingual adaptation.
+
+MULTILINGUAL AND MULTI-ASPECT HATE SPEECH (MLMA) (Ousidhoum et al., 2019) consists of tweets with various hate speech targets. We choose to work on its English part. We use the synthetic test provided in Dixon et al. (2018), generated by slotting a wide range of identity terms into manually defined templates.
+
+Table 1 reports statistics of the data sets. Alongside the size of train, test, and validation sets, we report also the percentage of hateful instances to show the class balance. Note that MLMA is highly unbalanced with  $88\%$  of instances associated with the hateful class. Note that the original MULTILINGUAL AND MULTI-ASPECT dataset comes in a multi-label, multiple class setting. Following Ousidhoum et al. (2021), we used the Hostility dimension of the dataset as target label and created a Hate binary from it as follows. We considered single-labeled "Normal" instances to be non-hate/non-toxic and all the other instances to be toxic.
+
+To further characterize our data sets, we explore the aspect of selection bias, reporting the measure  $B_{2}$  (Ousidhoum et al., 2020). The metric ranges from 0 to 1 and evaluates how likely topics of the data set are to contain keywords of the data collection. Values above 0.7 demonstrate high selection bias, implying the need for unbiaseding procedures.
+
+We report also the size and number of identity terms used in the synthetic test sets. The percentage of hateful content is perfectly balanced (50%) since each identity term should appear exactly in
+
+the same context as the others to measure the unintended bias. See Appendix B for the list of identity terms and further preprocessing details.
+
+# 3.1 Metrics
+
+We use the weighted and binary F1-score of the hateful class  $(\mathbf{F1}_{\mathbf{w}}$  and  $\mathbf{F1}_{\mathbf{hate}})$  as classification metrics. We consider both due to the class imbalance of test sets (see Table 1).
+
+We compute the unintended bias metrics from Dixon et al. (2018) and Borkan et al. (2019). They are computed from differences in the score distributions between instances mentioning a specific identity-term (subgroup distribution) and the rest (background distribution). The three per-term AUC-based bias scores are:
+
+1)  $AUC_{subgroup}$  calculates AUC only on the data subset of a given identity term. A low value means the model performs poorly in distinguishing between hateful and non-hateful comments that mention the identity term.  
+2) Background Positive Subgroup Negative  $(AUC_{bpsn})$  calculates AUC on the hateful background examples and the non-hateful subgroup examples. A low value means that the model confuses non-hateful examples that mention the identity term with hateful examples that do not.  
+3) Background Negative Subgroup Positive  $(AUC_{bnp})$  calculates AUC on the non-hateful background examples and the hateful subgroup examples. A low value means that the model confuses hateful examples that mention the identity with non-hateful examples that do not.
+
+We report the averaged metrics across identity terms, i.e.,  $\mathbf{AUC}_{\mathrm{subgroup}}$ ,  $\mathbf{AUC}_{\mathrm{bpsn}}$ , and  $\mathbf{AUC}_{\mathrm{bnsp}}$ .<sup>3</sup>
+
+# 3.2 Baselines
+
+We compare BERT+EAR against the following existing approaches: (1) BERT (Devlin et al., 2019), (2) BERT+SOC mitigation (Kennedy et al., 2020), where the authors modify BERT's loss to lower the importance weight of identity terms, computed with the Sampling-and-Occlusion (SOC) algorithm (Jin et al., 2019), (3) Nozza et al. (2019), a single-layer neural network architecture based on the Universal Sentence Encoder (USE) representation (Cer et al., 2018), (4) Lees et al. (2020), a multilingual BERT model fine-tuned on the training data, (5) Ousidhoum et al. (2021), a classifier based on TF-
+
+<table><tr><td rowspan="2"></td><td colspan="5">Unintended bias (synthetic)</td><td colspan="2">test</td></tr><tr><td>AUCsubgroup</td><td>AUCbnp</td><td>AUCbpsn</td><td>F1w</td><td>F1hate</td><td>F1w</td><td>F1hate</td></tr><tr><td>Nozza et al. (2019), no mitigation</td><td>49.83</td><td>49.83</td><td>49.83</td><td>49.97</td><td>51.33</td><td>72.29</td><td>71.62</td></tr><tr><td>Nozza et al. (2019), debiased</td><td>50.27</td><td>50.21</td><td>50.21</td><td>45.40</td><td>29.31</td><td>71.43</td><td>69.37</td></tr><tr><td>Zhang et al. (2020)</td><td>69.99</td><td>62.19</td><td>62.19</td><td>43.01</td><td>66.70</td><td>31.35</td><td>63.21</td></tr><tr><td>BERT, no mitigation</td><td>70.97</td><td>66.62</td><td>66.62</td><td>58.19</td><td>64.61</td><td>69.60</td><td>70.21</td></tr><tr><td>BERT+SOC mitigation</td><td>78.11</td><td>76.60</td><td>76.60</td><td>51.88</td><td>58.89</td><td>57.39</td><td>60.47</td></tr><tr><td>BERT+SOC mitigation, missing ITs</td><td>68.58</td><td>67.38</td><td>67.38</td><td>38.49</td><td>41.38</td><td>51.14</td><td>43.65</td></tr><tr><td>BERT+EAR</td><td>80.08</td><td>75.18</td><td>75.18</td><td>62.59▲</td><td>70.58▲</td><td>70.90▲</td><td>70.83▲</td></tr><tr><td>Lees et al. (2020), debiased</td><td>-</td><td>-</td><td>-</td><td>47.00</td><td>58.58</td><td>79.87</td><td>82.45</td></tr><tr><td>Zhang et al. (2020)</td><td>48.10</td><td>48.29</td><td>48.29</td><td>33.33</td><td>66.66</td><td>33.54</td><td>66.69</td></tr><tr><td>BERT, no mitigation</td><td>47.30</td><td>47.54</td><td>47.54</td><td>39.72</td><td>61.17</td><td>81.57</td><td>83.56</td></tr><tr><td>BERT+SOC mitigation, translated ITs</td><td>45.54</td><td>45.88</td><td>45.88</td><td>46.34</td><td>51.62</td><td>80.28</td><td>81.73</td></tr><tr><td>BERT+EAR</td><td>48.59</td><td>48.65</td><td>48.65</td><td>40.64</td><td>62.71▲</td><td>83.29▲</td><td>84.68▲</td></tr><tr><td>Ousidhoum et al. (2021), no mitigation</td><td>63.87</td><td>60.80</td><td>61.10</td><td>33.33</td><td>66.66</td><td>82.84</td><td>93.80</td></tr><tr><td>Zhang et al. (2020)</td><td>74.14</td><td>64.74</td><td>65.76</td><td>33.33</td><td>66.66</td><td>82.84</td><td>93.79</td></tr><tr><td>BERT, no mitigation</td><td>69.38</td><td>67.12</td><td>67.12</td><td>50.24</td><td>39.65</td><td>64.70</td><td>70.14</td></tr><tr><td>BERT+SOC mitigation</td><td>56.15</td><td>55.83</td><td>55.58</td><td>33.79</td><td>59.89</td><td>76.49</td><td>86.24</td></tr><tr><td>BERT+EAR</td><td>74.31</td><td>71.43</td><td>71.25</td><td>40.09</td><td>67.45▲</td><td>83.05▲</td><td>91.88▲</td></tr></table>
+
+Table 2: Results (in %) on MISOGYNY (EN) (top), MISOGYNY (ITA) (middle), and MLMA. Significance of BERT+EAR over BERT without mitigation ( $\bullet$ :  $p \leq 0.01$ ) and BERT with SOC mitigation ( $\triangle$ :  $p \leq 0.01$ ).
+
+IDF and Logistic Regression, and (6) Zhang et al. (2020), a debiasing training framework based on instance weighting.
+
+The debiased version proposed in Lees et al. (2020) is obtained by training the model on additional samples from Wikipedia articles (assumed to be non-hateful) to balance the distribution of specific identity terms. Nozza et al. (2019) extracted these additional non-hateful samples from an external Twitter corpus (Waseem and Hovy, 2016).
+
+To address the impact of different term lists, we also consider two different versions of BERT+SOC mitigation, one where we test the effect of missing identity terms and the other where the identity terms are translated for adapting to a new language.
+
+# 4 Experimental Results
+
+Table 2 shows classification and bias metrics on both synthetic and test set for the three corpora, i.e., MISOGNY (EN) (top), MISOGNY (ITA) (middle), and MLMA (bottom). The top rows in each table section report the performance of hate speech detection models specifically proposed for the respective dataset. The lower rows show the results of baselines and BERT+EAR. BERT+SOC mitigation uses the identity terms from Kennedy et al. (2020) (see Appendix C), unless a different identity terms lists is specified (e.g., "BERT+SOC mitigation, translated ITs").
+
+BERT+EAR obtains comparable and, in most
+
+cases, better performance on all three datasets than all state-of-the-art debiasing approaches, which are based on (i) the knowledge of identity terms and (ii) data augmentation techniques. However, identity terms are not always readily available, which severely limits the generalization of those approaches. Similarly, there are several drawbacks to data augmentation with (assumed) non-hateful samples containing the identity terms. 1) Data augmentation is expensive. It requires filtering a large dataset (usually Wikipedia) and retraining the model with a much larger set of instances. 2) Data augmentation with task-specific identity terms requires prior knowledge of those terms, and is therefore limited by the authors' knowledge. 3) The overlap between identity terms in the evaluation set and the augmented data inevitably (but somewhat unfairly) improves the performance on the synthetic dataset.
+
+BERT+EAR is overall the best debiasing model considering the proposed bias metrics. The only exception is MISOGNY (EN), for which BERT+EAR has lower  $\mathbf{AUC}_{\mathrm{bnp}}$  and  $\mathbf{AUC}_{\mathrm{bpsn}}$  than BERT+SOC mitigation. The latter's advantage, however, comes with high variability in the results. BERT+SOC mitigation seems more sensitive to random initialization. The standard deviation over 10 runs is  $37\%$ , compared to  $13\%$  of BERT+EAR. Figure 4 shows the AUCsubgroup metric separately by identity term on MISOGNY (EN). We compare
+
+![](images/024b2df0bd44550c7c2177a50f63dbcdf52a9335a791781582e368e501cc9b7b.jpg)  
+Figure 4:  $\mathrm{AUC}_{\mathrm{subgroup}}$  results broken down by identity term on MIsOGYNY (EN).
+
+BERT and BERT+EAR over 10 different initialization runs. EAR improves BERT across all identity terms
+
+Most existing models and AUC-based metrics for unintended bias focus only on the false positives (i.e., hateful instances wrongly recognized as non-hateful). While correctly recognizing hateful instances is important, we believe that the problem of false negatives is equally important. Since BERT+EAR does not rely on identity term lists, it regularizes terms that impact both the positive and negative class. BERT+EAR obtains an average decrease of  $15.04\%$  in false negative rate compared to BERT and BERT+SOC mitigation. Indeed, the performance difference between BERT+EAR vs. BERT and BERT+SOC is mainly due to nonhateful instances ( $\sim 95\%$  of the time). Reducing the impact of overfitting terms like  $f^{*}ck$  and  $p^{*}ssy$  in MISOGNY (EN) causes BERT+EAR to consider a larger context, and correctly labels them as non-hateful.
+
+# 4.1 Error Analysis
+
+Table 3 shows tweets from the MISOGNY (EN) data set which have been correctly predicted by BERT+EAR but misclassified by BERT or BERT+SOC. These tweets serve as qualitative examples of the effectiveness of forcing the model to attend to a wider context and not overfit to training-specific terms, exploiting the richness of information (Nozza et al., 2017). The examples are an excerpt of the most common cases where BERT+EAR classifies the non-hateful examples correctly: (1) when slurs or negative words (such as  $sk^{*}nk$ ) are used in a non-hateful context, like slang or lyrics, (2) when many words associated
+
+with misogyny appear in the sentence (e.g., rape, abuse) and (3) when the hateful target is male and the instance should not be classified as misogynous. The use of a wider context by BERT+EAR allows the model identify such non-misogynous instances compared to BERT and BERT+SOC. In particular, BERT+SOC is even more biased in these cases because its debiasing techniques overly rely on specific terms (e.g. woman) and increase overfitting to training-specific examples.
+
+# 4.2 Impact of predefined identity terms
+
+We also analyze the impact of predefined identity term lists on performance by evaluating the effect of (i) missing identity terms, and (ii) adapting to a new language where the list is unavailable.
+
+First, we remove every identity term of BERT+SOC from MISOGNY (EN) that appears at least once in the evaluation set, here women and woman out of 24 terms. This reflects the real-world case where the identity term list does not contain a specific group present in the data. The significant performance drop resulting from this case (Table 2, top, "missing ITs") highlights a strong weakness of term-based mitigation strategies.
+
+Second, we analyze the case where identity terms need to be adapted to a new language, e.g., Italian. We translated the English identity terms from BERT+SOC to Italian via Google Translate.4 Table 2 (middle, "translated ITs") shows that the performance is lower than BERT+EAR. A simple translation of predefined identity terms is therefore not an option for cross-lingual settings. This aligns with the findings by Nozza (2021), that demon
+
+<table><tr><td>text</td><td>Ground truth</td><td>BERT</td><td>BERT+SOC</td><td>BERT+EAR</td></tr><tr><td>I&#x27;m just a sk*nk for understanding the basics of life!</td><td>0</td><td>1</td><td>1</td><td>0</td></tr><tr><td>You&#x27;re such a f*cking hoe, I love it - the new Kanye and Lil Pump I kings make women feel comfortable about their sexuality.</td><td>0</td><td>1</td><td>1</td><td>0</td></tr><tr><td>GIRL, YOU&#x27;RE HYSTERICAL. I AM DANCING SO HAPPY FOR TODAY</td><td>0</td><td>0</td><td>1</td><td>0</td></tr><tr><td>#metoo I&#x27;m a victim of rape, abuse and harassment. Every woman who had any these experiences.</td><td>0</td><td>1</td><td>1</td><td>0</td></tr><tr><td>some people at school drive me insane. like cool b*tch! im depressed too!! doesn&#x27;t mean im a f*cking c*nt</td><td>0</td><td>1</td><td>1</td><td>0</td></tr><tr><td>@male_user And you are a hysterical k*nt.</td><td>0</td><td>0</td><td>1</td><td>0</td></tr><tr><td>@male_user F*ck you p*ssy</td><td>0</td><td>1</td><td>1</td><td>0</td></tr></table>
+
+Table 3: Examples of MISOGNY (EN) tweets misclassified by BERT or BERT+SOC, and correctly classified by BERT+EAR. Next to the tweet text, we report the ground truth label and the prediction of each model. Exact phrasing changed to protect privacy.
+
+strated that cross-lingual hate speech detection is limited by the use of non-hateful, language-specific taboo interjections that are not directly translatable.
+
+In sum, we demonstrated that relying on a predefined list of identity terms is a strong limitation for performance and generalizability of the model. In contrast, BERT+EAR's independence from any predefined terms makes it the ideal model in real-world scenarios.
+
+# 5 Extracting overfitting terms
+
+While being the core of EAR, attention entropy serves another purpose. Once standard fine-tuning is concluded (i.e., with no regularization involved), models have overfitted specific terms. We identify these terms using attention entropy.
+
+To extract the most indicative terms, we replicate training conditions. Specifically, we run inference using all the training data using a fine-tuned checkpoint and a standard BERT tokenizer. We collect attention entropy values for each term and average them over all training instances. Terms with lowest average entropy show the highest overfitting as the model learned them with a narrow context.
+
+Retrieving these terms after training allows us to gain insights into the domain and language-specific aspects driving the outcome.
+
+Table 4 shows the top 10 terms with highest lexical overfitting on the studied datasets extracted from the corresponding fine-tuned model. We extract terms strongly correlated with the positive
+
+class, e.g.,  $\text{womens} * \text{ck}$  (97%),  $\text{shut}$  (96%),  $\text{n} * \text{gger}$  (92%),  $\text{sb} * \text{rro}$  (97%),  $\text{c} * \text{lone}$  (95%). Note that these terms are not frequent in the corpus. Overfitting terms appear with an average document frequency of only 4.7%, while the most frequent terms have 32.5% average document frequency across datasets. These results suggest that the higher the class polarization of a token, the narrower the context BERT will use to learn its representation, and the higher the overfitting.
+
+# 6 Related Work
+
+The first works to study bias measurement and mitigation in neural representation aimed at removing implicit gender bias from word embeddings (Bolukbasi et al., 2016; Caliskan et al., 2017; Garg et al., 2018; Romanov et al., 2019; Ravfogel et al., 2020). More recently, researchers have started to focus on contextualized sentence representations and effective neural models for understanding the presence and resolution of bias (Nozza et al., 2021; Ousidhoum et al., 2021).
+
+While the majority of proposed approaches focus on data augmentation (Dixon et al., 2018; Nozza et al., 2019; Sharma et al., 2020; Bartl et al., 2020; de Vassimon Manela et al., 2021), different approaches have been proposed for bias mitigation intervening directly in the objective function. Kennedy et al. (2020) proposed to apply regularization during training to the explanation-based importance of identity terms, obtained with Sampling-and-Occlusion (SOC) explanations (Jin et al., 2019). Kaneko and Bollegala (2021) pro
+
+<table><tr><td>Dataset</td><td>Overfitting terms</td></tr><tr><td>MISOGNY (EN)</td><td>girls, womens*ck, hoes, c*ck, shut, stupid, hoe, p*ssy, trying, f*ck</td></tr><tr><td rowspan="2">MISOGNY (ITA)</td><td>pezzo, bel, bellissima, scoperei, p*ttanona, zitta, sb*tro, t*ttona, bella, c*lon</td></tr><tr><td>(piece, nice, very nice, I&#x27;d f*ck, sl*t, shut up, c*m, b*sty, beautiful, fat*ss)</td></tr><tr><td>MLMA</td><td>n*gger, n*gro, shut, chong, ching, d*ke, okay, sp*c, tw*t, f*ggot</td></tr></table>
+
+Table 4: Terms with highest lexical overfitting identified using attention entropy.
+
+posed a method for debiasing pre-trained contextual representation by retaining the learned semantic information for gender-related words (e.g., she, woman, he, man) and simultaneously removing any stereotypical biases in the pre-trained model. Zhou et al. (2021) exploited debiasing methods for natural language understanding (Clark et al., 2019a) to explicitly determine how much to trust the bias given the input. Vaidya et al. (2020) proposed a multi-task learning model for predicting the presence of identity terms alongside the toxicity of a sentence.
+
+The main drawback of all aforementioned works is their strict reliance on a set of predefined identity terms. This list can be either defined manually by experts or extracted a-priori from the data set. In both cases, the subsequent debiasing models will be strongly affected by these biased terms, limiting the applicability of the trained model to new data. This is a severe limitation, since it is not always possible to retrain a model on new data to reduce bias, resulting in limited use in real-world cases.
+
+# 7 Conclusion
+
+We introduce EAR, a regularization approach applicable to any attention-based model. Our approach does not require any a-priori knowledge of identity terms, e.g., lists. This feature (i) allows us to generalize to different languages and contexts, and (ii) avoids neglecting important terms. Thus, it prevents the introduction of further bias. As part of the training procedure, EAR also discovers the impact of relevant domain-specific terms. This automatic term extraction provides researchers with an analysis tool to improve data collection and bias mitigation approaches.
+
+EAR, applied to BERT, reliably classifies data with competitive performance and substantially improves various bias metrics. BERT+EAR generalizes better to new domains and languages than similar methods.
+
+In future work, we will apply EAR-based models to different downstream tasks to both improve bias mitigation and automatically extract biased terms.
+
+# Acknowledgments
+
+We would like to thank the anonymous reviewers and area chairs for their suggestion to strengthen the paper. This research is partially supported by funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation program (No. 949944, INTEGRATOR), and by Fondazione Cariplo (grant No. 2020-4288, MONICA). DN, and DH are members of the MilaNLP group, and of the Data and Marketing Insights Unit at the Bocconi Institute for Data Science and Analysis. EB is member of the DataBase and Data Mining Group (DBDMG) at Politecnico di Torino. GA did part of the work as a member of the DBDMG and is currently a member of MilaNLP. Computing resources were partially provided by the SmartData@PoliTO center on Big Data and Data Science.
+
+# Ethical Considerations
+
+In this paper, we propose term attention entropy as a proxy for unintended bias in attention-based architectures. Our approach allows us to extract, for a given classifier and data set, a list of terms that induce most of the bias in the model. While this list is intuitive and easy to obtain, we would like to point out some ethical dual-use considerations.
+
+The process of collecting the list is a data-driven approach, i.e., it is strongly dependent on the task, collected corpus, term frequencies, and the chosen model. Therefore, the list might lack specific terms or include terms that do not strictly perpetrate harm, but are prevalent in the sample. Because of these twin issues, the resulting lists should not be read as complete or absolute. We discourage users from developing new models based solely on the extracted terms. We want, instead, the terms to stand as a starting point for debugging and searching for potential bias issues in the task at hand, be it in data collection or model development.
+
+Further, while the probability is low, we can not exclude the possibility that future users run EAR on other tasks and data sets to derive private information or profile vulnerable groups.
+
+# References
+
+Giuseppe Attanasio and Eliana Pastor. 2020. PoliTeam @ AMI: Improving sentence embedding similarity with misogyny lexicons for automatic misogyny identification in Italian tweets. In Valerio Basile, Danilo Croce, Maria Maro, and Lucia C. Passaro, editors, EVALITA Evaluation of NLP and Speech Tools for Italian - December 17th, 2020, pages 48-54. Academia University Press.  
+Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.  
+Marion Bartl, Malvina Nissim, and Albert Gatt. 2020. Unmasking contextual stereotypes: Measuring and mitigating BERT's gender bias. In Proceedings of the Second Workshop on Gender Bias in Natural Language Processing, pages 1-16, Barcelona, Spain (Online). Association for Computational Linguistics.  
+Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai. 2016. Man is to computer programmer as woman is to homemaker? Debiasing word embeddings. In Proceedings of the 30th International Conference on Neural Information Processing Systems, NIPS'16, page 4356-4364, Red Hook, NY, USA. Curran Associates Inc.  
+Daniel Borkan, Lucas Dixon, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. 2019. Nuanced metrics for measuring unintended bias with real data for text classification. In *Companion Proceedings of The 2019 World Wide Web Conference*, WWW '19, page 491-500, New York, NY, USA. Association for Computing Machinery.  
+Gino Brunner, Yang Liu, Damian Pascual, Oliver Richter, Massimiliano Ciaramita, and Roger Wattenhofer. 2020. On identifiability in transformers. In International Conference on Learning Representations.  
+Aylin Caliskan, Joanna J. Bryson, and Arvind Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334):183-186.  
+Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St. John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, Brian Strope, and Ray Kurzweil. 2018. Universal sentence encoder for English. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 169-174, Brussels, Belgium. Association for Computational Linguistics.  
+Christopher Clark, Mark Yatskar, and Luke Zettlemoyer. 2019a. Don't take the easy way out: Ensemble based methods for avoiding known dataset biases. In Proceedings of the 2019 Conference on Empirical
+
+Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4069-4082, Hong Kong, China. Association for Computational Linguistics.  
+Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019b. What does BERT look at? an analysis of BERT's attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276-286, Florence, Italy. Association for Computational Linguistics.  
+Daniel de Vassimon Manela, David Errington, Thomas Fisher, Boris van Breugel, and Pasquale Minervini. 2021. Stereotype and skew: Quantifying gender bias in pre-trained and fine-tuned language models. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 2232-2242, Online. Association for Computational Linguistics.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Lucas Dixon, John Li, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. 2018. Measuring and mitigating unintended bias in text classification. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, AIES '18, page 67-73, New York, NY, USA. Association for Computing Machinery.  
+Yanai Elazar and Yoav Goldberg. 2018. Adversarial removal of demographic attributes from text data. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 11-21, Brussels, Belgium. Association for Computational Linguistics.  
+Elisabetta Fersini, Debora Nozza, and Paolo Rosso. 2018. Overview of the EVALITA 2018 Task on Automatic Misogyny Identification (AMI). In Proceedings of Sixth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2018). CEUR-WS.org.  
+Elisabetta Fersini, Debora Nozza, and Paolo Rosso. 2020. AMI @ EVALITA2020: Automatic Misogyny Identification. In Proceedings of the 7th evaluation campaign of Natural Language Processing and Speech tools for Italian (EVALITA 2020), Online. CEUR.org.  
+Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. 2018. Word embeddings quantify 100 years of gender and ethnic stereotypes. Proceedings of the National Academy of Sciences, 115(16):E3635-E3644.
+
+Hamidreza Ghader and Christof Monz. 2017. What does attention in neural machine translation pay attention to? In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 30-39, Taipei, Taiwan. Asian Federation of Natural Language Processing.  
+Alex Graves. 2013. Generating sequences with recurrent neural networks. CoRR, abs/1308.0850.  
+Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778.  
+Vijayasaradhi Indurthi, Bakhtiyar Syed, Manish Shrivastava, Nikhil Chakravartula, Manish Gupta, and Vasudeva Varma. 2019. FERMI at SemEval-2019 task 5: Using sentence embeddings to identify hate speech against immigrants and women in Twitter. In Proceedings of the 13th International Workshop on Semantic Evaluation, pages 70-74, Minneapolis, Minnesota, USA. Association for Computational Linguistics.  
+Xisen Jin, Zhongyu Wei, Junyi Du, Xiangyang Xue, and Xiang Ren. 2019. Towards hierarchical importance attribution: Explaining compositional semantics for neural sequence models. In International Conference on Learning Representations.  
+Masahiro Kaneko and Danushka Bollegala. 2021. Debiasing pre-trained contextualised embeddings. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1256-1266, Online. Association for Computational Linguistics.  
+Brendan Kennedy, Xisen Jin, Aida Mostafazadeh Davani, Morteza Dehghani, and Xiang Ren. 2020. Contextualizing hate speech classifiers with post-hoc explanation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5435-5442, Online. Association for Computational Linguistics.  
+Olga Kovaleva, Alexey Romanov, Anna Rogers, and Anna Rumshisky. 2019. Revealing the dark secrets of BERT. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4365-4374, Hong Kong, China. Association for Computational Linguistics.  
+Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, and Thomas Dandres. 2019. Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700.  
+Alyssa Lees, Jeffrey Sorensen, and Ian Kivlichan. 2020. Jigsaw@ AMI and HaSpeeDe2: Fine-Tuning a Pre-Trained Comment-Domain BERT Model. In Proceedings of the 7th evaluation campaign of Natural
+
+Language Processing and Speech tools for Italian (EVALITA 2020), Online. CEUR.org.  
+Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R. Bowman. 2020. CrowS-pairs: A challenge dataset for measuring social biases in masked language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1953-1967, Online. Association for Computational Linguistics.  
+Debora Nozza. 2021. Exposing the limits of zero-shot cross-lingual hate speech detection. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 907–914, Online. Association for Computational Linguistics.  
+Debora Nozza, Federico Bianchi, and Dirk Hovy. 2021. HONEST: Measuring hurtful sentence completion in language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2398-2406, Online. Association for Computational Linguistics.  
+Debora Nozza, Elisabetta Fersini, and Enza Messina. 2017. A multi-view sentiment corpus. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 273-280, Valencia, Spain. Association for Computational Linguistics.  
+Debora Nozza, Claudia Volpetti, and Elisabetta Fersini. 2019. Unintended bias in misogyny detection. WI '19, page 149-155, New York, NY, USA. Association for Computing Machinery.  
+Nedjma Ousidhoum, Zizheng Lin, Hongming Zhang, Yangqiu Song, and Dit-Yan Yeung. 2019. Multilingual and multi-aspect hate speech analysis. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4675-4684, Hong Kong, China. Association for Computational Linguistics.  
+Nedjma Ousidhoum, Yangqiu Song, and Dit-Yan Yeung. 2020. Comparative evaluation of label-agnostic selection bias in multilingual hate speech datasets. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2532-2542, Online. Association for Computational Linguistics.  
+Nedjma Ousidhoum, Xinran Zhao, Tianqing Fang, Yangqiu Song, and Dit-Yan Yeung. 2021. Probing toxic content in large pre-trained language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4262-4274, Online. Association for Computational Linguistics.
+
+Ji Ho Park, Jamin Shin, and Pascale Fung. 2018. Reducing gender bias in abusive language detection. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2799-2804, Brussels, Belgium. Association for Computational Linguistics.  
+Jing Qian, Mai ElSherief, Elizabeth Belding, and William Yang Wang. 2018. Hierarchical CVAE for fine-grained hate speech classification. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3550-3559, Brussels, Belgium. Association for Computational Linguistics.  
+Shauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton, and Yoav Goldberg. 2020. Null it out: Guarding protected attributes by iterative nullspace projection. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7237-7256, Online. Association for Computational Linguistics.  
+Alexey Romanov, Maria De-Arteaga, Hanna Wallach, Jennifer Chayes, Christian Borgs, Alexandra Chouldechova, Sahin Geyik, Krishnamaram Kenthapadi, Anna Rumshisky, and Adam Kalai. 2019. What's in a name? Reducing bias in bios without access to protected attributes. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4187-4195, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Sofia Serrano and Noah A. Smith. 2019. Is attention interpretable? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2931-2951, Florence, Italy. Association for Computational Linguistics.  
+C. E. Shannon. 1948. A mathematical theory of communication. The Bell System Technical Journal, 27(3):379-423.  
+Shubham Sharma, Yunfeng Zhang, Jesus M. Ríos Aliaga, Djallel Bouneffouf, Vinod Muthusamy, and Kush R. Varshney. 2020. Data augmentation for discrimination prevention and bias disambiguation. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, AIES '20, page 358-364, New York, NY, USA. Association for Computing Machinery.  
+Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. 2019. The woman worked as a babysitter: On biases in language generation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3407-3412, Hong Kong, China. Association for Computational Linguistics.  
+Anders Søgaard, Anders Johannsen, Barbara Plank, Dirk Hovy, and Hector Martínez Alonso. 2014.
+
+What's in a p-value in NLP? In Proceedings of the Eighteenth Conference on Computational Natural Language Learning, pages 1-10, Ann Arbor, Michigan. Association for Computational Linguistics.  
+Kaiser Sun and Ana Marasovic. 2021. Effective attention sheds light on interpretability. In *Findings of the Association for Computational Linguistics: ACLIJCNLP* 2021, pages 4126-4135, Online. Association for Computational Linguistics.  
+Ameya Vaidya, Feng Mai, and Yue Ning. 2020. Empirical analysis of multi-task learning for reducing identity bias in toxic comment detection. In Proceedings of the International AAAI Conference on Web and Social Media, volume 14, pages 683-693.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc.  
+Bertie Vidgen, Tristan Thrush, Zeerak Waseem, and Douwe Kiela. 2021. Learning from the worst: Dynamically generated datasets to improve online hate detection. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1667-1682, Online. Association for Computational Linguistics.  
+Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Yaron Singer, and Stuart Shieber. 2020. Investigating gender bias in language models using causal mediation analysis. In Advances in Neural Information Processing Systems, volume 33, pages 12388-12401. Curran Associates, Inc.  
+Zeerak Waseem and Dirk Hovy. 2016. Hateful symbols or hateful people? predictive features for hate speech detection on Twitter. In Proceedings of the NAACL Student Research Workshop, pages 88-93, San Diego, California. Association for Computational Linguistics.  
+Sarah Wegreffe and Yuval Pinter. 2019. Attention is not not explanation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 11-20, Hong Kong, China. Association for Computational Linguistics.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical
+
+Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.  
+Guanhua Zhang, Bing Bai, Junqi Zhang, Kun Bai, Conghui Zhu, and Tiejun Zhao. 2020. Demographics should not be the reason of toxicity: Mitigating discrimination in text classifications with instance weighting. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4134-4145, Online. Association for Computational Linguistics.  
+Xuhui Zhou, Maarten Sap, Swabha Swayamdipta, Yejin Choi, and Noah Smith. 2021. Challenges in automated debiasing for toxic language detection. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3143-3155, Online. Association for Computational Linguistics.
+
+# A Details on self-attention in Transformers
+
+The Transformer (Vaswani et al., 2017) is the building block of many recent neural language models. A Transformer model consists of two connected encoder and a decoder units which align a source and a target sequence. Differentiating from the original formulation, large language models, such as BERT, drop the encoder and use the remaining encoder to process a single input sequence.
+
+A transformer encoder consists of a multi-head self-attention block and a position-wise, fully connected feed forward neural network. Both the self-attention block and the feed forward network adopt a residual skip connection and batch normalization. We provide details for a standard forward pass in the encoder. In attention blocks, the multi-head output is computed with Scaled Dot-Product Attention between a set of queries and keys of dimension  $d_{k}$  and a set of values of dimension  $d_{v}$ . Let  $Q$ ,  $K$  and  $V$  be the respective matrix representations. The attention is then computed as
+
+$$
+\mathrm {A t t e n t i o n} (Q, K, V) = \mathrm {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right) V
+$$
+
+To improve expressiveness, the operation is performed on  $N$  different, independent linear projections of the same queries, keys and values, so that  $N$  attention heads are produced. The heads are then concatenated, projected back to the original input space, and finally fed through the fully connected neural network to produce the next layer embeddings. Let  $E = [e_0,\dots ,e_{d_s}]$  be the sequence of input embeddings $^6$ , with  $e_i\in \mathbb{R}^{d_m}$ . In the specific case of a transformer encoder, queries, keys and values correspond to the input embeddings - i.e.  $Q = K = V = E$ . As such, the output of the multi-head self-attention block is computed applying the previously presented Equation to the  $N$  token projections, concatenating and projecting back to the original space:
+
+$$
+\operatorname {M u l t i H e a d} (Q, K, V) = \left(\mathrm {o} _ {0} | | \dots | | \mathrm {o} _ {N}\right) W ^ {O}
+$$
+
+where
+
+$$
+\mathrm {o} _ {h} = \text {A t t e n t i o n} \left(Q W _ {h} ^ {Q}, K W _ {h} ^ {K}, V W _ {h} ^ {V}\right)
+$$
+
+and  $W^{O}$  and each  $W_{h}^{Q}, W_{h}^{K}, W_{h}^{V}$  are projection matrices.
+
+# B Experimental setup
+
+Hyper-parameters All our experiments use the Hugging Face transformers library (Wolf et al., 2020). We base our models and tokenizers on the bert-base-uncased checkpoint for English tasks and on the dbmdz/bert-base-italian-uncased checkpoint for Italian. We pre-process and tokenize our data using the standard pre-trained BERT tokenizer, with a maximum sequence length of 120 and right padding. We train all models with the following hyperparameters: batch size=64, learning rate=0.00002, weight decay=0.01, learning rate warmup steps=10%, full precision, maximum number of training epochs=30, and early stopping on non-improving validation loss after 5 epochs. Table 2 report results of BERT+EAR trained for 20 epochs with no early stopping, and regularization strength  $\alpha = 0.01$ . We chose the latter parameters with grid search on  $\alpha \in [0.0001, 0.001, 0.01, 0.1, 1]$  and epochs  $\in [10, 20, 30, 40, 50]$ . When fine-tuning on MULTILINGUAL AND MULTI-ASPECT, we use a weighted cross-entropy classification loss  $(\mathcal{L}_C)$  to discount class unbalance. Specifically, we normalize the loss for data points belonging to class  $C$  by the prior probability of  $C$ , evaluated as its relative frequency in the training set.
+
+For Kennedy et al. (2020), Nozza et al. (2019), Lees et al. (2020), and Ousidhoum et al. (2021), we kept all the parameters as specified by the respective authors. Please refer to our repository (https://github.com/g8a9/ear) for further details or the respective publications.
+
+We trained all models with 10 different initialization seeds per parameter configuration and averaged over them to obtain stable results and meaningfully compute significance.
+
+Statistical significance We compute the statistical significance of BERT+EAR over BERT and BERT with SOC mitigation via bootstrap sampling, following Søgaard et al. (2014), using  $\circ$  and  $\triangle$  (and their filled counterparts for a stronger significance) symbols, respectively. We use 1000 bootstrap samples and a sample size of  $20$ . For Hate Speech, significance can only be computed on F1-scores, since bias metrics require an assumption about the label distribution across identity terms that is not given.
+
+Selection bias We computed the  $\mathbf{B}_2$  metric following Ousidhoum et al. (2020). Specifically, we run the authors' code on each of our training dataset, using the query keywords used to sample each dataset. In case of queries composed of multiple words, we split and considered them separate keywords.
+
+Dataset preprocessing The original MULTILINGUAL AND MULTI-ASPECT dataset comes in a multi-label, multiple class setting. Following Ousidhoum et al. (2021), we used the Hostility dimension of the dataset as target label and created a Hate binary from it as follows. We considered single-labeled "Normal" instances to be non-hate/non-toxic and all the other instances to be toxic.
+
+Computation time We report NVIDIA Tesla V100 PCIE-16GB -equivalent computation time for the tested models. Averaging across the three presented data sets, training and evaluating 10 seeds of BERT+EAR (without early stop) requires 22 hours, compared to 72 hours for BERT+SOC and 7 hours for BERT. The regularization of attention entropy does not affect the computation time by a significant amount.
+
+$\mathrm{CO}_{2}$  emission Experiments were conducted using a private infrastructure, which has an estimated carbon efficiency of  $0.432\mathrm{kgCO}_{2}\mathrm{eq / kWh}$ . A cumulative of 319 hours of computation was performed on the hardware of type Tesla V100-PCIE-16GB (TDP of 300W). Total emissions are estimated to be  $41.34\mathrm{kgCO}_{2}\mathrm{eq}$ .
+
+Estimations were conducted using the Machine Learning Impact calculator presented in (Lacoste et al., 2019).
+
+# C List of identity terms
+
+In the following, we report the list of identity terms used in the considered data sets and methods.
+
+(Kennedy et al., 2020): muslim, jew, jews, white, islam, blacks, muslims, women, whites, gay, black, democrat, islamic,allah,jewish, lesbian, transgender race,brown,woman,mexican,religion, homosexual, homosexuality,africans
+
+(Nozza et al., 2019): woman, women, daughter, girl, girls, mother, she, wife, lady, ladies, girlfriend, sister
+
+(Fersini et al., 2020):onne, matrone, mamme, casalinghe, compagne, mo
+
+rose, femmine, donne, fidanzate, nonna, matrona, casalinga, morosa, femmina, mamma, donna, fidanzata, compagna
+
+(Dixon et al., 2018): lesbian, gay, bisexual, transgender, trans, queer, lgbt, lgbtq, homosexual, straight, heterosexual, male, female, nonbinary, african, african american, black, white, european, hispanic, latino, latina, latinx, mexican, canadian, american, asian, indian, middle eastern, chinese, japanese, christian, muslim, jewish, buddhist, catholic, protestant, sikh, Taoist, old, older, young, younger, teenage, millenial, middle aged, elderly, blind, deaf, paralyzed

entropybasedattentionregularizationfreesunintendedbiasmitigationfromlists/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:472c72c6f9a32b9fdb182eee8bc0efda730d0afc4f7c4fed6dee9db28e1df646
+size 389366

entropybasedattentionregularizationfreesunintendedbiasmitigationfromlists/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7b0f7e0401650f11496d133165ffecd4bf781811de0e4e8e3bf0c06eb7f0f02
+size 446570

eventtransitionplanningforopenendedtextgeneration/f6776941-6b61-4512-9543-e06c24f81d83_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9286ad4e3feab893818f81ae726b73d8cf889a882c24f4f80cbee236106eff69
+size 99203

eventtransitionplanningforopenendedtextgeneration/f6776941-6b61-4512-9543-e06c24f81d83_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:29008f7abcf4f2b1426696091938d778fd5679008cba37e9389e17cfa71e3008
+size 122651

eventtransitionplanningforopenendedtextgeneration/f6776941-6b61-4512-9543-e06c24f81d83_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62ae3a9124f6fa13deaa61f3f5a1b8bfccf21ffac980c6077de092e254beb82d
+size 705648

eventtransitionplanningforopenendedtextgeneration/full.md

@@ -0,0 +1,423 @@
+# Event Transition Planning for Open-ended Text Generation
+
+Qintong Li $^{\text{♥}}$  Piji Li $^{\text{♣}}$  Wei Bi $^{\text{♣}}$  Zhaochun Ren $^{\diamond}$  Yuxuan Lai $^{\text{♥}\dagger}$  Lingpeng Kong $^{\text{♥}\text{♣}}$
+
+Department of Computer Science, The University of Hong Kong
+
+Tencent AI Lab Shandong University
+
+$\spadesuit$ Shanghai Artificial Intelligence Laboratory
+
+qtli@connect.hku.hk
+
+{lipiji.pz, erutan.pkuicst}@gmail.com
+
+victoriabi@tencent.com, zhaochun. ren@sdu.edu.cn
+
+lpk@cs.hku.hk
+
+# Abstract
+
+Open-ended text generation tasks, such as dialogue generation and story completion, require models to generate a coherent continuation given limited preceding context. The open-ended nature of these tasks brings new challenges to the neural auto-regressive text generators nowadays. Despite these neural models are good at producing human-like text, it is difficult for them to arrange causalities and relations between given facts and possible ensuing events. To bridge this gap, we propose a novel two-stage method which explicitly arranges the ensuing events in open-ended text generation. Our approach can be understood as a specially-trained coarse-to-fine algorithm, where an event transition planner provides a "coarse" plot skeleton and a text generator in the second stage refines the skeleton. Experiments on two open-ended text generation tasks demonstrate that our proposed method effectively improves the quality of the generated text, especially in coherence and diversity. The code is available at: https://github.com/qtli/EventPlanforallGen.
+
+# 1 Introduction
+
+With the fast development of large-scale pre-trained models, considerable progress has been made in improving the quality of machine generated text (Radford et al., 2019; Rashkin et al., 2019a; Zhang et al., 2020b; Brown et al., 2020; Guan et al., 2021; Bakhtin et al., 2021). Today, machine learning models can do extremely well in generating text that looks human (Clark et al., 2021). The problem is still far from solved, however, as further reading of the machine-generated text often exposes defects such as self-contradiction and topic drifting (Bisk et al., 2020; Gao et al., 2020; Tan et al.,
+
+![](images/00796d78be30d5a866eac9c69efe38e0b5ec61a3a8327896b576cb154c50574a.jpg)  
+Figure 1: An illustration of our planning based framework in story completion task. Given story context, we extract corresponding event transition path, and use model EP to develop potential ensuing event transition paths. The planned paths accordingly guide the path-aware text generation model PG.
+
+2021; Fan et al., 2019; Dou et al., 2021; Dziri et al., 2021). These issues are particularly serious in open-ended text generation tasks (e.g., story completion), where the model is asked to produce a coherent continuation which often involves multiple events, given limited preceding context.
+
+To bridge this gap, we propose a two-stage method which explicitly models the event transitions in open-ended text generation. Multi-step generation has been adopted to control the generated content at a high level (Dong and Lapata, 2018; Ji et al., 2020; Xu, 2021). Different from previous works that rely on inflexible pattern retrieval, we leverage a generative model as an event transition planner in the first stage to boost the high-level coherence and causalities in open-ended text generation.
+
+Specifically, in stage one, an event transition planner (§3.1) outlines a transition path of events starting from the ones extracted from the input context. In stage two, this path is used to ensure a relevant and sound continuation from the actual text generator (§3.2). This method can be understood as a specially-trained coarse-to-fine algorithm, where
+
+<table><tr><td></td><td colspan="3">Dialogue Generation</td><td colspan="3">Story Completion</td></tr><tr><td>Input Context — Events</td><td colspan="3">[1] my husband lost a job but i&#x27;m hop-ing he can find a full time job soon. — my husband lost job, I hope he find job [2] He will, I have faith. — I have faith [3] thank you so much! — thank you</td><td colspan="3">[1] John got laid off from his company. — john get laid off [2] He was close to retirement age. —john is close retirement [3] John felt bored and listless his first week of unemployment. —john feel bored and listless [4] John decided to start a business of his own. —john decide start business</td></tr><tr><td>Target Output — Events</td><td colspan="3">No problem. What kind of work does he do? — what work he do</td><td colspan="3">He now has a flourishing online company. — john have a company</td></tr><tr><td rowspan="4">Event Transition Path</td><td rowspan="2">my husband lost job</td><td rowspan="2">XATTR</td><td rowspan="2">i hope he find job OREACT</td><td rowspan="2">john get laid off</td><td rowspan="2">XATTR</td><td rowspan="2">john is close to retirement XREACT</td></tr><tr></tr><tr><td rowspan="2">thank you OREACT</td><td rowspan="2" colspan="3">what work he do</td><td>XREACT</td><td>john feel bored and listless XREACT</td></tr><tr><td>decide start business XEFFECT</td><td>john have a company</td></tr></table>
+
+Table 1: Examples of event transition paths acquired from downstream tasks, i.e., dialogue generation and story completion. Events are marked in blue box.
+
+an event transition planner provides a "coarse" plot skeleton and a path-aware text generator refines the skeleton. Figure 1 shows an illustration of our approach.
+
+There are two main challenges in this method. First, the planer should produce high-quality and diverse paths that can generalize well to the unseen events at test time. For this challenge, we fine-tune a GPT-2 (Radford et al., 2019) on a large amount of event paths extracted from commonsense graphs (Sap et al., 2019), as well as from the training set of the specific task, aiming to extrapolate to event sequences that never appeared in these sources with the help of general knowledge stored in the large pre-trained model (Petroni et al., 2019; Lee et al., 2021).
+
+For the second challenge, the auto-regressive text generator need to work effectively under the supervision of the even transition path. We thus design an event query layer to absorb information from the planned paths and use the query layer to guide the text generation process.
+
+We validate our method thorough extensive experiments on two standard open-ended text generation tasks, dialogue generation (Rashkin et al., 2019b) and story completion (Mostafazadeh et al., 2016). Our two-stage approach outperforms a strong knowledge enhanced GPT-2 baseline (Guan et al., 2020a) in both automatic and human evaluation metrics. Further analysis shows that the improvements of the event transition planning model come in particular from the high-level consistency and diversity in long and difficult generation cases.
+
+# 2 Event Transition Path.
+
+In this work, the event transition path is defined as an alternating sequence between events and relations, where an event is a subject-verb phrase, a relation is chosen from a pre-defined label set (e.g., OREACT - object reaction; XATTR - subject attribute) of a commonsense atlas (Sap et al., 2019). Table 1 shows some text examples and their corresponding event transition paths. We collect event transition paths from a commonsense atlas ATOMIC (Sap et al., 2019), as well as from the training set of the specific task, to train an event transition planner.
+
+Sampling Paths from ATOMIC. We use everyday commonsense atlas ATOMIC (Sap et al., 2019) to acquire plenty event paths. ATOMIC is organized through 9 relations and 877k events (textual descriptions) of inferential commonsense, e.g., if "PersonX pays PersonY a compliment", then "PersonY will likely return the compliment". It has been demonstrated that ATOMIC is useful for open-ended text generation tasks, such as story generation (Guan et al., 2020b).
+
+Besides, to increase the flexibility, we introduce a reverse relation (e.g., _XATTR) for each original relation (e.g., XATTR) so that a sampled path can contain reverse triplets. The intuition is that, in open-ended text generation, the narrative maybe in a reverse order. After explaining the event A, the author may want to introduce the subsequent event B. Meanwhile, if the author introduce the event B first, she/he may want to describe the event A as an explanation for the reason/motivation.
+
+Finally, we collect sufficient event paths of variant lengths from ATOMIC via random walk
+
+![](images/03c422080a3fa2d5adccbf58fc63d82deb979a1570f3db099dd9ddabfad3f4aa.jpg)  
+Figure 2: Overall architecture of the proposed coarse-to-fine framework. It consists of two components. (1) Event Transition Planner: given a input context, it first extracts corresponding event path and then generates possible ensuing event path. The planner directly inherits the pre-trained parameters from GPT-2; (2) Event-path-aware Text Generator: another GPT-2-based generator is applied to generate a natural language sentence by attending to input context and explicit event transition path.
+
+ing<sup>1</sup>. We split the sampled paths into training/validation/test with the ratio of 18:1:1. We use these sampled paths to optimize the event transition planner which is responsible for generative event planning (see §3.1). The statistics of sampled paths are shown in Table 6 of Appendix A. We display several examples of the randomly sampled event transition paths in Table 7 of Appendix A.
+
+Extracting Paths from Specific Dataset. We use two kinds of event transition paths. A general kind is obtained from random walking on a daily commonsense graph, ATOMIC, as mentioned above. Another kind is extracted from the natural language instances of downstream datasets, which is used for the training and prediction stage of task-specific event planning. For example, given the inputs, "When the bride and groom entered, the audience cheered", the extracted event path is "bride and groom enter OFFECT audience cheer".
+
+In detail, for each sentence, to ensure the extracted events have complete semantics and keep a similar format with the events in ATOMIC, we use ASER event extractor tool $^2$  to distil events for all sentences of downstream datasets. We further predict the relations between these events, linking these isolated events as event transition paths. Specifically, we train a BERT (Devlin et al., 2019) classifier using event triples and relations in ATOMIC. The sizes of training/validation/test
+
+instances are 639,045/35,503/35,502, respectively. We finally achieve a accuracy score of  $85\%$  on the test set for the relation prediction.
+
+# 3 Methodology
+
+We focus on the conditional language modeling problem in open-ended text generation tasks. Formally, given an input context  $\mathbf{x}$ , models are required to generate a sentence  $\mathbf{y}$  that is consistent with input context and not contradicts itself.
+
+In this work, we propose a two-stage model for the generation process. In the first stage, we extract the starting event sequence  $\boldsymbol{r}_x$  from the input context and employ the event transition planer to generate subsequent event transition path  $\boldsymbol{r}_y$  based on  $\boldsymbol{r}_x$ . In the second stage, the output text is generated from an auto-regressive model conditioning on the path and the preceding context  $\boldsymbol{x}$ .
+
+Figure 2 gives an overview of our coarse-to-fine framework for open-ended text generation. In a nutshell, we first fine-tune a GPT-2 on event transition sequences as an event planner (i.e., a conditional generative model for event paths). This fine-tuning involves event transition sequences extracted from both commonsense graphs and the training set. We then build a path-aware text generator with an event query layer specifically designed to refer to the planned path when generating the output.
+
+# 3.1 Generative Event Transition Planner
+
+In this section, we describe the event transition planner which completes the partial event path
+
+given certain input context. Pre-trained language models can be good representation learners of relational knowledge (Petroni et al., 2019; Bosselut et al., 2019). In our model, we choose GPT-2 (Radford et al., 2019) as the backbone of our event transition planner.
+
+Specifically, we first fine-tune GPT-2 with large-scaled event transition paths sampled from ATOMIC (Sap et al., 2019). After that, we fine-tune the resulting model in addition on the event transitions extracted from the training corpus, so that the planner is aware of general transitions in the commonsense while focusing on the transitions in the specific domain in the meantime.
+
+In preliminary experiments, we find that directly running a full fine-tuning (i.e., updating all GPT-2 parameters) leads to a drop in the final performance. We suspect the reason is the full fine-tuning flushes out the original general knowledge from the largescale pre-training (Chen et al., 2019; Lee et al., 2020; Chen et al., 2020).
+
+To overcome this drawback, we pretend a trainable continuous event prompt  $\mathbf{z}$  to the input path  $\mathbf{r} = [r_x; r_y]$  of every transformer layer in event transition planner, as prefix-tuning (Li and Liang, 2021) does. A trainable matrix  $\mathbf{U}_{\theta}$  with parameters  $\theta$  is randomly initialized to embed event prompt  $\mathbf{z}$ . The aim is to use parameters  $\theta$  introduced by  $\mathbf{z}$  to store event transition patterns from ATOMIC. Then the representation of each input event transition path  $\mathbf{r}$  is prompted as  $\mathbf{r}' = [z; r]$ . To increase training speed and performance robustness, we apply an additional linear reparameterization function on  $\mathbf{U}_{\theta}$ .
+
+$$
+\mathbf {U} _ {\theta} = F F N _ {\theta} \left(\mathbf {U} _ {\theta} ^ {\prime}\right), \tag {1}
+$$
+
+where  $U_{\theta}^{\prime}$  is another randomly initialized matrix with smaller dimension,  $FFN$  is a large feedforward neural network (Vaswani et al., 2017). We perform gradient updates on the following log-likelihood objective:
+
+$$
+\begin{array}{l} \max  _ {\theta} \log \left(\boldsymbol {r} _ {y} \mid [ \boldsymbol {z}; \boldsymbol {r} _ {<   y} ]\right) = \\ \sum_ {y \in \mathbf {z} _ {\mathrm {i d x}}} \log E P _ {\phi , \theta} (\mathbf {r} _ {y} \mid \mathbf {h} _ {<   y}), \tag {2} \\ \end{array}
+$$
+
+where  $\phi$  denotes the pre-trained parameters from the backbone LM of event transition planner,  $\theta$  denotes the newly introduced parameters for the event prompt,  $z_{\mathrm{idx}}$  denotes the index sequence of the event prompt,  $EP$  is short for event transition
+
+planner, and  $\mathbf{h}_{< y}$  denotes the hidden states calculated by the trainable event prompt matrix and activation layers of the backbone LM:
+
+$$
+\mathbf {h} _ {y} = \left\{ \begin{array}{l l} \mathbf {U} _ {\theta} [ y, : ], & \text {i f} y \in \mathbf {z} _ {\mathrm {i d x}}, \\ \boldsymbol {L M} _ {\phi} (\boldsymbol {r} _ {y} \mid \mathbf {h} _ {<   y}) & \text {o t h e r w i s e .} \end{array} \right. \tag {3}
+$$
+
+Similar to the above event prompting technique, for the paths from downstream dataset, we pretend another event prompt  $z^{\prime}$  to the  $r^{\prime}$  and only optimize the parameters introduced by  $z^{\prime}$ . This effectively preserves the newly-learned event transition patterns from ATOMIC and continuously adapts the event transition planner to different downstream event transition patterns.
+
+# 3.2 Event-path-aware Text Generation
+
+Current state-of-the-art systems for open-ended text generation are based on fine-tuning pretrained language models with different downstream datasets. Although text generation fluency is usually not a crucial issue nowadays, topic-related mistakes (Dou et al., 2021) such as off-prompt and self-contradiction are common. We therefore integrate the event transition paths produced by the planner into the text generation model via an event query layer using the multi-head attention mechanism (MHA).
+
+The event query layer is built on top of the stacked transformer layers, aiming to explicitly induce the expected output with event transition paths. The input of the event query layer is the event transition path  $\mathbf{r}$  given the current input  $\mathbf{x}$ .  $\mathbf{r}$  not only summarizes the event transition in  $\mathbf{x}$ , also indicates possible event path following  $\mathbf{x}$ . The structure of the event query layer resembles the transformer layer. Its output serves as the key and value vectors in the multi-head attention mechanism, which computes another attention vector  $MHA(\mathbf{r})$ . We concatenate two multi-head attention vectors and derive the final event-path-aware attention vector  $\mathbf{m}$ :
+
+$$
+\mathbf {m} = M L P \left(\left[ M H A (\boldsymbol {x}); M H A (\boldsymbol {r}) \right]\right), \tag {4}
+$$
+
+where  $MHA(\boldsymbol{x})$  is the output from the multi-head attention function of the original transformer layer,  $MHA(\boldsymbol{r})$  is the output from the event query layer. The event-path-aware attention vector  $\mathbf{m}$  replaces the original multi-head attention vector  $MHA(\boldsymbol{x})$  and participates the remaining calculation of the language model.
+
+The optimization of the event-path-aware text generator is the standard cross-entropy objective: CrossEntropy  $(\mathbf{y}_j\mid \mathbf{y}_{< j},\mathbf{x},\mathbf{r})$
+
+# 3.3 Implementation Details
+
+We base our event planner and event-plan-aware text generator on pre-trained GPT-2-small models<sup>3</sup>. The event prompt length during training ATOMIC event transition paths are set to 5 according to pilot study. We inject and optimize the event query layer on the last layer of the stacked Transformers. When training the event-path-aware text generator, event path  $r_y$  is derived from the ground truth. During inference,  $r_y$  is the prediction from event transition planner given the input event transition path  $r_x$ . More details are elaborated in Appendix B.
+
+# 4 Experiments
+
+We conduct experiments on two open-ended text generation tasks, dialogue generation and story completion, to answer the following questions:
+
+- RQ1: How to develop a better event transition planner?  
+- RQ2: Whether the integration of event transition paths enhances the open-ended text generation?  
+- RQ3: How do the event transition paths benefit text generation?
+
+# 4.1 Evaluated Tasks
+
+- Story Completion requires models to complete a story given the first few sentences. We evaluated our framework on ROCSTORIES (Mostafazadeh et al., 2016), which contains 98k five-sentence stories. Our default setting is to predict the last sentence given the first four ones.  
+- Dialogue Generation aims to generate reasonable and human-like responses given the dialogue history. We evaluated our framework on EMPATHETICDIALOGUES (Rashkin et al., 2019b) which consists of  $25\mathrm{k}$  conversations grounded in prespecified situations.
+
+# 4.2 Event Transition Planning (RQ1)
+
+We compare our event transition planner, named as PLANGGeneration, with fine-tuned pre-trained GPT-2 (Radford et al., 2019) and several ablation settings, investigating how to develop a better event transition planner.
+
+Specifically, the compared settings include:
+
+- GPT-2 is a pre-trained GPT-2 model (Radford et al., 2019) directly fine-tuned on the event paths extracted from specific tasks, i.e., dialogue generation or story completion in our work.  
+- PLANGGeneration is our proposed event planning method, which explores a two-stage finetuning on event transition paths from ATOMIC (Sap et al., 2019) and the downstream task, equipping with the proposed event prompting module.  
+- w/o PROMPT is our proposed method without the event prompting module, but still using the two-stage fine-tuning strategy.  
+- w/o TUNING ON ATOMIC is our proposed method without the first-stage fine-tuning on the event paths extracted from external commonsense atlas ATOMIC.  
+- PLANRETRIEVAL is a retrieval based planning methods, which employs the BM25 ranking function (ROBERTSON et al., 1995) to retrieve from the paths extracted from the training sets according to the given context.
+
+Results. We use BLEU (Papineni et al., 2002) and DIST (Li et al., 2016) as the automatic metrics to evaluate the generated sentences in terms of the coherence and diversity, respectively. BLEU evaluates  $n$ -gram overlap between generation and ground truth. BLEU scores will become extremely low for large  $n$ . We thus experiment with  $n \in \{1,2,4\}$ . DIST measures the ratio of distinct  $n$ -grams to all the generated  $n$ -grams from the perspective of the generation diversity. For DIST metric, we adopt  $n \in \{1,2\}$ . The experimental results are shown in Table 2. The dataset needed in this section consists of event transition paths sampled from ATOMIC and extracted from downstream datasets. i.e., ROCSTORIES and EMPATHETICDIALOGUE. The details of event transition paths are shown in §2 and Appendix A.
+
+On both dialogue generation and story completion tasks, our proposed PLANGENERATION greatly outperforms baseline GPT-2 on event planning coherence (BLEU) and event path diversity (DIST). Specifically, on two downsteam tasks, our event transition planner PLANGENERATION surpasses the fine-tuned GPT-2 by 3.09 and 3.53 on BLEU-1, 0.31 and 0.30 on DIST-1. This improvement indicates that (1) the two-stage event prompting module could endow event transition planner powerful abilities on predicting the ensuing event paths; (2) enhanced with the large-scale event transition patterns from ATOMIC, our event transition
+
+<table><tr><td>Tasks</td><td>Methods</td><td>BELU-1</td><td>BLEU-2</td><td>BLEU-4</td><td>DIST-1</td><td>DIST-2</td></tr><tr><td rowspan="5">Dialogue Generation</td><td>GPT-2</td><td>23.43</td><td>11.50</td><td>3.31</td><td>1.57</td><td>4.18</td></tr><tr><td>PLANGENERATION (Ours)</td><td>26.52</td><td>12.38</td><td>3.29</td><td>1.88</td><td>5.52</td></tr><tr><td>w/o PROMPT</td><td>23.58</td><td>11.85</td><td>3.58</td><td>1.80</td><td>5.13</td></tr><tr><td>w/o TUNING ON ATOMIC</td><td>19.82</td><td>7.90</td><td>1.81</td><td>1.16</td><td>2.54</td></tr><tr><td>PLANRETRIEVAL</td><td>0.75</td><td>0.14</td><td>0.00</td><td>13.05</td><td>39.52</td></tr><tr><td rowspan="5">Story Completion</td><td>GPT-2</td><td>15.98</td><td>7.19</td><td>1.08</td><td>5.53</td><td>17.44</td></tr><tr><td>PLANGENERATION (Ours)</td><td>19.51</td><td>9.01</td><td>1.35</td><td>5.83</td><td>17.48</td></tr><tr><td>w/o PROMPT</td><td>13.64</td><td>6.14</td><td>1.12</td><td>4.71</td><td>15.77</td></tr><tr><td>w/o TUNING ON ATOMIC</td><td>12.74</td><td>4.61</td><td>0.47</td><td>6.08</td><td>12.27</td></tr><tr><td>PLANRETRIEVAL</td><td>1.28</td><td>0.15</td><td>0.00</td><td>11.88</td><td>37.70</td></tr></table>
+
+Table 2: Experimental results on event transition planning. For detailed description about the compared models, please refer to §4.2.
+
+planner becomes more creative and produces more diverse outcomes.
+
+Considering the ablation settings, without tuning on ATOMIC (w/o TUNING ON ATOMIC) or without the event prompting module (w/o PROMPT), the method performs worse on both tasks and across almost all metrics. The limited performance of w/o TUNING ON ATOMIC suggests the necessity and effectiveness of learning general event transition patterns from ATOMIC before optimizing on task-specific event paths. Tuning on ATOMIC event patterns could make event transition planner get familiar with the event-path-like language and generalize well on unseen event patterns. Compared to ablation model w/o PROMPT, EVENTPLANNING is comparatively more effective. This is because when optimizing on event paths of target tasks, the proposed event prompt protects the parameters of pre-trained language model from drastic change when training with event transition paths. This comparison confirms our intuition that event prompting module could improve event planning performance without destroying the eventual commonsense stored in pre-trained parameters. It provides a more flexible approach to blend the event transition patterns in both ATOMIC and specific tasks with the pre-trained GPT-2 model.
+
+We also attempt a variation of our PLANGENERATION method, i.e., PLANRETRIEVAL. We can see that the BELU scores of PLANRETRIEVAL are substantially lower than the generation based methods. The main reason is that the target event paths are flexible, infinite, and task-related. Many transition patterns are not seen in the training data or external commonsense graph.
+
+# 4.3 Event-path-aware Text Generation (RQ2)
+
+In this section, we compare our overall framework EP-PG with several baselines to investigate whether the integration of generative event transition paths benefits the open-ended text generation.
+
+We consider the following settings:
+
+- GPT-2 is a pre-trained GPT-2 model (Radford et al., 2019) fine-tuned on the task-specific dataset.  
+- GPT-2-CS-FT is a commonsense-enhanced GPT-2 model. By following Guan et al. (2020b), we conduct a first-stage post-training on the ATOMIC commonsense triples and then fine-tuning on task-specific dataset.  
+- EP-PG is our proposed framework, which is a fine-tuned GPT-2 model integrated with the event transition path produced from event transition planner PLANGeneration via an event query layer.  
+- R-EP-PG is another version of EP-PG to explore the proposed event query layer. The input event transition paths are produced by PLANRETRIEVAL in a retrieval way.
+
+Results. We consider the same evaluation metrics as in §4.2. As demonstrated in Table 3, EP-PG achieves the most satisfying performance among all settings on both tasks<sup>4</sup>. Integrated with the explicit guidance of the event transition paths, EP-PG produces more accurate open-ended generations with higher diversity.
+
+Particularly, our proposed framework EP-PG consistently and significantly improves GPT-2 baseline for all tasks on content quality (BLEU) and diversity (DIST), showcasing the advantage of injecting event query layer on fine-tuned GPT-2. Without the explicit modeling of event transition
+
+<table><tr><td>Tasks</td><td>Models</td><td>BLEU-1</td><td>BLEU-2</td><td>BLEU-4</td><td>DIST-1</td><td>DIST-2</td></tr><tr><td rowspan="4">Dialogue Generation</td><td>GPT-2</td><td>16.07</td><td>6.41</td><td>2.13</td><td>2.06</td><td>7.70</td></tr><tr><td>GPT-2-CS-FT (Guan et al.)</td><td>16.43</td><td>6.83</td><td>2.31</td><td>2.16</td><td>8.28</td></tr><tr><td>R-EP-PG</td><td>16.68</td><td>6.71</td><td>2.27</td><td>2.21</td><td>8.44</td></tr><tr><td>EP-PG (Ours)</td><td>16.74</td><td>6.94</td><td>2.39</td><td>2.19</td><td>8.25</td></tr><tr><td rowspan="4">Story Completion</td><td>GPT-2</td><td>25.03</td><td>9.58</td><td>2.70</td><td>8.38</td><td>31.33</td></tr><tr><td>GPT-2-CS-FT (Guan et al.)</td><td>25.09</td><td>9.64</td><td>2.72</td><td>8.07</td><td>30.68</td></tr><tr><td>R-EP-PG</td><td>24.72</td><td>9.27</td><td>2.63</td><td>7.01</td><td>26.49</td></tr><tr><td>EP-PG (Ours)</td><td>25.47</td><td>9.71</td><td>2.74</td><td>8.99</td><td>34.48</td></tr></table>
+
+Table 3: Results of experiments on open-ended text generations. For detailed information about each compared model, please refer to §4.3.
+
+paths, GPT-2-CS-FT which post-trains on commonsense triples only obtains a slight improvement or even performs comparable with GPT-2 model. EP-PG further improves generation performance from GPT-2-CS-FT across all metrics on the two tasks, highlighting the efficacy of long-range event planning via an additional event query layer.
+
+Particularly remarkable are the relative differences between R-EP-PG and EP-PG. Although R-EP-PG manages to bring generations more diversity, but in most cases, EP-PG is more effective on content planning and informativeness due to generative event transition patterns in higher qualities. Moreover, R-EP-PG performs even worse than GPT-2 on story completion. This implies that low-quality event paths even damage the generations. Thus, a reliable event path is a key guarantee for effective downstream text generation.
+
+# 4.4 Analysis: Event Transition Planning for Different Generation Scenarios (RQ3)
+
+To further investigate how do the event paths benefit text generation, we analyse the effectiveness of event paths on differently difficult levels of generation, i.e., token-level and sentence-level.
+
+Token-level. We first separate the test set into 5 groups according to the averaged target sentence lengths, and then observe the improvements of our proposed EP-PG over GPT-2 on BLUE-1 score. We find that our framework gains more on the longer instances in both story completion (from 0.4 to 1.3 on instances with more than 15 non-stop-words) and dialogue generation (from 0.3 to 0.9 on instances with more than 5 non-stop-words). We argue that the longer targets imply more sophisticated upcoming event transitions, where the guidance from the event transition planner becomes more important.
+
+![](images/4b11990dd3347470596cc7af19e8bb9d3b97ed17ea7929cbc63d236887d18036.jpg)  
+Figure 3: The log of BLEU-1 scores on story completion with different numbers of sentences as input.
+
+Sentence-level. For story completion on five-sentence story dataset ROCSTORIES, we further conduct experiments on EP-PG with various input sentences and output sentences, i.e., the numbers of input (output) sentence are 1 (4), 2 (3), 3 (2), and 4 (1), respectively. Figure 3 shows that, compared to GPT-2, the relative improvement proportion of EP-PG is nearly doubled on the most difficult setting where there is only one sentence as input. This improvement is much larger than the easiest situation where 4 sentences are input to the model. Despite less input context, EP-PG with event transition planning manages to perform better with smaller performance drop.
+
+# 4.5 Human Evaluation
+
+We set up a human evaluation as a complementary evaluation beyond automatic evaluation. For both tasks, we randomly select 100 samples from test set. For each sample, we compare three pairs of models: EP-PG versus GPT-2, GPT-2-CS-FT, and R-EP-PG. Each comparison is rated by three crowd workers, who are asked to give a preference (win, lose or tie) from two perspectives:
+
+<table><tr><td rowspan="2">Tasks</td><td rowspan="2">Models</td><td colspan="4">Coherence</td><td colspan="4">Diversity</td></tr><tr><td>Win</td><td>Lose</td><td>Tie</td><td>κ</td><td>Win</td><td>Lose</td><td>Tie</td><td>κ</td></tr><tr><td rowspan="3">Dialogue Generation</td><td>Ours vs. GPT-2</td><td>45%</td><td>11%</td><td>44%</td><td>0.290</td><td>71%</td><td>10%</td><td>19%</td><td>0.226</td></tr><tr><td>Ours vs. GPT-2-CS-FT</td><td>34%</td><td>10%</td><td>56%</td><td>0.286</td><td>54%</td><td>7%</td><td>39%</td><td>0.288</td></tr><tr><td>Ours vs. R-EP-PG</td><td>32%</td><td>8%</td><td>60%</td><td>0.472</td><td>67%</td><td>11%</td><td>22%</td><td>0.291</td></tr><tr><td rowspan="3">Story Completion</td><td>Ours vs. GPT-2</td><td>45%</td><td>12%</td><td>42%</td><td>0.397</td><td>59%</td><td>10%</td><td>31%</td><td>0.220</td></tr><tr><td>Ours vs. GPT-2-CS-FT</td><td>47%</td><td>17%</td><td>36%</td><td>0.387</td><td>56%</td><td>17%</td><td>27%</td><td>0.210</td></tr><tr><td>Ours vs. R-EP-PG</td><td>43%</td><td>17%</td><td>40%</td><td>0.393</td><td>61%</td><td>6%</td><td>33%</td><td>0.340</td></tr></table>
+
+Table 4: Manual evaluation results on downstream text generation. The scores indicate the percentages of Win, Lose or Tie when our model is compared with other baselines.  $\kappa$  denotes Fleiss' kappa (all are fair agreement or moderate agreement).
+
+- Coherence. It indicates whether the inference is natural, relevant, and follows logically from the given context.  
+- Diversity. Particularly, for baseline models, we use beam search decoding with a top- $k$  ( $k = 5$ ) sampling scheme (Fan et al., 2018) and a softmax temperature  $\tau$  ( $\tau = 0.7$ ) to generate three inferences per sample. For our method EP-PG, its event transition planner first predicts three paths via the same beam decoding, then its text generator uses greedy decoding based on the generated three paths to produce three inferences per sample. During pair-wise comparison, we ask annotators to evaluate which model's predictions contain more reasonable and coherent event transition patterns.
+
+The two aspects are independently evaluated and results are shown in Table 4. According to human evaluations, our proposed EP-PG significantly outperforms compared baselines in terms of both criteria on the test set of all datasets. Overall inter-rater agreement measured by Fleiss' kappa (Fleiss, 1971) and all the results show fair agreement  $(0.2 \leq k \leq 0.4)$  or moderate agreement  $(0.4 \leq k \leq 0.6)$ . The results indicate that explicit incorporating event transition patterns yields significant improvement in generating coherence texts given the input context. Specifically, with guidance from different event transition paths, our method could produce more diverse and reasonable inferences.
+
+# 4.6 Qualitative Study
+
+Table 5 illustrates how our model tends to produce more contentful and coherent predictions compared to the other systems. In this story completion case, the generated event path successfully captures the correlations between working out and pass physical test, which further helps our model produce the most reasonable output, Alex was able to pass
+
+# Story Context:
+
+Alex was in training to be a police officer.
+
+He was not in the best shape.
+
+Alex failed the physical assessment.
+
+Alex started working out.
+
+# Golden Event Path:
+
+XEFFECT he take the test again XEFFECT he pass
+
+# Retrieved Event Path:
+
+wants to be best police officer XWANT tells person to stop
+
+# Generated Event Path:
+
+XEFFECT Alex able get good shape XEFFECT Alex able pass physical test
+
+# Reference:
+
+He took the test again and passed.
+
+# PT-2:
+
+Alex was able to get a good job.
+
+# GPT-2-CS-FT:
+
+Alex made the squad.
+
+# R-EP-PG:
+
+Alex was able to become a police officer.
+
+# EP-PG:
+
+Alex was able to pass the physical exam.
+
+Table 5: Case study on story completion. The three sections from top to bottom are the input context, the event transition plans, and inferences from our model and baseline models, respectively.
+
+the physical exam. For the baseline without commonsense knowledge, GPT-2, is instead not related to the core context failed the physical assessment. Tuning on commonsense atlas ATOMIC, GPT-2-CSFT produces informative inference but contradicts the context. The retrieval-based model R-EP-PG searches a related event transition police officer. However, its flexibility is limited by search space and cannot maintain a long-range event path, which is easy to produce hallucinated inference. More case analysis are stated in the Appendix C.
+
+# 5 Related Work
+
+Recent advances in pre-trained language models have resulted in impressive performances on open-domain text generation, such as story com
+
+pletion (See et al., 2019; Yao et al., 2019; Fan et al., 2019; Ippolito et al., 2020), dialogue generation (Rashkin et al., 2019b; Zhang et al., 2020b; Li, 2020; Vulic et al., 2021), question generation (Cheng et al., 2021; Wang et al., 2021), and so on. For example, in dialogue generation, Zhang et al. (2020b) design a trainable generative pretrained transformer by training an autoregressive language model on large-scale Reddit context-response pairs with a maximum mutual information scoring function to improve diversity. Goldfarb-Tarrant et al. (2020) integrate semantic role labels and prompts into pre-trained BART (Lewis et al., 2020) during fine-tuning for prompt based story telling. In this paper, we focus on story completion and dialogue generation and build a generative coarse-to-fine method to generate open-ended text with explicit event transition paths.
+
+Despite the success of generative pre-trained language models on a series of open-ended text generation tasks, they still suffer in maintaining coherence throughout multiple sentences due to the left-to-right word-by-word generation style (Fan et al., 2019; Yu et al., 2020). To alleviate this problem, one research direction adopts coarse-to-fine progressive text generation (Tan et al., 2021). This generation paradigm has been studied in many text generation systems for specific tasks, such as data-to-text generation (Moryossef et al., 2019; Puduppully and Lapata, 2021), storytelling (Goldfarb-Tarrant et al., 2020; Orbach and Goldberg, 2020), and dialogue generation (Xu et al., 2020a). Our work adopts a generative event transition planner that is trained on a large amount of event transition paths, aiming to arrange the ensuing events in open-ended text generation.
+
+Another research direction incorporates external entities to guide the open-ended text generation (Guan et al., 2019; Zhang et al., 2020a; Dziri et al., 2021; Li et al., 2020; Peng et al., 2021). Ji et al. (2020) and Xu et al. (2020b) retrieve entities from knowledge bases to control the generated content. However, the retrieval-based methods also suffer from the sparsity problem and the domain shift between external sources and downstream tasks (Wang et al., 2020). Guan et al. (2020b) integrate entity relations into pre-trained language model via additional tuning on entity triples. Even with such specialized learning, the resulted model still often stuck in logical errors or repeats pieces of narratives (Guan et al., 2020b; Peng et al., 2021).
+
+This phenomenon demonstrates the need for an intact inductive bias on organizing event transition patterns for open-ended text generation. Different from using event triples as additional training instances, our method explicitly maintains generative event transition paths to make the generation process more explainable and improve the coherence.
+
+# 6 Conclusion
+
+In this paper, we propose a novel two-stage method to improve high-level consistency and diversity in open-ended text generation. We design a specialized event transition planner to explicitly arrange the ensuing events and introduce an event-path-aware text generator to exploit the event transition guidance for language generation. We investigate two open-ended text generation tasks, i.e., story completion and dialogue generation. Thorough experiments demonstrate that the explicit arrangement of event transition path indeed facilitate models to generate more coherent and diverse text in open-ended scenery. Besides, with the proposed event prompt and event query layer, our method could be extended to any other language models and open-ended generation tasks. A future line of investigation is to explore the effect of the proposed method on other open-ended tasks, such as commonsense question answering.
+
+# Acknowledgments
+
+We thank the anonymous reviewers whose suggestions helped clarify this work. This research is supported in part by the National Natural Science Foundation of China (Grant No. 62106105, 61902219), the Shanghai Committee of Science and Technology, China (Grant No. 21DZ1100100), and the Tencent AI Lab Rhino-Bird Focused Research Program.
+
+# References
+
+Anton Bakhtin, Yuntian Deng, Sam Gross, Myle Ott, Marc'Aurelio Ranzato, and Arthur Szlam. 2021. Residual energy-based models for text. J. Mach. Learn. Res., 22:40:1-40:41.  
+Yonatan Bisk, Ari Holtzman, Jesse Thomason, Jacob Andreas, Yoshua Bengio, Joyce Chai, Mirella Lapata, Angeliki Lazaridou, Jonathan May, Aleksandr Nisnevich, Nicolas Pinto, and Joseph Turian. 2020. Experience grounds language. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8718-8735, Online. Association for Computational Linguistics.
+
+Antoine Bosselut, Hannah Rashkin, Maarten Sap, Chaitanya Malaviya, Asli Celikyilmaz, and Yejin Choi. 2019. COMET: Commonsense transformers for automatic knowledge graph construction. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4762-4779, Florence, Italy. Association for Computational Linguistics.  
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.  
+Sanyuan Chen, Yutai Hou, Yiming Cui, Wanxiang Che, Ting Liu, and Xiangzhan Yu. 2020. Recall and learn: Fine-tuning deep pretrained language models with less forgetting. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7870-7881, Online. Association for Computational Linguistics.  
+Xinyang Chen, Sinan Wang, Bo Fu, Mingsheng Long, and Jianmin Wang. 2019. Catastrophic forgetting meets negative transfer: Batch spectral shrinkage for safe transfer learning. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 1906-1916.  
+Yi Cheng, Siyao Li, Bang Liu, Ruihui Zhao, Sujian Li, Chenghua Lin, and Yefeng Zheng. 2021. Guiding the growth: Difficulty-controllable question generation through step-by-step rewriting. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5968-5978, Online. Association for Computational Linguistics.  
+Elizabeth Clark, Tal August, Sofia Serrano, Nikita Hahuong, Suchin Gururangan, and Noah A. Smith. 2021. All that's 'human' is not gold: Evaluating human evaluation of generated text. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 7282-7296, Online. Association for Computational Linguistics.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of
+
+deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Li Dong and Mirella Lapata. 2018. Coarse-to-fine decoding for neural semantic parsing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 731-742, Melbourne, Australia. Association for Computational Linguistics.  
+Yao Dou, Maxwell Forbes, Rik Koncel-Kedziorski, Noah A Smith, and Yejin Choi. 2021. Scarecrow: A framework for scrutinizing machine text. ArXiv preprint, abs/2107.01294.  
+Nouha Dziri, Andrea Madotto, Osmar Zaiane, and Avishek Joey Bose. 2021. Neural path hunter: Reducing hallucination in dialogue systems via path grounding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2197-2214, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Angela Fan, Mike Lewis, and Yann Dauphin. 2018. Hierarchical neural story generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 889-898, Melbourne, Australia. Association for Computational Linguistics.  
+Angela Fan, Mike Lewis, and Yann Dauphin. 2019. Strategies for structuring story generation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2650-2660, Florence, Italy. Association for Computational Linguistics.  
+Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378.  
+Jianfeng Gao, Baolin Peng, Chunyuan Li, Jinchao Li, Shahin Shayandeh, Lars Liden, and Heung-Yeung Shum. 2020. Robust conversational AI with grounded text generation. CoRR, abs/2009.03457.  
+Seraphina Goldfarb-Tarrant, Tuhin Chakrabarty, Ralph Weischedel, and Nanyun Peng. 2020. Content planning for neural story generation with aristotelian rescoring. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4319-4338, Online. Association for Computational Linguistics.  
+Jian Guan, Fei Huang, Zhihao Zhao, Xiaoyan Zhu, and Minlie Huang. 2020a. A knowledge-enhanced pretraining model for commonsense story generation. Transactions of the Association for Computational Linguistics, 8:93-108.
+
+Jian Guan, Fei Huang, Zhihao Zhao, Xiaoyan Zhu, and Minlie Huang. 2020b. A knowledge-enhanced pretraining model for commonsense story generation. Transactions of the Association for Computational Linguistics, 8:93-108.  
+Jian Guan, Xiaoxi Mao, Changjie Fan, Zitao Liu, Wenbiao Ding, and Minlie Huang. 2021. Long text generation by modeling sentence-level and discourse-level coherence. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume I: Long Papers), pages 6379-6393, Online. Association for Computational Linguistics.  
+Jian Guan, Yansen Wang, and Minlie Huang. 2019. Story ending generation with incremental encoding and commonsense knowledge. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 6473-6480. AAAI Press.  
+Daphne Ippolito, David Grangier, Douglas Eck, and Chris Callison-Burch. 2020. Toward better storylines with sentence-level language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7472-7478, Online. Association for Computational Linguistics.  
+Haozhe Ji, Pei Ke, Shaohan Huang, Furu Wei, Xiaoyan Zhu, and Minlie Huang. 2020. Language generation with multi-hop reasoning on commonsense knowledge graph. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 725-736, Online. Association for Computational Linguistics.  
+Cheolhyoung Lee, Kyunghyun Cho, and Wanmo Kang. 2020. Mixout: Effective regularization to finetune large-scale pretrained language models. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net.  
+Kenton Lee, Kelvin Guu, Luheng He, Tim Dozat, and Hyung Won Chung. 2021. Neural data augmentation via example extrapolation. *ArXiv preprint*, abs/2102.01335.  
+Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics.
+
+Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110-119, San Diego, California. Association for Computational Linguistics.  
+Piji Li. 2020. An empirical investigation of pre-trained transformer language models for open-domain dialogue generation. ArXiv preprint, abs/2003.04195.  
+Qintong Li, Piji Li, Zhaochun Ren, Pengjie Ren, and Zhumin Chen. 2020. Knowledge bridging for empathetic dialogue generation. arXiv e-prints, pages arXiv-2009.  
+Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, Online. Association for Computational Linguistics.  
+Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.  
+Amit Moryossef, Yoav Goldberg, and Ido Dagan. 2019. Step-by-step: Separating planning from realization in neural data-to-text generation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2267-2277, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Nasrin Mostafazadeh, Lucy Vanderwende, Wen-tau Yih, Pushmeet Kohli, and James Allen. 2016. Story cloze evaluator: Vector space representation evaluation by predicting what happens next. In Proceedings of the 1st Workshop on Evaluating Vector-Space Representations for NLP, pages 24-29, Berlin, Germany. Association for Computational Linguistics.  
+Eyal Orbach and Yoav Goldberg. 2020. Facts2Story: Controlling text generation by key facts. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2329-2345, Barcelona, Spain (Online). International Committee on Computational Linguistics.  
+Sebastian Padó. 2006. User's guide to sigf: Significance testing by approximate randomisation.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia,
+
+Pennsylvania, USA. Association for Computational Linguistics.  
+Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. 2017. Automatic differentiation in pytorch.  
+Xiangyu Peng, Siyan Li, Sarah Wiegreffe, and Mark Riedl. 2021. Inferring the reader: Guiding automated story generation with commonsense reasoning. CoRR, abs/2105.01311.  
+Fabio Petroni, Tim Rocttäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China. Association for Computational Linguistics.  
+Ratish Puduppully and Mirella Lapata. 2021. Data-to-text generation with macro planning. Transactions of the Association for Computational Linguistics, 9:510-527.  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.  
+Hannah Rashkin, Eric Michael Smith, Margaret Li, and Y-Lan Boureau. 2019a. Towards empathetic open-domain conversation models: A new benchmark and dataset. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5370-5381, Florence, Italy. Association for Computational Linguistics.  
+Hannah Rashkin, Eric Michael Smith, Margaret Li, and Y-Lan Boureau. 2019b. Towards empathetic open-domain conversation models: A new benchmark and dataset. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5370-5381, Florence, Italy. Association for Computational Linguistics.  
+SE ROBERTSON, S WALKER, S JONES, MM HANCOCK-BEAULIEU, and M GATFORD. 1995. Okapi at trec-3. NIST special publication, (500225):109-123.  
+Maarten Sap, Ronan Le Bras, Emily Allaway, Chandra Bhagavatula, Nicholas Lourie, Hannah Rashkin, Brendan Roof, Noah A. Smith, and Yejin Choi. 2019. ATOMIC: an atlas of machine commonsense for if-then reasoning. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 3027-3035. AAAI Press.
+
+Abigail See, Aneesh Pappu, Rohun Saxena, Akhila Yerukola, and Christopher D. Manning. 2019. Do massively pretrained language models make better storytellers? In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 843-861, Hong Kong, China. Association for Computational Linguistics.  
+Bowen Tan, Zichao Yang, Maruan Al-Shedivat, Eric Xing, and Zhiting Hu. 2021. Progressive generation of long text with pretrained language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4313-4324, Online. Association for Computational Linguistics.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008.  
+Ivan Vulic, Pei-Hao Su, Samuel Coope, Daniela Gerz, Paweł Budzianowski, Inigo Casanueva, Nikola Mrkšić, and Tsung-Hsien Wen. 2021. ConvFiT: Conversational fine-tuning of pretrained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1151-1168, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Bingning Wang, Ting Yao, Weipeng Chen, Jingfang Xu, and Xiaochuan Wang. 2021. Multi-lingual question generation with language agnostic language model. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2262-2272, Online. Association for Computational Linguistics.  
+Peifeng Wang, Nanyun Peng, Filip Ilievski, Pedro Szekely, and Xiang Ren. 2020. Connecting the dots: A knowledgeable path generator for commonsense question answering. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 4129-4140, Online. Association for Computational Linguistics.  
+Jun Xu, Zeyang Lei, Haifeng Wang, Zheng-Yu Niu, Hua Wu, and Wanxiang Che. 2020a. Enhancing dialog coherence with event graph grounded content planning. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020, pages 3941-3947. ijcai.org.  
+Peng Xu, Mostofa Patwary, Mohammad Shoeybi, Raul Puri, Pascale Fung, Animashree Anandkumar, and Bryan Catanzaro. 2020b. Controllable story generation with external knowledge using large-scale language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2831-2845.
+
+Yang Xu. 2021. Global divergence and local convergence of utterance semantic representations in dialogue. In Proceedings of the Society for Computation in Linguistics 2021, pages 116-124, Online. Association for Computational Linguistics.
+
+Lili Yao, Nanyun Peng, Ralph M. Weischedel, Kevin Knight, Dongyan Zhao, and Rui Yan. 2019. Plan-and-write: Towards better automatic storytelling. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 7378-7385. AAAI Press.
+
+Changlong Yu, Hongming Zhang, Yangqiu Song, and Wilfred Ng. 2020. Cocolm: Complex commonsense enhanced language model. CoRR, abs/2012.15643.
+
+Houyu Zhang, Zhenghao Liu, Chenyan Xiong, and Zhiyuan Liu. 2020a. Grounded conversation generation as guided traverses in commonsense knowledge graphs. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2031-2043, Online. Association for Computational Linguistics.
+
+Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2020b. DIALOGPT: Large-scale generative pre-training for conversational response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270-278, Online. Association for Computational Linguistics.
+
+# A Event Transition Path
+
+The statistics about event transition paths sampled from ATOMIC are shown in Table 6. We display several examples of the sampled event transition paths in Table 7.
+
+# B Implementation Details
+
+For all the systems, including the event transition planner and text generator in our proposed method, we employ the small version of GPT-2 model which is a Transformer with 12-head, 12-layer, and hidden size of 768. The total parameter scalse is 117M. We use pre-trained GPT-2 Byte Pair Encoding (BPE) tokenizer with an extended vocabulary of 50,282 tokens to tokenize texts.
+
+The event prompt length during training ATOMIC event transition paths, EMPATHETICDI-LOGUES paths, and ROCSTORIES paths are 5,
+
+<table><tr><td>Total</td><td>Training</td><td>Validation</td><td>Test</td></tr><tr><td>4,016,468</td><td>3,614,981</td><td>200,752</td><td>200,735</td></tr></table>
+
+Table 6: Numbers of the sampled event transition paths from ATOMIC.  
+
+<table><tr><td>Sampled Event Transition Paths of Variant Lengths</td></tr><tr><td>[1] PersonX earns a bachelor&#x27;s degree XWANT Per-sonX wants to find a good job</td></tr><tr><td>[2] PersonX asks PersonY to join OWANT PersonY wants to be friends XREACT PersonY feels loved</td></tr><tr><td>[3] PersonX is inebriated _XATTR PersonX loses control of PersonX&#x27;s car XREACT PersonX feels scared _OREACT PersonY takes PersonX by force XREACT PersonY feels triumphant</td></tr></table>
+
+Table 7: Event transition paths sampled from daily commonsense reasoning atlas ATOMIC (Sap et al., 2019).
+
+5, and 10, respectively. The dimension of the randomly initialized smaller matrix  $\mathbf{U}_{\theta}^{\prime}$  in Eq.1 is 512.
+
+The batch size is 128 using AdamW optimizer (Loshchilov and Hutter, 2019) with a learning rate of 5e-5. We select the best checkpoint according to the perplexity on the development set of each task and apply early stopping on training where the patient value is set to 2. We adopt the pre-trained BERT-base model to train the event relation classifier. All experiments are implemented by PyTorch framework (Paszke et al., 2017) and run on NVIDIA V100 GPUs. The training time of the event transition planner and event-path-aware text generator are less than 5 hours and 3 hours with 8 GPUs.
+
+# C Case Study
+
+We qualitatively analyze our model predictions and find that although the proposed model outperforms the state-of-the-art baselines, many of predictions are still wrong. Table 8 shows several satisfying and unsatisfying predictions on the two datasets. One significant error originates from the weak alignment between event transition path and final prediction. For example, in the second case, despite "XEFFECT tommy be happy" is imperfect, the prediction "bought it" do not convey its information and makes co-reference mistake (the expected output is "bought them"). Another serious error type is event transition hallucination, where both the predicted event path and its corresponding
+
+inference fail to maintain the logic coherence, such as the fourth case. These problems could be alleviated if we design a better format of event transition path which is easier to learn or improve the relation modeling between events and sentences.
+
+<table><tr><td></td><td>Input</td><td>Corresponding Event Transition Path</td></tr><tr><td>Good Case on Story Completion</td><td>Context: 
+Our granddaughter is two. 
+Today she went to the doctor for a blood draw. 
+She did very well. 
+Our daughter sent a photo of her licking a lollypop afterward. 
+Target: 
+We were very proud of her. 
+Prediction: 
+We were amused by the photo.</td><td>Our granddaughter be two 
+XEFFECT she go doctor for draw 
+XEFFECT she do well 
+OREACT we daughter send a photo XEFFECT she lick lollypop 
+OEFFECT we be proud 
+XEFFECT we get good photo XEFFECT we be happy</td></tr><tr><td>Bad Case on Story Completion</td><td>Context: 
+Tommy wanted to buy a new computer. 
+After some research he decided to build a PC him-self. 
+He found a site that spelled out compatible partslists. 
+He shopped around for the cheapest parts he couldget. 
+Target: 
+The PC he made was more powerful than computers twice its price. 
+Prediction: 
+He finally found the perfect parts list and bought it.</td><td>tommy want buy new computer 
+XEFFECT After research decide build PC 
+XEFFECT he find site 
+XEFFECT he shop around part _XEFFECT he could get 
+XEFFECT he make powerful computer 
+XREACT he find part XEFFECT tommy be happy</td></tr><tr><td>Good Case on Dialogue Generation</td><td>Context: 
+Hi, I joined a firm 6 months ago and then I got apromotion for Junior Manager.Target: 
+Congratulations. That sounds like the fast track.Prediction: 
+That&#x27;s awesome! I bet you are excited!</td><td>i join firm ago XEFFECT i get promotion for manager 
+OREACT that like fast track 
+OREACT that be great XEFFECT you must be proud</td></tr><tr><td>Bad Case on Dialogue Generation</td><td>Context: 
+I got my four year old daughter her first tricycle yesterday.Target: 
+that&#x27;s so sweet of you.Prediction: 
+Wow, that&#x27;s a lot of fun. What kind of tricycle?</td><td>i get my daughter tricycle yesterday 
+OREACT that sweet 
+OEFFECT that be great XEFFECT what be tricycle</td></tr></table>
+
+Table 8: Summary table of issues found through a qualitative analysis of our model predictions. The errors that occur in our model predictions are colored in blue.

eventtransitionplanningforopenendedtextgeneration/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d8c1f7dcc29d41ba0434edcceed0080a85b4ddf5dc2a01952c60c31d362fcc4
+size 629337

eventtransitionplanningforopenendedtextgeneration/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2c708e95fa94cc0bd883ce3fea06ef824bea1127020aa428d715de4eb6f3148
+size 435422

explainingclassesthroughstablewordattributions/7a59d2d8-3f9e-4bb1-a7af-4f4928f35dae_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb72fd80bcb1b5350a35a5004679ba31b6c27a372f5f32d418e69e4de7e11c1a
+size 82119

explainingclassesthroughstablewordattributions/7a59d2d8-3f9e-4bb1-a7af-4f4928f35dae_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:706607177dd40f853186eba9007dc543de4f84e8ba25aebc0fcafedde78faaed
+size 97227