Add Batch 8263ede4-2ce3-43eb-97c6-88d7a993ee10

acomprehensivecomparisonofwordembeddingsinevententitycoreferenceresolution/bfb93e37-ebe9-4a41-a3d3-d2562c21f9ce_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ad7d22ed5f8b38c7cbac5065fe06abcac8762dc58b9760f9a2b76e69f6d9e82
+size 67684

acomprehensivecomparisonofwordembeddingsinevententitycoreferenceresolution/bfb93e37-ebe9-4a41-a3d3-d2562c21f9ce_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9076ccec24527d03023cafd8283151dd851ba4cbfbb58d74e31517886f3961f
+size 82633

acomprehensivecomparisonofwordembeddingsinevententitycoreferenceresolution/bfb93e37-ebe9-4a41-a3d3-d2562c21f9ce_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:443bfaf7a7f3ac4d582e5272ab29083b16c48590973835930b1dacfab137f038
+size 435987

acomprehensivecomparisonofwordembeddingsinevententitycoreferenceresolution/full.md

@@ -0,0 +1,285 @@
+# A Comprehensive Comparison of Word Embeddings in Event & Entity Coreference Resolution.
+
+Judicael Poumay
+
+ULiege/HEC Liege
+
+Rue Louvrex 14, 4000 Liege, Belgium
+
+judicael.poumay@uliege.be
+
+Ashwin Ittoo
+
+ULiege/HEC Liege
+
+Rue louvrex 14, 4000 Liege, Belgium
+
+ashwin.ittoo@uliege.be
+
+# Abstract
+
+Coreference Resolution is an important NLP task and most state-of-the-art methods rely on word embeddings for word representation. However, one issue that has been largely overlooked in literature is that of comparing the performance of different embeddings across and within families in this task. Therefore, we frame our study in the context of Event and Entity Coreference Resolution (EvCR & EnCR), and address two questions: 1) Is there a trade-off between performance (predictive & run-time) and embedding size? 2) How do the embeddings' performance compare within and across families? Our experiments reveal several interesting findings. First, we observe diminishing returns in performance with respect to embedding size. E.g. a model using solely a character embedding achieves  $86\%$  of the performance of the largest model (Elmo, GloVe, Character) while being  $1.2\%$  of its size. Second, the larger model using multiple embeddings learns faster overall despite being slower per epoch. However, it is still slower at test time. Finally, Elmo performs best on both EvCR and EnCR, while GloVe and FastText perform best in EvCR and EnCR respectively.
+
+# 1 Introduction
+
+Coreference Resolution (CR) is an important NLP task. It can be subdivided into Event and Entity Coreference Resolution (EvCR and EnCR). These tasks serves as the basis for several downstream applications such as information extraction, text summarization, machine translation and text mining (Humphreys et al., 1997; Azzam et al., 1999; Miculicich Werlen and Popescu-Belis, 2017; Su et al., 2008).
+
+State-of-the-art methods for CR(Barhom et al., 2019; Lee et al., 2017; Joshi et al., 2019) rely on various word embeddings for word representation. These embeddings are organized into three families: static, contextual and character embeddings
+
+(Almeida and Xexéo, 2019; Liu et al., 2020; dos Santos and Zadrozny, 2014), each differing in size. Contextual embeddings are larger (1024) compared to the other families (usually 300 for static and 50 for character). They also tend to outperform the other families in most tasks but lead to larger and heavier models (Devlin et al., 2019; Peters et al., 2018). We are thus confronted with a trade-off of performance (predictive & run-time) vs. dimensionality. Moreover, embeddings also differ within families which also leads to differences in predictive performance.
+
+Several studies investigated how different embeddings influence the predictive performance in different tasks (Berardi et al., 2015; Gromann and Declerck, 2018; Joshi et al., 2019; Li et al., 2018). However, the two aforementioned issues of the performance vs. dimensionality trade-off and performance variations within and across embedding families have been overlooked to a large extent, especially in coreference resolution. Literature is still unclear about which embeddings perform best in which tasks, and whether larger, more expressive embeddings should also be preferred or whether some predictive performance can be compromised for improved run time.
+
+Thus, we seek to address two questions in the context of CR: 1) Is there a trade-off between performance (predictive & run-time) and embedding size? 2) How do the embeddings' performance compare within and across families? The current state-of-the-art in EvCR (Barhom et al., 2019) rely on three families of embeddings for word representation, and thus provides a suitable frameworks for addressing our research questions. Starting from the original model of Barhom et al. (2019), we performed various experiments and ablative studies across and within each family of embeddings, resulting in 16 different models. We compared
+
+their predictive performance, size (number of parameters), run-time and memory usage.
+
+We discovered high level of diminishing returns in term of predictive performance per embedding. The smallest model (using solely a character embedding (dos Santos and Zadrozny, 2014)) achieves  $86\%$  of the performance of the largest model (GloVe (Pennington et al., 2014), ELMo (Peters et al., 2018), Character embedding) with  $1.2\%$  of its size. Hence, incorporating additional embeddings leads to diminishing returns in terms of predictive performance. In addition, we found that size and run-time are weakly correlated: larger (more complex) models can converge faster (number of epochs and total training time) than smaller ones. In terms of predictive performance, we found GloVe and FastText perform best in EvCR and EnCR respectively in their family with ELMo being the best overall. Moreover, we found that the smallest aforementioned model outperforms Word2Vec ( $\sim +10$  F1), yielding predictive performance close to the previous state-of-the-art (Kenyon-Dean et al., 2018) in EvCR (68.43 vs 69 F1). Our results can have important implications for practitioners in implementing CR and other NLP models in real-life applications.
+
+# 2 Background and Related work
+
+# 2.1 Word embeddings families
+
+Literature generally distinguishes between three families: static, contextual and character embeddings (Almeida and Xexéo, 2019; Liu et al., 2020; dos Santos and Zadrozny, 2014).
+
+Static embeddings, such as word2vec, FastText, and GloVe, create a one-to-one mapping between words and their vector representations. Word2vec (Mikolov et al., 2013) learns through a language modelling task by either learning to predict a word given its context (CBOW) or predict the context given a word (Skip-gram). FastText (Bojanowski et al., 2017) learns sub-words embeddings which are then combined for each word. Finally, GloVe (Pennington et al., 2014) relies on word cooccurrence information. Both Glove and FastText are trained on a Skip-gram task.
+
+Contextual embeddings take into account the context of a given word, i.e. their vector representations changes depending on surrounding words. ELMo is a Bi-LSTM trained on a language modelling task. GPT-2 is similar except that it is unidirectional. Finally, BERT is based on a transformer
+
+architecture and trained on a masked language modelling task.
+
+Lastly, character embeddings learn vectors based on character sequences (dos Santos and Zadrozny, 2014).
+
+Since their development, word embeddings have been very largely studied (Tan et al., 2015; Chen et al., 2018; Wang et al., 2018; Clark et al., 2019; Tenney et al., 2019) and a complete literature review is out of the scope of our work. Hence, we will focus on studies closest to ours. First, we will review studies on embeddings' performance regardless of the task. Then, we move to our task of interest which is coreference resolution.
+
+# 2.2 Studies on Embeddings' Performance
+
+Gromann and Declerck (2018) found that FastText (0.812 F1) outperformed Polyglot (0.675 F1) and Word2Vec (0.750 F1) for ontology alignment. They used two ontologies: Global Industry Classification Standard and Industry Classification Benchmark. They also demonstrated the ability of FastText to better handle out-of-vocabulary words.
+
+Berardi et al. (2015) found that Word2Vec (Accuracy (ACC)  $43.63\%$  ) outperformed polyglot (ACC  $4\%$  ) and GloVe (ACC  $30.21\%$  ) on a word analogy test using Wikipedia and a collection of Italian books (mostly novels) as datasets.
+
+Joshi et al. (2019) found that BERT significantly outperformed ELMo on EnCR (+11.5 F1) on the GAP and OntoNotes datasets.
+
+Li et al. (2018) found that GloVe outperformed FastText and Word2Vec on a tweet classification task, especially when trained on specific corpora, viz.CrisisLexT6, CrisisLexT26, and 2CTweets.
+
+# 2.3 Word embeddings in Coreference Resolution.
+
+Event Coreference Resolution and Entity Coreference Resolution (EvCR and EnCR respectively) are concerned with clustering Event and Entity mentions that refer to the same reality (Barhom et al., 2019; Lee et al., 2017). Figure 1 depicts two event mentions with the same meaning.
+
+SpaceX launched a South Korean Military satellite
+
+South Korea's first military satellite was delivered by SpaceX
+
+Figure 1: Two coreferent event mentions with colors indicating associated coreferent entity mentions.
+
+Events mentions refer to textual representations
+
+of real-life events. As can be seen from Figure 1, events generally consist of a trigger word (most often a verb), such as "launched", and a set of arguments, such as "SpaceX" and "a South Korean Military satellite". Four argument types are generally distinguished: Arg0, Arg1, location, and time, as defined in Barhom et al. (2019), where Arg0 (resp. Arg1) is the closest entity on the left (resp. right) of the trigger word. These arguments are optional and often referred to as entities. The goal of EvCR (and EnCR) is to identify which events (and entities) are coreferent with each other and to cluster them.
+
+We now briefly review studies using word embeddings for EnCR and EvCR.
+
+EnCR: Lee et al. (2017) used GloVe as word representation allied with a Bi-LSTM and attention mechanisms. Their model achieved state-of-the-art (68.8 F1) on the the CoNLL-2012 corpus. As already mentioned, Joshi et al. (2019) reported higher EnCR performance when using BERT compared to ELMo: +3.9 F1 in OntoNotes and +11.5 F1 in GAP.
+
+EvCR : Choubey and Huang (2017) relied on GloVe for EvCR using the ECB+ corpus (Cybulska and Vossen, 2014). They used a joint modelling approach to perform within and cross document EvCR and achieved state-of-the-art performance. The same corpus was employed by Barhom et al. (2019), who proposed an EvCR/EnCR model based on ELMo (Peters et al., 2018), GloVe (Pennington et al., 2014) as well as a fine-tuned character embedding. Similarly, it jointly performs EnCR and EvCR. Their model yielded performance of 79.5 F1 in EvCR.
+
+# 3 Methodology
+
+# 3.1 Original model
+
+Our approach is based on the state-of-the-art model of Barhom et al. (2019), which we refer to as the ORIGINAL  ${}^{2}$  model. This model consists of two neural networks, which jointly resolve entities and events coreferences. Figure 2 shows the input of both networks. The two event (resp. entity) mentions embeddings are in blue and the green box represents an element-wise multiplication of the mentions. Finally, binary features indicate whether the two encoded mentions have coreferent arguments. The constituents of each mention, i.e. trigger, Arg0, Arg1, Location and time, are represented
+
+by a static (GloVe) and a character embedding. The trigger is also represented by a contextual embedding (ELMo). Furthermore, the character embedding is fine tuned during training while the contextual and static embeddings are not.
+
+![](images/aa4405ec70275e31eac32d0fef08a2d4792cba2a20a148c6a964f406db654c0a.jpg)  
+Figure 2: Original input structure of Barhom et al. (2019)'s model.
+
+The input dimensionality is  $3^{*}(1024 + 5^{*}(300 + 50)) + 200 = 8522$ , where 1024, 300 and 50 are the dimensions of ELMo, GloVe and the character embeddings, and 200 corresponds to the size of the binary features. This input is then fed into two subsequent ReLU layers with dimensions equal to half the input dimension (4261 neurons each). Since the number of parameters is proportional to the square of the input dimension, we have a model size exceeding 54 million parameters, computed as  $\left(\frac{\text{input}^2}{2} + \left(\frac{\text{input}}{2}\right)^2 + \frac{\text{input}}{2}\right)$ .
+
+# 3.2 Derived models
+
+The gist of our methodology involves substituting and/or removing specific embeddings from Barhom et al. (2019)'s original model (which uses 3 embeddings : static=GloVe, contextual=ELMo and character), resulting in 16 different models shown in Table 1. In the first group of models, one, two, or three (of the three) embeddings are removed from the original model. In the second group, the static embedding is changed to Word2Vec (Skip-gram) or FastText (other embeddings are either left unchanged or removed). Similarly, in the third group the contextual embedding is changed to BERT or GPT-2 (other embeddings are either left unchanged or removed). Note: in Table 1, gray rows denote identical models.
+
+We implemented our models using Pytorch. Models were trained and tested following Barhom et al. (2019)'s procedure. Pre-trained vectors and models were used for the embeddings. Our code is
+
+available online  ${}^{3}$  .
+
+<table><tr><td>Model</td><td>Stat.</td><td>Ctx.</td><td>Char.</td></tr><tr><td colspan="4">Group 1: Across family study</td></tr><tr><td>Original (2019)</td><td>GloVe</td><td>ELMo</td><td>✓</td></tr><tr><td>Contextual/Static</td><td>GloVe</td><td>ELMo</td><td>X</td></tr><tr><td>Contextual/Char</td><td>X</td><td>ELMo</td><td>✓</td></tr><tr><td>Static/Char</td><td>GloVe</td><td>X</td><td>✓</td></tr><tr><td>Static</td><td>GloVe</td><td>X</td><td>X</td></tr><tr><td>Contextual</td><td>X</td><td>ELMo</td><td>X</td></tr><tr><td>Char</td><td>X</td><td>X</td><td>✓</td></tr><tr><td>No word embed</td><td>X</td><td>X</td><td>X</td></tr><tr><td colspan="4">Group 2: Within family study: Static</td></tr><tr><td>GloVe</td><td>GloVe</td><td>ELMo</td><td>✓</td></tr><tr><td>Word2Vec</td><td>Word2Vec</td><td>ELMo</td><td>✓</td></tr><tr><td>FastText</td><td>FastText</td><td>ELMo</td><td>✓</td></tr><tr><td>Only GloVe</td><td>GloVe</td><td>X</td><td>X</td></tr><tr><td>Only FastText</td><td>Word2Vec</td><td>X</td><td>X</td></tr><tr><td>Only Word2Vec</td><td>FastText</td><td>X</td><td>X</td></tr><tr><td colspan="4">Group 3: Within family study: Contextual</td></tr><tr><td>ELMo</td><td>GloVe</td><td>ELMo</td><td>✓</td></tr><tr><td>BERT</td><td>GloVe</td><td>BERT</td><td>✓</td></tr><tr><td>GPT-2</td><td>GloVe</td><td>GPT-2</td><td>✓</td></tr><tr><td>Only ELMo</td><td>X</td><td>ELMo</td><td>X</td></tr><tr><td>Only BERT</td><td>X</td><td>BERT</td><td>X</td></tr><tr><td>Only GPT-2</td><td>X</td><td>GPT-2</td><td>X</td></tr></table>
+
+Table 1: List of trained and tested model and their components. Ctx. = Contextual; Stat. = Static; Char. = Character; X/√ indicate absence/presence of an input.
+
+# 4 Experimentation setup
+
+# 4.1 Dataset
+
+The dataset we use for our study is ECB+ (Cybulska and Vossen, 2014). Together with EECB (Lee et al., 2012), it is one of the largest datasets for within and cross document EvCR and EnCR (Lee et al., 2012; Barhom et al., 2019). Both EECB and  $\mathrm{ECB + }$  are extensions of ECB (Bejan and Harabagiu, 2010) and consist of English Google News documents clustered into topics and annotated for coreference. For more details on the  $\mathrm{ECB + }$  corpus statistics, please refer to Barhom et al. (2019).
+
+Other dataset for coreference resolution exist: GAP, OntoNotes, CoNLL 2012, ACE, TAC KBP and MUC. However, the definition of coreference resolution in these corpora do not suits our study
+
+and model. For example, GAP is a corpus of ambiguous pronoun-name pairs while ECB+ defines mentions cluster for events and their entities (Joshi et al., 2019). OntoNotes annotates coreferences but does not indicate which mentions is an event and which is an entity. MUC, ACE, and TAC KBP do not provide cross document coreferences(Lu and Ng, 2018). Finally, while CoNLL 2012 defines an event coreference task, events represent only a small portion of the all the coreferent mentions and again it does not provide cross document coreferences (Pradhan et al., 2012). In-depth reviews of the listed datasets are provided in (Stylianou and Vlahavas, 2021; Lu and Ng, 2018; Sukthanker et al., 2018).
+
+# 4.2 Experiments
+
+We performed three sets of experiments. The first set concerns models of Group 1 (see Table 1). We investigated the impact of removing one, two, or three (of the three) embeddings from the original model. Our aim was to determine the contribution of the different embeddings (static, contextual and character) on the predictive performance of the ORIGINAL model. Thus, the models will have varying sizes, translating into varying run-time and memory requirements. Therefore, for this set of experiments, we also report on model size (number of parameters), run-time (seconds) and memory usage (RAM).
+
+The second (third) set concerns models of Group 2 (Group 3) (see Table 1) and aim at investigating the contributions of static (contextual) embeddings.
+
+For the latter two experiments, we do not consider model size as all possible sizes would have been investigated in group 1. For all experiments, we will report the predictive performance achieved by the various models with the CoNLL F1 and MUC F1 metrics (Moosavi and Strube, 2016).
+
+Following Barhom et al. (2019)'s original paper, we can claim that a difference of 1 point between any two models is significant with a p-value  $< 0.001$ . This confirms that our results are statistically sound and not due to randomness.
+
+# 5 Results
+
+# 5.1 Results 1: All Embedding Families
+
+As mentioned earlier, our aim was to investigate the contributions of the static (Glove), contextual (ELMo) and character embedding to the original model's performance via an ablative study. The
+
+predictive performance scores (CoNLL/MUC F1) of Group 1 models are in Figure 3, respectively from left to right.
+
+A first observation is that the baseline performance differs between the two measures (CoNLL & MUC F1). This is due to the mention identification effect (Moosavi and Strube, 2016) which makes CoNLL F1 more optimistic than it should be for low performing models. Interestingly, CoNLL seems more pessimistic than MUC for high performing models. Moreover, Barhom et al. (2019)'s model is helped by using gold cluster for within-document entity coreference. This explains the non-zero MUC F1 performance of the baseline on the entity coreference resolution task.
+
+Another important observation is that, when using only two embedding, the STATIC/CHAR model is the one experiencing the largest drop in performance (CoNLL & Event MUC). At the same time, when using only one embedding, the CONTEXTAL model performs best. It even outperforms the aforementioned model with two embeddings: STATIC/CHAR. These results lead us to conclude that the contextual embeddings is the most expressive for this task. This is not surprising since contextual embeddings take context into account while static and character do not.
+
+More interestingly, we note that removing either the static or contextual embedding results in an average performance drop of  $\sim 2.5$  and  $\sim 4$  CoNLL points respectively (see model CONTEXTUAL/CHAR and STATIC/CHAR). However, when both are removed simultaneously, the performance drops by  $\sim 10$  CoNLL points (see model CHAR). That is, the sum of the losses incurred by removing either one of these embeddings ( $\sim 6.5$ ) is smaller than the loss ( $\sim 10$ ) incurred when both are simultaneously removed. Similarly, adding any one embedding to the baseline NO WORD EMBEDDING model significantly improves the latter's performance, in the range of  $\sim [+27,5$  to  $+34,7]$ . However, if any one embedding is removed from the ORIGINAL model, then the latter's performance drops by a much smaller amount,  $\sim [-1,1$  to  $-4]$ . That is, removing an embedding from the ORIGINAL model does not impact performance in a comparable way as adding an embedding to the baseline model. But performance does drop significantly when all embeddings are removed. In other words, we face diminishing returns in terms of performance per embeddings.
+
+# Impact of Dimensionality on Model Size
+
+As mentioned earlier, the model size is related to the square of the input, resulting in more than 54 million parameters in the ORIGINAL model. Thus, an important question is that of whether the gains in performance of such large models outweigh the corresponding increase in size. Our observations in this respect are in Figure 4, depicting the model's respective size and predictive performance. We observed similar diminishing returns when considering performance relative to size, i.e. increasing the model size by incorporating larger, more complex embeddings results in modest performance gains.
+
+The CONTEXTUAL and CHAR models are particularly interesting. The former achieves  $96\%$  of the performance of the ORIGINAL model with  $14.7\%$  of its size. While the latter, i.e. CHAR, achieves  $86\%$  of the performance of the ORIGINAL model's performance, with only  $1.2\%$  of its size. Its performance (68.43 F1) is even comparable to that of the previous event coreference resolution state-of-the-art in EvCR (69 F1) (Kenyon-Dean et al., 2018).
+
+# Model Size & Run-Time
+
+Our investigations on the influence of model size on run-time and memory usage revealed paradoxical results. They are presented in Figure 5. For the run-time and memory analysis, we focus only on the largest and smallest models to have a better idea of the magnitude of differences and to avoid overcrowding the Figures.
+
+As can be seen, the huge difference in model size (54 Million vs. 0.67 Million), does not translate into equally large the differences in run-time (training & testing) - the run-time reductions afforded by the CHAR model are relatively modest. While the actual reasons deserve further investigation, we can posit that this could be attributed to hardware and software optimization, enabling a high level of parallelization such that larger models run comparably to smaller ones.
+
+Paradoxically, however, the larger ORIGINAL model trains in fewer epochs than the smaller CHAR model (14 vs. 24 respectively). In consequence, it is  $21\%$  faster to train overall (68924.8 sec. vs 87587.28 sec. or about 19h9 vs 24h19). These results confirm the observation of Li et al. (2020)
+
+![](images/86c3271a6219e383688b3635ce546d0d878145ea4ef44c3840d34bfa2291053b.jpg)  
+Ablative study  
+Figure 3: Comparing the predictive performance of the original model (using 3 embeddings) with models where we removed one, two or all three embeddings.
+
+![](images/61630e71d70977eae7082a6d0a8bcb1b882d2b05806e1ab1907becd7f43a01a3.jpg)  
+Ablative study : Model size  
+Figure 4: Comparing the size and predictive performance of the original model (using 3 embeddings) with models where we removed one, two or all three embeddings. The size of each model is the number of neural connections.
+
+that larger models tend to converge faster. One possible explanation could be that larger models have to optimize a error surface of higher dimensionality, leading to more possible paths for gradient descent, some of which might lead to convergence more rapidly. Thus, although adding more embedding in the model results in diminishing returns in term of predictive performance, it can lead to faster training. However, more experiments are needed to investigate this issue.
+
+Concerning memory usage, we found that, as expected, the smaller CHAR model required substantially smaller amounts of memory, especially during training as evidence by Figure 6. Note that, the RAM usage of the ORIGINAL model is mostly due to GloVe pre-trained vectors.
+
+# 5.2 Results 2: Static Embeddings
+
+We now focus on the second set of experiments, focusing our attention to static embeddings. The models concerned are from Group 2 of Table 1.
+
+First, we varied the static embedding (GloVe, Word2Vec, FastText), while keeping the same contextual embedding and character embedding as in the ORIGINAL model. It can be seen in Figure 7 that, when used with other embeddings (contextual and character), all static embeddings show comparable performance. The average performance rang-
+
+![](images/e9cfbe33ec4112e0b7626dcbefac0c9c353c7b0f81b835683dc0cce77a844495.jpg)  
+Figure 5: Run-time between the largest (54M weights) and smallest (677k weights) models. The total training time is associated with the right axis while the other measures are associated with the left axis.
+
+ing from 77.12 (GLOVE) to 75.59 (WORD2VEC). This corroborates with our earlier findings of section 5.1 whereby the model with only contextual and character embeddings, i.e. CONTEXTUAL/CHAR, achieved comparable performance to the ORIGINAL (static/contextual/char) model, indicating that the specific static embedding chosen contribute only marginally to the model's performance.
+
+However, when used alone (see Figure 8), we see a drastic difference in performance between them; with the average performance ranging from 72.73 (GLOVE) to 51.56 (WORD2VEC).
+
+Thus, it is only when studied alone that static embeddings show their differences. Once we iso
+
+![](images/f961b289964e2b8ba2aada2f63ded8195c691d094f66a56a6ab0e8136f3bd8bb.jpg)  
+Figure 6: Memory usage between the largest (54M weights) and smallest (677k weights) models.
+
+![](images/f9ededface23c3fb02a30f25a7b43a45e79c3b34d6736bb2a62fce9d9a463b06.jpg)  
+Figure 7: Comparing the predictive performance of static embeddings when used with other embeddings (ELMo and Character)
+
+late static embeddings, we see GloVe works best for EvCR. However, for EnCR, the FASTTEXT model shows significantly higher MUC. The better performance of GloVe and FastText with respect to word2vec can be explained by their construction. Compared to Word2Vec, GloVe takes words co-occurrence information into account. If coreferent event mentions are more likely to share co-occurring words, it would explain parts of the performance gain. FastText also outperforms Word2Vec; here the difference is that FastText takes sub-word information into account which can be advantageous for coreferent entity mentions. E.g. in Figure 1, "Korea" and "Korean" have similar sub-word information.
+
+What is most surprising is that Word2Vec is significantly outperformed by a simple character embedding as we can see on Figure 9. Moreover, in term of dimension Word2Vec has 300 and the character embedding has 50. Thus, the resulting model is not only more accurate but also  $\sim 24$  times smaller (Figure 9). This could indicate that the internal structure of a word (char embedding) contains more information about possible coreferences than its usual entourage (Word2Vec).
+
+![](images/8f114007d33c6094f9627d43cb55594009a56b80ba863a5b2f1fc775df247229.jpg)  
+Figure 8: Comparing the predictive performance of static embeddings when used alone
+
+![](images/ffc90b47f0c02d3df6627e08b9a7eb80c4135cf0ca8b3ec3171d45cc54af14c7.jpg)  
+Figure 9: Comparing the predictive performance of solely Word2Vec vs solely a character embedding
+
+# 5.3 Results 3: Contextual Embeddings
+
+We now focus on the third set of experiments about contextual embeddings. The models concerned are from Group 3 of Table 1.
+
+Similarly to the previous section, we present the performance of different contextual embeddings when used in tandem with the static (GloVe) and character embedding of the original model (Figure 10) or when used alone (Figure 11). We see the same as in the previous section, i.e. the difference in performance between the contextual embeddings is clearer when they are used alone versus when they are used with GloVe and a character embedding. Thus, we will only focus on the Figure 11 which better represent the differences between ELMo, BERT, and GPT-2.
+
+A first observation is that BERT both outperforms and is outperformed by GPT-2 on both tasks. Specifically, BERT performs better in EvCR while GPT-2 performs better in EnCR.
+
+A second observation is that ELMo clearly outperforms GPT-2 and BERT on both tasks. This result contradicts Joshi et al. (2019) who found that BERT greatly outperforms ELMo on EnCR (+11.5 F1 on the GAP benchmark). Such disparity may be indicative of differences in the model and dataset.
+
+![](images/a791991aa1628797feb67fcd245ec840dcfe15924175a60ccbe2ac60047f45b4.jpg)  
+Figure 10: Comparing the predictive performance of contextual embeddings when used with other embeddings (GloVe and Character embedding)
+
+![](images/4350dc20397b95bd459b2f6061ab66f80aa617be7605558635ee513b7354a493.jpg)  
+Figure 11: Comparing the predictive performance of contextual embeddings when used alone
+
+Joshi et al. (2019) uses a span ranking approach which asks, for each mention, which is the most likely antecedent. This implicitly produces a tree which clusters coreferent mentions. Such method only takes local information between two mention into account while the method used in Barhom et al. (2019) uses global information between two entity clusters and related event clusters. Moreover,  $\mathrm{ECB + }$  or EventCorefBank  $^+$  is an EvCR dataset first and foremost and only defines EnCR to support EvCR; you could argue that the EnCR tasks is more about argument than entities. GAP on the other hand is a corpus of ambiguous pronoun-name pairs (Joshi et al., 2019).
+
+Thus, while an EnCR task is defined by both dataset, they are significantly different. We argue that both the task definition and the use of global versus local information play a major role in the disparity between the performance reported by Joshi et al. (2019) and our study. Further confirming these findings would require evaluating Barhom et al. (2019)'s model on GAP and Joshi et al. (2019)'s on  $\mathrm{ECB + }$ . However, these models are not interchangeable because the datasets and the task they define differs.
+
+# 6 Conclusion
+
+We used the state-of-the-art in EvCR (Barhom et al., 2019) as a framework to investigate the complexity-performance trade-off and compare the predictive performance of word embeddings across and within the three families.
+
+We observed that the smallest model using solely a character embedding yielded  $86\%$  of the performance of the original (largest) model (using Elmo, GloVe and a character embeddings) despite being only  $1.2\%$  of its size. In fact, that smallest model achieves similar performance (68.43 F1) to the previous state-of-the-art in EvCR (69 F1) (Kenyon-Dean et al., 2018).
+
+Paradoxically, we found that the largest model converged faster during training (by  $21\%$  in overall run-time) as it took only 14 epochs vs 24 for the character model. Overall, we found size and runtime to be weakly correlated.
+
+In addition, our experiments revealed that augmenting the model with additional embeddings does not substantially improve the performance, leading to diminishing returns in term of predictive performance per embedding.
+
+Concerning predictive performance, one of our most interesting result is that the model using solely a character embedding significantly outperformed  $(\sim +10$  F1) a larger model using solely a static embedding (Word2Vec) while being radically smaller (4% of its size). Hence, while character embeddings have often been used as supplementary embeddings, they can actually compete with other embeddings' families in terms of predictive performance per size.
+
+Finally, our experiments lead us to conclude that for the task of Event and Entity Coreference Resolution, GloVe, FastText and Elmo yielded the best predictive performance. GloVe and FastText performed best in EvCR and EnCR respectively in their family while Elmo performs best overall.
+
+Future directions include working on other comprehensive study of embeddings in other tasks and experimenting with CR models using different embeddings for different tasks to improve performance. E.g. GloVe and FastText in EvCR and EnCR respectively.
+
+# 7 Ethical considerations
+
+We trained 16 models over a two months period, estimated cost ranges from  $350\mathrm{kWh}$  to  $400\mathrm{kWh}$ . The estimated carbon impact ranges from  $105\mathrm{Kg}$
+
+to  $120\mathrm{Kg}$  of CO2 based on local data (300g CO2/kWh). We believe no other ethical considerations are raised by the content of this paper.
+
+# Acknowledgments
+
+This research was funded by KPMG Belgium & Luxembourg through the HEC Digital Lab/HEC-Liège/ULiège.
+
+# References
+
+Felipe Almeida and Geraldo Xexéo. 2019. Word embeddings: A survey. CoRR, abs/1901.09069.  
+Saliha Azzam, Kevin Humphreys, and Robert Gaizauskas. 1999. Using coreference chains for text summarization. In Coreference and Its Applications.  
+Shany Barhom, Vered Shwartz, Alon Eirew, Michael Bugert, Nils Reimers, and Ido Dagan. 2019. Revisiting joint modeling of cross-document entity and event coreference resolution. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4179-4189, Florence, Italy. Association for Computational Linguistics.  
+Cosmin Bejan and Sanda Harabagiu. 2010. Unsupervised event coreference resolution with rich linguistic features. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1412-1422, Uppsala, Sweden. Association for Computational Linguistics.  
+Giacomo Berardi, Andrea Esuli, and Diego Marcheggiani. 2015. Word embeddings go to italy: A comparison of models and training datasets. In IIR.  
+Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information. Transactions of the Association for Computational Linguistics, 5:135-146.  
+Juntian Chen, Yubo Tao, and Hai Lin. 2018. Visual exploration and comparison of word embeddings. Journal of Visual Languages & Computing, 48:178 - 186.  
+Prafulla Kumar Choubey and Ruihong Huang. 2017. Event coreference resolution by iteratively unfolding inter-dependencies among events. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2124-2133, Copenhagen, Denmark. Association for Computational Linguistics.  
+Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. 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.
+
+Agata Cybulska and Piek Vossen. 2014. Using a sledgehammer to crack a nut? lexical diversity and event coreference resolution. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14), pages 4545-4552, Reykjavik, Iceland. European Language Resources Association (ELRA).  
+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.  
+Cicero Nogueira dos Santos and Bianca Zadrozny. 2014. Learning character-level representations for part-of-speech tagging. In Proceedings of the 31th International Conference on Machine Learning, ICML 2014, Beijing, China, 21-26 June 2014, volume 32 of JMLR Workshop and Conference Proceedings, pages 1818-1826. JMLR.org.  
+Dagmar Gromann and Thierry Declerck. 2018. Comparing pretrained multilingual word embeddings on an ontology alignment task. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).  
+Kevin Humphreys, Robert Gaizauskas, and Saliha Azzam. 1997. Event coreference for information extraction. In *Operational Factors in Practical, Robust Anaphora Resolution* for Unrestricted Texts.  
+Mandar Joshi, Omer Levy, Luke Zettlemoyer, and Daniel Weld. 2019. BERT for coreference resolution: Baselines and 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 5803-5808, Hong Kong, China. Association for Computational Linguistics.  
+Kian Kenyon-Dean, Jackie Chi Kit Cheung, and Doina Precup. 2018. Resolving event coreference with supervised representation learning and clustering-oriented regularization. In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, pages 1-10, New Orleans, Louisiana. Association for Computational Linguistics.  
+Heeyoung Lee, Marta Recasens, Angel Chang, Mihai Surdeanu, and Dan Jurafsky. 2012. Joint entity and event coreference resolution across documents. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 489-500, Jeju Island, Korea. Association for Computational Linguistics.
+
+Kenton Lee, Luheng He, Mike Lewis, and Luke Zettle-moyer. 2017. End-to-end neural coreference resolution. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 188-197, Copenhagen, Denmark. Association for Computational Linguistics.  
+Hongmin Li, Xukun Li, Doina Caragea, and Cornelia Caragea. 2018. Comparison of word embeddings and sentence encodings as generalized representations for crisis tweet classification tasks. Proceedings of ISCRAM Asia Pacific.  
+Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joseph E. Gonzalez. 2020. Train large, then compress: Rethinking model size for efficient training and inference of transformers.  
+Qi Liu, Matt J. Kusner, and Phil Blunsom. 2020. A survey on contextual embeddings.  
+Jing Lu and Vincent Ng. 2018. Event coreference resolution: A survey of two decades of research. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, July 13-19, 2018, Stockholm, Sweden, pages 5479-5486. ijcai.org.  
+Lesly Miculicich Werlen and Andrei Popescu-Belis. 2017. Using coreference links to improve Spanish-to-English machine translation. In Proceedings of the 2nd Workshop on Coreference Resolution Beyond OntoNotes (CORBON 2017), pages 30-40, Valencia, Spain. Association for Computational Linguistics.  
+Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space.  
+Nafise Sadat Moosavi and Michael Strube. 2016. Which coreference evaluation metric do you trust? a proposal for a link-based entity aware metric. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 632-642, Berlin, Germany. Association for Computational Linguistics.  
+Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar. Association for Computational Linguistics.  
+Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 2227-2237, New Orleans, Louisiana. Association for Computational Linguistics.
+
+Sameer Pradhan, Alessandro Moschitti, Nianwen Xue, Olga Uryupina, and Yuchen Zhang. 2012. CoNLL-2012 shared task: Modeling multilingual unrestricted coreference in OntoNotes. In Joint Conference on EMNLP and CoNLL - Shared Task, pages 1-40, Jeju Island, Korea. Association for Computational Linguistics.  
+Nikolaos Stylianou and Ioannis Vlahavas. 2021. A neural entity coreference resolution review. Expert Systems with Applications, 168:114466.  
+Jian Su, Xiaofeng Yang, Huaqing Hong, Yuka Tateisi, and Jun'ichi Tsujii. 2008. Coreference Resolution in Biomedical Texts: a Machine Learning Approach. In Ontologies and Text Mining for Life Sciences: Current Status and Future Perspectives, number 08131 in Dagstuhl Seminar Proceedings, Dagstuhl, Germany. Schloss Dagstuhl - Leibniz-Zentrum für Informatik, Germany.  
+Rhea Sukthanker, Soujanya Poria, Erik Cambria, and Ramkumar Thirunavukarasu. 2018. Anaphora and coreference resolution: A review.  
+Luchen Tan, Haotian Zhang, Charles Clarke, and Mark Smucker. 2015. Lexical comparison between Wikipedia and Twitter corpora by using word embeddings. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 657-661, Beijing, China. Association for Computational Linguistics.  
+Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593-4601, Florence, Italy. Association for Computational Linguistics.  
+Yanshan Wang, Sijia Liu, Naveed Afzal, Majid Rastegar-Mojarad, Liwei Wang, Feichen Shen, Paul Kingsbury, and Hongfang Liu. 2018. A comparison of word embeddings for the biomedical natural language processing. Journal of Biomedical Informatics, 87:12-20.

acomprehensivecomparisonofwordembeddingsinevententitycoreferenceresolution/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7367cf2972093a78ede49fcc5cd3e37968587747966aba45cd2848145261987
+size 335262

acomprehensivecomparisonofwordembeddingsinevententitycoreferenceresolution/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b80e4b66394d2d28c43e70bbcf6950b9b05d627a4e6ba36f818ab72ff57a66f9
+size 303910

acomputationalexplorationofpejorativelanguageinsocialmedia/d49c7cbd-60a9-4972-933e-dbbe723e6ed5_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ffa44dd631b3d74b990f202a850ea9a6f09074607ec161b423f651183f6163d
+size 41310

acomputationalexplorationofpejorativelanguageinsocialmedia/d49c7cbd-60a9-4972-933e-dbbe723e6ed5_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4914b472fa9a6b67129ec6bf2ade7cbb1723e49e68ab942c2b27dc76a7d48e6
+size 51073

acomputationalexplorationofpejorativelanguageinsocialmedia/d49c7cbd-60a9-4972-933e-dbbe723e6ed5_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1cce81c502d6e955421066cc93b91400df8c88841ad9da3575cd7af428731f47
+size 191388

acomputationalexplorationofpejorativelanguageinsocialmedia/full.md

@@ -0,0 +1,161 @@
+# A Computational Exploration of Pejorative Language in Social Media
+
+Liviu P. Dinu<sup>1</sup>, Ioan-Bogdan Iordache<sup>1</sup>, Ana Sabina Uban<sup>1</sup>, Marcos Zampieri<sup>2</sup>
+
+<sup>1</sup>University of Bucharest, Romania
+
+$^{2}$ Rochester Institute of Technology, USA
+
+ldinu@fmi.unibuc.ro, iordache.bogdan1998@gmail.com
+
+ana.uban@gmail.com, mazgla@rit.edu
+
+# Abstract
+
+In this paper we study pejorative language, an under-explored topic in computational linguistics. Unlike existing models of offensive language and hate speech, pejorative language manifests itself primarily at the lexical level, and describes a word that is used with a negative connotation, making it different from offensive language or other more studied categories. Pejorativity is also context-dependent: the same word can be used with or without pejorative connotations, thus pejorativity detection is essentially a problem similar to word sense disambiguation. We leverage online dictionaries to build a multilingual lexicon of pejorative terms for English, Spanish, Italian, and Romanian. We additionally release a dataset of tweets annotated for pejorative use. Based on these resources, we present an analysis of the usage and occurrence of pejorative words in social media, and present an attempt to automatically disambiguate pejorative usage in our dataset.
+
+# 1 Introduction
+
+With the increase of social media usage, the issue of toxic language has become an important problem in our society. Automatic methods are needed to help mitigate this problem, and for this reason the study of toxic speech in NLP has become very popularity in recent years. Different categories and definitions have been proposed, including hate speech (Schmidt and Wiegand, 2017; Vashistha and Zubiaga, 2021), offensive language (Zampieri et al., 2019; Bucur et al., 2021), aggression (Kumar et al., 2018, 2020), as well as further sub-categories depending on the targets, such as women, migrants, etc. (Basile et al., 2019). From a computational perspective, the problem is usually approached as a classification task at the post level, where a classifier is trained to predict whether a social media post contains offensive/toxic language.
+
+In this paper we address the question of pejorative words. Pejorative words are words or phrases that have negative connotations or that are intended to disparage or belittle<sup>1</sup>. Pejorativity is closely related to the notion of slurs or insults: “as noun phrases, ‘insult’ and ‘slur’ refer to symbolic vehicles designed by convention to derogate targeted individuals or groups” (Anderson and Lepore, 2013). While pejorative language is often used in offensive speech (Castroviejo et al., 2020), they are not identical categories. There are offensive posts that do not use pejorative words (e.g. “Women belong in the kitchen”), and pejorative uses of words that are not harmful (“What a shitty chair”) because the offensive content is not targeted at a person or a group as described in the popular annotation taxonomy of the Offensive Language Identification Dataset (OLID) (Zampieri et al., 2019).
+
+Words can have a negative meaning in one context and not in others (such as the figurative meanings of "trash" or "pussy"); or be pejorative in one language or culture, and not in others (such as the Romanian "cioara" (literally, "crow") - a slur for people of color). Slurs can also lose their pejorative meaning through semantic change (e.g. the word "queer" went through semantic amelioration over the years - it used to be a slur and is losing its negative connotation (Brontsema, 2004)). Recognizing the complexity of the phenomenon, with its linguistic subtleties as well as the variability related to culture and context, are important to successfully recognize pejorative words and by extension offensive posts and hate speech.
+
+Pejorative language is still largely underexplored in computational linguistics. There are very few studies addressing or taking pejorative language into account (Wiegand et al., 2018; Mendelsohn et al., 2020; Palmer et al., 2017; Eder et al., 2019; Castroviejo et al., 2020). A few related works
+
+to ours include Palmer et al. (2017) who focused on pejorative connotations for nominalized adjectives and Mendelsohn et al. (2020) who built a lexicon of vulgar terms (and vulgarity scores) for German based on derogatory terms found in Wiktionary.
+
+In this study, we address this important gap by leveraging dictionaries to build a multilingual lexicon of pejorative language for four languages. We compare the occurrence of pejorativity in social media with other established categories of toxic language, relying on existing hate speech corpora. Unlike most existing studies in hate speech and offensive language identification, our paper focuses on the lexical level and approaches the issue of ambiguity in toxic language, formulating the problem of pejorativity detection as a word sense disambiguation (WSD) task. The main contributions of this work are the following:
+
+1. We create a multilingual lexicon of pejorative words in four languages: English, Spanish, Italian, and Romanian.  
+2. We present several experiments to automatically distinguish pejorative from nonpejorative uses of words relying on state-of-the-art word sense representations based on contextual embeddings.  
+3. We release annotated datasets containing pejorative words in English and Spanish tweets.
+
+# 2 Pejorative Lexicon
+
+# 2.1 Data Collection
+
+We started by gathering a pejorative lexicon for four languages: English, Spanish, Italian and Romanian. For each language, we assembled a list of words that can be used with a pejorative sense according to existing language resources. We focused on providing a lexicon consisting of words that can be used pejoratively on their own, rather than words that are part of pejorative expressions or idioms. In order to collect these terms for English, Spanish, and Italian we used Wiktionary $^2$ , and collected the terms that were part of the "derogatory terms" category. For Romanian, we used another online-available dictionary, dexonline $^3$ , and selected all of the words that had a pejorative definition and where the definition was intended for the word not for an expression built around the word.
+
+# 2.2 Lexicon Description
+
+For each language's lexicon, we computed the frequency of each word, based on occurrence across different large corpora including Wikipedia and social media datasets, using the wordfreq Python library (Speer et al., 2018). We used the WordNet (Miller, 1995) to count the number of senses a word can have (by counting the number of synsets that they are contained in) as well as their parts of speech. Statistics are shown in Table 1. The distribution across parts of speech is illustrated in Figure 1. For a given word, we counted all its possible parts of speech according to WordNet.
+
+<table><tr><td>Lang.</td><td>Words</td><td>WF cover.</td><td>WN cover.</td><td>Senses</td></tr><tr><td>EN</td><td>2903</td><td>28.97%</td><td>25.56%</td><td>3.07</td></tr><tr><td>ES</td><td>881</td><td>51.99%</td><td>18.05%</td><td>3.05</td></tr><tr><td>IT</td><td>149</td><td>53.02%</td><td>49.66%</td><td>1.87</td></tr><tr><td>RO</td><td>770</td><td>12.34%</td><td>32.21%</td><td>2.41</td></tr></table>
+
+Table 1: Number of words for each language, coverage in wordfreq, WordNet coverage, and average number of senses for words in WordNet.
+
+![](images/3be3eeb4b0d594bb549a3e70413c8e1da66b72ba5c20add0ac375aac1a4d0389.jpg)  
+Figure 1: Distribution of parts of speech for the collected words for each language in WordNet.
+
+# 3 Pejorative Tweet Dataset
+
+For building a data set of English texts containing words that are used pejoratively, we started by looking at three datasets of hate speech on Twitter: (Davidson et al., 2017), (Basile et al., 2019). (Waseem and Hovy, 2016), and selected the tweets that contain words from our pejorative lexicon (after normalizing words to their stems). For each data set, we extracted pairs of words and tweets where they occur.
+
+The dataset published by Davidson et al. (2017) contains tweets annotated with one of three classes (hateful, offensive and neither). For each label, the
+
+number of pejorative words found in the tweets is the following: 1, 114 out of 1, 430 hateful tweets, 8, 358 out of 19, 190 offensive tweets, and 2, 221 among the remaining 4, 163 tweets were found to contain pejorative words. The hate speech dataset published as part of the HatEval shared task (Basile et al., 2019) contains tweets annotated with labels for hateful and aggressive speech. Out of the 4, 210 hateful tweets, 1, 985 contain words from our lexicon, while from 1, 763 aggressive tweets, 822 were selected. Finally, the dataset by Waseem and Hovy (2016) contains tweets annotated for racist and sexist speech. 8 tweets out of the 1, 970 racist tweets, and 897 from 3, 378 sexist tweets, contain pejorative words.
+
+For Spanish, we employed the same technique of filtering tweets. We looked at the Spanish tweets data set provided by Basile et al. (2019) and considered only the binary label for hate speech classification. Out of the total of 5,000 tweets, we have extracted 1,621 hateful examples and 1,667 non-hateful examples that contain words from our Spanish pejorative lexicon.
+
+# 3.1 Annotation
+
+We then built a data set of English tweets annotated for pejorative usage of words, by selecting tweets from the HatEval data set (Davidson et al., 2017), which we chose given the large number of unique pejorative words it contains (1, 77 for hate, 3, 95 for offensiveness and 2, 77 for none). We extracted two separate data sets in two different ways.
+
+The first data set (PEJOR1) was built by selecting a fixed percentage of tweets from each class, in order to obtain a balanced dataset with respect to the three labels (keeping only words that are represented at least once in each class). In this way, we attempt to conserve the relative distribution of the pejorative stems across the three classes.
+
+The second data set (PEJOR2) was built to be balanced with regard to both the words' distribution and the original labels. For each pejorative stem we extracted a fixed number of pairs from each of the three classes.
+
+The selected tweet-word pairs extracted for both of the data sets were then annotated with binary valued labels, denoting whether the word in the pair is used pejoratively (label 1) or not (label 0) in the tweet. We used the Wiktionary definitions in order to label words as pejorative only when used with senses marked as "derogatory" in Wiktionary.
+
+The Table 2 shows statistics for the two datasets, while Figure 2 illustrates the distribution of labels for words in PEJOR2. Data was annotated by specialists in linguistics. We used two annotators for each datapoint, and used a third one where there was disagreement. The obtained Cohen's  $k$  agreement score was 0.933.
+
+![](images/5ad5d33848efdeea45bbd566a12ad90a79965f7b34f8fc3c3714cd6c2714099d.jpg)  
+Figure 2: Distribution of labels for the PEJOR2 English dataset.
+
+<table><tr><td colspan="4">PEJOR1</td><td colspan="4">PEJOR2</td></tr><tr><td></td><td>pairs 944</td><td>words 23</td><td>label 1 49.7%</td><td></td><td>pairs 313</td><td>words 11</td><td>label 1 51.4%</td></tr><tr><td></td><td>hate</td><td>offensive</td><td>neither</td><td></td><td>hate</td><td>offensive</td><td>neither</td></tr><tr><td>0</td><td>8.04%</td><td>21.59%</td><td>20.74%</td><td>0</td><td>12.46%</td><td>15.34%</td><td>20.77%</td></tr><tr><td>1</td><td>27.20%</td><td>14.07%</td><td>8.36%</td><td>1</td><td>21.09%</td><td>17.89%</td><td>12.46%</td></tr></table>
+
+Table 2: Number of tweet-word pairs in the datasets, number of unique words, and the frequency of the 1 label. Overlap with (Davidson et al., 2017) labels.
+
+For Spanish, we built a pejorative data set by selecting tweets from the (Basile et al., 2019) data set, following the same approach used for extracting the PEJOR2 English examples. We annotated a small subset of the tweets, consisting of 12 pejorative words with 10 tweets each (balanced between hateful and non-hateful tweets).
+
+# 4 Classification Experiments
+
+The classification task we approached was inferring the 0/1 label for tweet-word pairs. Namely, given a word and a tweet, where the word appears in the tweet, we want to be able to say if the word was used pejoratively or not in that tweet.
+
+In order to prepare our data, for each tweet-word pair, the tweet was tokenized and the position of the occurrence of the word was found among the tokens. Then, we generated a contextual embedding (Devlin et al., 2019) for that occurrence, by employing various BERT models, pre-trained on
+
+English texts, provided by the huggingface Python library (Wolf et al., 2019). The embedding obtained for the specified position is computed by summing the 768-dimensional hidden states generated for that position by each of the 12 layers of the BERT architecture. We note that, for out-of-vocabulary words, the BERT tokenizer provided by the huggingface library splits them into sub-words. In this case we chose to generate the embeddings for each of the sub-words of our word occurrence and then average them to obtain the final 768-dimensional embedding.
+
+Figure 3 illustrates an example of uses of a pejorative word ("cracker") in the PEJOR2 dataset, by representing its embeddings reduced to two dimensions using PCA. We can see that most of the similar labelled examples are clustered together.
+
+![](images/c9989ad05d665b239abc781c58bfcc7d30f1c02a167bc500aa4ba5ef0b97c022.jpg)  
+Figure 3: 2D plot of the contextual embeddings generated for the word 'cracker' in the PEJOR2 data set, for each of its occurrences in the tweets, using a pretrained BERT model. Embeddings were reduced to two dimensions using PCA.
+
+For classification on our English data set, we grouped the pairs by the pejorative word contained in the tweet, and independently for each group, we fitted a classifier on the contextual embeddings (Liu et al., 2020). For extracting the embeddings we used various transformer models (BERT base (Devlin et al., 2019), BERTweet (Nguyen et al., 2020), RoBERTa (Liu et al., 2019), Multilingual BERT (Devlin et al., 2019)) and for the classification algorithm we used K-Nearest Neighbors, Support Vector Machines (SVM), Multilayer Perceptron (MLP). For K-Nearest Neighbors, we considered the cosine similarity as the distance function and found through hyper-parameter tuning that neighborhoods of size 4 were the best performing setting.
+
+For evaluation, we employed a 5-fold cross-validation. Performance metrics were computed for
+
+each word independently, measuring the capacity of distinguishing the pejorative and non-pejorative usage of the word in different contexts. We report, for each metric, the value resulted by averaging over the scores obtained for all of the word groups. We leave out from this averaging the words that appear with only one label in the whole data set (only pejorative or only non-pejorative), since they will be always classified correctly regardless of the contextual embeddings. We also employed a baseline that based on the training data it learns to predict only the most frequent label. Table 3 shows the obtained results. The appendix contains a table with nearest neighbors found for example tweets.
+
+We notice a promising performance of the classifiers in distinguishing pejorative usage, of up to 0.86 F1-score. Following the best performing models for each data set, overall 107 samples were misclassified in the PEJOR1 dataset, while for PEJOR2 there were 37. Words in PEJOR2 seem slightly easier to classify, which might be expected given the dataset is more balanced in positive and negative examples.
+
+<table><tr><td rowspan="2">Dataset Embeddings</td><td rowspan="2">Classifier</td><td colspan="2">PEJOR1</td><td colspan="2">PEJOR2</td></tr><tr><td>Acc</td><td>F1</td><td>Acc</td><td>F1</td></tr><tr><td>—</td><td>baseline</td><td>67.7%</td><td>0.604</td><td>67.3%</td><td>0.694</td></tr><tr><td>BERT base</td><td>4-NN</td><td>76.9%</td><td>0.776</td><td>81.1%</td><td>0.841</td></tr><tr><td>BERT base</td><td>SVM</td><td>79.2%</td><td>0.768</td><td>80.3%</td><td>0.837</td></tr><tr><td>BERT base</td><td>MLP</td><td>79.8%</td><td>0.801</td><td>82.5%</td><td>0.864</td></tr><tr><td>RoBERTa</td><td>4-NN</td><td>72.6%</td><td>0.724</td><td>67.7%</td><td>0.716</td></tr><tr><td>RoBERTa</td><td>SVM</td><td>72.1%</td><td>0.654</td><td>68.9%</td><td>0.692</td></tr><tr><td>RoBERTa</td><td>MLP</td><td>76.4%</td><td>0.781</td><td>77.2%</td><td>0.802</td></tr><tr><td>BERTweet</td><td>4-NN</td><td>80.4%</td><td>0.797</td><td>75.4%</td><td>0.776</td></tr><tr><td>BERTweet</td><td>SVM</td><td>78.0%</td><td>0.760</td><td>77.9%</td><td>0.793</td></tr><tr><td>BERTweet</td><td>MLP</td><td>81.9%</td><td>0.802</td><td>78.1%</td><td>0.803</td></tr><tr><td>Multilg. BERT</td><td>4-NN</td><td>71.0%</td><td>0.714</td><td>74.2%</td><td>0.784</td></tr><tr><td>Multilg. BERT</td><td>SVM</td><td>73.0%</td><td>0.657</td><td>74.3%</td><td>0.786</td></tr><tr><td>Multilg. BERT</td><td>MLP</td><td>76.9%</td><td>0.750</td><td>75.1%</td><td>0.796</td></tr></table>
+
+Table 3: Performance scores for various contextual embeddings and classifiers on the PEJOR1 and PEJOR2 English data sets
+
+For the Spanish pejorative data set, since most of the examples were not labelled, we tried an unsupervised clustering approach. For each group of example pairs defined by the common pejorative word, we extracted contextual embeddings using the same previously explained method. Using KMeans clustering, we grouped those embeddings into two classes. We then computed, using the annotated examples, the amount of overlap between those two clusters and the pejorative la
+
+bels. The overlap was computed as the accuracy and the macro-F1 score of the clusters when used for predicting the labels. We averaged the scores computed for all of the groups where there was at least one positively and one negatively labelled example. The results obtained using various embeddings (BETO (Cañete et al., 2020) and Multilingual BERT (Devlin et al., 2019)) can be found in table 4. For reference, we have used the random chance of assigning the clusters as a baseline.
+
+<table><tr><td>Method</td><td>Accuracy</td><td>F1 score</td></tr><tr><td>random chance</td><td>50.0%</td><td>0.488</td></tr><tr><td>BETO</td><td>68.9%</td><td>0.573</td></tr><tr><td>Multilingual BERT</td><td>65.0%</td><td>0.503</td></tr></table>
+
+Table 4: Overlap score for unsupervised clustering on the Spanish pejorative data set
+
+# 5 Conclusions
+
+We have addressed an important but under-explored lexical category in the intersection of lexical semantics and toxic speech: pejorativity. We released a public lexicon of pejorative words in four languages (including a low-resource language), as well as dataset of tweets annotated for pejorative uses of words.<sup>4</sup> We have modelled pejorativity detection as a problem of disambiguation, and performed experiments using state-of-the-art contextual embeddings in order to automatically distinguish pejorative from non-pejorative uses of words, obtaining promising results. In the future, we would like to explore modelling the problem of pejorativity detection as a sequence labelling task.
+
+At the application level, integrating pejorativity detection into hate speech detection systems, for example, would be a promising area for future research. From a linguistic perspective, it would be interesting to analyze occurrence and pejorative value cross-lingually taking advantage of large pretrained cross-lingual models as in Ranasinghe and Zampieri (2020, 2021) for offensive language identification. We expect pejorative connotations to be difficult to translate and not transfer well across languages, which could also have practical implications. We would also like to extend our dataset of social media posts to cover more pejorative terms, as well as other languages.
+
+# Ethical Considerations
+
+Our dataset of tweets was obtained by sampling existing hate and offensive speech datasets cited in this paper, complying with the terms of use of each of these datasets. All datasets were anonymized, no usernames or any of their demographics are included in the data used to train our models.
+
+# Acknowledgments
+
+We warmly thank our annotators Laurentia Nodit and Laurentiu Zoicaş for their time. We would like to thank the anonymous EMNLP reviewers for their insightful feedback.
+
+This research is supported in part by the Romanian Ministry of Research, Innovation and Digitization, CNCS/CCCDI UEFISCDI, project number 411PED/2020 and project number 108PCE/2021, within PNCDI III.
+
+# References
+
+Luvell Anderson and Ernie Lepore. 2013. What did you call me? slurs as prohibited words setting things up. Analytic Philosophy, 54(3):350-63.  
+Valerio Basile, Cristina Bosco, Elisabetta Fersini, Nozza Debora, Viviana Patti, Francisco Manuel Rangel Pardo, Paolo Rosso, Manuela Sanguinetti, et al. 2019. SemEval-2019 task 5: Multilingual detection of hate speech against immigrants and women in twitter. In Proceedings of SemEval.  
+Robin Brontsema. 2004. A queer revolution: Reconceptualizing the debate over linguistic reclamation. Colorado Research in Linguistics, 17.  
+Ana-Maria Bucur, Marcos Zampieri, and Liviu P. Dinu. 2021. An exploratory analysis of the relation between offensive language and mental health. In Findings of the ACL.  
+Elena Castroviejo, Katherine Fraser, and Agustín Vicente. 2020. More on pejorative language: Insults that go beyond their extension. Synthese, pages 1-26.  
+Jose Cañete, Gabriel Chaperon, Rodrigo Fuentes, JouHui Ho, Hojin Kang, and Jorge Pérez. 2020. Spanish pre-trained bert model and evaluation data. In PML4DC at ICLR 2020.  
+Thomas Davidson, Dana Warmsley, Michael Macy, and Ingmar Weber. 2017. Automated hate speech detection and the problem of offensive language. In Proceedings ICWSM.
+
+bidirectional transformers for language understanding. In Proceedings of NAACL.  
+Elisabeth Eder, Ulrike Krieg-Holz, and Udo Hahn. 2019. At the lower end of Language—Exploring the vulgar and obscene side of German. In Proceedings of the ALW.  
+Ritesh Kumar, Atul Kr Ojha, Shervin Malmasi, and Marcos Zampieri. 2018. Benchmarking aggression identification in social media. In Proceedings of TRAC.  
+Ritesh Kumar, Atul Kr. Ojha, Shervin Malmasi, and Marcos Zampieri. 2020. Evaluating aggression identification in social media. In Proceedings of TRAC.  
+Jerry Liu, Nathan O'Hara, Alexander Rubin, Rachel Draelos, and Cynthia Rudin. 2020. Metaphor detection using contextual word embeddings from transformers. In Proceedings of Fig-Lang.  
+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.  
+Julia Mendelsohn, Yulia Tsvetkov, and Dan Jurafsky. 2020. A framework for the computational linguistic analysis of dehumanization. Frontiers in Artificial Intelligence, 3:55.  
+George A Miller. 1995. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41.  
+Dat Quoc Nguyen, Thanh Vu, and Anh Tuan Nguyen. 2020. Bertweet: A pre-trained language model for english tweets. arXiv preprint arXiv:2005.10200.  
+Alexis Palmer, Melissa Robinson, and Kristy K. Phillips. 2017. Illegal is not a noun: Linguistic form for detection of pejorative nominalizations. In Proceedings of ALW.  
+Tharindu Ranasinghe and Marcos Zampieri. 2020. Multilingual Offensive Language Identification with Cross-lingual Embeddings. In Proceedings of EMNLP.  
+Tharindu Ranasinghe and Marcos Zampieri. 2021. MUDES: Multilingual Detection of Offensive Spans. In Proceedings of NAACL.  
+Anna Schmidt and Michael Wiegand. 2017. A survey on hate speech detection using natural language processing. In Proceedings of SocialNLP.  
+Robyn Speer, Joshua Chin, Andrew Lin, Sara Jewett, and Lance Nathan. 2018. Luminosoin-sight/wordfreq: v2.2.  
+Neeraj Vashistha and Arkaitz Zubiaga. 2021. Online multilingual hate speech detection: experimenting with Hindi and english social media. Information, 12(1):5.
+
+Zeerak Waseem and Dirk Hovy. 2016. Hateful symbols or hateful people? predictive features for hate speech detection on twitter. In Proceedings of NAACL SRW.  
+Michael Wiegand, Josef Ruppenhofer, Anna Schmidt, and Clayton Greenberg. 2018. Inducing a lexicon of abusive words - a feature-based approach. In Proceedings of NAACL.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Fun-towicz, et al. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771.  
+Marcos Zampieri, Shervin Malmasi, Preslav Nakov, Sara Rosenthal, Noura Farra, and Ritesh Kumar. 2019. Predicting the type and target of offensive posts in social media. In Proceedings of NAACL.

acomputationalexplorationofpejorativelanguageinsocialmedia/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d71a27c27126cfc0f790dfd61d35dbd6c8dee5b395aee1fe128c2c19d8f0f44
+size 153010

acomputationalexplorationofpejorativelanguageinsocialmedia/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1749f302523dbc3a52e6a9e5c290e6b0fcf89e8717ac1a9bdbb78cc52d70ba0d
+size 167001

aconditionalgenerativematchingmodelformultilingualreplysuggestion/ba86f020-96cc-42e7-b587-426a976040ea_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9e8ae34d206f7a84b2ab31b66eac27e80e0c34fa60fc9fb7879f7e9ba72a3ed
+size 145360

aconditionalgenerativematchingmodelformultilingualreplysuggestion/ba86f020-96cc-42e7-b587-426a976040ea_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:acd1bfc04d8f338653ac6da8a362fe5b94106b3e8fc8d8fcda4296e12d0ac732
+size 164751

aconditionalgenerativematchingmodelformultilingualreplysuggestion/ba86f020-96cc-42e7-b587-426a976040ea_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b322ab188ee76cdb88fc3618887d9c809dfdd1ac208862d6f620cdc6554fdcaa
+size 3895236

aconditionalgenerativematchingmodelformultilingualreplysuggestion/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cff5e0e2e9f07fbcda6f522ddc7774d1d6d4996b8e00ba5d870feaea2face68d
+size 1693367

aconditionalgenerativematchingmodelformultilingualreplysuggestion/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:630456bf5c0ffed4bd6d22e72ebdc7e19b127e6f904877ccae096ee7c83d96e8
+size 530876

acorpusbasedsyntacticanalysisoftwotermedunlikecoordination/d08cbc6b-abd4-494d-889b-908ee6cbc4ae_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:644cd35a31afcfdc23dc9b27132a5e82bc8392066d1b89a57809e0892fb1825f
+size 75727

acorpusbasedsyntacticanalysisoftwotermedunlikecoordination/d08cbc6b-abd4-494d-889b-908ee6cbc4ae_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ebcaa1ef183befa03f13df1131b2f31f369d12213252e028c59a33fa2be134e
+size 97631

acorpusbasedsyntacticanalysisoftwotermedunlikecoordination/d08cbc6b-abd4-494d-889b-908ee6cbc4ae_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8da80caae55dcbc001170635d95aaf2b0ace40fdb26d380e30095b28a4c6697c
+size 419705

acorpusbasedsyntacticanalysisoftwotermedunlikecoordination/full.md

@@ -0,0 +1,367 @@
+# A Corpus-based Syntactic Analysis of Two-termed Unlike Coordination
+
+Julie Kallini and Christiane Fellbaum
+
+Department of Computer Science
+
+Princeton University
+
+{jkallini, fellbaum}@princeton.edu
+
+# Abstract
+
+Coordination is a phenomenon of language that conjoins two or more terms or phrases using a coordinating conjunction. Although coordination has been explored extensively in the linguistics literature, the rules and constraints that govern its structure are still largely elusive and widely debated amongst linguists. This paper presents a study of two-termed unlike co-ordinations in particular, where the two conjuncts of the coordination phrase form valid constituents but have distinct categories. We conducted a syntactic analysis of the phrasal categories that can be conjoined in such unlike co-ordinations through a computational corpus-based approach, utilizing the Corpus of Contemporary American English (COCA) as the main data source, as well as the Penn Treebank (PTB). The results show that the two conjuncts within unlike co-ordinations display different properties based on their position, supporting an antisymmetric view of the structure of coordination. This research provides new data and perspectives through the use of statistical techniques that can help shape future theories and models of coordination.
+
+# 1 Introduction
+
+# 1.1 Motivation
+
+Coordination is a phenomenon of language that conjoins two or more terms or phrases. The terms or phrases that are grouped in coordination phrases are normally called conjuncts, and they are often conjoined by a coordinating conjunction, such as and, or, but, or nor. A common assumption in the linguistics literature is that two elements may only be coordinated if they share the same syntactic category, as in (1).
+
+(1) a. [NP The chicken] and [NP the rice] go well together.  
+b. The president will  $[\mathrm{VP}$  understand the criticism] and  $[\mathrm{VP}$  take action].
+
+For example, in (1a), the two conjuncts being coordinated are "the chicken" and "the rice," which share the same syntactic category of noun phrase (NP). The assumption that the conjuncts of a coordination phrase will always have the same category is known as the Law of the Coordination of Likes (LCL) (Williams, 1981). The LCL explains why many instances of coordination are ungrammatical, such as the coordination of a prepositional phrase (PP) and a clause (CP) shown in (2) (Prazmowska, 2015).
+
+(2) a. The scene of the movie was in Chicago.  
+b. The scene that I wrote was in Chicago.  
+c. *The scene [PP of the movie] and [CP that I wrote] was in Chicago.
+
+Even though the prepositional phrase and the clause are both grammatical when standing alone within the context sentence, as in (2a) and (2b), their co-ordination in (2c) is ungrammatical, supposedly because of the LCL.
+
+However, several examples of syntactically unlike coordination can be found in English, such as the examples in (3) (Sag et al., 1985).
+
+(3) a. Pat is [NP a Republican] and [AP proud of it].  
+b. John is [AP healthy] and [PP in good shape].  
+c. That was [NP a rude remark] and [PP in very bad taste].
+
+In the above examples, the two conjuncts within each coordination phrase do not share the same syntactic category. In these cases, the LCL seems to be too restrictive.
+
+Yet, there are also cases in which the LCL is not restrictive enough—a coordination phrase can still be ungrammatical even if its conjuncts have the same syntactic category (Prazmowska, 2015).
+
+(4) a. *John ate [PP with his mother] and [PP with good appetite].  
+b. \*John [AdvP probably] and [AdvP unwillingly] went to bed.
+
+Example (4a) contains the coordination of two prepositional phrases, and (4b) contains the coordination of two adverbs. Despite the two conjuncts having like categories, these examples result in ungrammatical sentences. Semantics seems to play a role in the acceptability of coordinations as well; a stronger version of the LCL requires that conjuncts must also be alike in their semantic function. For example, in (4a), the first prepositional phrase "with his mother" expresses accompaniment, whereas the second "with good appetite" expresses manner (Prazmowska, 2015). However, identifying and articulating rigorous rules that predict all grammatical possibilities of coordination has been a difficult task for linguists, and as a result, the underlying syntactic structure of coordination phrases has been elusive.
+
+# 1.2 Goal
+
+The goal of this project is to explore and answer questions about the syntax of coordination phrases through a quantitative corpus analysis. By analyzing a large corpus of naturally-occurring spoken and written language using natural language processing and statistical techniques, we will investigate the patterns of syntactic categories found in unlike coordinations. An overarching goal for this project is to share data that may inform linguistic hypotheses about the underlying structure of coordination.
+
+By taking a computational approach, we can explore a larger and deeper set of questions regarding coordination, such as:
+
+- What combinations of syntactic categories are attested in English data, and which appear most frequently?  
+- Does this depend on the genre of the text or the type of conjunction (and, or, but, nor)?
+
+This paper begins by introducing the relevant problem background and related work. We then detail our corpus-based approach and implementation, which utilizes the Corpus of Contemporary American English (COCA), the Penn Treebank (PTB), and the Berkeley Neural Parser. We then follow with a presentation of the results and provide an in-depth discussion of the significant findings.
+
+![](images/bfd0c0c97618ecdf605e200fbb61638a153471d4b1b4724f120f525d6081d57e.jpg)  
+Figure 1: Flat multi-headed proposal for the structure of coordination.
+
+# 2 Background and Related Work
+
+Capturing the structure of coordination has been a difficult problem in many theories of syntax. A flat, multi-headed structure was proposed in earlier theories, in which two or more lexical heads share the same phrase-level projection, as in the templates shown in Figure 1 (Progovac, 1998a; Chomsky, 1981). This theory captures the intuitive idea that the coordination of two NPs is an NP, that the coordination of two VPs is a VP, etc. An example of a two-termed coordination of NPs is provided in (5). We use  $CC$  as the name for the functional category of coordinating conjunctions, which is also the label used in the PTB.
+
+(5) [NP the cat] and [NP the dog]
+
+![](images/3229fb228b91e860f539de5272867a50c731efa6924d7e45d4b5f24deba7eb31.jpg)
+
+There are several problems with this view, but the problem we are most concerned with relates to the aforementioned counterexamples to the LCL, restated below in (6).
+
+(6) a. Pat is [NP a Republican] and [AP proud of it].  
+b. John is [AP healthy] and [PP in good shape].  
+c. That was [NP a rude remark] and [PP in very bad taste].
+
+In fact, the LCL was formulated due to this proposal for the syntax of coordination. Coordination was said to denote a relation between two (or more) elements that are "hierarchically equal" in that neither of the elements is more prominent than the
+
+other, leading to a symmetrical and flat vision of coordination structures (Prazmowska, 2015). Since conjuncts were assumed to be symmetrical and equal in status, it followed that they must share the same syntactic category to be grammatically coordinated.
+
+One proposal that seems to address the existence of the unlike category coordinations seen in (6) is Bowers's Pred (predicate) functional category (Bowers, 1993). On top of the NPs, APs, and PPs being coordinated in these sentences, there is another level of structure. Bowers suggests that a null Pred head selects an NP, AP, or PP as its complement, forming a predicate phrase (PredP). Thus, unlike coordinations are actually like coordinations in disguise—all conjuncts have the category of PredP. PredPs are complements of the copula  $be$  in these sentences, as made apparent in (7).
+
+(7) a. Pat is  $\left[\mathrm{PredP}\emptyset\right]$  [NP a Republican] and  $\left[\mathrm{PredP}\emptyset\right]$  [AP proud of it].  
+b. John is  $\left[\mathrm{PredP}\emptyset \left[\mathrm{AP}\right.\right]$  healthy] and  $\left[\mathrm{PredP}\emptyset \right.$  [pp in good shape] ].  
+c. That was  $[\mathrm{PredP}\emptyset [\mathrm{NP}$  a rude remark] and  $[\mathrm{PredP}\emptyset [\mathrm{PP}$  in very bad taste] ].
+
+However, Bowers's proposal does not account for cases where the coordinated strings are not predicates, such as in (8). In each of these examples, the coordination phrase is an adjunct of VP rather than a predicate complement of VP, and the conjuncts semantically serve the purpose of adverbial modification.
+
+(8) a. The surgeon operated [AdvP slowly] and [PP with great care].  
+b. Alice will visit home [AdvP tomorrow] or [PP on the weekend].
+
+Other proposals dodge the problem of unlike coordination entirely by making the coordinating conjunction the head of its own coordination phrase (CCP). One example of such a theory is shown in (9).
+
+(9) [NP a Republican] and [AP proud of it]
+
+![](images/5efe41d68b0df6e5d3e8fc23808f9d573414764da01cc8d71b47eec15a3c3e5a.jpg)
+
+Here, conjuncts are specifiers and complements of the head conjunction (Johannessen, 1998; Zoerner, 1995). With such a construction, the categories of the conjuncts by themselves do not pose a restriction on the possibility of coordination. Thus, such theories do not have anything to say about the LCL, but they are still problematic in that they over-generate; no combinations of categories are prohibited.
+
+# 3 Approach
+
+We approached the task of capturing the structure of two-termed coordination by conducting a computational syntactic analysis on a large quantity of corpus data. Our primary data source is the Corpus of Contemporary American English (COCA) (Davies, 2015), and our additional data source is the Penn Treebank (PTB) augmented with Ficler and Goldberg's PTB coordination annotation extension (Ficler and Goldberg, 2016). We extracted coordination phrases from both of these datasets and performed a quantitative syntactic analysis using the constituency parses of the sentences within both texts.
+
+This approach has a few advantages over previous work. Much of the research that has shaped current theories of coordination have relied on the acceptability judgments of a few individuals, usually the author(s). By using corpus data, we gain an understanding of coordination on a much larger scale and emphasize empirical rather than intuitive judgments. We can also investigate differences in the patterns we identify based on the genre from which a coordination was found or the conjunction it contains.
+
+# 3.1 Corpus Data
+
+The Corpus of Contemporary American English (COCA) is a large, genre-balanced corpus of American English containing more than 450 million words of text (Davies, 2015). The COCA contains text from five genres: academic, fiction, magazine, newspaper, and spoken texts. Each genre includes 20 million words each year from 1990-2012. A balanced corpus, especially one that includes spoken data, was important for this project, as there may be variations in the coordinations found across different genres.
+
+In addition to COCA data, we use the Penn Treebank (PTB), a collection of 2,499 stories from the Wall Street Journal gathered over a three-year pe
+
+riod (Marcus et al., 1993). Sentences from the PTB are already tokenized and annotated with phrase structure, unlike the COCA. However, coordination annotations in the PTB are often inconsistent, include errors, and lack internal structure in many cases. For this reason, we make use of Ficler and Goldberg's PTB coordination annotation extension, which improves the coordination annotation in the PTB (Ficler and Goldberg, 2016). This extension provides an annotation that explicitly marks coordination phrases and the role of each element in coordination structures (i.e., conjuncts, markers, connectives, and shared elements are all identified and marked).
+
+# 3.2 Syntactic Analysis
+
+The main task of our syntactic analysis involves the detection and extraction of coordination phrases from our corpus data. Since the COCA is provided in a raw text format, we use the Berkeley Neural Parser to produce syntax trees of sentences in the COCA. This is a state-of-the-art constituency parser that generates syntax trees in the style of the Penn Treebank (Kitaev and Klein, 2018). To implement a good search algorithm for coordinations within parsed COCA data, we studied several sentence parse trees containing coordinations and identified three patterns in the way that the Berkeley Neural Parser most often represents the structure of coordination phrases, as shown in Figure 2.
+
+Since the PTB is already annotated as phrase structure trees, the possible problems of using a constituency parser on novel text are eliminated. The identification of coordination phrases is made much simpler here with the help of the coordination annotation extension. The explicit function markers allow for the straightforward detection and isolation of conjuncts and conjunctions from other tangential elements that may be contained within a coordination phrase, such as modifiers and connectives. Figure 3 shows an example of a PTB phrase structure tree with the extension's additional function marking.
+
+For our syntactic analysis, we include coordinations of six types of PTB phrasal category labels: noun phrases (NP), verb phrases (VP), prepositional phrases (PP), adjective phrases (ADJP), adverb phrases (ADVP), and subordinate clauses (SBAR, often called complementizer phrases (CP) in more recent syntax literature). We have chosen this set of labels because they correspond to the
+
+![](images/2d0a3371918177e086c21e3d2b6ef3d980edd77fa616e78f11ca8cdf80c0c2e4.jpg)  
+(a) Simple ternary-branching pattern.
+
+![](images/552582ea7058e3bd60166faa0138c175e4f97165e1e2dba8595b7bcf2d0f81a3.jpg)  
+(b) Neither-nor pattern.
+
+![](images/4225a4908c84ff1b3b1d819a69f5e4af3e4f05bba5a4a79f9a3a277781a007a3.jpg)  
+(c) Verb-complement pattern.  
+Figure 2: Three patterns used to detect two-termed co-ordination phrases in parsed COCA data. X, Y, and Z may be any PTB constituent tags.
+
+most frequent phrasal categories in the data. Once coordination phrases have been identified, we run statistical tests on the frequencies of their different attributes, such as the categories of the conjuncts, the type of conjunction used, and the genre from which the coordination was found.
+
+# 4 Results
+
+In our analyses, we employ the chi-square  $(\chi^2)$  tests, which determine whether a set of observed frequencies deviate significantly from a set of expected frequencies. We consider  $p$  -values less than 0.05 to be statistically significant. Since our sam
+
+![](images/53403b07356ae5274d8973b8354f4da5ed9bc4f42639b0ee53c3d98353450b51.jpg)  
+Figure 3: A tree containing the explicit function marking from the PTB coordination annotation extension.
+
+<table><tr><td>V</td><td>Association</td></tr><tr><td>0.00–0.05</td><td>negligible</td></tr><tr><td>0.05–0.10</td><td>weak</td></tr><tr><td>0.10–0.15</td><td>moderate</td></tr><tr><td>0.15–0.25</td><td>strong</td></tr><tr><td>0.25–1.00</td><td>very strong</td></tr></table>
+
+![](images/110ea757b0f4fb5fa7b6d5715f1db1d5f5c20e28c4385e2a6cfe7f822001440c.jpg)  
+Figure 4: Most frequent unlike category combinations in the COCA data. Frequencies are relative to all unlike coordinations.
+
+ple sizes are very large, we conduct additional posttests to accompany any statistically significant results. We use Cramer's  $V$  to measure strength of association (Table 1) (Akoglu, 2018).
+
+# 4.1 Most Frequent Unlike Coordinations
+
+We performed an analysis of the most frequent unlike coordinations in the COCA data. Figure 4 displays the top ten most common unlike coordinations found in all of the COCA data we parsed along with their relative frequencies, and Table 2 contains examples. We found a significant difference in the distribution of unlike category coordinations, with a moderate tendency toward the most common coordination combination,  $\mathsf{NP} + \mathsf{SBAR}$ ,  $\chi^2 (9,N = 24456) = 3142.0$ ,  $p < .001$ ,  $V = .119$ .
+
+# 4.1.1 By COCA Genre
+
+We also performed an analysis of the most frequent unlike coordinations in each of the five COCA genres. In each genre, a significant difference was found in the distribution of unlike category coordinations. Table 3 summarizes the results of the chi-square tests and Cramer's  $V$  for each COCA genre, and Appendix B contains figures displaying the top
+
+Table 1: Interpretation of strength of association/tendency based on Cramer's  $V$  
+
+<table><tr><td>Coordination</td><td>Example Sentence</td></tr><tr><td>NP+SBAR*</td><td>You&#x27;d get to watch two adults talk about [NP America] and [SBAR what they would do to lead it].</td></tr><tr><td>NP+VP</td><td>Voids are [NP a nightmare] and [VP initialed by the employee and his supervisor].</td></tr><tr><td>ADJP+VP*</td><td>It was [ADJP emotionally manipulative] and [VP designed to scare people into faith].</td></tr><tr><td>ADVP+PP*</td><td>The phenomenon fell into place [ADVP organically] and [PF with ease].</td></tr><tr><td>NP+ADJP*</td><td>He&#x27;s [NP a free spirit] and [ADJP playful], prompting managers and teammates to shake their heads and proclaim he&#x27;s Manny being Manny.</td></tr><tr><td>PP+VP</td><td>In Gaza, meanwhile, Hamas leaders insist that they are still [PP in charge] and [VP leading the Palestinian authority].</td></tr><tr><td>PP+ADVP*</td><td>A big question many taxpayers face is whether to file [PF by paper] or [ADVP electronically].</td></tr><tr><td>NP+PP</td><td>I called him a liar again, and then I punched him [NP a lot of times] and [PP with all my might].</td></tr><tr><td>PP+NP</td><td>More Americans work [PP out of the house] and [NP longer hours], so we&#x27;ve become more dependent on meals we don&#x27;t cook ourselves.</td></tr><tr><td>VP+NP</td><td>Erosion and years of neglect have left the brick structure [VP crumbling] and [NP a clear safety hazard].</td></tr></table>
+
+* Also in the top ten unlike coordinations in the PTB.
+
+Table 2: Examples extracted from the COCA for each of the top ten most common unlike coordinations.  
+
+<table><tr><td>Genre</td><td>χ2</td><td>N</td><td>p</td><td>V</td></tr><tr><td>Academic</td><td>450.59</td><td>5105</td><td>&lt; .001</td><td>.099</td></tr><tr><td>Fiction</td><td>693.22</td><td>4358</td><td>&lt; .001</td><td>.133</td></tr><tr><td>Magazine</td><td>583.69</td><td>5324</td><td>&lt; .001</td><td>.110</td></tr><tr><td>Newspaper</td><td>616.91</td><td>4851</td><td>&lt; .001</td><td>.118</td></tr><tr><td>Spoken</td><td>2391.3</td><td>5095</td><td>&lt; .001</td><td>.228</td></tr></table>
+
+Table 3: Summary of chi-square test and Cramer's  $V$  results for the frequency difference among the top ten unlike category combinations in each COCA genre.
+
+unlike coordinations in each genre. In the academic genre, there was a weak tendency toward the most common combination, NP+SBAR (Figure 7). In the fiction genre, a moderate tendency was found toward the most common combination, ADJP+VP (Figure 8). In the magazine genre, we also found a moderate tendency toward the most common combination, which was again NP+SBAR, as in the academic genre (Figure 9). In the newspaper genre, an indication of a moderate tendency toward the most common combination was found once again, with NP+VP being the most common combination (Figure 10). In the spoken genre, there is a notable indication of a strong tendency toward the most common combination, which was NP+SBAR, as in the academic and magazine genres (Figure 11).
+
+<table><tr><td>Conjunction</td><td>x²</td><td>N</td><td>p</td><td>V</td></tr><tr><td>and</td><td>2933.0</td><td>19621</td><td>&lt;.001</td><td>.129</td></tr><tr><td>or</td><td>752.87</td><td>4317</td><td>&lt;.001</td><td>.139</td></tr><tr><td>but</td><td>73.893</td><td>1042</td><td>&lt;.001</td><td>.089</td></tr><tr><td>nor</td><td>14.333</td><td>45</td><td>.111</td><td>-</td></tr></table>
+
+# 4.1.2 By Conjunction
+
+We also performed an analysis of the most frequent unlike category combinations based on the type of coordinating conjunction used to conjoin them. Table 4 summarizes the results of the chi-square tests and Cramer's  $V$  for each type of conjunction, and Appendix B again contains figures displaying the top unlike coordinations for each conjunction. For the conjunctions and, or, and but, a significant difference was found in the distribution of unlike category coordinations. For unlike coordinations containing and, there was a moderate tendency toward the most common combination, which was  $\mathrm{NP + SBAR}$  (Figure 12). For unlike coordinations containing or, we also found a moderate tendency toward the most common combination, which was again  $\mathrm{NP + SBAR}$  (Figure 13). For unlike coordinations containing but, there was a weak tendency toward the most common combination, ADJP+VP (Figure 14). For unlike coordinations containing nor, no significant difference was found in the distribution of unlike category coordinations (Figure 15).
+
+# 4.1.3 In the PTB
+
+We performed an analysis of the most frequent unlike coordinates in the PTB as well. Figure 5 displays the top ten most common unlike coordinates in the PTB data, along with their relative frequencies. We found a significant difference in the distribution of unlike category coordinations with a moderate tendency toward the most common combinations,  $\chi^2 (9,N = 216) = 22.981$ ,  $p = .006$ ,  $V = .109$ . The most common unlike coordination in the PTB was ADVP+PP.
+
+# 4.2 Differences Between Conjunct Positions
+
+In addition to the most frequent combinations of categories, we conducted an analysis of the categories for each conjunct independently. We first
+
+![](images/d2acc96adafec52cf079975f0aa9a94b853523b354cd0877f56106142d8f72e3.jpg)  
+Figure 5: Most frequent unlike category combinations in the PTB. Frequencies are relative to all unlike co-ordinations.
+
+Table 4: Summary of chi-square test and Cramer's  $V$  results for the frequency difference among the most common unlike coordinations based on the coordinating conjunction used to conjoin them (from COCA data).  
+
+<table><tr><td>Category</td><td>1st Conjunct</td><td>2nd Conjunct</td><td>x²</td><td>N</td><td>p</td><td>V</td></tr><tr><td>NP</td><td>70.75%</td><td>29.24%</td><td>3200.7</td><td>18582</td><td>&lt; .001</td><td>.415</td></tr><tr><td>VP</td><td>32.42%</td><td>67.58%</td><td>1764.4</td><td>14277</td><td>&lt; .001</td><td>.352</td></tr><tr><td>PP</td><td>53.47%</td><td>46.53%</td><td>68.789</td><td>14248</td><td>&lt; .001</td><td>.069</td></tr><tr><td>ADJP</td><td>55.73%</td><td>44.27%</td><td>125.57</td><td>9566</td><td>&lt; .001</td><td>.114</td></tr><tr><td>ADVP</td><td>48.71%</td><td>51.29%</td><td>5.076</td><td>7645</td><td>.024</td><td>.026</td></tr><tr><td>SBAR</td><td>23.97%</td><td>76.03%</td><td>2385.8</td><td>8800</td><td>&lt; .001</td><td>.521</td></tr></table>
+
+Table 5: Summary of chi-square test and Cramer's  $V$  results for the frequency difference between the two conjunct positions for each type of phrasal category from COCA data.
+
+report the results based on frequencies from the COCA. Table 5 summarizes the results of the chi-square tests and Cramer's  $V$  for each of the six phrasal categories. For NPs, a very strong tendency was found toward the first conjunct position; for VPs, a very strong tendency was found toward the second conjunct position; for PPs, only a weak tendency was found toward the first conjunct position; for ADJPs, a moderate tendency was found toward the first conjunct position; for ADVPs, only a negligible tendency was found toward the second conjunct position; and for SBARs, a very strong tendency was found toward the second conjunct position.
+
+Next, we report the results based on frequencies from the PTB. Table 6 summarizes the results of the chi-square tests and Cramer's  $V$  for each of the six phrasal categories. For NPs, a very strong tendency was found toward the first conjunct position; for VPs, PPs, ADJPs, and ADVPs, no significant difference was found in the distribution of conjunct positions; and for SBARs, a very strong tendency was found toward the second conjunct position.
+
+<table><tr><td>Category</td><td>1st Conjunct</td><td>2nd Conjunct</td><td>x²</td><td>N</td><td>p</td><td>V</td></tr><tr><td>NP</td><td>65.57%</td><td>34.43%</td><td>11.836</td><td>122</td><td>&lt;.001</td><td>.311</td></tr><tr><td>VP</td><td>38.18%</td><td>61.82%</td><td>3.073</td><td>55</td><td>.080</td><td>-</td></tr><tr><td>PP</td><td>50.34%</td><td>49.66%</td><td>.0069</td><td>145</td><td>.934</td><td>-</td></tr><tr><td>ADJP</td><td>54.17%</td><td>45.83%</td><td>.8333</td><td>120</td><td>.361</td><td>-</td></tr><tr><td>ADVP</td><td>44.25%</td><td>55.75%</td><td>1.496</td><td>113</td><td>.221</td><td>-</td></tr><tr><td>SBAR</td><td>26.09%</td><td>73.91%</td><td>10.522</td><td>46</td><td>.001</td><td>.478</td></tr></table>
+
+Table 6: Summary of chi-square test and Cramer's  $V$  results for the frequency difference between the two conjunct positions for each type of phrasal category from the PTB data.
+
+# 5 Evaluation
+
+A portion of the data we have presented in the previous section was gathered through the use of a constituency parser to identify coordination phrases. While the Berkeley Neural Parser is state-of-the-art, no parser is perfect, especially concerning coordination disambiguation. Furthermore, there are additional types of coordination structures that we do not consider, including non-constituent coordination and gapping. In non-constituent coordination, each conjunct in a coordination phrase does not form its own constituent under traditional theories of clause structure, as shown in example (10).
+
+(10) The girl from California walked [into the room at 9 PM] and [out of the room at 10 PM].
+
+Gapping is the phenomenon in which a phrase is coordinated with another phrase that seems to be missing some material, as shown in (11).
+
+(11) [Mary ate beans] and [John potatoes].
+
+While this paper only seeks to analyze the coordination of constituents and does not consider these additional types of coordination, they still pose challenges in the identification and labeling of coordination phrases by parsers. We have conducted an evaluation plan in which human raters manually assessed a random sample of unlike coordinations to estimate an error rate for each type of category combination.
+
+Each type of unlike coordination was assigned a score based on the judgments of three independent raters. A single rater contributes to the score by providing the percentage of samples in which they agreed with the parser's labels. The overall score for that type of coordination is then assigned by taking the mean of the three raters' scores. The scores for each type of unlike coordination are enumerated in Table 7, along with the sample size, confidence level, and margin of error used for sampling.
+
+<table><tr><td></td><td>NP</td><td>VP</td><td>PP</td><td>ADJP</td><td>ADVP</td><td>SBAR</td></tr><tr><td>NP</td><td>-</td><td>50.9% (103)</td><td>72.3% (100)</td><td>72.4% (101)</td><td>61.2% (96)</td><td>83.4% (103)</td></tr><tr><td>VP</td><td>61.7% (99)</td><td>-</td><td>69.0% (96)</td><td>70.3% (95)</td><td>62.0% (85)</td><td>63.0% (105)</td></tr><tr><td>PP</td><td>61.7% (100)</td><td>64.4% (101)</td><td>-</td><td>80.0% (90)</td><td>80.3% (101)</td><td>70.7% (97)</td></tr><tr><td>ADJP</td><td>77.6% (97)</td><td>80.0% (102)</td><td>89.6% (97)</td><td>-</td><td>66.6% (78)</td><td>65.5% (64)</td></tr><tr><td>ADVP</td><td>55.5% (86)</td><td>66.0% (89)</td><td>85.3% (101)</td><td>56.5% (76)</td><td>-</td><td>63.7% (96)</td></tr><tr><td>SBAR</td><td>79.5% (96)</td><td>50.0% (93)</td><td>75.3% (81)</td><td>58.0% (36)</td><td>56.5% (46)</td><td>-</td></tr></table>
+
+Table 7: Average agreement with the Berkeley Neural Parser's labeling of each type of unlike coordination phrase, based on the judgments of the three raters. Rows correspond to the first conjunct's category, and columns correspond to the second conjunct's category. A  $90\%$  confidence level and  $\pm 8\%$  margin of error were used for sampling each category combination. The sample size,  $n$ , is reported in parentheses.
+
+<table><tr><td>κ</td><td>Agreement</td></tr><tr><td>0.00–0.20</td><td>poor</td></tr><tr><td>0.20–0.40</td><td>fair</td></tr><tr><td>0.40–0.60</td><td>moderate</td></tr><tr><td>0.60–0.80</td><td>substantial</td></tr><tr><td>0.80–1.00</td><td>near perfect</td></tr></table>
+
+Table 8: Interpretation of strength of agreement based on the Cohen's Kappa Coefficient.
+
+Fleiss' Kappa showed that, among the three raters, there was fair agreement in their judgments,  $\kappa = .291$  (95% CI [.271, .312]),  $p < .001$ . The strength of agreement is determined based on the Cohen's Kappa Coefficient (Table 8) (McHugh, 2015).
+
+# 6 Discussion
+
+# 6.1 Most Frequent Unlike Coordinations
+
+The results of the analysis of the most frequent unlike coordinations in the COCA data indicate that NP+SBAR is the most common unlike coordination. It was also the most frequent unlike coordination in three of the five genres (academic, magazine, and spoken). Some examples from the COCA are shown in (12) below.
+
+(12) a. Be sure to tell us  $[\mathrm{NP}$  your full name] and [SBAR where you live].  
+b. I support [NP the president] and [SBAR what he did].
+
+c. The zone's size depends on [NP the weather] and [SBAR how much flow the Mississippi brings each year].
+
+One possible explanation for the high frequency of NP+SBAR coqindinations is that subordinate clauses have very similar syntactic distributions to noun phrases in other contexts as well. In particular, subordinate clauses, which are called complementizer phrases (CP) in the syntax literature, can be the subjects of sentences. When a CP occupies the subject position of a sentence, it is called a sentential subject (Lohndal, 2014). Sentential subjects can be headed by a variety of different complementizers; (13) shows a few examples using that, whether, what, and how.
+
+(13) a. [CP That Joe fell asleep in the meeting] disappointed us.  
+b. [CP Whether she shows up or not] doesn't matter.  
+c. [CP What a huge scandal it was] didn't emerge until later.  
+d. [CP How we got here] is a total mystery.
+
+Some linguists have theorized that sentential subjects and more typical nominal subjects have the same syntactic category. Much like Bowers's predicate phrase analysis discussed in Section 2, sentential subjects may be analyzed as having a null determiner head that forms a determiner phrase (DP) from a CP (Lohndal, 2014).
+
+(14)  $[\mathrm{DP}\emptyset [\mathrm{CP}\mathrm{That}\mathrm{Mary}\mathrm{left}\mathrm{early}]]$  disappointed us.
+
+Although we do not expound on the arguments for DPs, most of the constituents that we have been treating as noun phrases for simplicity in this paper are often analyzed as DPs instead. The heads of determiner phrases may be overt, such as the determiners the and  $a$  in phrases like [DP the dog] or [DP a child], or they may be null, as in the case of plural nouns like [DP  $\varnothing$  dogs] or [DP  $\varnothing$  children]. The argument for clauses as DPs would posit that the same null determiner head that plays a role in the formation of plural DPs could also play a role in the formation of DPs from subordinate clauses. The data collected in this project provide more evidence through coordination that DPs and CPs have very similar syntactic distributions.
+
+While NP+SBAR was also within the top ten unlike coordinations in the PTB, ADVP+PP and
+
+PP+ADVP were the most common in the PTB. Examples from the PTB are presented in (15).
+
+(15) a. Beauregard was mentioned twice—although [ADVP very briefly] and [PP in passing].  
+b. A huge production system built [PP in the sea off Santa Barbara] and [ADVP ashore] is sitting idle.
+
+ADVP+PP and PP+ADVP were within the top co-ocbinations from the COCA data as well. Their frequent co-occurrence likely has to do with ADVP's and PP's shared purpose of adverbial modification in adjunct position. A null functional morpheme could be used to explain this coordination, and this idea would be quite similar to Bowers's Pred (predicate) proposal but applied to adjuncts of verbs instead of complements.
+
+# 6.2 Differences Between Conjunct Positions
+
+When considering each phrasal category in isolation and controlling for their different total frequencies, in both the COCA and the PTB, NPs had a very strong tendency toward being in the first conjunct position, and SBARs had a very strong tendency toward the second conjunct position. In the COCA data, VPs had a very strong tendency toward the second conjunct position, and ADJPs had a moderate tendency toward the first conjunct position.
+
+It seems like phrasal categories that can be very short, like NPs, are more likely to appear as the first conjunct, but longer phrases, like CPs or VPs, are more likely to be the second conjunct. This may be related to a phenomenon called heavy NP shift, in which a noun phrase appears to the right of its expected canonical position due to its "weight" (Kayne, 1994, Chapter 7). Example (16) explores heavy NP shift through prepositional dative constructions, where the recipient of a ditransitive verb (in this case, "Jen") is the object of the preposition to (Colleman et al., 2010).
+
+(16) a. I gave [NP the large book of poems] [PP to Jen].  
+b. I gave [PP to Jen] [NP the large book of poems].
+
+All of the constituents in (16a) appear in their canonical, expected positions; the direct object noun phrase appears closest to the verb, and the prepositional phrase containing the recipient is after the NP. In (16b), heavy NP shift moves the direct
+
+object NP into a position after the PP. Shifting can only occur if the NP is long and complex; when it is short, shifting is prohibited, as (17) shows.
+
+(17) a. I gave [NP it] [PP to Jen].  
+b. * I gave [PP to Jen] [NP it].
+
+Shifting can also target syntactic categories other than noun phrases. In (18a), the complement and adjunct of the noun "statue" appear in their expected positions, with the complement  $\left[\mathrm{PP}\right.$  of him] closer to the noun. In (18b), the complement is heavier than the adjunct and thus appears further to the right.
+
+(18) a. the statue [PP of him] [PP in the park]  
+b. the statue [pp in the park] [PP of that old musician from the 19th century]
+
+The main idea behind shifting can be applied to coordination and the trends that we observed in the results section regarding asymmetry in conjunct positions. If heavier constituents undergo shifting to appear after lighter constituents within phrases, this would explain why longer and more complex conjuncts tend to appear in the second conjunct position of coordination phrases. Example (19) shows this intuition through the like coordination of two NPs with different lengths.
+
+(19) a. I bought [NP apples] and [NP some strange looking fruits I found in the produce aisle].  
+b. ? I bought [NP some strange looking fruits I found in the produce aisle] and [NP apples].
+
+We can also observe heavier constituents appearing in the second conjunct position in unlike coordinations, as shown by example (20). Although (20b) is not ungrammatical, (20a) sounds a bit more natural.
+
+(20) a. John is [AP healthy] and [PP in the best shape of his life].  
+b. John is [PP in the best shape of his life] and [AP healthy].
+
+# 6.3 Limitations
+
+One shortcoming of this paper lies in the evaluation plan: the human reviewers were not blind to the labels given to coordination phrases by the parser. With more resources, a future iteration of this project could include the creation of a small gold standard dataset of coordinations and use the
+
+more formal precision, recall, and F1 metrics to gauge the parser's accuracy in the identification of coordinations. Still, the raters' evaluations reveal the limitations of an analysis that utilizes an existing constituency-based parser on raw COCA data, which includes a size of parse errors. We acknowledge the drawbacks of such an approach and have supplemented the analysis of COCA data with data from the Penn Treebank for this purpose, which is not processed using a parser. These data sources together provide more concrete examples of the possibilities of unlike constituent coordination.
+
+# 7 Conclusion
+
+This paper approached the problem of understanding the syntax of two-termed coordination phrases through a computational corpus analysis. Previous research has not attempted a thorough analysis of coordination based on English corpora, instead relying on intuitive acceptability judgments to inform their theories. We conducted a syntactic analysis by extracting coordination phrases from the Corpus of Contemporary American English and the Penn Treebank, and we investigated the most common unlike coordinations and the syntactic categories that appeared in either of the two conjunct positions.
+
+Some of the findings from this project have interesting implications for coordination and syntax as a whole. The high frequency of coordinations of noun phrases with subordinate clauses provides further proof that noun phrases and clauses share similar syntactic distributions and may be structurally defined as determiner phrases. The tendency for first conjuncts to be shorter constituents and second conjuncts to be longer ones might suggest that shifting occurs in coordination structures as well. One of the main takeaways from these results is that there are evident syntactic distinctions between the two conjuncts of a coordination phrase, which support theories that posit an antisymmetric account for the structure of coordination.
+
+# Acknowledgments
+
+We would like to thank Srinivas Bangalore for his suggestions and feedback as the second reader of this paper, as well as the students of his Introduction to Machine Translation class, who helped complete the project's evaluation plan. We also appreciate the three anonymous reviewers' careful reading of our paper and their constructive comments.
+
+# References
+
+Haldun Akoglu. 2018. User's guide to correlation coefficients. Turkish Journal of Emergency Medicine, 18.  
+John Bowers. 1993. The syntax of predication. Linguistic Inquiry, 24(4):591-656.  
+Noam Chomsky. 1981. Lectures On Government and Binding. Foris Publications.  
+Timothy Colleman, Bernard De Clerck, and Magda Devos. 2010. Prepositional dative constructions in english and dutch: A contrastive semantic analysis. Neophilologische Mitteilungen, 111(2):131-152.  
+Mark Davies. 2015. Corpus of Contemporary American English (COCA).  
+Jessica Ficler and Y. Goldberg. 2016. Coordination annotation extension in the penn tree bank. ArXiv, abs/1606.02529.  
+Janne Bondi Johannessen. 1998. Coordination. Oxford University Press.  
+Richard Kayne. 1994. The Antisymmetry of Syntax. MIT Press.  
+Nikita Kitaev and Dan Klein. 2018. Constituency parsing with a self-attentive encoder. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Melbourne, Australia. Association for Computational Linguistics.  
+Terje Lohndal. 2014. Sentential subjects in english and norwegian. Syntax and Semantics, 15:81-113.  
+Mitchell P. Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. 1993. Building a large annotated corpus of english: The penn treebank. Comput. Linguist., 19(2):313-330.  
+Mary L. McHugh. 2015. Interrater reliability: the kappa statistic. Biochemia medica, 22(3):276-282.  
+Anna Prazmowska. 2015. Is unlike coordination against the law (of the coordination of likes)?  
+Ljiljana Progovac. 1998a. Structure for coordination: Part 1. In GLOT International 3.7.  
+Ivan Sag, Gerald Gazdar, Thomas Wasow, and Steven Weisler. 1985. Coordination and how to distinguish categories. *Natural Language and Linguistic Theory*, 3:117-171.  
+Edwin S. Williams. 1981. Transformationless grammar. In Linguistic Inquiry 12, pages 645-653.  
+Cyril Edward Zoerner. 1995. Coordination: The Syntax of & P. UCI dissertations in linguistics. University of California, Irvine.
+
+![](images/c2d1c9199d7400df98da0cae49a926badb0dc7e1b82c008258d37bd81435312e.jpg)  
+Figure 6: Heatmap displaying raw frequencies of all unlike category combinations (from COCA data).
+
+# A Heatmap of Unlike Coordinations in COCA
+
+For completion, we include the frequency distribution of unlike coordinations for all 30 combinations of categories in the COCA data. Figure 6 visualizes these data in the form of a heatmap.
+
+# B Top Unlike Coordinations by Genre and Conjunction
+
+The figures is in this appendix display the most frequent unlike category coordinations for each COCA genre and for each type of coordinating conjunction (and, or, but, nor) from the COCA data. Figures 7, 8, 9, 10, and 11 correspond to each of the five COCA genres, and the coordination frequencies are taken relative to all unlike coordinations within that genre. Figures 12, 13, 14, and 15 correspond to each of the four coordinating conjunctions, and the coordination frequencies are taken relative to all unlike coordinations that use the given conjunction.
+
+![](images/e0aaac6aab53d64574523dc03b33f87716f6c93318decc317b9a68edcd64f214.jpg)  
+Figure 7: Academic genre.
+
+![](images/3f091bdbea1b375e4ef934542046636022c946ec7ab72d2d4f468a20a3edbc2c.jpg)  
+Most Common Unlike Coordinations in Fiction Genre
+
+![](images/690e89ba7c1a80be2ddc22134a1b3df8bedb1d87a464983e7406fb7966895455.jpg)  
+Figure 8: Fiction genre.  
+Most Common Unlike Coordinations in Magazine Genre  
+Figure 9: Magazine genre.
+
+![](images/f290cb7303fc647c460e5e53c147c7edcfa19c23aa7ad1e014975434c3804bfb.jpg)  
+Most Common Unlike Coordinations in Newspaper Gen  
+Figure 10: Newspaper genre.
+
+![](images/f13cc17b8e43e171ec3bbca29ce07132a9832b3e18572a00f55fd5fe468b1599.jpg)  
+Most Common Unlike Coordinations in Spoken Genre  
+Figure 11: Spoken genre.
+
+![](images/2a949aa39329c93c6a69bea4d0351827499e3574ff089c58229fc5ce1594c377.jpg)  
+Most Common Unlike Coordinations Containing 'and'  
+Figure 12: Unlike coordinations using and.
+
+![](images/479df33c217c507d77e852eb8ad37527ba907c087619b7bf25e8b40d80db23a9.jpg)  
+Most Common Unlike Coordinations Containing 'or'  
+Figure 13: Unlike coordinations using or.
+
+![](images/805bbed6b54906fa2f92efa2d32752119eb4339e2a2ea4c72bc9e2ed05a7f98c.jpg)  
+Most Common Unlike Coordinations Containing 'but'  
+Figure 14: Unlike coordinations using but.
+
+![](images/c94c27ef8c172a727b302197973111f9b74742aa8f4b56cc41599e1cd49df99d.jpg)  
+Most Common Unlike Coordinations Containing 'nor'  
+Figure 15: Unlike coordinations using nor.

acorpusbasedsyntacticanalysisoftwotermedunlikecoordination/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da17db397b0c69f54c1ddd94f4a2e014f1ea610f9f14ba59aaca3d92640a0271
+size 615432

acorpusbasedsyntacticanalysisoftwotermedunlikecoordination/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1910131a3afb8d6bc6e4c9607c0933cf1f79b8a992d80e9897aee2581e106166
+size 389525

activelearningforrumoridentificationonsocialmedia/03db8d81-f471-42d6-8235-95dc7714b4f7_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7273985444ef32d3dc16b237a67f61fdae13bddcd83b5615fd2a2b4039874389
+size 67098

activelearningforrumoridentificationonsocialmedia/03db8d81-f471-42d6-8235-95dc7714b4f7_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a3413034734a9e17b1b35698bcd4934c05526a8754ea30a550db23e3c548cb00
+size 78532

activelearningforrumoridentificationonsocialmedia/03db8d81-f471-42d6-8235-95dc7714b4f7_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56fdc22be6ae113c9ddb796ce8c3d97c0c09354f28056704bdab9ff28c030d1b
+size 345083

activelearningforrumoridentificationonsocialmedia/full.md

@@ -0,0 +1,255 @@
+# Active Learning for Rumor Identification on Social Media
+
+Parsa Farinneya<sup>1</sup>, Mohammad Mahdi Abdollah Pour<sup>1</sup>, Sardar Hamidian<sup>2</sup> and Mona Diab<sup>2,3</sup>
+
+<sup>1</sup>Dep. of Computer Engineering, Amirkabir University of Technology
+
+$^{2}$ Dep. of Computer Science, The George Washington University
+
+$^{3}$ Facebook AI Research
+
+{p FAR, mabdollahpour}@aut.ac.ir, sardar@gwu.edu, mdiab@fb.com
+
+# Abstract
+
+Social media has emerged as a key channel for seeking information. Online users spend several hours reading, posting, and searching for news on microblogging platforms daily. However, this could act as a double-edged sword especially when not all information online is reliable. Moreover, the inherently unmoderated nature of social media renders identifying unverified information ever more challenging. Most of the existing approaches for rumor tracking are not scalable because of their dependency on a significant amount of labeled data. In this work, we investigate this problem from different angles. We design an Active-Transfer Learning (ATL) strategy to identify rumors with a limited amount of annotated data. We go beyond that and investigate the impact of leveraging various machine learning approaches in addition to different contextual representations. We discuss the impact of multiple classifiers on a limited amount of annotated data followed by an interactive approach to gradually update the models by adding the least certain samples (LCS) from the pool of unlabeled data. Our proposed Active Learning (AL) strategy achieves faster convergence in terms of the F-score while requiring fewer annotated samples (42% of the whole dataset for the best model).
+
+# 1 Introduction
+
+Rumor detection in social networks is the task of identifying if a post's remark is unverifiable. This detection can help stop the spread of misinformation/dis-information that could potentially cause harm and distress. When a rumor about a subject emerges, there are thousands of posts shared about that subject. Ahsan (2019) show that having abundant in-domain labeled data can significantly impact the accuracy of the rumor detection model on Tweets by more than  $30\%$  improvement. This also points to the impact of out-of-domain/topic training on rumor detection per
+
+formance. However in a real world scenarios for rumor detection, in domain human-annotated data is typically missing in early stages of rumor propagation, resulting in mediocre accuracy levels for such models. A viable solution for this problem would ideally be a framework that yields decent accuracy despite the absence of in-domain manually annotated training data. To this end, this paper proposes a semi-supervised framework based ATL for rumor detection in social media, specifically for Twitter data. There are three main variables for the proposed framework: the representation of the Tweets, the estimator, and the Active Learning strategy. Other experimental variables will be discussed in the following sections. As we evaluate all the different variables, we observe that TweetBERT, linear regression and least confidence strategy yield comparable results as non-Active Learning methods yet with a fraction of human-annotated data needed in non-Active Learning based methods. Further for robustness of our proposed models, we experiment with using an exploration method by choosing some random queries in each loop to prevent the model from overfitting. We also reach an approximation of the minimum labeled data needed for a decent classification in this task with the proposed method.
+
+# 2 Related Work
+
+Qazvinian et al. (2011) ran experiments to examine the effect of in-domain labeled data on rumor detection accuracy. They conducted learning curve experiments injecting their training models with labeled data, going from 400 to almost 2000 training examples. The experiments exhibit rapid performance improvement plateauing at an accuracy of  $80\%$ . Hamidian and Diab (2016) introduced the Tweet Latent Vector (TLV) feature, which is a 100-d vector that was created by a mixing Twitter features and network-specific features such as Hashtags, URL, Re-Tweets, and Content features such
+
+as POS and content n-grams, as well as pragmatic features representing Named Entities, Sentiment. In 2019 ACL RumorEval shared task on rumor detection and verification, Derczynski et al. (2017) used a subset of the PHEME dataset in two subtask to identify the stance of comments as well as measure the veracity of the subset of rumor posts. The best models for this task utilized contextualized word embedding such as BERT. Additionally, the models used were mostly deep neural networks with the exception of the best performing model (Li et al., 2019), which was an ensemble of Support Vector Machine, Random Forrest and Logistic Regression. Bhattacharjee et al. (2017) proposed a simple, yet efficient, learning method for fake news detection in a weakly supervised scenario. The proposed method in this work improved generalization ability through interactive human participation by annotating a small amount of relevant samples that provide the most insightful information on the data. Their model was based on GloVe word embeddings and a CNN-based embedding model on the character level with fully connected layers for classification. They evaluated their models on the KDnugget's Fake News dataset<sup>1</sup>, Liar Dataset, (Wang, 2017) and Harvard Dataverse Twitter Collection. Hasan et al. (2020) proposed an Active Learning framework for fake news detection based on entropy sampling. In this approach, by using just  $4\%$  to  $28\%$  of available training data, the model achieves a comparable performance to supervised learning with all available labeled training data. Inspired by this latter work, despite inherent differences in the task at hand (rumor detection vs. fake news detection), we believe that similar principles would hold for rumor detection. Accordingly, we propose a novel method for rumor detection that will reduce the need for human annotation in this task.
+
+# 3 Problem Definition and Approach
+
+# 3.1 Problem Definition
+
+We cast the problem of rumor detection as a binary classification task. Tweets are classified as either rumors or non-rumor. We propose a human in the loop annotation strategy. When Tweets about a subject start spreading, and it is not clear whether it is a rumor, the proposed human-in-loop framework gradually trains a classification model specific for the emerged Tweet's subject. We propose a
+
+framework combining Active Learning with Transfer Learning. In each iteration of the proposed ATL pipeline, a batch of unlabeled Tweets that are the most informative for the model are passed to a human for annotation (similar to an Oracle in the Active Learning literature). This loop continues until the annotation budget is exhausted.
+
+# 3.2 Active Learning
+
+In this work, we leverage the most common Active Learning scenario that is the unlabeled pool scenario. This approach is also the most similar to real-life problems. In this scenario, there is a large pool of unlabeled data. The model is at first trained on a small subset of pre-annotated data. Then the framework queries for a batch of unlabeled data to be labeled by a human (oracle) and added to the train set on each iteration. Since annotation may be expensive or time-consuming, it is preferable to run this process as few times as possible. The sample queries are chosen among unused unlabelled data based on their score, and the scoring function is called the strategy in the Active Learning literature. This step is repeated until the annotation budget is exhausted. The algorithm is described in Algorithm1.
+
+Various strategies are proposed in the literature for data selection in an Active Learning pipeline. Selection based on prediction uncertainty is the most popular approach, which is also applied in this work.
+
+Least Confidence (LC) Least Confidence (LC) is a strategy based on prediction uncertainty. LC tries to find data samples that the model is not certain about, as a proxy for the model having trouble classifying that data. Certainty is measured as confidence in most likely label as defined in the equation 1 by  $\max(y)$ .  $y$  is probabilities predicted by the model given  $x$  as input and  $score(x)$  is the uncertainty measure.
+
+$$
+\operatorname {s c o r e} (x) = 1 - \max  (y) \tag {1}
+$$
+
+Query by Committee (QBC) In Query by Committee (QBC) strategy, instead of measuring the uncertainty of a single model, we train an ensemble of models. For a given sample, disagreement between models is taken as a measure of uncertainty. There are also two special cases of QBC: bagging (BAG) and boosting (BOOST). In BOOST, we bootstrap random samples with replacement from the available initial data for the committee members. In
+
+# Algorithm 1: Pool-based Active Learning
+
+Input:  $D_{i}, D_{p}, D_{te}$ , batch size, strategy, estimator, annotation budget
+
+Output: Model, metrics
+
+$D_{i}$  Initial data;
+
+$D_{p}$  Pool data;
+
+$D_{te}$  Test data;
+
+Instantiate  $D_{tr}$  as empty, Train data;
+
+Instantiate model;
+
+Add  $D_{i}$  to  $D_{tr}$
+
+model  $=$  estimator.train  $(D_{tr})$
+
+while annotation budget is not over do
+
+$D_{q} = \mathrm{Query}(D_{p},\mathrm{batch}$  size,strategy, model);
+
+Remove  $D_q$  from  $D_p$ ;
+
+Annotate  $D_q$ ;
+
+Add  $D_{q}$  to  $D_{tr}$ ;
+
+train model on  $D_{tr}$  from scratch;
+
+Compute and save metrics;
+
+end
+
+BAG, we perform bootstrapping for both initial and train data.
+
+Ranked Batch (Batch-LC) We also use the Ranked Batch strategy (Batch-LC) as proposed in Cardoso et al. (2017) which uses a scoring function as in Equation 2 to find a ranked list of query data.
+
+$$
+\operatorname {s c o r e} = \alpha (1 - \Phi (x, X _ {\text {l a b e l e d}})) + (1 - \alpha) U (x) \tag {2}
+$$
+
+In Equation 2,  $X_{\text{labeled}}$  is the labeled dataset,  $U(x)$  is the uncertainty of predictions for  $x$ , and  $\Phi$  is a similarity function, for instance, cosine similarity. This latter function measures how well the feature space is explored near  $x$ .  $\alpha$  is also computed by Equation 3.
+
+$$
+\alpha = \frac {\left| X _ {u n l a b e l e d} \right|}{\left| X _ {l a b e l e d} \right| + \left| X _ {u n l a b e l e d} \right|} \tag {3}
+$$
+
+After score computation for each sample, the highest scoring sample is removed, and scores are recalculated until the desired number of examples are available to send for the query.
+
+Epsilon-Greedy (EG) In order to find a balance between exploration and exploitation, we use a method inspired by  $\epsilon$ -greedy (EG) strategy in Reinforcement Learning. We implement this approach in two ways: inter-batch and intra-batch. Inter-batch EG (EG-inter) selects query data at each iteration randomly with probability  $\epsilon$  otherwise chooses
+
+the data based on LC  $(1 - \epsilon)$ . Intra-batch (EG-Intra) dedicates  $\epsilon \%$  of query data to RND and  $1 - \epsilon \%$  of that to LC. We use  $\epsilon = 0.2$  (20% in EG-intra) in our experiments.
+
+# 3.3 Cross Topic Transfer
+
+Data from other domains can be beneficial to improve performance on the target domain. Therefore, we design another type of experiment in which Tweets from other topics are considered in the initial feed to the model (zero shot). The model queries the pooled data at each iteration from the target topic. This is the setting that usually appears in real-world problems. There are datasets from previous topics that can not generalize well to the target topic. However, by choosing a minimum number of data through Active Learning, the model can adapt to the target domain.
+
+# 4 Experimental Setup
+
+We compare each experiment with Least Confidence (LC) and random strategy (RND) leveraging different representations, and learning algorithms. In random strategy, data samples are chosen uniformly random at each iteration. To mitigate the effect of randomness in both strategies, training algorithms and data splits, for each experiment, we randomly split the dataset into two sections, first one for initial and pool data and the second one as test data. We do this 5 times, and average the results (in a cross validation type evaluation strategy). For each of these 5 runs, the splits are the same among different experiments.
+
+At each iteration, Multi-Layer Perceptron (MLP) is retrained on the batch of data annotated in the current iteration since training from scratch would be computationally expensive. However, other models are trained from scratch on data that was obtained in the current and previous iterations. In all experiments, we train the models on 20 randomly chosen samples as the initial dataset and query for 50 samples at each iteration, i.e. batch size  $= 50$ .
+
+After examining all the models and representation settings with LC and RND strategy, the best setting is utilized for further experiments leveraging other sampling strategies such as Query batched committee (QBC), Epsilon-Greedy (EG), and Ranked Batch (Batch-LC).
+
+# 4.1 Representation
+
+We use SOTA representations for this task, such as BERT (Devlin et al., 2019) and TweetBERT (Qudar and Mago, 2020). TweetBERT, a domain-specific BERT based language model trained specifically on social media data. TweetBERT was trained on about 680 million Tweets. We also use earlier representations such as GloVe (Pennington et al., 2014). For each sentence, we average the GloVe vector representation of all tokens in the sentence and use it as the input to the models. We use Twitter GloVe, which is consistent with the domain of our work. Twitter GloVe was trained over by 2B Tweets, 27B tokens, and 1.2M vocab. A dimensionality of 200 was determined empirically to yield best results. The representations are frozen during training.
+
+# 4.2 Model
+
+We also examine different models that have been mainly used for short text classification tasks, namely, MLP (Hinton, 1990), Support Vector Machines (SVM) (Platt et al., 1999), Random Forests (RF) (Breiman, 2001), Logistic Regression (LR) (Cramer, 2002), Ada boosted decision trees (Ada) (Freund and Schapire, 1997), K-Nearest Neighbors (KNN) (Fix, 1985), Gaussian Process Classifier (GP) (Rasmussen, 2003), Linear Discriminant Analysis (LDA) (Cohen et al., 2003), and Quadratic Discriminant Analysis (QDA) (Tharwat, 2016).
+
+We used Radial Basis Function (RBF) kernel for SVM with inverse regularization term  $C = 1$ . For Random Forest, we used 100 estimators and a max depth of 1000 with the Gini criterion. For Logistic Regression, we used  $l_{2}$  penalty and LBFGS (Liu and Nocedal, 1989) solver with a maximum of 100 iterations. For Ada boosted decision tree, we used an ensemble of 50 trees with SAMME.R real boosting (Freund and Schapire, 1997). The MLP had a hidden layer of size 128 and a drop-out layer after the hidden layer with  $p = 0.3$  and it was trained by adam optimizer (Kingma and Ba, 2015). We use two KNN models: one with 5 neighbors (KNN5) and the other with 3 neighbors (KNN3). GP is used with an RBF kernel and optimized with the L-BFGS-B (Byrd et al., 1995) algorithm. An SVD solver was used for both LDA and QDA.
+
+Machine Learning Tools There are some tools and libraries used to build the experimental pipeline. MLP was implemented in Tensorflow $^4$  and for other models (RF, SVM, LR, Ada, KNN, GP, LDA and QDA) we used Scikit-Learn (Pedregosa et al., 2011) package. Active Learning workflows were developed using modAL (Danka and Horvath, 2018) framework.
+
+# 4.3 Data
+
+The PHEME dataset is curated from highly retweeted Tweets associated with newsworthy events (Zubiaga et al., 2016). It includes five cases of breaking news: Ferguson unrest, Ottawa shooting, Sydney Siege, Charlie Hebdo shooting, and Germanwings plane crash. It also includes four specific rumors: Prince to play in Toronto, Gurlitt collection, Putin missing, and Michael Essien contracted Ebola. This dataset consists of 6425 Tweets comprising 2402 rumors, and 4023 non-rumors. In this study, we work on Charlie Hebdo, Ferguson, and Sydney Siege since they have the highest number of annotated Tweets in the dataset (more than 1000 Tweets each). In topics with a small number of Tweets, it is not possible to have an unbiased test set and examine the effect of AL on choosing a minimum number of data. For example for a topic with only 100 labeled tweets in dataset, we can not have a reliable test set (at least 1000 samples) and a big pool dataset (100 samples are consumed by AL in 2 iterations)
+
+Preprocessing The texts of Tweets were processed by changing ""t" to "not", for privacy and generalization changing usernames to "Username", removing punctuation except question marks, removing special characters, removing trailing white space, and changing URLs to "Link". Table 1 shows some samples of this data set.
+
+# 4.4 Metrics
+
+We evaluate the performance of models by F1 score instead of accuracy since the test data does not come from a distribution with balanced labels. Moreover, we examine the effectiveness of Active Learning through some additional metrics. For each setting, we compute the F1 score variation in Active Learning loops. Namely, we compute F1 score on points which account for using  $0\%$ ,  $25\%$ ,  $50\%$ ,  $75\%$ , and  $100\%$  of the pool data.
+
+<table><tr><td colspan="2">Charlie Hebdo</td></tr><tr><td>Rumors</td><td>Non-rumors</td></tr><tr><td>#Charlie Hebdo witness - Gunmen told me to tell the media they were Al-Qaeda in Yemen</td><td>Just arrived at scene of massacre
+#Paris #charliehebdo</td></tr><tr><td>According to #Charlie Hebdo\u2019s lawyer four well-known French cartoonists were killed by the masked gunmen: Cabu, Wolinski, Charb et Tignous.</td><td>Anybody who wants to talk about what Charlie Hebdo might have done to \&#x27;&#x27;provoke\&quot; this should probably shut up, forever</td></tr></table>
+
+Table 1: Tweet samples of PHEME dataset
+
+In order to determine the minimum amount of data needed for each experiment to achieve a promising result, we consider a minimum number of data samples needed to achieve at least  $f_{max} - 1\%$  where  $f_{max}$  is the maximum F1 score reached in that experiment.
+
+# 5 Experimental Results
+
+# 5.1 Baselines
+
+We compare the models against two baselines: RANDOM and Majority based on training data observations. RANDOM is simply random prediction. Majority is simply the majority of labels observed in the training data at each iteration are predicted for all samples in the test set.
+
+# 5.2 Estimator selection
+
+The proposed method of Active Learning has a base estimator that estimates pool data and predicts the test set. Topics that weren't used in Active Learning loops due to having few Tweets, Ottawa shooting, Germanwings, were used for hyper-parameter tuning in a greedy search base method.
+
+# 5.3 Results
+
+Tables 2 shows F1 score at points of using  $0\%$ ,  $25\%$ ,  $50\%$ ,  $75\%$ , and  $100\%$  of pool data for each experiment. In each column scores go from red to white and green as models consume more data showing how rapidly the model improves.
+
+# 5.3.1 Model and Representation Comparison
+
+By examining the results, Logistic Regression yields the best performance among our models, and TweetBERT is the best representation. This is expected since TweetBERT is pretrained on the tweets genre. Interestingly, GloVe representations outperform BERT representations, despite the fact that BERT is known for its more sophisticated architecture yielding contextualized embed
+
+dings. However, our version of GloVe embeddings is trained on Twitter data. This observation suggests that the genre of the training data has a larger impact on performance than the representation model complexity. TweetBERT+LR with LC strategy achieves the best scores. Experiments with LC strategy perform better than RND. TweetBERT+LR with LC strategy also achieves the best performance with only  $25\%$  of data.
+
+# 5.3.2 Strategy Comparison
+
+We examine the performance gain of LC in more detail by comparing the difference of F1 score for RND strategy and LC strategy of a fixed setting at points of using  $0\%$ ,  $25\%$ ,  $50\%$ ,  $75\%$ , and  $100\%$  of pool data for each experiment. Table 2 illustrates some of these observations. For instance, rows  $0\%$  and  $100\%$  in the table shows where the model has access to same portion of data, whether using LC or RND. Subtracting values in RND column from LC column for LR with TweetBERT yields 0, 2.07, 2.2, 1.33 and 0.03 for each row, respectively. Similarly for rows  $25\%$ ,  $50\%$ , and  $75\%$ , we observe the effect of Active Learning such that the differences for most of model-representation pairs would be positive, indicating an improvement over the RND strategy. RF, SVM, and GP get the benefit the most from LC strategy.
+
+Table 4 compares the performance of uncertainty strategies using best representation-model pairs. Except for QBC, other strategies are very close. Based on our results, Ranked Batch (Batch-LC), boosting (BOOST), and bagging (BAG) yield the best performance, respectively. QBC fails to make a diverse ensemble but when used with BOOST and BAG there are more diverse voters. In each row scores go from red to white and green as models consume more data showing improvement of models.
+
+Figure 1, 2 and 3 show F1 score for Tweet-
+
+<table><tr><td></td><td>Model</td><td colspan="2">Ada</td><td colspan="2">GP</td><td colspan="2">KNN3</td><td colspan="2">KNN5</td><td colspan="2">LDA</td></tr><tr><td></td><td>Stra.</td><td>LC</td><td>RND</td><td>LC</td><td>RND</td><td>LC</td><td>RND</td><td>LC</td><td>RND</td><td>LC</td><td>RND</td></tr><tr><td rowspan="5">TweetBERT</td><td>0%</td><td>58.6</td><td>58.83</td><td>60.6</td><td>60.6</td><td>64.3</td><td>64.3</td><td>62.73</td><td>62.73</td><td>59.83</td><td>59.83</td></tr><tr><td>25%</td><td>70.37</td><td>70.43</td><td>72.57</td><td>63.5</td><td>70.2</td><td>69</td><td>71.57</td><td>70.27</td><td>68.8</td><td>67.83</td></tr><tr><td>50%</td><td>72.2</td><td>72.6</td><td>76.87</td><td>68.2</td><td>73.6</td><td>73.33</td><td>75.03</td><td>74.3</td><td>68.4</td><td>67</td></tr><tr><td>75%</td><td>74.1</td><td>73.6</td><td>75.37</td><td>70.67</td><td>75.33</td><td>74.97</td><td>76.1</td><td>74.93</td><td>64.53</td><td>62.6</td></tr><tr><td>100%</td><td>74.17</td><td>74.07</td><td>72.3</td><td>72.23</td><td>75.87</td><td>75.83</td><td>76.2</td><td>76</td><td>63.93</td><td>63.77</td></tr><tr><td rowspan="5">GloVe</td><td>0%</td><td>60.37</td><td>59.53</td><td>58.33</td><td>58.33</td><td>64.7</td><td>64.7</td><td>64.23</td><td>64.23</td><td>63.6</td><td>63.6</td></tr><tr><td>25%</td><td>70.2</td><td>70.53</td><td>73.47</td><td>66.33</td><td>72.3</td><td>71.27</td><td>73.03</td><td>70.9</td><td>64.7</td><td>61.23</td></tr><tr><td>50%</td><td>73.4</td><td>72.9</td><td>77.97</td><td>71.83</td><td>75.5</td><td>75.33</td><td>75.87</td><td>75.03</td><td>66.1</td><td>67.23</td></tr><tr><td>75%</td><td>74.7</td><td>73.77</td><td>76.2</td><td>74.37</td><td>76.33</td><td>76</td><td>76.67</td><td>76.33</td><td>73.37</td><td>73.1</td></tr><tr><td>100%</td><td>73.9</td><td>74.07</td><td>75.37</td><td>75.4</td><td>76.83</td><td>76.97</td><td>77.37</td><td>77.53</td><td>76.8</td><td>76.67</td></tr><tr><td rowspan="5">BERT</td><td>0%</td><td>56.93</td><td>57.53</td><td>58.67</td><td>58.67</td><td>55.73</td><td>55.73</td><td>52.27</td><td>52.27</td><td>55.67</td><td>55.67</td></tr><tr><td>25%</td><td>67.6</td><td>65.97</td><td>64.63</td><td>66.97</td><td>66.47</td><td>66.77</td><td>65.8</td><td>66.73</td><td>70.57</td><td>71.07</td></tr><tr><td>50%</td><td>70.43</td><td>68.93</td><td>68.1</td><td>68.23</td><td>69.87</td><td>69.67</td><td>70.27</td><td>69.87</td><td>70.67</td><td>69.43</td></tr><tr><td>75%</td><td>71.23</td><td>70.97</td><td>71.4</td><td>70.03</td><td>71.03</td><td>70.57</td><td>71.63</td><td>71.53</td><td>69.53</td><td>66.5</td></tr><tr><td>100%</td><td>70.5</td><td>71.67</td><td>71.33</td><td>71.27</td><td>71.57</td><td>71.33</td><td>71.83</td><td>71.77</td><td>63.5</td><td>63.33</td></tr><tr><td></td><td>Model</td><td colspan="2">LR</td><td colspan="2">MLP</td><td colspan="2">QDA</td><td colspan="2">RF</td><td colspan="2">SVM</td></tr><tr><td></td><td>Stra.</td><td>LC</td><td>RND</td><td>LC</td><td>RND</td><td>LC</td><td>RND</td><td>LC</td><td>RND</td><td>LC</td><td>RND</td></tr><tr><td rowspan="5">TweetBERT</td><td>0%</td><td>60.77</td><td>60.77</td><td>41.77</td><td>41.77</td><td>50.17</td><td>50.17</td><td>56.3</td><td>56.3</td><td>43.1</td><td>43.1</td></tr><tr><td>25%</td><td>76.6</td><td>74.53</td><td>56.47</td><td>55.63</td><td>43.83</td><td>40.2</td><td>74.33</td><td>67.73</td><td>43.53</td><td>41.73</td></tr><tr><td>50%</td><td>78.6</td><td>76.4</td><td>75.63</td><td>69.43</td><td>28.13</td><td>32.73</td><td>75.13</td><td>70.27</td><td>54.53</td><td>55.17</td></tr><tr><td>75%</td><td>78.93</td><td>77.6</td><td>69.57</td><td>71.87</td><td>34.93</td><td>35.83</td><td>73.5</td><td>71.37</td><td>63.7</td><td>61.03</td></tr><tr><td>100%</td><td>78.5</td><td>78.47</td><td>71.67</td><td>74.17</td><td>37.43</td><td>38.13</td><td>72.03</td><td>72.4</td><td>65.93</td><td>66.03</td></tr><tr><td rowspan="5">GloVe</td><td>0%</td><td>59.4</td><td>59.4</td><td>45.37</td><td>42.97</td><td>51.67</td><td>51.67</td><td>58.87</td><td>58.87</td><td>50.17</td><td>50.17</td></tr><tr><td>25%</td><td>75.9</td><td>73.13</td><td>56.93</td><td>57.77</td><td>47.5</td><td>44.17</td><td>74.4</td><td>67.67</td><td>69.17</td><td>61.37</td></tr><tr><td>50%</td><td>78.47</td><td>76.1</td><td>74</td><td>70.37</td><td>38.9</td><td>46.03</td><td>76.23</td><td>71.13</td><td>75.83</td><td>71.2</td></tr><tr><td>75%</td><td>78.2</td><td>77.53</td><td>75.07</td><td>74.13</td><td>42.27</td><td>42.33</td><td>74.67</td><td>72.7</td><td>76.3</td><td>75.43</td></tr><tr><td>100%</td><td>78.27</td><td>78.27</td><td>76.83</td><td>76.7</td><td>47.97</td><td>47.7</td><td>74.3</td><td>74.03</td><td>77.47</td><td>77.3</td></tr><tr><td rowspan="5">BERT</td><td>0%</td><td>58.27</td><td>58.27</td><td>46.63</td><td>43.83</td><td>49.7</td><td>49.7</td><td>53.47</td><td>53.47</td><td>43.1</td><td>43.1</td></tr><tr><td>25%</td><td>73.33</td><td>72.5</td><td>64.73</td><td>65.73</td><td>51.9</td><td>50.57</td><td>70</td><td>58.3</td><td>65.07</td><td>45.53</td></tr><tr><td>50%</td><td>76.87</td><td>74.57</td><td>72.03</td><td>69.97</td><td>50.6</td><td>51.53</td><td>68.87</td><td>63.27</td><td>62.7</td><td>54.87</td></tr><tr><td>75%</td><td>76.77</td><td>75.53</td><td>73.5</td><td>72.93</td><td>52.37</td><td>52.37</td><td>66.9</td><td>64.8</td><td>63.2</td><td>60.67</td></tr><tr><td>100%</td><td>76.83</td><td>76.77</td><td>73.4</td><td>73.57</td><td>49.2</td><td>50.03</td><td>66.6</td><td>66.57</td><td>64.13</td><td>64.1</td></tr></table>
+
+Table 2: F1 score at points of using  $0\%$ ,  $25\%$ ,  $50\%$ ,  $75\%$ , and  $100\%$  of pool data for each experiment with all representations and model representations. The scores are averaged over the three chosen topics in the PHEME dataset. Baseline RANDOM prediction baseline achieves  $26.98\%$  F1 score, and Baseline Majority prediction baseline achieves  $40.90 \pm 3.3\%$ . Intensity of color green shows high F1 scores, red show low F1 scores and white for inbetween F1 scores.  
+
+<table><tr><td>Approach</td><td>0%</td><td>25%</td><td>50%</td><td>75%</td><td>100%</td></tr><tr><td>Few Shot</td><td>60.767</td><td>76.6</td><td>78.6</td><td>78.933</td><td>78.5</td></tr><tr><td>Zero Shot</td><td>50.1</td><td>57.5</td><td>65.867</td><td>69.6</td><td>71.033</td></tr></table>
+
+Table 3: F1 score for TweetBERT+LR with LC strategy using  $0\%$ ,  $25\%$ ,  $50\%$ ,  $75\%$  and  $100\%$  of pool dataset. The scores are averaged over three chosen topics in PHEME dataset. In the Zero Shot setting, the initial dataset includes other topics, as opposed to the Few Shot setting, in which, initial training data consists of 20 samples of the target topic. Intensity of color green shows high F1 scores, red show low F1 scores and white for inbetween F1 scores
+
+<table><tr><td>Representation</td><td>Model</td><td>Strategy</td><td>0%</td><td>25%</td><td>50%</td><td>75%</td><td>100%</td></tr><tr><td rowspan="7">TweetBERT</td><td rowspan="7">LR</td><td>QBC</td><td>64.7</td><td>66.8</td><td>70.3</td><td>69.3</td><td>68.3</td></tr><tr><td>BAG</td><td>60.7</td><td>75.4</td><td>79.067</td><td>79.533</td><td>79</td></tr><tr><td>Batch_LC</td><td>64.767</td><td>76.4</td><td>78.8</td><td>79.267</td><td>79.167</td></tr><tr><td>BOOST</td><td>60.7</td><td>75.4</td><td>79.3</td><td>79.767</td><td>79.033</td></tr><tr><td>EG_intra</td><td>64.9</td><td>77.533</td><td>78.167</td><td>78.3</td><td>78.133</td></tr><tr><td>EG_inter</td><td>52.7</td><td>75.2</td><td>77.133</td><td>78.267</td><td>78</td></tr><tr><td>LC</td><td>60.767</td><td>76.6</td><td>78.6</td><td>78.933</td><td>78.5</td></tr><tr><td rowspan="5">GloVe</td><td rowspan="5">LR</td><td>QBC</td><td>65.2</td><td>65.3</td><td>68.2</td><td>67.3</td><td>67.6</td></tr><tr><td>BAG</td><td>59.933</td><td>75.1</td><td>78.733</td><td>78.6</td><td>78.767</td></tr><tr><td>Batch_LC</td><td>64.133</td><td>75.6</td><td>78.967</td><td>78.867</td><td>78.8</td></tr><tr><td>BOOST</td><td>59.933</td><td>75.1</td><td>78.7</td><td>78.6</td><td>78.767</td></tr><tr><td>LC</td><td>59.4</td><td>75.9</td><td>78.467</td><td>78.2</td><td>78.267</td></tr></table>
+
+Table 4: F1 score for advanced strategies of a fixed setting at points of using  $0\%$ ,  $25\%$ ,  $50\%$ ,  $75\%$  and  $100\%$  of pool data for best representation-estimator pairs. The scores are average over three chosen topics in PHEME dataset. Intensity of color green shows high F1 scores, red show low F1 scores and white for inbetween F1 scores.
+
+BERT+LR with different uncertainty strategies. The diagrams indicate that most models plateau with 100-200 data samples and are able to achieve decent performance with a small amount of data. Most models have a large gain with 100-200 (well-chosen with AL) data samples and there is a small gain after having more than 200 samples. Our best model (TweetBBERT+LR with BATCH-LC) achieves at least  $f_{max} - 1\%$  with 250, 300 and 250 data samples from pool data for each topic. ( $f_{max}$  being maximum of F1 score reached in that experiment) On average, it achieves at least  $f_{max} - 1\%$  with  $42\%$  of pool data (There are 1039, 571, 610 samples in pool dataset of each topic).
+
+# 5.3.3 Cross-Topic Evaluation
+
+Table 3 compares best performing model-representation pair with LC strategy starting from two different initial training datasets. The initial training dataset in zero-shot approach is all data for all topics except the target topic. The initial training dataset of few-shot approach contains a minimal number of in topic in domain samples. We experiment with 20 samples of the target topic. We observe that only a few topic-related samples perform much better than a large dataset of samples, namely, the few shot setting outperforms the zero shot setting as observed in the  $0\%$  of pool data column in Table 3. Data from other domains/topics causes a high variance, which takes many related samples for the model to converge onto reasonable performance. The results demonstrated that Tweets from other rumor topics can add some bias to the model and make the model degrade in performance.
+
+![](images/6a60f44174d18b81118fc3ba10d24510598a352c9e2f86f3ca3e9fbb953e1dfb.jpg)  
+Figure 1: Performance of different strategies with TweetBERT+LR on the Charlie Hebdo topic. The Verticals axis shows F1 score and the horizontal axis shows number of data samples used for training during Active Learning.
+
+# 6 Error Analysis
+
+The confusion matrix for TweetBERT+LR with EG-intra strategy on the Sydney Siege topic for different steps is shown in Table 5. We see that the performance gain is majorly the result of decreasing false negatives. Confusion matrix for other topics also showed a similar behaviour. The model ability to detect rumors improves with the amount of data compared to ability to detect non-rumors. Since, model is able to encode better boundaries for rumors, while non-rumors might be diverse.
+
+![](images/01f0a769ece7b94668264c9d98b4c9950b36d74ee8fa481838a623bb1549b6a1.jpg)  
+Figure 2: Performance of different strategies with TweetBERT+LR on the Sydney Siege topic. The Verticals axis shows F1 score and the horizontal axis shows number of data samples used for training during Active Learning.
+
+<table><tr><td></td><td>TN</td><td>FP</td><td>FN</td><td>TP</td></tr><tr><td>0%</td><td>44.42</td><td>12.14</td><td>19.40</td><td>24.02</td></tr><tr><td>25%</td><td>45.32</td><td>11.24</td><td>10.88</td><td>32.54</td></tr><tr><td>50%</td><td>46.58</td><td>9.98</td><td>10.38</td><td>33.04</td></tr><tr><td>75%</td><td>47.25</td><td>9.31</td><td>10.21</td><td>33.21</td></tr><tr><td>100%</td><td>47.45</td><td>9.11</td><td>10.88</td><td>32.54</td></tr></table>
+
+Table 5: Confusion matrix on Sydney siege topic, averaged over 5 runs. Numbers in the columns are percentages of True Negatives (NT), False Positives (FP), False Negatives (FN) and True Positives (TP), respectively. These numbers of average of number in 5 runs. The First column shows percentage of pool data consumed by the model.
+
+# 7 Conclusion & Future Directions
+
+We proposed an active-transfer learning framework for the rumor detection task. In our proposed framework, we examined different word representations, estimators and Active Learning strategies. More than 300 experimental setups were run and each setup was fine tuned to yield the best results. Our experiments indicate multiple new findings: 1. The approximate minimum number of labeled in-domain data needed for a decent rumor detection model with our proposed method is around 200; 2. In-genre pretrained (contextualized) LMs have the biggest impact on model performance; 3. We investigate and empirically show how epsilon
+
+![](images/b5ac668dbe90c7c68bc913c92fc5daa4aaabafcab4abf8a0e35dfaa2df7c01ba.jpg)  
+Figure 3: Performance of different strategies with TweetBERT+LR on the Ferguson topic. The Verticals axis shows F1 score and the horizontal axis shows number of data samples used for training during Active Learning.
+
+greedy inspired methods that joins randomness and uncertainty in query selection could prevent the model from over-fitting; and, 4. We also showed that naive use of Tweets relating to other topics can degrade the performance (the zero shot setting). Although the method proposed in this paper did not show improvement in using data from other topics, information from different topics can be exploited by incorporating other techniques such as weighting the data samples or meta-learning few-shot domain adaptation. Diverse initial datasets may yield an initial model with better uncertainty scores and earlier convergence. The next step for this method would be incorporating metadata such as reply stances, user information, network propagation information, etc. Finally, another method that could improve our proposed model would be using an ensemble of different representations and different models to generalize better.
+
+# 8 Ethical Considerations
+
+# 8.1 NLP Application
+
+Misuse Potential and Failure Mode When used as intended, applying the strategy described in this paper can help to use the minimum amount of labeled data to identify new emerging rumors online. However, the annotation volume might be inconsistent in some rumors with high variants. This may lead to Failure and high bias. Further research is needed to address the rumor identification issues for emerging rumors, as this issue is present among all current methodologies.
+
+Environmental Cost The experiments described in the paper use a single CPU for all the machine learning models except MLP, which used GPUs. The experiments may take several hours. Several dozen experiments were run due to parameter search for all the models, and future work should experiment with distilled models for more lightweight training. We note that while our work required extensive experiments to draw sound conclusions, future work will be able to draw on these insights and need not run as many large-scale comparisons. Models in production may be trained once for use using the most promising settings.
+
+# References
+
+Mohammad Ahsan. 2019. Detection of context-varying rumors on twitter through deep learning. 128:45-58.  
+Sreyasee Das Bhattacharjee, Ashit Talukder, and Bala Venkatram Balantrapu. 2017. Active learning based news veracity detection with feature weighting and deep-shallow fusion. In 2017 IEEE International Conference on Big Data, BigData 2017, Boston, MA, USA, December 11-14, 2017, pages 556-565. IEEE Computer Society.  
+Leo Breiman. 2001. Random forests. Machine learning, 45(1):5-32.  
+Richard H Byrd, Pei Huang Lu, Jorge Nocedal, and Ciyou Zhu. 1995. A limited memory algorithm for bound constrained optimization. SIAM Journal on scientific computing, 16(5):1190-1208.  
+Thiago NC Cardoso, Rodrigo M Silva, Sérgio Canuto, Mirella M Moro, and Marcos A Gonçalves. 2017. Ranked batch-mode active learning. Information Sciences, 379:313-337.  
+Jacob Cohen, Patricia Cohen, Stephen G West, and Leona S Aiken. 2003. Applied multiple regression. Correlation Analysis for the Behavioral Sciences, 3.
+
+Jan Salomon Cramer. 2002. The origins of logistic regression.  
+Tivadar Danka and Peter Horvath. 2018. modAL: A modular active learning framework for Python. Available on arXiv at https://arxiv.org/abs/1805.00979.  
+Leon Derczynski, Kalina Bontcheva, Maria Liakata, Rob Procter, Geraldine Wong Sak Hoi, and Arkaitz Zubiaga. 2017. SemEval-2017 task 8: RumourEval: Determining rumour veracity and support for rumours. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), pages 69–76, Vancouver, Canada. 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.  
+Evelyn Fix. 1985. Discriminatory analysis: nonparametric discrimination, consistency properties, volume 1. USAF school of Aviation Medicine.  
+Yoav Freund and Robert E Schapire. 1997. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of computer and system sciences, 55(1):119-139.  
+Sardar Hamidian and Mona Diab. 2016. Rumor identification and belief investigation on Twitter. In Proceedings of the 7th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, pages 3-8, San Diego, California. Association for Computational Linguistics.  
+Md Saqib Hasan, Rukshar Alam, and Muhammad Abdullah Adnan. 2020. Truth or lie: Pre-emptive detection of fake news in different languages through entropy-based active learning and multi-model neural ensemble. pages 55-59.  
+Geoffrey E Hinton. 1990. Connectionist learning procedures. artificial intelligence, 40 1-3: 185 234, 1989. reprinted in j. carbonell, editor,". Machine Learning: Paradigms and Methods", MIT Press.  
+Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.  
+Quanzhi Li, Qiong Zhang, and Luo Si. 2019. eventAI at SemEval-2019 task 7: Rumor detection on social media by exploiting content, user credibility and propagation information. In Proceedings of the 13th International Workshop on Semantic Evaluation, pages 855-859, Minneapolis, Minnesota, USA. Association for Computational Linguistics.
+
+Dong C Liu and Jorge Nocedal. 1989. On the limited memory bfgs method for large scale optimization. Mathematical programming, 45(1):503-528.  
+F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830.  
+Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar. Association for Computational Linguistics.  
+John Platt et al. 1999. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61-74.  
+Vahed Qazvinian, Emily Rosengren, Dragomir R. Radev, and Qiaozhu Mei. 2011. Rumor has it: Identifying misinformation in microblogs. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1589-1599, Edinburgh, Scotland, UK. Association for Computational Linguistics.  
+Mohiuddin Md Abdul Qudar and Vijay Mago. 2020. Tweetbert: A pretrained language representation model for twitter text analysis. arXiv preprint arXiv:2010.11091.  
+Carl Edward Rasmussen. 2003. Gaussian processes in machine learning. In Summer school on machine learning, pages 63-71. Springer.  
+Alaa Tharwat. 2016. Linear vs. quadratic discriminant analysis classifier: a tutorial. International Journal of Applied Pattern Recognition, 3(2):145-180.  
+William Yang Wang. 2017. "liar, liar pants on fire": A new benchmark dataset for fake news detection. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 422-426, Vancouver, Canada. Association for Computational Linguistics.  
+Arkaitz Zubiaga, Maria Liakata, Rob Procter, Geraldine Wong Sak Hoi, and Peter Tolmie. 2016. Analysing how people orient to and spread rumours in social media by looking at conversational threads. PLoS ONE, 11(3).

activelearningforrumoridentificationonsocialmedia/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ced1a9bac05b4583dff4ed95f4b3dcf0ee77b9c89688e6da95f45b5b2ac3ed5
+size 621465

activelearningforrumoridentificationonsocialmedia/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f99ffa1ece1a27f8c7ee4ed5462d2d0e85747c983850d6354909a387bb69c88c
+size 308576

adaptingentitiesacrosslanguagesandcultures/b83b9238-122e-4491-b8bb-76c1f417091d_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61c158aea3e8acb7a5580906840139f6cfe0e49a24009040b6258432e6cff626
+size 129427

adaptingentitiesacrosslanguagesandcultures/b83b9238-122e-4491-b8bb-76c1f417091d_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ae27ed69c14ee594497a5998b0f73b122588e6f101960ae2d58c2f4690fb760
+size 149960

adaptingentitiesacrosslanguagesandcultures/b83b9238-122e-4491-b8bb-76c1f417091d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0674e79b3982da5fe7467893bab51f5d5a8b826a7a407600a6d77b570a3f94f3
+size 879890

adaptingentitiesacrosslanguagesandcultures/full.md

@@ -0,0 +1,478 @@
+# Adapting Entities across Languages and Cultures
+
+Denis Peskov
+
+Computer Science University of Maryland
+
+dpeskov@umd.edu
+
+Viktor Hangya
+
+Center for Information and Language Processing LMU Munich
+
+hangyav@cis.lmu.de
+
+Jordan Boyd-Graber
+
+CS, LSC, UMIACS, and iSchool
+
+University of Maryland
+
+jbg@umiacs.umd.edu
+
+Alexander Fraser
+
+Center for Information and Language Processing
+
+LMU Munich
+
+fraser@cis.lmu.de
+
+# Abstract
+
+How would you explain Bill Gates to a German? He is associated with founding a company in the United States, so perhaps the German founder Carl Benz could stand in for Gates in those contexts. This type of translation is called adaptation in the translation community (Vinay and Darbelnet, 1995). Until now, this task has not been done computationally. Automatic adaptation could be used in natural language processing for machine translation and indirectly for generating new question answering datasets and education. We propose two automatic methods and compare them to human results for this novel NLP task. First, a structured knowledge base adapts named entities using their shared properties. Second, vector arithmetic and orthogonal embedding mappings identify better candidates, but at the expense of interpretable features. We evaluate our methods through a new dataset<sup>1</sup> of human adaptations.
+
+# 1 When Translation Misses the Mark
+
+Imagine reading a translation from German, "I saw Merkel eating a Berliner from Dietsch on the ICE". This sentence is opaque without cultural context.
+
+An extreme cultural adaptation for an American audience could render the sentence as "I saw Biden eating a Boston Cream from Dunkin' Donuts on the Acela", elucidating that Merkel is in a similar political post to Biden; that Dietsch (like Dunkin' Donuts) is a mid-range purveyor of baked goods; both Berliners and Boston Creams are filled, sweet pastries named after a city; and ICE and Acela are slightly ritzier high-speed trains. Human translators make this adaptation when it is appropriate to the translation (Gengshen, 2003).
+
+# Bill Gates
+
+Top Adaptations:
+
+WikiData
+
+F. Zeppelin
+
+Günther Jauch
+
+N. Harnoncourt
+
+3CosAdd
+
+constar
+
+Alnatura
+
+GMX
+
+Human
+
+A. Bechtolsheim
+
+Dietmar Hopp
+
+Carl Benz
+
+Table 1: WikiData and unsupervised embeddings (3CosAdd) generate adaptations of an entity, such as Bill Gates. Human adaptations are gathered for evaluation. American and German entities are color coded.
+
+Because adaptation is understudied, we leave the full translation task to future work. Instead, we focus on the task of cultural adaptation of entities: given an entity in a source, what is the corresponding entity in English? Most Americans would not recognize Christian Drosten, but the most efficient explanation to an American would be to say that he is the "German Anthony Fauci" (Loh, 2020). We provide top adaptations suggested by algorithms and humans for another American involved with the pandemic response, Bill Gates, in Table 1.
+
+Can machines reliably find these analogs with minimal supervision? We generate these adaptations with structured knowledge bases (Section 3) and word embeddings (Section 4). We elicit human adaptations (Section 5) to evaluate whether our automatic adaptations are plausible (Section 5.3).
+
+# 2 Wer ist Bill Gates?
+
+We define cultural adaptation and motivate its application for tasks like creating culturally-centered training data for QA. Vinay and Darbelnet (1995) define adaptation as translation in which the relationship not the literal meaning between the receiver and the content needs to be recreated.
+
+You could formulate our task as a tradi-
+
+tional analogy Drosten::Germany as Fauci::United States (Turney, 2008; Gladkova et al., 2016), but despite this superficial resemblance (explored in Section 4), traditional approaches to analogy ignore the influence of culture and are typically within a language. Hence, analogies are tightly bound with culture; humans struggle with analogies outside their culture (Freedle, 2003).
+
+We can use this task to identify named entities (Kasai et al., 2019; Arora et al., 2019; Jain et al., 2019) and for understanding other cultures (Katan and Taibi, 2004).
+
+# 2.1 ... and why Bill Gates?
+
+This task requires a list of named entities adaptable to other cultures. Our entities come from two sources: a subset of the top 500 most visited German/English Wikipedia pages and the non-official characterization list (Veale, 2016, NOC), “a source of stereotypical knowledge regarding popular culture, famous people (real and fictional) and their trade-mark qualities, behaviours and settings”. Wikipedia contains a plethora of singers and actors; we filter the top 500 pages to avoid a pop culture skew. We additionally select all Germans and a subset of Americans from the Veale NOC list as it is human-curated, verified, and contains a broader historical period than popular Wikipedia pages. Like other semantic relationships (Boyd-Graber et al., 2006), this is not symmetric. Thus, we adapt entities in both directions; while Berlin is the German Washington, DC, there is less consensus on what is the American Berlin, as Berlin is both the capital, a tech hub, and a film hub. A full list of our entities is provided in Appendix D.
+
+# 3 Adaptation from a Knowledge Base
+
+We first adapt entities with a knowledge base. We use WikiData (Vrandecic and Krötzsch, 2014), a structured, human-annotated representation of Wikipedia entities that is actively developed. This resource is well-suited to the task as features are standardized both within and across languages.
+
+Many knowledge bases explicitly encode the nationality of individuals, places, and creative works. Entities in the knowledge base are a discrete sparse vector, where most dimensions are unknown or not applicable (e.g., a building does not have a spouse).
+
+For example, Angela Merkel is a human (instance of), German (country of citizenship), politician (occupation), Rotarian (member of), Lutheran (religion), 1.65 meters tall (height), and has a PhD (academic degree). How would we find the "most similar" American adaptation to Angela Merkel? Intuitively, we should find someone whose nationality is American.
+
+Some issues immediately present themselves; contemporary entities will have more non-zero entries than older entities. Some characteristics are more important than others: matching unique attributes like "worked as journalist" is more important than matching "is human".
+
+Each entity in WikiData has "properties", which we can think about as the dimension of a sparse vector and "values" that those properties can take on. For example, Merkel has the properties "occupation" and "academic degree". Values for those properties are that her "occupation" is "politician" and her "academic degree" is a "doctorate". To match entities across cultures, we focus on matching properties rather than values; many of the values are more relevant inside a culture. For example, we cannot find American politicians who belong to the Christian Democratic Union, but we can find politicians who have an academic degree and a dissertation title.
+
+As a toy example, if Beethoven, Merkel, and Bach all have only two properties: Beethoven has an "occupation" and "genre", Merkel has an "Erdős number" and "political party", and Bach has a "occupation" and "genre", then Beethoven and Bach has a distance of zero and are the closest entities while Merkel has a distance of two since {"Erdős number", "political party"} is two away from {"occupation", "genre)}.
+
+First, we bifurcate WikiData into two sets: an American set  $\mathcal{A}$  for items which contain the value "United States of America" and a German set  $\mathcal{D}$  for those with German values. This is a liberal approximation, but it successfully excludes roughly seven out of the eight million items in WikiData. Then we explore the properties from WikiData. We create entity vectors with dimensions corresponding to frequently-occurring properties.
+
+The properties are discrete and categorical; Merkel either has an "occupation" or she does not. Each entity then has a sparse vector. We calculate the similarity of the vectors with Faiss's  $L_{2}$  distance (Johnson et al., 2021) and for each vector in  $\mathcal{A}$  find the closest vector in  $\mathcal{D}$  and vice versa.
+
+So who is the American Angela Merkel? One possible answer is Woodrow Wilson, a member of a "political party", who had a "doctoral advisor" and a "religion", and ended up with "awards". This answer may be unsatisfying as it was Barack Obama who sat across from Merkel for nearly a decade. To capture these more nuanced similarities, we turn to large text corpora in Section 4.
+
+# 4 An Alternate Embedding Approach
+
+While the classic NLP vector example (Mikolov et al., 2013c) isn't as magical as initially claimed (Rogers et al., 2017), it provides useful intuition. We can use the intuitions of the cliché:
+
+$$
+\overrightarrow {\text {K i n g}} - \overrightarrow {\text {M a n}} + \overrightarrow {\text {W o m a n}} = \overrightarrow {\text {Q u e e n}} \tag {1}
+$$
+
+to adapt between languages.
+
+This, however, requires relevant embeddings. First, we use the entire Wikipedia in English and German, preprocessed using Moses (Koehn et al., 2007). We follow Mikolov et al. (2013b) and use named entity recognition (Honnibal et al., 2020) to tokenize entities such as Barack_Obama.
+
+We use word2vec (Mikolov et al., 2013b), rather than FastText (Bojanowski et al., 2017), as we do not want orthography to influence the similarity of entities. Angela Merkel in English and in German have quite different neighbors, and we intend to keep it that way by preserving the distinction between languages.
+
+However, the standard word2vec model assumes a single monolingual embedding space. We use unsupervised Vecmap (Artetxe et al., 2018), a leading tool for creating cross-lingual word embeddings, to build bilingual word embeddings. We propose two approaches for adaptation.
+
+3CosAdd We follow the word analogy approach of  $3\mathrm{CosAdd}^4$  (Levy and Goldberg, 2014; Köper et al., 2016). American  $\rightarrow$  German adaptation takes the source entity's  $(v)$  embedding in the English vector space and looks for its adaptation  $(u^{*})$  based on embeddings in the German space. This is like the word analogy task, i.e., what entity has the
+
+role in the German culture as  $v$  does in American culture. As an example, Merkel has a similar role in the German culture as Biden. Formally, the adaptation of the English entity  $v$  into German is
+
+$$
+\vec {a} \equiv \operatorname {a v g} \left(\overrightarrow {E ^ {e n}} _ {\text {U n i t e d S t a t e s}}, \overrightarrow {E ^ {d e}} _ {\text {U S A}}\right) \tag {2}
+$$
+
+$$
+\vec {d} \equiv \operatorname {a v g} \left(\overrightarrow {E ^ {e \vec {n}}} _ {\text {G e r m a n y}}, \overrightarrow {E ^ {d e}} _ {\text {D e u t s c h l a n d}}\right) \tag {3}
+$$
+
+$$
+u ^ {*} = \underset {u \in V ^ {d e}} {\arg \max } \operatorname {s i m} \left(\overrightarrow {E _ {u} ^ {d e}}, \overrightarrow {E _ {v} ^ {e n}} - \overrightarrow {a} + \overrightarrow {d}\right), \tag {4}
+$$
+
+where  $\overrightarrow{E_w^l}$  is the embedding of word  $w$  in language  $l$ ,  $V^{de}$  is the German vocabulary and sim is the cosine similarity. The American anchor word  $\vec{a}$  and German anchor  $\vec{d}$  represent the American and German cultures. We average the English and German embeddings of the individual word types for robust anchor vectors. In standard analogies, as in Equation 1, the  $\vec{a}$  and  $\vec{d}$  vectors are different for each test pair; here they are the same for each example, as we always are pivoting between the two cultures.
+
+Learned adaptation To eliminate the need for manual anchor selection for both cultures, our second approach learns the adaptation as a linear transformation of source embeddings to the target culture given a few adaptation examples. Specifically, we use the human adaptations sourced for the Wikipedia entities as training for the Veale NOC ones. We follow the work of Mikolov et al. (2013a) and learn a transformation matrix  $\mathbf{W}_{en\rightarrow de}$  for American  $\rightarrow$  German by minimizing the  $L_{2}$  distance of  $\mathbf{W}_{en\rightarrow de}\overrightarrow{E}_{vi}^{en}$  and  $\overrightarrow{E}_{ui}^{de}$  over gold adaptation  $v_{i},u_{i = 1}^{n}$  entity pairs. The adaptation of a source entity  $v$  is  $u^{*} = \mathbf{W}_{en\rightarrow de}\overrightarrow{E}_{v}^{en}$ . Likewise, we learn the reverse mapping  $\mathbf{W}_{de\rightarrow en}$  for German  $\rightarrow$  American adaptation. This requires supervised training data—but not much (Conneau et al., 2018)—which we collect in Section 5.
+
+# 5 Comparing Automation to Human Judgment
+
+The automated methods can generate entities at scale, but humans have to evaluate their relevance.
+
+# 5.1 Adaptation by Locals
+
+Since quality control is difficult for generation (Peskov et al., 2019), we need users who
+
+will answer the task accurately. We recruit five American citizens educated at American universities and five German citizens educated at German ones. These human annotations serve as a gold standard against which we can compare our automated approaches. To improve the user experience, we create an interface that provides a brief summary of each source entity from Wikipedia and asks the users to select a target adaptation that autocompletes Wikipedia page titles (all entities; targets are not limited to the lists in Section 2) in a text box  $a$  la answer selection in Wallace et al. (2019). The annotation task requires two hours for our users to complete. Obviously, German annotators are more familiar with German culture than the Americans, and vice-versa. Annotators translate into their native language. Since we are focusing on popular entities, they are often known despite the cultural divide, but the introductory paragraph from Wikipedia reminds users if not.
+
+# 5.2 Are the Adaptations Plausible?
+
+To validate and compare all our adaptation strategies' precision, five German translators who understand American culture assess the adaptations. The top five adaptations from WikiData, 3CosAdd, learned adaptation, and humans—as well as five randomly selected options from the human pool—are evaluated for plausibility on a five-level Likert scale. $^{7}$  Fleiss' Kappa (0.382) and Krippendorf's Alpha (0.381) assess interannotator Agreement; this "fair" agreement suggests that vetting an adaptation is challenging and sometimes subjective, even for translators.
+
+# 5.3 Why Adaptation is Difficult
+
+Embedding adaptations are better than Wikidata's, and human adaptations are better still (Figure 1). Thus, we use human adaptations as the gold standard for evaluating recall. Only the learned embedding method uses training data, so we use human adaptations from Wikipedia to train the projection matrix and evaluate (for all methods) using human adaptations the NOC list. Given that the task is subjective, we take our results with a grain of salt given cultural variation (e.g., some people view Angela Merkel's conservatism as a defining characteristic, while others focus on her science pedigree).
+
+![](images/1f34691a853b4c6ef7e6456e417a8ea3b7b4e2254652c2ac91374dee5f2f93ae.jpg)  
+Figure 1: We validate adaptation strategies with expert translators on a five-point Likert scale. The human-generated adaptations are rated best—between "related" (3) and "similar" (4). These human adaptations become the reference for evaluation in Table 2.
+
+<table><tr><td>Data</td><td>Metric</td><td>WikiData</td><td>3CosAdd</td><td>Learned</td></tr><tr><td colspan="5">American→German</td></tr><tr><td rowspan="3">Wikipedia</td><td>Rec@5</td><td>7.5%</td><td>14.2%</td><td>-</td></tr><tr><td>Rec@100</td><td>34.4%</td><td>52.8%</td><td>-</td></tr><tr><td>MRR</td><td>0.05</td><td>0.10</td><td>-</td></tr><tr><td rowspan="3">Veale NOC</td><td>Rec@5</td><td>3.0%</td><td>22.9%</td><td>28.6%</td></tr><tr><td>Rec@100</td><td>42.4%</td><td>51.4%</td><td>45.7%</td></tr><tr><td>MRR</td><td>0.03</td><td>0.17</td><td>0.24</td></tr><tr><td colspan="5">German→American</td></tr><tr><td rowspan="3">Wikipedia</td><td>Rec@5</td><td>3.1%</td><td>17.2%</td><td>-</td></tr><tr><td>Rec@100</td><td>15.4%</td><td>40.5%</td><td>-</td></tr><tr><td>MRR</td><td>0.01</td><td>0.12</td><td>-</td></tr><tr><td rowspan="3">Veale NOC</td><td>Rec@5</td><td>0.0%</td><td>25.0%</td><td>25.0%</td></tr><tr><td>Rec@100</td><td>25.0%</td><td>70.0%</td><td>55.0%</td></tr><tr><td>MRR</td><td>0.02</td><td>0.12</td><td>0.15</td></tr></table>
+
+Table 2: If we consider human adaptations as correct, where do they land in the ranking of automatic adaptation candidates? In this recall-oriented approach, learned mappings (which use a small number of training pairs), rate highest.
+
+We use the mean reciprocal rank (Voorhees, 1999, MRR) to measure how high the gold adaptations are ranked by our other adaptation strategies. Since MRR decreases geometrically and our gold standard is not exhaustive, the Recall@5, and @100 metrics are more intuitive. We calculate Recall@n by measuring what fraction of the correct adaptations of a source entity is retrieved in the top n predictions. Table 2 validates that the human annotations are near the top of the automatic adaptations; the precision-oriented evaluation (Figure 1) validates whether the top of the list is reasonable. All human annotations and a sample of the automatic adaptations are provided in Appendix D.
+
+# 5.4 Qualitative Analysis
+
+There is no single answer to what makes a good adaptation. Let us return to the question of who Bill Gates is, which underlines how there is often no one right answer to this question but several context-specific possibilities. The human adaptations show the range of plausible adaptations, each appropriate for a particular facet of the position Bill Gates has in US society. As previously mentioned, Carl Benz represents a larger than life founder who created an entire industry with his company. However, Carl Benz made cars, not computers.
+
+Even within technology, different adaptations highlight different aspects of Bill Gates. Like the implementer of the BASIC programming language, Konrad Zuse contributed to computers that were more than single-purpose machines. Just as as Bill Gates's Microsoft is seen as a stodgy tech giant, Dietmar Hopp founded SAS, a giant German tech company that is more often discussed in board rooms than in living rooms. And because the epicenter of modern tech is America's West Coast, Andreas von Bechtolsheim represents a German founder of Sun Microsystems and early Google investor that made his way to Silicon Valley.
+
+Other times, there is more consensus: a majority of raters declare Angela Merkel is the German Hilary Clinton, and Joseph Smith is the American Martin Luther. There are even some unanimous adaptations: Bavaria is the German California. Adaptations of fictional characters seem particularly difficult, although this may represent the supremacy of American popular culture; Superman and Homer Simpson are so well known in Germany that there are no clear adaptations; Till Eulenspiegel, Maverick, Bibi Blocksberg are not superheroes from a dying world and Heidi is not a dumb, bald everyman.
+
+# 6 A New Computational Task
+
+We formally introduce entity adaptation as a new computational task. Word2vec embeddings and WikiData can be used to figuratively—not just literally—translate entities into a different culture. Humans are better at generating candidates for this task than our computational methods (Figure 1). These methods are well-motivated, but have room for improvement. Knowledge bases improve over time and increased coverage of entities—as well as improved information about each entity—would improve the method. Alternate
+
+word embedding approaches—perhaps those that discard orthography—may provide better candidates. Even humans occasionally disagree with other humans on this task, so evaluation for this task is nontrivial.
+
+Our new dataset of machine-generated adaptations, human adaptations, and human evaluation of these adaptations can serve as an evaluation for future automatic methods.
+
+People need NLP systems that reflect their language and culture, but datasets are lacking: adaptation can help. There has been an explosion of English-language QA datasets, but other languages continue to lag behind. Several approaches try to transfer English's bounty to other languages (Lewis et al., 2020; Artetxe et al., 2019), but most of the entities asked about in major QA datasets are American (Gor et al., 2021). Adapting entire questions will require not just adapting entities and non-entities in tandem but will also require integration with machine translation (Kim et al., 2019; Hangya and Fraser, 2019). Our automatic methods did not create precise adaptations, but the alternative "incorrect" adaptations may be useful for low-precision tasks, such as generating numerous simple open-ended questions or gauging the popularity of a entity.
+
+Given the existence of robust datasets in high resource languages can we adapt, rather than literally translate, them to other cultures and languages?
+
+# 7 Acknowledgments
+
+Peskov was supported by a DAAD Research Fellowship and by the wonderful faculty and students of Ludwig Maximilians Universität München. Fraser is supported by the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement No. 640550) and by German Research Foundation (DFG; grant FR 2829/4-1). Boyd-Graber is supported by NSF Grant IIS-1822494. Any opinions, findings, conclusions, or recommendations expressed here are those of the authors and do not necessarily reflect the view of the sponsors.
+
+We thank Sander Schulhoff for his web development expertise and Pedro Rodriguez for his help with processing WikiData. Thanks to Phillip Dufter, Benno Krojer, Stephan Huber, Janika Linke, Andrew Guo, Adam Visokay, Christian Rice, Connor Knight, and others for help with adaptations.
+
+# Ethics
+
+We worked with human participants to collect our data. They are all adults who participated of their own volition and no payment was made. No personal data was collected or used for the dataset. For evaluation of the adaptations, we hired translators through Upwork. They were paid $40 for a task that took roughly between one and two hours.
+
+The broad motivation of this work is to spread cultural understanding. Humans must be kept in the-loop for making claims about cultural relevance. Having multiple diverse opinions is necessary for supporting any cultural claim. Like with language, nationality is often correlated with culture, but is not synonymous. Large countries contain multitudes, while some nationalities (e.g., Kurds) lack a de jure nation but span many nations. We elide this detail and focus on information often available in knowledge bases.
+
+These lists contain figures that are controversial. From a research perspective, research datasets should reflect the real world and prior work, thus we include prominent entities as identified by Veale NOC and Wikipedia. Any list may contain biases in the collection processes, and this should not be thought of as an exclusive and definitive list, but as a start that can be refined and ultimately expanded to other cultures.
+
+# References
+
+Ravneet Arora, Chen-Tse Tsai, Ketevan Tsereteli, Prabhanjan Kambadur, and Yi Yang. 2019. A semi-Markov structured support vector machine model for high-precision named entity recognition. In Proceedings of the Association for Computational Linguistics.  
+Mikel Artetxe, Gorka Labaka, and Eneko Agirre. 2018. A robust self-learning method for fully unsupervised cross-lingual mappings of word embeddings. In Proceedings of the Association for Computational Linguistics.  
+Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2019. On the cross-lingual transferability of monolingual representations. CoRR, abs/1910.11856.  
+Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information. In Proceedings of the Association for Computational Linguistics.  
+Jordan Boyd-Graber, Christiane Fellbaum, Daniel Osherson, and Robert Schapire. 2006. Adding dense, weighted, connections to WordNet. In Proc. Global
+
+WordNet Conference 2006. Global WordNet Association.  
+Alexis Conneau, Guillaume Lample, Marc'Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou. 2018. Word translation without parallel data. In Proceedings of the International Conference on Learning Representations.  
+Roy Freedle. 2003. Correcting the SAT's ethnic and social-class bias: A method for reestimating sat scores. Harvard Educational Review, 73(1):1-43.  
+Hu Gengshen. 2003. Translation as adaptation and selection. Perspectives: Studies in Translatology, 11(4):283-291.  
+Anna Gladkova, Aleksandr Drozd, and Satoshi Matsuoka. 2016. Analogy-based detection of morphological and semantic relations with word embeddings: what works and what doesn't. In Proceedings of the NAACL Student Research Workshop.  
+Maharshi Gor, Kellie Webster, and Jordan Boyd-Graber. 2021. Toward deconfounding the influence of subject's demographic characteristics in question answering. In Proceedings of Empirical Methods in Natural Language Processing.  
+Abigail Green. 2003. Representing Germany? the zollverein at the world exhibitions, 1851-1862. The Journal of Modern History, 75(4):836-863.  
+Viktor Hangya and Alexander Fraser. 2019. Unsupervised parallel sentence extraction with parallel segment detection helps machine translation. In Proceedings of the Association for Computational Linguistics.  
+Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. 2020. spaCy: Industrial-strength Natural Language Processing in Python.  
+Alankar Jain, Bhargavi Paranjape, and Zachary C. Lipton. 2019. Entity projection via machine translation for cross-lingual NER. In Proceedings of the Association for Computational Linguistics.  
+Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2021. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547.  
+David Jurgens, Mohammad Taher Pilehvar, and Roberto Navigli. 2014. SemEval-2014 Task 3: Cross-level semantic similarity. In Proceedings of the Workshop on Semantic Evaluation. Association for Computational Linguistics.  
+Jungo Kasai, Kun Qian, Sairam Gurajada, Yunyao Li, and Lucian Popa. 2019. Low-resource deep entity resolution with transfer and active learning. In Proceedings of the Association for Computational Linguistics.
+
+David Katan and Mustapha Taibi. 2004. Translating cultures: An introduction for translators, interpreters and mediators. Routledge.  
+Yunsu Kim, Yingbo Gao, and Hermann Ney. 2019. Effective cross-lingual transfer of neural machine translation models without shared vocabularies. In Proceedings of the Association for Computational Linguistics.  
+Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, et al. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of the Association for Computational Linguistics.  
+Maximilian Köper, Sabine Schulte im Walde, Max Kisselew, and Sebastian Padó. 2016. Improving zero-shot-learning for german particle verbs by using training-space restrictions and local scaling. In Proceedings of the Fifth Joint Conference on Lexical and Computational Semantics.  
+Omer Levy and Yoav Goldberg. 2014. Linguistic regularities in sparse and explicit word representations. In Conference on Computational Natural Language Learning.  
+Patrick Lewis, Barlas Oğuz, Rudy Rinott, Sebastian Riedel, and Holger Schwenk. 2020. MLQA: Evaluating cross-lingual extractive question answering. In Proceedings of the Association for Computational Linguistics.  
+Tim Loh. 2020. Germany has its own Dr. Fauci—and actually follows his advice. Bloomberg.  
+Tomas Mikolov, Quoc V. Le, and Ilya Sutskever. 2013a. Exploiting Similarities among Languages for Machine Translation. CoRR, abs/1309.4.  
+Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013b. Distributed representations of words and phrases and their compositionality. In Proceedings of Advances in Neural Information Processing Systems.  
+Tomáš Mikolov, Wen-tau Yih, and Geoffrey Zweig. 2013c. Linguistic regularities in continuous space word representations. In Conference of the North American Chapter of the Association for Computational Linguistics.  
+Denis Peskov, Nancy Clarke, Jason Krone, Brigi Fodor, Yi Zhang, Adel Youssef, and Mona Diab. 2019. Multi-domain goal-oriented dialogues (MultiDoGO): Strategies toward curating and annotating large scale dialogue data. In Proceedings of Empirical Methods in Natural Language Processing.  
+Anna Rogers, Aleksandr Drozd, and Bofang Li. 2017. The (too many) problems of analogical reasoning with word vectors. In Proceedings of the Joint Conference on Lexical and Computational Semantics.
+
+Hagen Schulze. 1991. The Course of German Nationalism: From Frederick the Great to Bismarck 1763-1867. Cambridge University Press.  
+Peter D Turney. 2008. A uniform approach to analogies, synonyms, antonyms, and associations. In Proceedings of International Conference on Computational Linguistics.  
+Tony Veale. 2016. Round up the usual suspects: Knowledge-based metaphor generation. In Proceedings of the Fourth Workshop on Metaphor in NLP.  
+Jean-Paul Vinay and Jean Darbelnet. 1995. Comparative stylistics of French and English: A methodology for translation, volume 11. John Benjamins Publishing.  
+Ellen M Voorhees. 1999. The TREC-8 question answering track report. In Proceedings of the Text Retrieval Conference, volume 99.  
+Denny Vrandecic and Markus Krötzsch. 2014. Wiki-data: a free collaborative knowledgebase. Communications of the ACM, 57(10):78-85.  
+Eric Wallace, Pedro Rodriguez, Shi Feng, Ikuya Yamada, and Jordan Boyd-Graber. 2019. Trick me if you can: Human-in-the-loop generation of adversarial question answering examples. Transactions of the Association of Computational Linguistics, 10.
+
+# A Appendix
+
+Our appendix contains our entire human-collected dataset, as well as a sample of our WikiData and embedding approaches for adaptation.
+
+Figure 2 shows our collection tool. Table 3 shows German  $\rightarrow$  American Veale NOC items. Table 4 shows American  $\rightarrow$  German Veale NOC items. Table 5 shows German  $\rightarrow$  American Veale NOC items. Table 6 shows American  $\rightarrow$  German Veale NOC items.
+
+Table 7 shows our WikiData predictions, Table 8 shows our 3CosAdd predictions. and Table 9 shows our Learned Adaptations predictions. We pose several background questions about Wikipedia and WikiData as well:
+
+# B Wikipedia Analysis
+
+Are the Wikipedia pages in German and English visited from the associated country? Yes; the Wikipedias for the respective languages are most used by visitors located in those countries:  $63\%$  of German wikipedia was visited from Germany and  $32\%$  of English Wikipedia was visited from the United States in the past year.[9]
+
+Are the top Wikipedia topics notably different across languages? Yes; less than a quarter of top 500 searches for 2019 are identical across English and German.
+
+Does WikiData cover areas outside of the United States? Wikipedia coverage does not mean that WikiData annotations are conducted equally across German and American entities. Analyzing WikiData<sup>10</sup> reveals a discrepancy in coverage of Germans and Americans.
+
+Out of 8,126,559 titles, 1,030,762 include a reference to the United States in any capacity. However, only 184,692 contain a reference to (broader) Germany. This imbalance is significant but has enough German items for our methodology. As WikiData is a maintained resource, there is room for future additional coverage and standardization of fields.
+
+Countries use different names throughout history. While the United States of America is straightforward, Germany includes several variations, such as: German Empire, the Kingdom of Bavaria, the Kingdom of Prussia, etc. The WikiData feature-based approach can be used for other countries as well (... or anything that is consistently coded). For example, there are 65,957 Russian, 152,701 French, and 48,026 Chinese items in WikiData.[11]
+
+Are the top Wikipedia topics necessarily belonging to the culture? No; the top 10 most visited German Wikipedia includes a cultural potpurri: Germany, Greta Thurnberg, Asperger Syndrome, Game of Thrones, and Freddie Mercury. While there are uniquely German entities in the longer list—ZDF, Capital Bra, The Cratez, Niki Lauda—we cannot conclude that all top entities in a language belong culturally to a given country. Therefore, we need a stricter methodology.
+
+Where does one find entities? We rely on a human-sourced dataset: Veale's Non-Official Characterization list (Veale, 2016). This list contains 1031 people, real and fictional, such as Daniel Day-Lewis, Anton Chekhov, and Bridget Jones. These people are annotated with properties, one of which is conveniently their address. There are 25 people with a German location and 575 with an American one. Removing fictional characters written by non-nationals causes the German leaves the list with 20 entities. An American author filters the list of Americans down to 35 iconic ones with achievements that span politics, music, activism, athletics, and pop culture.
+
+Wikipedia provides another avenue for gauging popular topics in a language. We manually filter the top 500 German/English Wikipedia topics to remove non-German/non-American entities; Game of Thrones and Unix-Shell are popular in the German Wikipedia, but they are not culturally idiosyncratic. For the 2019 German Wikipedia we are left with roughly 200 items, which we further reduce down to 120
+
+after putting a cap on pop culture entities. For the American counterpart, over 300 items are culturally American. We add a three-year filter to remove pop items to make it comparable to the German one.
+
+# C Interfaces
+
+We are studying cultural differences between German and American wikipedia. These are entities that are top 500 entities from Wikipedia for the German language. Please type whichever AMERICAN entity you think is most similar to the provided German entity. If you are unfamiliar with the entity, you may reference an outside source.
+
+The following German Entity is most similar to which American Entity:
+
+# Deutschland
+
+Germany (German: Deutschland, German pronunciation: ['doytflant]), officially the Federal Republic of Germany (German: Bundesrepublik Deutschland, listen), is a country in Central and Western Europe. Covering an area of 357,022 square kilometres (137,847 sq mi), it lies between the Baltic and North seas to the north, and the Alps to the south. It borders Denmark to the north, Poland and the Czech Republic to the east, Austria and Switzerland to the south, and France, Luxembourg, Belgium and the Netherlands to the west. Various Germanic tribes have inhabited the northern parts of modern Germany since classical antiquity. A region named Germania was documented before AD 100.
+
+![](images/a22eb23719aef1608d488425ec17324f035fe2c5b2410edc70bdb8806d965576.jpg)  
+Figure 2: Our interface provides users with information about the entity and asks them to select an option from possible Wikipedia pages
+
+Compare the below German entities to this American entity: Abraham Lincoln / Abraham Lincoln was an American statesman and lawyer who served as the 16th president of the United States from 1861 until his assassination in 1865.
+
+Click for Instructions
+
+Konrad Adenauer / Konrad Hermann Joseph Adenauer was a German statesman who served as the first Chancellor of the Federal Republic of Germany from 1949 to 1963.
+
+Helmut Schmidt / Helmut Heinrich Waldemar Schmidt was a German politician and member of the Social Democratic Party of Germany, who served as Chancellor of the Federal Republic of Germany from 1974 to 1982.
+
+Willy Brandt / Willy Brandt was a German politician and statesman who was leader of the Social Democratic Party of Germany from 1964 to 1987 and served as Chancellor of the Federal Republic of Germany from 1969 to 1974.
+
+Helmut Kohl / Helmut Josef Michael Kohl was a German statesman and politician of the Christian Democratic Union who served as Chancellor of Germany from 1982 to 1998 and as chairman of the CDU from 1973 to 1998.
+
+![](images/154b640cde17faf0d2201edda8c4b4dd279e8a7cc23d3f774a3c9378cfb0b0a9.jpg)  
+Figure 3: Our Qualtrics survey
+
+# D Data
+
+<table><tr><td>Entity</td><td>Human Adaptation: NOC German→American</td></tr><tr><td>Adolf Eichmann</td><td>Andrew Jackson, Andrew Jackson, Franklin D. Roosevelt, Nathan Bedford Forrest, Steve Bannon</td></tr><tr><td>Angela Merkel</td><td>Barack Obama, Donald Trump, Hillary Clinton, Hillary Clinton, Hillary Clinton, Hillary Clinton, Joe Biden</td></tr><tr><td>Baron Munchausen</td><td>Captain America, Daniel Bolger, Joseph Smith, Paul Bunyan, Robert Jordan , Yankee Doodle</td></tr><tr><td>Carl von Clausewitz</td><td>Alfred Thayer Mahan, Dwight D. Eisenhower, Henry Knox, Robert E. Lee, Ulysses S. Grant</td></tr><tr><td>Friedrich Nietzsche</td><td>Ayn Rand, Henry David Thoreau, Henry Thoreau, Jordan Peterson, William James</td></tr><tr><td>Henry Kissinger</td><td>Henry Kissinger, Henry Kissinger, John Kerry, Madeleine Albright, Richard Nixon</td></tr><tr><td>Immanuel Kant</td><td>Benjamin Franklin, John Dewey, John Locke, John Rawls, Robert Nozick</td></tr><tr><td>Johann Sebastian Bach</td><td>Aaron Copland, Elvis Presley, Elvis Presley, Irving Berlin, Johnny Cash, Scott Joplin</td></tr><tr><td>Johann Wolfgang von Goethe</td><td>Edgar Allan Poe, Ernest Hemingway, Walt Whitman</td></tr><tr><td>Johannes Gutenberg</td><td>Benjamin Franklin, Bill Gates, Eli Whitney, Thomas Edison</td></tr><tr><td>Joseph Goebbels</td><td>David Duke, Franklin D. Roosevelt, George Rockwell, Rupert Murdoch, david duke</td></tr><tr><td>Karl Lagerfeld</td><td>Anna Wintour, Anna Wintour, Marc Jacobs, Ralph Lauren, Ralph Lauren, Ralph Lauren</td></tr><tr><td>Karl Marx</td><td>Angela Davis, Beck, Bernie Sanders, John Jay, John Rawls, John Rawls</td></tr><tr><td>Leni Riefenstahl</td><td>DW Griffith, David Wark Griffith, Frank Capra, Judy Garland</td></tr><tr><td>Ludwig van Beethoven</td><td>Aaron Copland, Aaron Copland, Aaron Copland, Elvis Presley, Frank Sinatra, George Gershwin, George Gershwin, Scott Joplin</td></tr><tr><td>Marlene Dietrich</td><td>Bette Davis, Clara Bow, Elizabeth Taylor, Marilyn Monroe, William Tecumseh Sherman</td></tr><tr><td>Martin Luther</td><td>Barry Goldwater, Brigham Young, Joseph Smith, Joseph Smith, Joseph Smith</td></tr><tr><td>Otto von Bismarck</td><td>Abraham Lincoln, George Washington, George Washington, George Washington, George Washington, Ulysses S. Grant</td></tr><tr><td>Pope Benedict XVI</td><td>Billy Graham, Billy Graham, Brigham Young, John Carroll , Seán Patrick O’Malley</td></tr><tr><td>Richard Wagner</td><td>Charles Ives, Frank Sinatra, Leonard Bernstein, Philip Glass</td></tr></table>
+
+Table 3: Veale NOC German→American adaptations.
+
+Entity  
+Human Adaptation: NOC American  $\rightarrow$  German adaptations  
+
+<table><tr><td>Abraham Lincoln</td><td>Helmut Kohl, Konrad Adenauer, Wilhelm Friedrich Ludwig von Preußen, Willy Brandt, Willy Brandt</td></tr><tr><td>Al Capone</td><td>Adolf Leib, Carlos Lehder-Rivas, Jan Marsalek, Nasser Abou-Chaker, Nasser About-Chaker</td></tr><tr><td>Alfred Hitchcock</td><td>Bernd Eichinger, Bernd Eichinger, Michael Bully Herbig, Roland Emmerich, Wim Wenders</td></tr><tr><td>Benedict Arnold</td><td>Hansjoachim Tiedge, Otto von Bismarck, Otto von Bismarck, Robert Blum</td></tr><tr><td>Bill Gates</td><td>Andreas von Bechtolsheim, Carl Benz, Dietmar Hopp, Konrad Zuse</td></tr><tr><td>Britney Spears</td><td>Helene Fischer, Herbert Grönemeyer, Jeanette Biedermann, Nena, Til Schweiger</td></tr><tr><td>Charles Lindbergh</td><td>Ferdinand von Richthofen, Heinrich Horstman, Karl Wilhelm Otto Lilienthal, Ludwig Hofmann, Wernher von Braun</td></tr><tr><td>Donald Trump</td><td>Adolf Hitler, Adolf Hitler, Carsten Maschmeyer, Christian Lindner</td></tr><tr><td>Elvis Presley</td><td>Peter Kraus, Rammstein, The Scorpions, Udo Lindenberg, Udo Lindenberg</td></tr><tr><td>Ernest Hemingway</td><td>Günter Grass, Hermann Hesse, Johann Wolfgang von Goethe, Karl May, Martin Walser</td></tr><tr><td>Frank Lloyd Wright</td><td>Gerhard Richter, Hugo Häring, Karl Lagerfeld, Max Dudler, Walter Gropius</td></tr><tr><td>George Washington</td><td>Friedrich II, Heinrich I, Konrad Adenauer, Otto I. der Groß, Otto von Bismarck</td></tr><tr><td>Henry Ford</td><td>Carl Benz, Carl Benz, Carl Benz, Ferdinand Porsche, Gottlieb Wilhelm Daimler</td></tr><tr><td>Hillary Clinton</td><td>Angela Merkel, Angela Merkel, Angela Merkel, Kramp-Karrenbauer, Sahra Wagenknecht</td></tr><tr><td>Homer Simpson</td><td>Alf, Heidi, Pumuckl, Werner, Werner - Beinhart!</td></tr><tr><td>Jack The Ripper</td><td>Armin Meiwes, Der Bulle von Tölz, Joachim Kroll, Karl Denke, Rudolf Pleil</td></tr><tr><td>Jay Z</td><td>Capital Bra, Marteria, Sido, Sido, Sido</td></tr><tr><td>Jimi Hendrix</td><td>Bela B., Gisbert zu Knyphausen, Herbert Grönemeyer, Rudolf Schenker, Spider Murphy Gang</td></tr><tr><td>John F. Kennedy</td><td>Hanns Martin Schleyer, Willy Brandt, Willy Brandt, Wolfgang Schäuble</td></tr><tr><td>Kim Kardashian</td><td>Carmen Geiss, Gina-Lisa Lohfink, Heidi Klum, Heidi Klum, Sarah Connor</td></tr><tr><td>Louis Armstrong</td><td>Günter Sommer, Helmut Brandt, Jan Delay, Michael Abene, Mozart</td></tr><tr><td>Marilyn Monroe</td><td>Heidi Klum, Ingrid Steeger, Marlene Dietrich, Micaela Schäfer, Uschi Glas</td></tr><tr><td>Michael Jordan</td><td>Dirk Nowitzki, Dirk Nowitzki, Dirk Nowitzki, Franz Beckenbauer, Michael Schuhmacher</td></tr><tr><td>Neil Armstrong</td><td>Alexander Gerst, Sigmund Jahn, Sigmund Jahn, Ulf Merbold, Wernher von Braun</td></tr><tr><td>Noam Chomsky</td><td>Helmut Glück, Juergen Habermas, Jürgen Habermas, Ludwig Wittgenstein, Wilhelm Röttgen</td></tr><tr><td>Oprah Winfrey</td><td>Anne Will, Arabella Kiesbauer, Maybrit Illner, Thomas Gottschalk, Thomas Gottschalk</td></tr><tr><td>Orville Wright</td><td>Carl Benz, Gustav Otto, Gustav Weißkopf, Otto Lilienthal, Wern-her von Braun</td></tr><tr><td>Richard Nixon</td><td>Franz Josef Strauss, Helmut Kohl, Ludwig Erhard, Ludwig Erhard, Richard von Weizsäcker</td></tr><tr><td>Rosa Parks</td><td>Anne Wizorek, Marie Juchacz, Sophie Scholl, Sophie Scholl, Vera Lengsfeld</td></tr><tr><td>Serena Williams</td><td>Andrea Petkovic, Boris Becker, Sabine Lisicki, Steffi Graf, boris becker</td></tr><tr><td>Steve Jobs</td><td>Carl Benz, Dietmar Hopp, Dietmar Hopp, Karl Lagerfeld</td></tr><tr><td>Steven Spielberg</td><td>Michael Bully Herbig, Roland Emmerich, Roland Emmerich, Roland Emmerich, Wim Wenders</td></tr><tr><td>Superman</td><td>Bibi Blocksberg, Fix and Foxi, Maverick, Superman, Till Eulenspiegel</td></tr><tr><td>Tiger Woods</td><td>Boris Becker, Martin Kaymer, Martin Kaymer, Michael Schumacher, Serge Gnabry</td></tr><tr><td>Walt Disney</td><td>Axel Springer, Christian Becker, Franz Mack, Gerhard Hahn, Röttger Feldmann</td></tr></table>
+
+Table 4: Veale NOC American  $\rightarrow$  German adaptations.
+
+# Entity
+
+# Human Adaptation: Wikipedia German→American
+
+<table><tr><td>ARD</td><td>NPR, PBS, PBS</td></tr><tr><td>Adolf Hitler</td><td>Donald Trump, Donald Trump, Franklin D. Roosevelt, Franklin D. Roosevelt, Franklin D. Roosevelt</td></tr><tr><td>Airbus</td><td>Boeing, Boeing, Boeing, Boeing, Lockheed Martin</td></tr><tr><td>Albert Einstein</td><td>Carl Sagan, J. Robert Oppenheimer, J. Robert Oppenheimer, John Forbes Nash Jr., Thomas Edison</td></tr><tr><td>Alice Merton</td><td>Ariana Grande, Elle King, K.T. Tunstall, P!NK, Vanessa Carlton</td></tr><tr><td>Alternative für Deutschland</td><td>Libertarian Party , Republican Party, Tea Party movement</td></tr><tr><td>Andrea Nahles</td><td>Elizabeth Warren, Hillary Clinton, Nancy Pelosi, Tammy Duckworth</td></tr><tr><td>Andrej Mangold</td><td>Kawhi Leonard, Kevin Durant, Kris Humphries, Yao Ming</td></tr><tr><td>Annalena Baerbock</td><td>Al Gore, Al Gore, Alexandria Ocasio-Cortez, Bernie Sanders, Jill Stein</td></tr><tr><td>Anne Frank</td><td>Anna Green Winslow, Clara Barton, Emmett Till, Kunta Kinte</td></tr><tr><td>Annegret Kramp-Karrenbauer</td><td>Condoleezza Rice, Hillary Clinton</td></tr><tr><td>AnnenMayKantereit</td><td>Guns N' Roses, Milky Chance, Polar Bear Club, Red Hot Chili Peppers</td></tr><tr><td>Apache 207</td><td>Fetty Wap, Tekashi 69, XXXTentacion, Zayn Malik</td></tr><tr><td>Arnold Schwarzenegger</td><td>Chuck Norris, Dwayne Johnson, Ronnie Coleman, Sylvester Stallone, Sylvester Stallone</td></tr><tr><td>BMW</td><td>Cadillac, Cadillac, Chevrolet, Chrysler</td></tr><tr><td>Babylon Berlin</td><td>Game of Thrones, Man From U.N.C.L.E., Peaky Blinders , The Americans, Turn</td></tr><tr><td>Baden-Württemberg</td><td>California, Chicago metropolitan area, San Diego, Southern United States, Texas</td></tr><tr><td>Bastian Yotta</td><td>Chad Johnson, Colton Underwood, Dan Bilzerian</td></tr><tr><td>Bauhaus</td><td>Frank Lloyd Wright</td></tr><tr><td>Bayerischer Rundfunk</td><td>NPR, National Public Radio, National Public Radio, national public ra</td></tr><tr><td>Bayern</td><td>Florida, New York, The Confederacy</td></tr><tr><td>Benjamin Piwko</td><td>Bruce Lee, Colton Underwood, Derek Hough</td></tr><tr><td>Berlin</td><td>New York City, Portland Oregon, Washington D.C., Washington D.C., Washington D.C.</td></tr><tr><td>Berliner Mauer</td><td>Border Patrol Police, Mason-Dixon line, Mason-Dixon line, US-Mexican border</td></tr><tr><td>Bertolt Brecht</td><td>Tennessee Williams, Tennessee Williams</td></tr><tr><td>Björn Höcke</td><td>Lindsey Graham, Mike Pence</td></tr><tr><td>Borussia Dortmund</td><td>Golden State Warriors, New England Patriots, New England Patriots</td></tr><tr><td>Brandenburg</td><td>Maryland, New York, Northeastern United States, Richmond Virginia, Virginia</td></tr><tr><td>Bruno Ganz</td><td>Clint Eastwood, Ethan Hawke, Marlon Brando, Robert De Niro, Robert De Niro</td></tr><tr><td>Bundespresident</td><td>First Lady, President of the United States, Speaker of the House</td></tr><tr><td>Bundeswehr</td><td>Department of Defense , US military, United States Armed Forces, United States Army</td></tr><tr><td>Capital Bra</td><td>Drake, Eminem, Eminem, Kanye West, Kendrick Lamar</td></tr><tr><td>Carola Rackete</td><td>American Civil Liberties Union, Dawn Wooten, Rosa Parks, Whale Wars</td></tr><tr><td>Carolin Kebekus</td><td>Amy Schumer, Sarah Silverman, Tina Fey, Tina Fey</td></tr><tr><td>Charité</td><td>Call the Midwife, Grey's Anatomy, Grey's Anatomy, The Queen's Gambit</td></tr><tr><td>Chris Töpperwien</td><td>Gordon Ramsey , Guy Fieri, Jeff Probst</td></tr><tr><td>Christoph Waltz</td><td>Anthony Hopkins, Christoph Waltz, Denzel Washington</td></tr><tr><td>Dark</td><td>Stranger Things, Stranger Things</td></tr><tr><td>Deutsche Bahn</td><td>Amtrack, Norfolk Southern Railway, Union Pacific Corporation</td></tr><tr><td>Deutsche Demokratische Republik</td><td>Confederate States of America, Confederate States of America, Texas, The Confederacy, The Confederate States of America</td></tr><tr><td>Deutsche Nationalhymne</td><td>Born in the U.S.A., Lazy Eye , Star Spangled Banner, The Star Spangled Banner</td></tr><tr><td>Deutschland</td><td>America, America, Continental United States, USA, United States, United States</td></tr><tr><td>Dieter Bohlen</td><td>Billy Joel, Blake Shelton, Daryl Hall, Paula Abdul, Ryan Seacrest</td></tr><tr><td>Dirk Nowitzki</td><td>LeBron James, Michael Jordan, Shaquille O'Neal</td></tr><tr><td>Doreen Dietel</td><td>Jessica Alba, Lisa Kudrow, Warrick Brown</td></tr><tr><td>Dreißigähriger Krieg</td><td>American Civil War, American Civil War, American Indian Wars, Civil war</td></tr><tr><td>Elisabeth von Österreich-Ungarn</td><td>Edith Roosevelt, Hillary Clinton, Jackie Kennedy</td></tr><tr><td>Elyas M'Barek</td><td>Adam Sandler, Adam Sandler, Chris Pine</td></tr><tr><td>Europawahl in Deutschland 2019</td><td>2018 United States elections, American presidential election 2020, Us election 2018</td></tr><tr><td>Europäisches Parliament</td><td>North Atlantic Council, Representative of the United States of America to the European Union, United Nations, United States Congress</td></tr><tr><td>Evelyn Burdecki</td><td>Hannah Brown, Kaitlyn Bristowe, Kim Kardashian, Kim Kardashian</td></tr><tr><td>FC Bayern München</td><td>Dallas Cowboys, Dc United, New York Yankees, New York Yankees, New York Yankees</td></tr><tr><td>Falco</td><td>David Bowie, Frederick William Schneider III, MC Hammer, Michael Jackson</td></tr><tr><td>Ferdinand Sauerbruch</td><td>Ben Carson, Ben Carson, Cornelius P. Rhoads, Jonas Salk, Virginia Apgar</td></tr><tr><td>Flughafen Berlin Brandenburg</td><td>Cincinnati Subway, DCA , John F. Kennedy International Airport, LaGuardia Airport</td></tr><tr><td>Frankfurt am Main</td><td>Chicago, Los Angeles, Los Angeles, New York City, Washington D.C.</td></tr><tr><td>Fritz Honka</td><td>Ted Bundy, Ted Bundy, Ted Bundy, Zodiac</td></tr><tr><td>Hamburg</td><td>Chicago, Chicago, Los Angeles, New York, Philadelphia</td></tr><tr><td>Hannalore Elsner</td><td>Elizabeth Taylor, Jane Lynch, Julia Roberts</td></tr><tr><td>Heidi Klum</td><td>Chrissy Teigen, Cindy Crawford, Gigi Hadid, Karlie Kloss, Tyra Banks</td></tr><tr><td>Heinz-Christian Strache</td><td>Anthony Weiner, Ben Carson, Donald J. Trump, Rob Ford, Roger Stone</td></tr><tr><td>Helene Fischer</td><td>Beyoncé, Kelly Clarkson, Taylor Swift, Taylor Swift</td></tr><tr><td>Hessen</td><td>Arizona, Illinois, Mid-Atlantic , Napa County California</td></tr><tr><td>Holocaust</td><td>Chattel Slavery, Japanese interned in American camps, Slavery in the United States</td></tr><tr><td>Ich bin ein Star – Holt mich hier raus!</td><td>Survivor, Survivor</td></tr><tr><td>Jürgen Kollopp</td><td>Bill Belichick, Bill Belichick, John Wooden</td></tr><tr><td>Kevin Kühnert</td><td>Bernie Sanders, Bernie Sanders, Bernie Sanders, Pete Buttigieg</td></tr><tr><td>Klaus Kinski</td><td>Christopher Lee, Clark Gable, John Wayne, Robert Pattinson, Robert Pattinson</td></tr><tr><td>Kontra K</td><td>50 Cent, Eminem, Eminem, Jesus Is King, Travis Scott</td></tr><tr><td>Köln</td><td>Boston, Chicago, Chicago, Houston</td></tr><tr><td>Leila Lowfire</td><td>Paris Hilton, Sasha Grey, Zendaya</td></tr><tr><td>Leipzig</td><td>Denver, Detroit, Miami, San Diego</td></tr><tr><td>Lena Meyer-Landrut</td><td>Ariana Grande, Kelly Clarkson, Kelly Clarkson, Meghan Trainor, Selena Gomez</td></tr><tr><td>Liechtenstein</td><td>Connecticut, Mexico, Philippines, Victoria British Columbia</td></tr><tr><td>Lisa Martinek</td><td>Julie Benz, Katherine Heigl, Mandy Moore, Meryl Streep</td></tr><tr><td>Ludwig van Beethoven</td><td>Aaron Copland, Aaron Copland, Aaron Copland, Aaron Copland, Elvis Presley, Frank Sinatra, George Gershwin, George Gershwin, Scott Joplin</td></tr><tr><td>Lufthansa</td><td>Delta, United, United Airlines, United Airlines</td></tr><tr><td>Luxemburg</td><td>Canada, Connecticut, Mexico, Victoria British Columbia</td></tr><tr><td>Mark Forster</td><td>Bruno Mars, Post Malone</td></tr><tr><td>Mero</td><td>DaBaby, Fetty Wap, Lil Nas X, Lil Nas X, Post Malone</td></tr><tr><td>Michael Schumacher</td><td>Dale Earnhardt, Dale Earnhardt, James Gordon, Jeff Gordon, Tiger Woods</td></tr><tr><td>München</td><td>Chicago, Los Angeles, New York City, New York City, Washington D.C.</td></tr><tr><td>Nico Santos</td><td>Harry Styles, Justin Bieber, Shawn Mendes</td></tr><tr><td>Niki Lauda</td><td>Dale Earnhardt, Dale Earnhardt Jr., Jeff Gordon, Jeff Gordon, Tiger Woods</td></tr><tr><td>Norddeutscher Rundfunk</td><td>NPR, NPR, National Public Radio, PBS, Sirius XM</td></tr><tr><td>Nordrhein-Westfalen</td><td>California, California</td></tr><tr><td>Philipp Amthor</td><td>Alexandria Ocasio-Cortez, Ben Shapiro</td></tr><tr><td>RAF Camora</td><td>Bad Bunny, Drake, Drake , Eminem, Future</td></tr><tr><td>Rammstein</td><td>Green Day, Metallica, Metallica, Metallica, Sum 41</td></tr><tr><td>Rhein</td><td>Mississippi, Mississippi River, Mississippi River</td></tr><tr><td>Robert Habeck</td><td>Al Gore, Bernie Sanders, Jill Stein, Ralph Nader</td></tr><tr><td>Rudi Assauer</td><td>Dave Roberts, Gregg Berhalter, Tom Flores, Vince Lombardi, Vince Lombardi</td></tr><tr><td>Sahra Wagenknecht</td><td>Alexandria Ocasio-Cortez, Elizabeth Warren, Elizabeth Warren, Elizabeth Warren, Nancy Pelosi</td></tr><tr><td>Sarah Connor</td><td>Beyoncé, Britney Spears, Mariah Carey</td></tr><tr><td>Schweiz</td><td>Canada, Canada, Iowa, Mexico, United States</td></tr><tr><td>Sebastian Kurz</td><td>Alexandria Ocasio-Cortez, Greg Abbott, Justin Trudeau, Justin Trudeau, Mitch McConnell</td></tr><tr><td>Serge Gnabry</td><td>Clint Dempsey, JuJu Smith-Schuster, Phillip Rivers, Stephen Curry, Zion Williamson</td></tr><tr><td>Sido</td><td>Eminem, Eminem, Macklemore</td></tr><tr><td>The Cratez</td><td>DJ Khaled, Drake , Twenty One Pilots</td></tr><tr><td>Thüringen</td><td>Iowa, Midwestern United States, Tennessee, Tennessee</td></tr><tr><td>Till Lindemann</td><td>James Hetfield, James Hetfield, James Hetfield, Ozzy Osborne</td></tr><tr><td>Tom Kaulitz</td><td>Adam Levine, Blink-182, Chris Martin, Green Day, Maroon 5</td></tr><tr><td>UEFA Champions League</td><td>Major League Soccer, NFC, NFL, National Football League, NCAA</td></tr><tr><td>Udo Jürgens</td><td>Aretha Franklin, Billy Joel, Elton John, Michael Jackson, Rolling Stone, Tom Lehrer</td></tr><tr><td>Udo Lindenberg</td><td>Johnny Cash, Mick Jagger, Roger Taylor , Travis Barker</td></tr><tr><td>Ursula von der Leyen</td><td>Condoleezza Rice, Hillary Clinton, Mike Pence, Sarah Palin, Susan Rice</td></tr><tr><td>Volkswagen AG</td><td>Ford Motor Company, Ford Motor Company, Ford Motor Company, Ford Motor Company, Ford Motor Company</td></tr><tr><td>Walter Lübecke</td><td>Harvey Milk, John F. Kennedy, John Roll, Steve Scalise</td></tr><tr><td>Weimarer Republik</td><td>America, Confederation Period, Congress of the Confederation, Counterculture of the 1960s, The Confederate States of America</td></tr><tr><td>Westdeutscher Rundfunk Köln</td><td>ABC News, NBC, NPR</td></tr><tr><td>Wien</td><td>Austin Texas, Richmond Virginia, Toronto, Washington D.C.</td></tr><tr><td>Wilhelm II.</td><td>William Howard Taft, Woodrow Wilson, Woodrow Wilson</td></tr><tr><td>Wolfgang Amadeus Mozart</td><td>Alan Menken, Elvis Presley, Leonard Bernstein</td></tr><tr><td>ZDF</td><td>NPR, NPR, National Public Radio, PBS, PBS</td></tr><tr><td>Österreich</td><td>Canada, Mexico, Texas, Texas, United States</td></tr><tr><td>Ötzi</td><td>Spirit Cave mummy, Spirit Cave mummy, Spirit Cave mummy, Sue</td></tr></table>
+
+Table 5: Top Wikipedia German→American adaptations.
+
+Entity  
+Human Adaptation: Wikipedia American  $\rightarrow$  German  
+
+<table><tr><td>13 Reasons Why</td><td>Club der roten Bändner, Gute Zeiten schlechte Zeiten, Lammbock, Türkisch für Anfänger</td></tr><tr><td>Albert Einstein</td><td>Albert Einstein, Albert Einstein, Albert Einstein, Max Planck, Max Planck</td></tr><tr><td>Alexander Hamilton</td><td>Konrad Adenauer, Max Weber, Otto von Bismarck, Otto von Bismarck</td></tr><tr><td>American Civil War</td><td>Deutscher Krieg, Dreizigjähriger Krieg, German Revolution of 1918–1919, German revolutions of 1848–1849</td></tr><tr><td>American Horror Story</td><td>Dark, Der goldene Handschuh, Good Bye Lenin!, Tintenherz</td></tr><tr><td>Angelina Jolie</td><td>Barbara Schöneberger, Franka Potente, Marlene Dietrich, Romy Schneider, Veronica Maria Cächilia Ferres</td></tr><tr><td>Apple Inc.</td><td>BMW, Fujitsu, SAP, Siemens</td></tr><tr><td>Ariana Grande</td><td>Lena Meyer-Landrut, Lena Meyer-Landrut, Lena Meyer-Landrut, Sarah Connor, Sarah Connor</td></tr><tr><td>Arnold Schwarzenegger</td><td>Arnold Schwarzenegger, Karl Lauterbach, Matthias Steiner, Peter Maffay, Ralf Rudolf Möller</td></tr><tr><td>Ashton Kutcher</td><td>Florian David Fitz, Matthias Schweighöfer, Til Schweiger, Til Schweiger</td></tr><tr><td>Australia</td><td>Australia, Russia, Schweiz, South Africa, Österreich</td></tr><tr><td>Avengers Infinity War</td><td>Das Arche Noah Prinzip, Fack ju Göhte, Fantastic Four, Who Am I</td></tr><tr><td>Barack Obama</td><td>Angela Merkel, Angela Merkel, Angela Merkel, Helmut Schmidt, Helmut Schmidt</td></tr><tr><td>Beyoncé</td><td>Helene Fischer, Sarah Connor, Veronica Ferres, Xavier Naidoo, Yvonne Catterfeld</td></tr><tr><td>Black Mirror</td><td>Dark, Dark, DieCOMMenden Tage, Krabat</td></tr><tr><td>Blake Lively</td><td>Josefine Preuß, Maria Furtwängler, Maria Furtwängler, Til Schweiger</td></tr><tr><td>Brad Pitt</td><td>Florian David Fitz, Frederick Lau, Til Schweiger, Til Schweiger, Til Schweiger</td></tr><tr><td>Bruce Lee</td><td>Götz Georg, Henry Maske, Julian Jacobi, Max Schmeling, no one is like Bruce Lee</td></tr><tr><td>Caitlyn Jenner</td><td>Kristin Otto, Magdalena Neuner, Magdalena Neuner, Niklas Kaul, Ulrike Meyfarth</td></tr><tr><td>California</td><td>Bavaria, Bavaria, Bayern, Bayern</td></tr><tr><td>Camila Cabello</td><td>Helene Fischer, Lena Meyer-Landrut, Lena Meyer-Landrut, Nadja Benaissa</td></tr><tr><td>Canada</td><td>Austria, Italy, Schweiz, Sweden, Österreich</td></tr><tr><td>Cardi B</td><td>Ace Tee, Pamela Reif, Sabrina Setlur, Sarah Connor, Schwester Ewa</td></tr><tr><td>Charles Manson</td><td>Andreas Baader, Issa Rammo, Papst benedikt xvi, Paul Schäfer</td></tr><tr><td>Charlize Theron</td><td>Baran bo Odar, Josefine Preuß, Josefine Preuß, Veronica Ferres, Veronica Maria Cächilia Ferres</td></tr><tr><td>Cher</td><td>Marlene Dietrich, Nena, Nena, Nena</td></tr><tr><td>Chris Pratt</td><td>Elyas M&#x27;Barek, Jan Josef Liefers, Matthias Schweighöfer, Ralf Moeller, Til Schweiger</td></tr><tr><td>Clint Eastwood</td><td>Heinz Erhardt, Klaus Kinski, Mario Adorf, Til Schweiger, Wim Wenders</td></tr><tr><td>Darth Vader</td><td>Adolf Hitler, Belzebub, Hagen von Tronje, Jens Maul</td></tr><tr><td>Donald Glover</td><td>Elyas M&#x27;Barek, Helge Schneider, Money Boy, Stefan Raab</td></tr></table>
+
+Drake
+
+Dwayne Johnson
+
+Elon Musk
+
+Eminem
+
+Facebook
+
+Friends
+
+Game of Thrones
+
+Google
+
+Harry Potter
+
+Heath Ledger
+
+It
+
+Jason Momoa
+
+Jeff Bezos
+
+Jeffrey Dahmer
+
+Jennifer Aniston
+
+Jennifer Lawrence
+
+Jennifer Lopez
+
+John Cena
+
+Johnny Cash
+
+Johnny Depp
+
+Julia Roberts
+
+Justin Bieber
+
+Keanu Reeves
+
+Kylie Jenner
+
+Lady Gaga
+
+LeBron James
+
+Leonardo DiCaprio
+
+Lisa Bonet
+
+Madonna
+
+Mark Wahlberg
+
+Martin Luther King Jr.
+
+Bushido, Cro, Falco, Fler
+
+Alexander Wolfe, Arnold Schwarzenegger, Peter Alexander, Tim Wiese, Tim Wiese
+
+Alexander Samwer, August Horch, Carl Benz, Herbert Diess, Werner von Siemens
+
+Bushido, Kollegah, Sido, Sido, Sido
+
+Das Erste, Lokalisten, Lokalisten, Schüler VZ, StudiVZ, StudiVZ
+
+Gute Zeiten schlechte Zeiten, Gpsz, Lindenstraße, Stromberg
+
+Babylon Berlin, Babylon Berlin, Babylon Berlin, Die unendliche Geschichte, Krakat
+
+Ecosia, Fastbot, SAP, SAP, i.d.k.
+
+Die Unendliche Geschichte, Die unendliche Geschichte, Harry Potter und ein Stein, Meggie Folchart
+
+Christoph Waltz, Florian David Fitz, Henry Blanke, Matthias Schweighöfer, Tilman Valentin Schweiger
+
+Dark, Der goldene Handschuh, Die Wolke, Pandora
+
+Arnold Schwarzenegger, Benno FÜRmann, Christoph Waltz, Elyas M'Barek, Elyas M'Barek, Elyas M'Barek
+
+Alexander Samwer, Beate Heister, Martin Winterkorn, Oliver Samwer
+
+Armin Meiwes, Fritz Haermann, Joachim Kroll, Karl Denke, Karl Denke
+
+Barbara Schöneberger, Diane Kruger, Diane Kruger, Franka Potente, Iris Berben
+
+Iris Berben, Josefine Preuß, Karoline Herfurth, Ruby O. Fee
+
+Heidi Klum, Helene Fischer, Jeanette Biedermann, Mandy Capristo, Sarah Connor
+
+Arnold Schwarzenegger, Max Schmeling, Max Schmeling, Ralf Möller
+
+Fantasticen vier, Helge Schneider, Peter Maffay, Peter Maffay, Christoph Maria Herbst, Christoph Waltz, Cro, Til Schweiger, Xavier Naidoo
+
+Karoline Herfurth, Maria Furtwängler, Marlene Dietrich, Marlene Dietrich
+
+Cro, Felix Jaehn, Lukas Rieger, McFittie, Mike Singer
+
+Daniel Bruhl, Mario Adorf, Til Schweiger, til schweiger
+
+Barbara Schöneberger, Heidi Klum, Karoline Einhoff, Sarah Connor, Stefanie Giesinger
+
+Helene Fischer, Nena, Nena, Nina Hagen, Sarah Lombardi
+
+Dirk Nowitzki, Dirk Nowitzki, Dirk Nowitzki, Dirk Nowitzki, Toni Kroos
+
+Matthias Schweighöfer, Moritz Bleibtreu, Til Schweiger, Til Schweiger, Til Schweiger
+
+Franka Potente, Iris Berben, Karoline Herfurth, Maria Furtwangler
+
+Blümchen, Helene Fischer, Helene Fischer, Helene Fischer, Sarah Connor
+
+Florian David Fitz, Til Schweiger, Tilman Valentin Schweiger, Alexei Alexejewitsch
+
+Hans Scholl, Hans Scholl, Helmut Palmer, Robert Blum, Sophie Scholl
+
+<table><tr><td>Marvel Cinematic Universe</td><td>Bavaria Film, Havelstudios, Phantásien, Rat Pack Filmproduktion, Tatort</td></tr><tr><td>Michael Jackson</td><td>Herbert Grönemeyer, Nena, Udo Jürgens, Xavier Naidoo, Xavier Naidoo</td></tr><tr><td>Mila Kunis</td><td>Josefine Preuß, Matthias Schweighöfer, Vanessa Mai</td></tr><tr><td>Miley Cyrus</td><td>Lena Meyer-Landrut, Lukas Rieger, Nena, Sarah Connor, Yvonne Catterfeld</td></tr><tr><td>Muhammad Ali</td><td>Alexander Abraham, Boris Becker, Max Schmeling, Max Schmel-ing, Sven Ottke</td></tr><tr><td>Natalie Portman</td><td>Barbara Schöneberger, Diane Kruger, Franka Potente, Iris Berben</td></tr><tr><td>New York City</td><td>Berlin, Berlin, Berlin, Berlin, Frankfurt</td></tr><tr><td>Nicole Kidman</td><td>Evelyn Hamann, Franka Potente, Senta Berger, iris berben</td></tr><tr><td>Peaky Blinders</td><td>Dark, Dieter Schwarz, Im Westen Nichts Neues, Tatort, Tatort</td></tr><tr><td>Philippines</td><td>Greece, Griechenland, Mallorca, Mallorca</td></tr><tr><td>Post Malone</td><td>Bushido, Bushido, Cro, Cro, Kollegah</td></tr><tr><td>Rianna</td><td>Helene Fischer, Lena Meyer-Landrut, Lena Meyer-Landrut, Nena</td></tr><tr><td>Riverdale</td><td>Babylon Berlin, Berlin Tag und Nacht, Neues vom Südhof, Türkisch für Anfänger</td></tr><tr><td>Robert Downey Jr.</td><td>Christoph Waltz, Günter Strack, Martin Semmelrogge, Moritz Bleibtreu, Til Schweiger</td></tr><tr><td>Robin Williams</td><td>Hape Kerkeling, Heinz Erhardt, Peter Maffay, Silvia Seidel, Tim Bendzko</td></tr><tr><td>Ronald Reagan</td><td>Helmut Schmidt, Konrad Adenauer, Konrad Adenauer, Konrad Adenauer</td></tr><tr><td>Ryan Reynolds</td><td>Daniel Brühl, Florian David Fitz, Matthias Schweighöfer, Til Schweiger, Til Schweiger</td></tr><tr><td>Scarlett Johansson</td><td>Lena Gercke, Romy Schneider, Sarah Connor, Sarah Connor, Veronica Ferres</td></tr><tr><td>Selena Gomez</td><td>Lena Meyer-Landrut, Lena Meyer-Landrut, Nena, Nora Tschirner</td></tr><tr><td>September 11 attacks</td><td>Anschlag im OEZ, Dresden Bombing, Mauerfall, RAF-Attentate, Terroranschlag Olympia 1972</td></tr><tr><td>Shaquille O’Neal</td><td>Dirk Nowitzki, Dirk Nowitzki, Mehmet Scholl, Niklas Sülle</td></tr><tr><td>Star Wars</td><td>Dark, Metropolis, Traumschiff Surprise – Periode 1, Who Am I?, i.d.k</td></tr><tr><td>Stephen Curry</td><td>Dirk Nowitzki, Dirk Nowitzki, Dirk Nowitzki, Manuel Neuer</td></tr><tr><td>Stranger Things</td><td>8 Tage, Babylon Berlin, Dark, Tatort, Tatort</td></tr><tr><td>Sylvester Stallone</td><td>Henry Blanke, Jan Josef Liefers, Michael Bully Herbig, Michael Fassbender, Til Schweiger</td></tr><tr><td>Taylor Swift</td><td>Lena Meyer-Landrut, Lena Meyer-Landrut, Sarah Connor, Sarah Connor, Yvonne Catterfeld</td></tr><tr><td>Ted Bundy</td><td>Joachim Kroll, Josef Fritzl, Niels Högel, Rudolf Pleil, Rudolf Pleil</td></tr><tr><td>The Big Bang Theory</td><td>Doctor's Diary, Stromberg, Stromberg, der Tatortreiniger</td></tr><tr><td>The Crown</td><td>Babylon Berlin, Deutschland 83, Die Deutschen, Karl der Groß</td></tr><tr><td>The Handmaid’s Tale</td><td>Dark, Dark, Der Pass, Die Wanderhure, Er ist wieder da</td></tr><tr><td>The Walking Dead</td><td>Dark, Dark, Der goldene Handschuh, Zombies From Outer Space</td></tr><tr><td>Tom Brady</td><td>Franz Beckenbauer, Michael Ballack, Oliver Kahn, Thomas Müller, Uli Stein</td></tr><tr><td>Tom Cruise</td><td>Benno Fürmann, Benno Fürmann, Christoph Waltz, Elyas M’Barek, Matthias Schweighöfer</td></tr><tr><td>Tom Hanks</td><td>Christoph Waltz, Christoph Waltz, Daniel Brühl, Til Schweiger</td></tr><tr><td>Tom Hardy</td><td>Bruno Ganz, Michael Herbig, Til Schweiger, Wotan Wilke Möhring</td></tr><tr><td>Tom Holland</td><td>Daniel Brühl, Frederick Lau, Matthias Schweighöfer, Matthias Schweighöfer, Til Schweiger</td></tr><tr><td>Tupac Shakur</td><td>Farid Bang, Haftbefehl, Kollegah, Kristoffer Klauß, Peter Fox</td></tr><tr><td>United States</td><td>BRD, Bundesrepublik Deutschland, Deutschland, Germany, Germany</td></tr><tr><td>Vietnam War</td><td>Berlin Wall, First world war, Kosovokrieg, World War II</td></tr><tr><td>Wikipedia</td><td>Brockhaus, Brockhaus Enzyklopädie, Brockhaus Enzyklopädie, Duden, dict.cc</td></tr><tr><td>Will Smith</td><td>Daniel Brühl, Elyas M'Barek, Sascha Reimann, Sido, Til Schweiger</td></tr><tr><td>X-Men</td><td>Abwärts, Fantastic Four, Freaks, Krabat, Who Am I</td></tr><tr><td>YouTube</td><td>Lokalisten, MyVideo, MyVideo, ProSieben, lokalisten</td></tr><tr><td>Zac Efron</td><td>Frederick Lau, Lukas Rieger, Peter Kraus, Walter Sedlmayr</td></tr><tr><td>Zendaya</td><td>Franka Potente, Iris Berben, Lena Meyer-Landrut, Lena Meyer-Landrut, Yvonne Catterfeld</td></tr></table>
+
+Table 6: Top Wikipedia American  $\rightarrow$  German adaptations.
+
+# Entity
+
+# Top Five WikiData Adaptations
+
+<table><tr><td>Abraham Lincoln</td><td>Victor Adler, Johann Joachim Christoph Bode, Willem Barentsz, Hermann Wagener, Robert von Mohl</td></tr><tr><td>Al Capone</td><td>Hans H. Zerlett, Fritz Thyssen, Adam Rainer, Franz Winkelmeier, Christian Louis, Duke of Brunswick-Lüneburg</td></tr><tr><td>Alfred Hitchcock</td><td>Edgar Reitz, Jan Josef Liefers, Mario Adorf, Max Frisch, Armin Mueller-Stahl</td></tr><tr><td>Benedict Arnold</td><td>Hans-Georg Hess, Isabelle Eberhardt, Günther Heydemann, Max Schreck, Louis Blenker</td></tr><tr><td>Bill Gates</td><td>Ferdinand von Zeppelin, Günther Jauch, Nikolaus Harmoncourt, Sepp Blatter, Alfred Grosser</td></tr><tr><td>Britney Spears</td><td>Herta Müller, Günter Grass, Joachim Gauck, Hans-Dietrich Genscher, Koča Popović</td></tr><tr><td>Donald Trump</td><td>Max Frisch, Thomas Gottschalk, Jan Josef Liefers, Rainer Werner Fassbinder, Christa Wolf</td></tr><tr><td>Elvis Presley</td><td>Reinhard Lakomy, James Last, Herbert Achternbusch, Fritz Hauser, Hans-Peter Pfammmatter</td></tr><tr><td>Ernest Hemingway</td><td>Karlheinz Böhm, Ricardo Huch, Michael Ballhaus, Arnold Zweig, Michael Fassbender</td></tr><tr><td>Frank Lloyd Wright</td><td>Ferdinand Hodler, Johan Zoffany, Hans Thoma, Arne Jacobsen, Lucas Cranach the Younger</td></tr><tr><td>George Washington</td><td>Friedrich Wilhelm von Seydlitz, Dagobert Sigmund von Wurmser, Heinz Guderian, Ernst Gideon von Laudon, George Olivier, count of Wallis</td></tr><tr><td>Henry Ford</td><td>Heinz Sielmann, Wieland Schmied, Manfred Krug, Paul Maar, Armin Mueller-Stahl</td></tr><tr><td>Hillary Clinton</td><td>Pope Benedict XVI, Willy Brandt, Angela Merkel, Helmut Schmidt, Kurt Biedenkopf</td></tr><tr><td>Homer Simpson</td><td>Elizabeth Lavenza, Hans Fugger, Baron Strucker, Herbert of Wetterau, Prince Johannes of Liechtenstein</td></tr><tr><td>Jimi Hendrix</td><td>Marius Müller-Westernhagen, Karl Richter, Reinhard Lakomy, Michael Cretu, Paul van Dyk</td></tr><tr><td>Kim Kardashian</td><td>Erika Mann, Frank Wedekind, Til Schweiger, Fritz von Opel, Carmen Electra</td></tr><tr><td>Marilyn Monroe</td><td>Gerhart M. Riegner, Viktor de Kowa, Otto Sander, Hans Hass, Dorothee Sölle</td></tr><tr><td>Michael Jordan</td><td>Jean-Claude Juncker, Richard von Weizsäcker, Herta Müller, Konrad Adenauer, Helmut Kohl</td></tr><tr><td>Louis Armstrong</td><td>Herbert Prikopa, Till Lindemann, Nico, Klaus Voormann, Jakob Adlung</td></tr><tr><td>Neil Armstrong</td><td>Stefan Hell, Franz-Ulrich Hartl, Reinhard Genzel, Charles Weiss-mann, Harald zur Hausen</td></tr><tr><td>Noam Chomsky</td><td>Günter Grass, Herta Müller, Heinrich Böll, Peter Handke, Juli Zeh</td></tr><tr><td>Oprah Winfrey</td><td>Günter Grass, Peter Scholl-Latour, Elfriede Jelinek, Juli Zeh, Christa Wolf</td></tr><tr><td>Orville Wright</td><td>Frank Thiess, Jessica Hausner, Elmar Wepper, Wolf Jobst Siedler, Marc Rothemund</td></tr><tr><td>Richard Nixon</td><td>Heinrich von Brentano, Ernst Benda, Gustav Heinemann, Heiner Geißler, Heinrich Albertz</td></tr><tr><td>Superman</td><td>Magneto, Nightcrawler, Sinterklaas, Silent Night, Victor Frankenstein</td></tr><tr><td>Steve Jobs</td><td>Victor Klemperer, Joschka Fischer, Jürgen Kuczynski, Joachim Fest, Dieter Hallervorden</td></tr><tr><td>Steven Spielberg</td><td>Herta Müller, Jean-Claude Juncker, Hans-Dietrich Genscher, Joachim Gauck, Koča Popović</td></tr><tr><td>Tiger Woods</td><td>Charles Dutoit, Shania Twain, Lise Meitner, Michael Haneke, Otto Hahn</td></tr><tr><td>Walt Disney</td><td>Shania Twain, Charles Dutoit, Lise Meitner, Otto Hahn, Michael Haneke</td></tr><tr><td>John F. Kennedy</td><td>Bernhard von Bülow, Otto von Habsburg, Hans-Jochen Vogel, Prince Henry of Prussia, Frederick Augustus III of Saxony</td></tr><tr><td>Charles Lindbergh</td><td>Pina Bausch, Ferdinand von Zeppelin, Nikolaus Harmoncourt, Jan Josef Liefers, Wolf Biermann</td></tr><tr><td>Rosa Parks</td><td>Hermann Lenz, Wilhelm Feldberg, Horst Tappert, Peter Stein, Gert Jonke</td></tr><tr><td>Serena Williams</td><td>Charles Dutoit, Lise Meitner, Michael Haneke, Richard von Coudenhove-Kalergi, Klaus Clusius</td></tr></table>
+
+Table 7: We show top-5 predictions out of the top-100 for American  $\rightarrow$  German adaptations on the Veale NOC subset using WikiData. These are compared to our human annotations in our results.
+
+<table><tr><td>Entity</td><td>Top Five 3CosAdd Adaptations: American→German adapta-tions on the Veale NOC</td></tr><tr><td>Abraham Lincoln</td><td>Napoleon, Napoléon Bonaparte, Erzherzog Johann, Otto von Bismarck, Kaiser Wilhelm II.</td></tr><tr><td>Al Capone</td><td>Nazis, SA-Mann, Verhaftungswellen, Judenverfolgung, Fluchthilfe</td></tr><tr><td>Alfred Hitchcock</td><td>Fritz Lang, Helmut Käutner, Willi Forst, Emil Jannings, Heinz Ruhmann</td></tr><tr><td>Benedict Arnold</td><td>Russlandfeldzug 1812, Schlacht bei Roßbach, Jean-Victor Moreau, schwedischen Armee, Alexander Wassiljewitsch Suworow</td></tr><tr><td>Bill Gates</td><td>congenstar, Alnatura, GMX, ChessBase, Gardeur</td></tr><tr><td>Britney Spears</td><td>Glasperlenspiel, Unheilig, Helene Fischer, Christina Aguilera, Herbert Grönenemeyer</td></tr><tr><td>Charles Lindbergh</td><td>Segelflieger, Flugpioniere, Zeppelinins, Adolf Hitler, Caproni</td></tr><tr><td>Donald Trump</td><td>Deutschland, Österreich, Trump, Strache, Bundestagswahlkampf</td></tr><tr><td>Elvis Presley</td><td>Udo Jürgens, Elvis Presley, Hits, den Beatles, der Beatles</td></tr><tr><td>Ernest Hemingway</td><td>Stefan Zweig, Franz Werfel, Joachim Ringelnatz, Hermann Hesse, Gottfried Benn</td></tr><tr><td>Frank Lloyd Wright</td><td>Adolf Loos, Le Corbusier, Bruno Schmitz, Entwurfen, Fritz Höger</td></tr><tr><td>George Washington</td><td>Napoléon Bonaparte, Friedrich dem Großen, Napoleon, Friedrich der Große, Napoleon Bonaparte</td></tr><tr><td>Henry Ford</td><td>Ferdinand Porsche, Büssing, Krupp, Ettore Bugatti, Steyr-Daimler-Puch</td></tr><tr><td>Hillary Clinton</td><td>Deutschland, Bundestagswahlkampf, Österreich, Sarkozy, Strache</td></tr><tr><td>Homer Simpson</td><td>Eingangsszene, verulkt, Schlusssequenz, Off-Stimme, Muminfam-ilie</td></tr><tr><td>Jack The Ripper:Ripper</td><td>Tat, Werwolf, Täter, Dritten Reich, Mörder</td></tr><tr><td>Jay Z</td><td>Xavier Naidoo, D-Bo, Sido, Rosenstolz, David Guetta</td></tr><tr><td>Jimi Hendrix</td><td>Udo Jürgens, Tangerine Dream, Jimi Hendrix, Pink Floyd, De-peche Mode</td></tr><tr><td>John F. Kennedy</td><td>Adolf Hitler, Bundeskanzlers, Adolf Hitlers, Adolf Hitler, Hitler</td></tr><tr><td>Kim Kardashian</td><td>Kaas, gotv, Frank Zander, Herbert Grünemeyer, Roland Kaiser</td></tr><tr><td>Louis Armstrong</td><td>Richard Tauber, Django Reinhardt, Udo Jürgens, Sidney Bechet, Jazzorchester</td></tr><tr><td>Marilyn Monroe</td><td>Marlene Dietrich, Lil Dagover, Elisabeth Bergner, Brigitte Bardot, Romy Schneider</td></tr><tr><td>Michael Jordan</td><td>Powerplay, Xavi, Predrag Mijatović, NHL-Historie, Franck Ribéry</td></tr><tr><td>Neil Armstrong</td><td>Juri Gagarin, Vorbeiflag, Weltraum, Raumstation Mir, Raumfahrer</td></tr><tr><td>Noam Chomsky</td><td>Jürgen Habermas, Hans-Ulrich Wehler, Carl Schmitt, Theodor W. Adorno, Norbert Elias</td></tr><tr><td>Oprah Winfrey</td><td>Harald Schmidt, Thomas Gottschalk, Satiresendung, ORF-Sendung, Hape Kerkeling</td></tr><tr><td>Orville Wright</td><td>Parseval, Luft Hansa, Hugo Junkers, Ernst Heinkel, Claude Dornier</td></tr><tr><td>Richard Nixon</td><td>Österreich, Deutschland, Bundeskanzler, Bundeskanzlers, Bundes-spráidenten</td></tr><tr><td>Rosa Parks</td><td>NS-Militärjustiz, Franz Jägerstätter, NS-Opfer, Bücherverbn-nung, Baum-Gruppe</td></tr><tr><td>Serena Williams</td><td>Dick Jaspers, Philipp Kohlschreiber, Semifinale, Achtelfinale, Do-minic Thiem</td></tr><tr><td>Steve Jobs</td><td>Steve Jobs, Sony, Electronic Arts, Netscape, Atari</td></tr><tr><td>Steven Spielberg</td><td>Hörspielproduktion, Helmut Käutner, Fellini, Oliver Hirschbiegel, Kinofilm</td></tr><tr><td>Superman</td><td>Superman, Batman, Superhelden, Monster, Spider-Man</td></tr><tr><td>Tiger Woods</td><td>Rekordeuropameister, Österreich, spanische Team, ÖFB-Cupsieger, Deutschland</td></tr><tr><td>Walt Disney</td><td>Fritz Lang, Sascha-Film, Fellini, UFA, "Das Cabinet des Dr. Cali-gari"</td></tr></table>
+
+Table 8: We show top-5 predictions out of the top-100 for American  $\rightarrow$  German adaptations on the Veale NOC subset using 3CosAdd. These are compared to our human annotations in our results.
+
+<table><tr><td>Entity</td><td>Top Five Learned Adaptations: American→German adapta-tions on the Veale NOC</td></tr><tr><td>Abraham Lincoln</td><td>Konrad Adenauer, Helmut Schmidt, Willy Brandt, Helmut Kohl, Adenauer</td></tr><tr><td>Al Capone</td><td>Andreas Baader, Leo Katzenberger, Paul Schäfer, Strippel, Hermann Langbein</td></tr><tr><td>Alfred Hitchcock</td><td>Helmut Käutner, Til Schweiger, Mario Adorf, Paul Verhoeven, Dennis Hopper</td></tr><tr><td>Benedict Arnold</td><td rowspan="2">Otto von Bismarck, Bismarcks, Bismarck, Preußens, Kaiserreiches Martin Winterkorn, Volkswagen AG, DaimlerChrysler, Robert Bosch GmbH, Volkswagen AG</td></tr><tr><td>Bill Gates</td></tr><tr><td>Britney Spears</td><td>Sarah Connor, Nena, Helene Fischer, Lena Meyer-Landrut, Moses Pelham</td></tr><tr><td>Charles Lindbergh</td><td>Chaim Weizmann, Tomás Garrigue Masaryk, Ferdinand Sauer-bruch, Fritz Haber, Chaim Arlosoroff</td></tr><tr><td>Donald Trump</td><td>Helmut Schmidt, Angela Merkel, Gerhard Schröder, Helmut Kohl, Bundesaußenminister</td></tr><tr><td>Elvis Presley</td><td>Udo Jürgens, Peter Maffay, Cliff Richard, Achim Reichel, Lou Reed</td></tr><tr><td>Ernest Hemingway</td><td>Paul Schlenther, Marcel Reich-Ranicki., Timothy Leary, Erwin Leiser, Alice Walker</td></tr><tr><td>Frank Lloyd Wright</td><td>Albert Einstein, Max Planck, Max Born, Hermann von Helmholtz, Arnold Sommerfeld</td></tr><tr><td>George Washington</td><td>Otto von Bismarck, Otto von Bismarck, Konrad Adenauer, Engelbert Dollfuß, Joseph Wirth</td></tr><tr><td>Henry Ford</td><td>Ernst Abbe, Carl Duisberg, Bubbe, Aby Warburg, Sybel</td></tr><tr><td>Hillary Clinton</td><td>Angela Merkel, Angela Merkel, Helmut Schmidt, Gerhard Schröder, Bundesinnenminister</td></tr><tr><td>Homer Simpson</td><td>Rolf Hochhuth, Carl Bernstein, Uwe Tellkamp, Wolfgang Völz, Richard Gere</td></tr><tr><td>Jack The Ripper:Ripper</td><td>Sarah Connor, Spike Jonze, Timberlake, "Das Urteil", "Nichts als die Wahrheit"</td></tr><tr><td>Jay Z</td><td>will.i.am, Moses Pelham, Silbermond, Xavier Naidoo, Kanye West</td></tr><tr><td>Jimi Hendrix</td><td>Peter Maffay, Udo Lindenberg, Depeche Mode, Xavier Naidoo, Die Toten Hosen</td></tr><tr><td>John F. Kennedy</td><td>Konrad Adenauer, Helmut Schmidt, Willy Brandt, Helmut Kohl, Bundeskanzler</td></tr><tr><td>Kim Kardashianian</td><td>Heidi Klum, Ruth Moschner, Ellen DeGeneres, Circus HalliGalli, Oliver Pocher</td></tr><tr><td>Louis Armstrong</td><td>Peter Maffay, Radioaufnahmen, Udo Lindenberg, Achim Reichel, Helge Schneider</td></tr><tr><td>Marilyn Monroe</td><td>Walter Giller, Jessica Tandy, Liv Ullmann, Edgar Selge, Betty White</td></tr><tr><td>Michael Jordan</td><td>Dirk Nowitzki, Toni Kroos, Zlatan Ibrahimović, Xavi, Zinédine Zidane</td></tr><tr><td>Neil Armstrong</td><td>Max von Laue, Albert Einstein, Chaim Weizmann, Johannes R. Becher, Ernst Abbe</td></tr><tr><td>Noam Chomsky</td><td>Albert Einstein, Nobelpreisträger, Max Planck, American Psychological Association, Hans Bethe</td></tr><tr><td>Oprah Winfrey</td><td>Anja Kling, "Forsthaus Falkenau", Uschi Glas, "Saturday Night Live", Anke Engelke</td></tr><tr><td>Orville Wright</td><td>Kawaishi, Rjabuschinski, Monistenbund, Dethmann, Leo Baeck 
+Instituts</td></tr><tr><td>Richard Nixon</td><td>Helmut Schmidt, Konrad Adenauer, Willy Brandt, Helmut Kohl, 
+Gerhard Schröder</td></tr><tr><td>Rosa Parks</td><td>Sophie Scholl, Die letzten Tage, Emil Jannings., Ruth Wilson, 
+Monica Bleibtreu</td></tr><tr><td>Serena Williams</td><td>Max Schmeling, Wilfried Dietrich, Gottfried von Cramm, Henry 
+Maske, László Kubala</td></tr><tr><td>Steve Jobs</td><td>DaimlerChrysler, Volkswagen, Siemens, Sanyo, Fujitsu</td></tr><tr><td>Steven Spielberg</td><td>Til Schweiger, Ethan Hawke, Matthias Schweighöfer, Samuel L. 
+Jackson, Ryan Reynolds</td></tr><tr><td>Superman</td><td>Jabberwocky, Freaks, Scarface, Leatherface, Krabat</td></tr><tr><td>Tiger Woods</td><td>Dirk Nowitzki, deutschen U21-Nationalmannschaft, MTV Gießen, 
+Mats Hummels, Franz Beckenbauer</td></tr><tr><td>Walt Disney</td><td>Helmut Dietl, Peter Ustinov, David Mamet, Rainer Werner Fass- 
+binder, Sonke Wortmann</td></tr></table>
+
+Table 9: We show top-5 predictions out of the top-100 for American  $\rightarrow$  German adaptations on the Veale NOC subset with our Learned Adaptation approach. These are compared to our human annotations in our results.

adaptingentitiesacrosslanguagesandcultures/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f60b97a4a1dbad2f54fb6ea9259a298e8f080b718786f6c2aa4f5920747e6c91
+size 4185207

adaptingentitiesacrosslanguagesandcultures/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be3d86f8d80f14cdfe3473f47505370dd0617b9627c62bccff24e10eb6d7626b
+size 482143

adaptinglanguagemodelsforzeroshotlearningbymetatuningondatasetandpromptcollections/d68800a5-6161-4aaa-bd0d-24773b0dfaa3_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7381e2ce02c56122efacd1ffd839cbaba61bb7fdfb92502e66e1f114b46a32c2
+size 156858

adaptinglanguagemodelsforzeroshotlearningbymetatuningondatasetandpromptcollections/d68800a5-6161-4aaa-bd0d-24773b0dfaa3_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46826453d1a6e3b8d67785fe64b168b3b6c69c9f88529d2c567fb10291aafc40
+size 196738

adaptinglanguagemodelsforzeroshotlearningbymetatuningondatasetandpromptcollections/d68800a5-6161-4aaa-bd0d-24773b0dfaa3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04dcb0dd62a9e0a84cfd101c116334689dc5ae9702f4b81610925de70973f7da
+size 1160400

adaptinglanguagemodelsforzeroshotlearningbymetatuningondatasetandpromptcollections/full.md

@@ -0,0 +1,699 @@
+# Adapting Language Models for Zero-shot Learning by Meta-tuning on Dataset and Prompt Collections
+
+Ruiqi Zhong Kristy Lee* Zheng Zhang* Dan Klein Computer Science Division, University of California, Berkeley {ruiqi-zhong, kristylee, zhengzhang1216, Klein} @berkeley.edu
+
+# Abstract
+
+Large pre-trained language models (LMs) such as GPT-3 have acquired a surprising ability to perform zero-shot learning. For example, to classify sentiment without any training examples, we can "prompt" the LM with the review and the label description "Does the user like this movie?", and ask whether the next word is "Yes" or "No". However, the next word prediction training objective is still misaligned with the target zero-shot learning objective. To address this weakness, we propose meta-tuning, which directly optimizes the zero-shot learning objective by fine-tuning pre-trained language models on a collection of datasets. We focus on classification tasks, and construct the meta-dataset by aggregating 43 existing datasets and annotating 441 label descriptions in a question-answering (QA) format. When evaluated on unseen tasks, meta-tuned models outperform a sames-sized QA model and the previous SOTA zero-shot learning system based on natural language inference. Additionally, increasing parameter count from 220M to 770M improves AUC-ROC scores by  $6.3\%$ , and we forecast that even larger models would perform better. Therefore, measuring zero-shot learning performance on language models out-of-the-box might underestimate their true potential, and community-wide efforts on aggregating datasets and unifying their formats can help build models that answer prompts better.
+
+# 1 Introduction
+
+The goal of zero-shot classification (ZSC) is to classify textual inputs using label descriptions without any examples (Yin et al., 2019). Large language models - whose only training objective is to predict the next word given the context - have acquired a surprising ability to perform ZSC (Radford et al., 2019; Brown et al., 2020; Le Scao and Rush, 2021). For example, to classify whether the sentence "This movie is amazing!" is positive, we
+
+can prompt the language model with the context "Review: This movie is amazing! Positive Review? _", and check whether the next word is more likely to be "Yes" or "No" (Zhao et al., 2021). To convert ZSC into a language modeling (LM) task that an LM model is likely to perform well, many recent works focus on finding better prompts (Shin et al., 2020; Schick and Schütze, 2020a,b; Gao et al., 2021).
+
+However, the LM training objective is correlated but still misaligned with the target objective to answer prompts. Our work addresses this weakness by directly optimizing the zero-shot classification objective through fine-tuning (Section 4). This requires us to 1) unify different classification tasks into the same format, and 2) gather a collection of classification datasets and label descriptions (prompts) for training (Section 2). Since we fine-tune our model on a meta-dataset, we name our approach meta-tuning.
+
+We focus on binary classification tasks and unify them into a "Yes"/"No" QA format (Clark et al., 2019; McCann et al., 2018), where the input is provided as the context and the label information is provided in the question (Figure 1 (a)). Using this format, we gathered a diverse set of classification datasets from 43 different sources listed on Kaggle, SemEval, HuggingFace, and other papers. These tasks range from hate speech detection, question categorization, sentiment classification to stance classification, etc, and the genre ranges from textbooks, social media, to academic papers, etc. In total, these datasets contain 204 unique labels, and we manually annotated 441 label descriptions (Figure 2).
+
+To evaluate ZSC, we need to define what counts as a task that the model has not seen during training time. While prior work considers different notions of "unseen" by disallowing the same label or the same dataset to appear during training, our work defines "unseen" more harshly by dis
+
+![](images/c222e854fb40f521572f839480789306f70a351cf9640136274a953a22e013dd.jpg)  
+(a) Task Conversion
+
+![](images/825953518b3cf2888346ef317f1dd2c7aafeb45bc3dc4c9170279414eb10e836.jpg)  
+(b) Meta-tuning and Evaluation  
+(c) Results  
+Figure 1: (a) We convert the format to question answering. We manually annotate label descriptions (questions) ourselves (Section 2). (b) We finetune the UnifiedQA (Khashabi et al., 2020) model (with  $770\mathrm{M}$  parameters) on a diverse set of tasks (Section 4), and evaluate its 0-shot classification (ZSC) performance on an unseen task. (c) For each label description (question) we evaluate the AUC-ROC score for the "Yes" answer, and each dot represents a label description (Section 3). The  $x$ -value is the ZSC performance of UnifiedQA; the  $y$ -value is the performance after meta-tuning. In most cases, the  $y$ -value improves over the  $x$ -value (above the red line) and is better than random guesses (above the black line) by a robust margin (Section 5).
+
+allowing similar datasets. For example, we consider AG News topic classification dataset (Zhang et al., 2015) and the topic classification dataset from Yin et al. (2019) to be similar, even though their sources and label spaces are different.
+
+Meta-tuning improves ZSC over UnifiedQA for most labels (Figure 1 (c)). Moreover, larger models are better, and hence we forecast that metatuning would work for even larger models. We also find that the performance can be slightly improved by training on datasets similar to the test dataset, ensembling different label descriptions, or initializing with a QA model (Section 5.1). All of our findings reliably hold under different robustness checks (Section 5.2), and our approach outperforms the previous SOTA Yin et al. (2019) using the same pre-training method (Section 5.3).
+
+Our results suggest two promising future directions (Section 6). First, large language models' (e.g. GPT-3) potential for zero-shot learning, as currently measured by context-prompting, might have been broadly underestimated; metatuning might significantly improve their performance. Second, community-wide efforts on aggregating and unifying datasets can scale up training and evaluation for zero-shot learning models. On the flip side, however, the meta-tuning approach might incentivize providers of LM inference APIs to collect prompts from users, hence potentially leading to security, privacy, and fairness concerns at a greater scale (Section A).
+
+Contributions To summarize, we 1) curate a dataset of classification datasets with expert an
+
+notated label descriptions. 2) demonstrate a simple approach to train models to perform zero-shot learning, and 3) identify several factors that improve performance; in particular, larger pretrained models are better.
+
+# 2 Data
+
+We gather a wide range of classification datasets and unify them into the "Yes"/"No" question answering format for binary classification. Then we group similar datasets together to determine what counts as unseen tasks during evaluation.
+
+Gathering classification datasets We collect classification datasets from Kaggle $^2$ , Huggingface (Wolf et al., 2020), SemEval $^3$ , and other papers. We looked through these sources and only considered English classification datasets. We also skipped the tasks that we felt were already better represented by other datasets in our collection. Then we manually examined a few examples in each remaining dataset to make sure it seemed plausibly clean.
+
+The goals of these classification datasets include, but are not limited to sentiment classification (IMDB Reviews, Maas et al. (2011a)), topic classification (AG News, Zhang et al. (2015)), grammaticality judgement (CoLA, Warstadt et al. (2018)), paraphrase detection  $(\mathrm{QQP}^4)$ , definition
+
+![](images/d02a5751140968370359418e283afbf0f5af08a6c3f021959b939ae568c2c92f.jpg)  
+Figure 2: For each dataset, we annotate 1-3 descriptions for each label in the form of questions, and associate it with a set of property tags. The question answering format can be seen in Figure 1 (a).  
+Figure 3: Some example manually annotated label descriptions (questions). Three of the authors manually wrote 441 questions in total, and each of them is proofread by at least another author.
+
+detection (SemEval 2020 Task 6, Spala et al. (2019)), stance classification (SemEval 2016 Task 6, Mohammad et al. (2016)), etc. The genre includes academic papers, reviews, tweets, posts, messages, articles, and textbooks. The comprehensive list of datasets is in Appendix B. Overall, we aim for a high diversity of tasks and genres by building upon what the broader research community has studied. Our approach is complementary to that of Weller et al. (2020), which asks turkers to generate tasks, and that of Mishra et al. (2021), which generates tasks by decomposing existing templates used to construct reading comprehension datasets. The concurrent work of Bragg et al. (2021) unifies the evaluation for few-shot learning; their zero-shot evaluation setup is the closest to ours, and they used templates and verbalizers (Schick and Schütze, 2020a) to specify the semantics of a task.
+
+Some of our datasets are noisy and not peer reviewed, or contain tasks that are too complicated (e.g. Multi-NLI, Williams et al. (2018)) for ZSC. To make our evaluation more informative, we only include them for training but not testing. We make these decisions before running our experiments in Section 5 to prevent selection bias.
+
+Unifying the dataset format We convert each classification dataset into a "Yes"/"No" question answering format and provide label information in the question. For each label, we annotate 1-3 questions. If the label is null (for example, a text that does not express a particular emotion in an emotion classification dataset), we skip this label. Three of the authors<sup>5</sup> manually annotated 441 questions for 204 unique labels, and each question
+
+Are these two questions asking for the same thing? Does the tweet contain irony?
+
+Is this news about world events?
+
+Does the text contain a definition?
+
+Is the tweet an offensive tweet?
+
+Is the text objective?
+
+Does the question ask for a numerical answer?
+
+Is the tweet against environmentalist initiatives?
+
+Is this abstract about Physics?
+
+Does the tweet express anger?
+
+Does the user dislike this movie?
+
+Is the sentence ungrammatical?
+
+Is this text expressing a need for evacuation?
+
+Is this text about Society and Culture?
+
+Is this a spam?
+
+is proofread by at least another author. See Figure 2 for a concrete example, and Figure 3 for some representative label descriptions.
+
+Additionally, some datasets contain thousands of labels (Chalkidis et al., 2019; Allaway and McKeown, 2020). In this case, we use templates to automatically synthesize label descriptions and exclude them from evaluation.
+
+Grouping similar datasets Our goal is to test the models' ability to generalize to tasks that are different enough from the training tasks. Therefore, at test time, we need to exclude not only the same dataset that appeared in the meta-tuning phase, but also ones that are similar.
+
+This poses a challenge: whether two datasets perform the same task involves subjective opinion, and there is no universally agreed definition. On one extreme, most datasets can be counted as dissimilar tasks, since they have different label spaces and input distributions. On the other extreme, all datasets can be considered the same task, since they can all be unified into the question answering format.
+
+To tackle this challenge, we create a set of tags, each describing a dataset property. The set of tags includes domain classification, article, emotion, social-media, etc, and the full set of them can be seen in Appendix C. Then we define the
+
+![](images/bcf964c006ad74e792156fb0160967fde0d25f8d225a61a5e5a479686a197d57.jpg)  
+Figure 4: Example dataset groups based on tags. We never train and test on datasets from the same group, e.g. train on hotel review and test on movie review.
+
+two datasets to be similar if they are associated with the same set of tags, and prohibit the model to learn from one and test on the other. For example, our work considers the topic classification datasets from Zhang et al. (2015) (AG News) and Yin et al. (2019) to be similar since they both classify topics for articles, even though their sources and label spaces are different. Some example dataset groups can be seen in Figure 4.
+
+Nevertheless, our procedure is not bullet-proof and one can argue that our notion of unseen tasks, though harsher than prior works (Yin et al., 2019; Pushp and Srivastava, 2017), is still lenient. Therefore, as additional robustness checks, for each dataset we evaluate, we manually identify and list the most relevant dataset that is allowed during training in Appendix F. For example, the most relevant dataset to the IMDB review sentiment classification dataset is the emotion classification dataset from Yin et al. (2019), which classifies the input text into 9 emotions, such as "joy", "surprise", "guilt", etc. We consider the emotion classification dataset to be relevant, since sentiment classification often involves identifying emotions. However, one can also argue that they are different tasks: their input and label spaces are different, and sadness can be caused by a great tragedy, or a bad movie that wastes the users' time. The comprehensive list of label descriptions grouped by dataset similarity is in Appendix D.
+
+In total, we spend around 200 hours to collect this dataset. This time estimate includes skimming through the dataset repos and recent NLP papers, writing programs to download the datasets and unify their format, annotating label descriptions, performing quality controls, and documenting the collection process.
+
+# 3 Metrics
+
+To reliably aggregate performance across different datasets and present as much information as possible, we report a set of descriptive statistics and provide visualizations whenever we compare two models. We generally do not reduce a model's performances on different datasets into one scalar quantity and compare this number only.
+
+Descriptive statistics For each label description (question), we calculate the AUC-ROC score  $^6$  by treating the "Yes" answer as the positive class. After calculating the AUC-ROC score for each label, we calculate the following set of descriptive statistics to compare two models. Suppose that model  $Y$  is hypothetically better than  $X$ . Denoting  $\Delta$  as the change of AUC-ROC of a label description from  $X$  to  $Y$ , we can summarize how  $\Delta$  is distributed across the set of label descriptions with the following statistics:
+
+-  $\mathbb{E}[\Delta]$ : the average change in AUC-ROC.  
+-  $\mathbb{P}[\Delta > t]$ : the fraction of label descriptions where the change is over the threshold  $t$ .  
+-  $\mathbb{P}[\Delta < -t]$ : the fraction of label descriptions where the change is less than  $-t$ .  
+-  $Std[\Delta]$ : the standard deviation of the change.
+
+In the main paper, we weight each label description equally in this distribution to calculate the above statistics. We may also weight each label or dataset equally, and the corresponding results are in Appendix E. To make sure our conclusions are robust, we consider one model to be better only when  $\mathbb{E}[\Delta] > 0$  and  $\mathbb{P}[\Delta > t] > \mathbb{P}[\Delta < -t]$  for all  $t \in \{1\%, 5\%, 10\}$ , under all three types of weighting. In other words, we claim that one model is better than the other only when 12 conditions simultaneously hold.
+
+Visualizations We use scatter plots to visualize and compare the performance of two models, where each dot represents a label description, its x-value represents the AUC-ROC score of the model  $X$ , and its y-value represents that of  $Y$ . If most dots are above the identity line  $y = x$ , the model  $Y$  is better than  $X$ .
+
+The descriptive statistics and the visualizations are explained in Figure 5.
+
+![](images/62a9a3da3e9fba938a13033df2d68feb0888438845e3b4bfa4309e421e62b1c1.jpg)  
+Figure 5: Each dot represents a label description, and its  $x / y$ -value each represents the performance of model  $X / Y$  (measured by AUC-ROC score). For example, on label description  $D1$ , model  $X / Y$  has AUC-ROC score 0.5/0.65. If the dot is above the black line ( $y = 0.5$ ), model  $Y$  is performing better than random guesses. If the dot is above the red line ( $y = x$ ), model  $Y$  is better than model  $X$ . Since one out of two dots are above  $y = x + 0.05$ , we have  $\mathbb{P}[\Delta > 5\%] = 0.5$ .
+
+# 4 Model
+
+Architecture We format the inputs to the model in the same way as UnifiedQA (Khashabi et al., 2020), which concatenates the context to the question and adds a "[SEP]" token in between. Then we feed the concatenated input into the T5 encoder and produce the answer score by normalizing the "Yes"/"No" probability of the first decoded token. Unless otherwise noted, we initialize our model with T5-Large (770 Million parameters). We sometimes compare to or initialize with the UnifiedQA model (Khashabi et al., 2020), which is trained on a wide range of question answering datasets. For a fair comparison, we use the UnifiedQA model initialized with T5-Large as well. To meta-tune non-Seq2Seq pretrained models, such as BERT (Devlin et al., 2019) or RoBERTa (Liu et al., 2019), we add an MLP layer on top of the pooled output/[CLS]" token to classify between "Yes"/"No". We leave the improvement on model architectures (Ye and Ren, 2021; Li and Liang, 2021; Lester et al., 2021) and training objectives (Murty et al., 2021; Yin et al., 2020) for future work.
+
+Meta-tuning We create a training distribution that balances between datasets, label descriptions, and "Yes"/“No” answers. To create the next training datapoint for meta-tuning, we select a
+
+dataset from the training split uniformly at random (u.a.r.); then we select a label description (question) u.a.r. and with  $50\%$  probability select a textual input with the answer "Yes"/"No". To prevent over-fitting, we do not train on any combination of label description and textual input twice. Unless otherwise noted, we meta-tune the model for 5000 steps and use batch size 32. We did not tune any hyper-parameters or training configurations since they work well during our first attempt. To evaluate ZSC performance on each dataset, we leave out one group of similar datasets as the evaluation set and train on the rest. Altogether, the experiments take around 250 GPU hours on Quadro 8000.
+
+# 5 Results
+
+# 5.1 Hypotheses and Conclusions
+
+We investigate and validate the following hypotheses, sorted by importance in descending order.
+
+Meta-tuned models outperform general question answering models in zero-shot classification.  
+- Larger pre-trained models are better.  
+- Pre-training does the heavy lifting.  
+- Performance can be improved by training on similar datasets, initializing with a QA model, or assembling label descriptions.  
+- Early stopping is crucial to performance.
+
+Meta-tuned models are better. We compare a meta-tuned T5-Large model (770 M parameters) with the same-sized UnifiedQA model (Khashabi et al., 2020) out of the box. Relevant descriptive statistics can be seen in the first row of Table 1 and Figure 6 (a). Adapting the model for ZSC improves the average AUC-ROC by  $3.3\%$ .
+
+Larger pre-trained models are better. We compare T5-Base (220 Million parameters) against T5-Large (770 M). The statistics can be seen in the second row of Table 1 and Figure 6 (b). Increasing the model size from 220 M to 770M improves the average AUC-ROC by  $6.3\%$ .
+
+<table><tr><td></td><td>E[Δ]</td><td>P[Δ &gt; 1%]</td><td>P[Δ &lt; -1%]</td><td>Std(Δ)</td></tr><tr><td>Meta-tuned vs. UnifiedQA</td><td>3.3%</td><td>59.5%</td><td>28.1%</td><td>9.5%</td></tr><tr><td>Larger</td><td>6.3%</td><td>75.1%</td><td>15.1%</td><td>8.1%</td></tr><tr><td>Pre-trained vs. Random</td><td>23.8%</td><td>95.7%</td><td>3.2%</td><td>14.0%</td></tr><tr><td>Train on Similar</td><td>0.7%</td><td>43.8%</td><td>20.5%</td><td>3.2%</td></tr><tr><td>Ensemble Descriptions</td><td>0.7%</td><td>28.9%</td><td>16.8%</td><td>3.1%</td></tr><tr><td>Initialize with UnifiedQA</td><td>1.1%</td><td>54.1%</td><td>24.3%</td><td>6.9%</td></tr></table>
+
+Table 1: The statistics used to compare two models, introduced in Section 3. The larger  $\mathbb{E}[\Delta]$  and the difference between  $\mathbb{P}[\Delta > 1\%]$  and  $\mathbb{P}[\Delta < -1\%]$ , the better. Row 1 finds that a meta-tuned model is better than UnifiedQA; row 2 finds that the larger model is better; row 3 finds that pre-training does the heavy lifting; row 4, 5, and 6 finds that the performance can be improved by training on similar datasets, assembling label descriptions, and initializing with a UnifiedQA model. Note that  $Std(\Delta)$  is the standard deviation of individual descriptions, not the standard deviation of the estimated mean. Due to space constraint we only show  $t = 1\%$  in this table.
+
+Pre-training does the heavy lifting. In Figure (c) and the third row of Table 1, we compare pretrained and random initializations, where the latter cannot beat the random baseline (average AUC-ROC 0.503). Hence, meta-tuning alone is far from enabling the model to perform ZSC. An intuitive interpretation is that the model already "knows" how to perform ZSC after pre-training under the LM objective, and learns how to use this knowledge during meta-tuning.
+
+Training on similar datasets improves performance. Unlike before, we no longer avoid training on similar datasets from the same group. Instead, we perform straightforward leave-one-out cross-validation. The statistics can be seen in the fourth row of Table 1 and Figure 6 (d), and it improves the average AUC-ROC by  $0.7\%$ . The performance gain is not as significant as increasing the model size or adapting for ZSC. We conjecture that it is because we have not collected enough datasets; otherwise, there might be more similar datasets, hence improving ZSC performance.
+
+Ensembling label descriptions improves performance. Instead of asking the model a single question for each label and obtain the probability of the answer being "Yes", we can average the probability obtained by asking multiple questions with the same meaning. This approach is different from traditional ensembling, which typically needs to store/train multiple models to average across them. The fifth row of Table 1 and Figure 6 (e) verifies that ensembling descriptions improves performance slightly (0.7% AUC-ROC score).
+
+Initializing with UnifiedQA improves performance. Figure 6 (f) and the sixth row of Table 1
+
+compare the UnifiedQA against against the T5 initialization. Initializing with UnifiedQA improves average AUC-ROC by  $1.1\%$ .
+
+Early stopping is crucial to performance. If we train the model for too long, the model might simply "memorize" that certain label descriptions correspond to certain training tasks, and the performance on unseen tasks may drop. To explore this possibility, we meta-tune our models for 100K steps, which is 20 times as long as our default setting and encourages the model to memorize the training tasks. We then evaluate them on the three benchmark zero-shot classification datasets by Yin et al. (2019) (which we describe in more details in the next section). We calculate the average AUC-ROC across all label descriptions for each of the 3 datasets, and plot them in Figure 7.
+
+The performance decreases  ${}^{8}$  as training continues. On the other hand, however, the performance drop of  $3\%$  in AUC-ROC is not fatal and the model's performance is still much better than random guesses.
+
+# 5.2 Robustness Checks
+
+We examine a series of additional results to make sure our conclusions are robust. The observed improvements in Table 1 and Figure 6 might be caused by the improvement of a small number of labels that are annotated with more descriptions, or by the improvement on a dataset with more distinct labels. Appendix E.1 compares the performance by assigning equal weights to each label/datasets.
+
+To provide additional supporting evidence for
+
+![](images/8a290466952a2840c089b373f85e5c54d3ed2f4dc86f731e5e8ef1ed4ed9c2e4.jpg)  
+(a)
+
+![](images/6f09035c3555da673ceb851fa09c1a3e06a2caf88d1a0a162b7658b02ebff5eb.jpg)  
+(b)
+
+![](images/2a7e4d058bc7b74a1d31f20661a02a859feb857d1aee45dec6b6903eae6ef8fe.jpg)  
+(c)
+
+![](images/9e8d07883e89c9685f87577d1d78d4ec7042903023c4bd8ae1dbc364d44e08ff.jpg)  
+(d)
+
+![](images/7993e0581a0600f8bcdd887d2c18d7c4d442a2d774c958591889ff5fcb93b823.jpg)  
+(e)
+
+![](images/1373831d65a6b3eb404b4e2099b16c43d0dc2447c7375a1224589d947805e58a.jpg)  
+(f)  
+Figure 6: The interpretation of these figures can be seen in Figure 5. (a) compares a meta-tuned model  $(y)$  against UnifiedQA  $(x)$ ; (b) compares T5-Large  $(770\mathrm{M}$  parameters) against T5-base  $(220\mathrm{M})$ ; (c) compares the T5 pretrained initialization against the random initialization; (d), (e), and (f) investigate whether performance can be improved by training on similar datasets, assembling different label descriptions (questions), and initializing with UnifiedQA. Conclusion: Since most dots are above the red line  $y = x$  for all 6 figures and above the random guess baseline  $(y = 0.5)$  by a robust margin, all conclusions listed at the beginning of Section 5 hold.
+
+![](images/f9e0a2fa7f7f2a7e97b946da712f07a53c8749f4795862555f6ae34937fa4d01.jpg)  
+Figure 7: Each curve corresponds to the models' performance on a dataset from Yin et al. (2019).  $x$ -value is the number of training steps;  $y$ -value is the average AUC-ROC score across all label descriptions, relative to the value at step 5000. Training for too long decreases performance on unseen tasks.
+
+our forecast that larger models are better, Appendix E.2 compares a 60M-parameter model against a 220M-parameter model, and finds that the latter is much better. One concern, however, is that our models are initialized with T5 (Raffel et al., 2019), which is trained on the open web and might have seen the datasets we gathered. There-
+
+<table><tr><td>Model</td><td>emotion</td><td>situation</td><td>topic</td></tr><tr><td>Yin et al. (2019)</td><td>25.2</td><td>38.0</td><td>52.1</td></tr><tr><td>Meta-tuned</td><td>28.2</td><td>48.4</td><td>54.3</td></tr></table>
+
+Table 2: "Prior" means the best performing system from Yin et al. (2019) for each dataset; "Meta-tuned" means meta-tuning on RoBERTa. Our approach is better on all three datasets.
+
+fore, larger models might be better simply because they are better at memorization (Sagawa et al., 2020). Appendix E.3 addresses this by showing that larger models are also better with BERT initialization (Devlin et al., 2019), which is trained on Wikipedia and Book Corpus (Zhu et al., 2015).
+
+We also report the models' performance on each dataset for readers' reference in Appendix G.
+
+# 5.3 Comparison with Yin et al. (2019)
+
+This section shows that our approach has higher performance than the zero-shot classification system built by Yin et al. (2019). Their system ensembles several natural language inference models based on RoBERTA-Large (355M parameters, Liu et al. (2020)), and another model trained to categorize Wikipedia articles. It was evaluated on three classification datasets:
+
+- topic (10-way): classifies article domains, such as family & relationship, education, sports, etc. The metric is accuracy.  
+- emotion (10-way): classifies emotion types, such as joy, anger, guilt, shame, etc. The metric is label-weighted F1.  
+- situation (12-way): classifies disaster situations, e.g. regime change, crime & violence, and the resource they need, e.g. search & rescue. The metric is label-weighted F1.
+
+We use the exact same evaluation metrics as in Yin et al. (2019), and the same label resolution strategy when the model answers "Yes" for multi-label classification. Concretely, when the model predicts "Yes" on multiple labels, the one with the highest probability is selected. For a fair comparison, we meta-tune RoBERTa of the same size and compare it with the highest performing model in Yin et al. (2019) for each of the three datasets.
+
+The results are in Table 2, and our model has higher performance across all 3 datasets using the same pre-training method.
+
+# 6 Discussion and Future Directions
+
+Main takeaways We construct a dataset of classification datasets to adapt the language model for zero-shot classification via meta-tuning. The adapted model outperforms a general-purpose question answering model and the prior state of the art based on natural language inference. We forecast that meta-tuning would be more effective on larger models, and the current engineering ceiling for zero-shot learning might have been broadly under-estimated.
+
+Aggregating and unifying datasets The main bottleneck of our research is to manually gather a wide range of datasets and unify their format. The difficulties are: 1) we need to brainstorm and review the NLP literature extensively to decide what new tasks to look for; 2) different datasets encode their data in different formats, and we need to write programs manually for each of them to convert to the desired format; 3) it is hard to tell the quality of a dataset purely by its provenance, and sometimes we need to examine the dataset manually. If we as a community can aggregate and unify datasets better, we could potentially train and evaluate zero-shot learning models at a larger scale.
+
+Meta-tuning as a probe There is a growing interest in measuring the intelligence (Hendrycks et al., 2021a,b) or the few-shot learning ability (Brown et al., 2020) of large language models like GPT-3. However, since these models are not adapted to answer those prompts (Holtzman et al., 2021), we suspect that its knowledge and true potential to perform few-shot learning is much higher than reported. Since pre-training does the heavy lifting and meta-tuning is unlikely to provide additional ZSC ability to the model, we can potentially first use meta-tuning as a probe to make them adapted to answering prompts before measuring their performance.
+
+Still, to make this methodology rigorous, interpreting and controlling the strength of the probes will be an important future direction (Hewitt and Liang, 2019). For example, if the training set contains a prompt that is too similar to the prompt to be tested, the probe will be meaningless.
+
+Beyond Shallow Correlations One possibility is that the model only learns shallow statistical correlations from meta-tuning rather than "more sophisticated reasoning skills". For example, the word "exciting" might occur in positive reviews more. This is unlikely, given that larger models are consistently better than smaller or randomly initialized ones. To explain this performance gap, larger models must have learned to use more complicated features during meta-tuning.
+
+Relation to Meta/Multitask-Learning Our method is closely related to, but different from meta-learning (Yin, 2020; Murty et al., 2021) and multi-task learning (Ye et al., 2021; Aghajanyan et al., 2021). Both meta-learning and multitask-learning typically involve at least a few examples from the target task; in our setup, however, the model does not learn from any target task examples. The "meta" in our name does not mean "meta-learning", but reflects the fact that our model learns from a meta-dataset of tasks.
+
+Nevertheless, our framework can be easily adapted to a few-shot learning setup, which enables the language model to learn to learn from incontext examples (see below). Since this approach models the learning process as a sequence classification problem, it can be seen as a form of meta-learning similar to (Ravi and Larochelle, 2016).
+
+Annotating Prompts Three of our authors annotated the label descriptions. Since they are all
+
+Computer Science major students who understand machine learning and natural language processing, they might not be representative of the final user population of this ZSC application. Annotating prompts that match the target user distribution will be an important research direction.
+
+Additionally, shorter and more natural descriptions sometimes fail to capture the exact semantics of the label. For example, in Yin et al. (2019), the description of the label "medical" is "people need medical assistance"; or alternatively, it can be longer but more accurate: "people need an allied health professional who supports the work of physicians and other health professionals". How to scalably generate more accurate and detailed label descriptions without expert efforts will be another future direction.
+
+Optimizing Prompts Our work is complementary to recent works that optimize the prompts to achieve better accuracy. Even if our metatuned model is specialized in answering prompts, it might still react very differently towards different prompts. For example, in the stance classification dataset (Barbieri et al., 2020), we annotated two label descriptions (prompts) for the same label: "Does this post support atheism?" and "Is the post against having religious beliefs?". They have similar meanings, but the former has much lower accuracy than the later. We conjecture that this is because the model cannot ground abstract concepts like "atheism".
+
+Other extensions We conjecture that metatuning can be extended to more diverse tasks beyond zero-shot binary classification. To extend to multi-label classification, we need to develop a procedure to resolve the labels when the model predicts positive for more than one labels. To extend to few-shot learning, we need to increase the context length to fit several training examples into the input, which requires a larger context window and hence more computational resources. To extend to other sequence generation tasks, we need to collect a wide range of diverse sequence generation tasks to meta-tune the model, such as machine translation, summarization, free-form question answering, grammar correction, etc.
+
+# Acknowledgements
+
+We thank Eric Wallace for his feedbacks throughout the project. We thank Steven Cao, David
+
+Gaddy, Haizhi Lai, Jacob Steinhardt, Kevin Yang and anonymous reviewers for their comments on the paper.
+
+# References
+
+Armen Aghajanyan, Anchit Gupta, Akshit Shrivastava, Xilun Chen, Luke Zettlemoyer, and Sonal Gupta. 2021. Muppet: Massive multi-task representations with pre-finetuning. arXiv preprint arXiv:2101.11038.  
+Emily Allaway and Kathleen McKeown. 2020. Zero-Shot Stance Detection: A Dataset and Model using Generalized Topic Representations. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8913-8931, Online. Association for Computational Linguistics.  
+Tiago Almeida, José María Gómez Hidalgo, and Tiago Pasqualini Silva. 2013. Towards sms spam filtering: Results under a new dataset. International Journal of Information Security Science, 2(1):1-18.  
+Francesco Barbieri, Jose Camacho-Collados, Luis Espinosa Anke, and Leonardo Neves. 2020. TweetEval: Unified benchmark and comparative evaluation for tweet classification. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 1644–1650, Online. Association for Computational Linguistics.  
+Valerio Basile, Cristina Bosco, Elisabetta Fersini, Debora Nozza, Viviana Patti, Francisco Manuel Rangel Pardo, Paolo Rosso, and Manuela Sanguinetti. 2019. SemEval-2019 task 5: Multilingual detection of hate speech against immigrants and women in Twitter. In Proceedings of the 13th International Workshop on Semantic Evaluation, pages 54–63, Minneapolis, Minnesota, USA. Association for Computational Linguistics.  
+Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc.  
+Jonathan Bragg, Arman Cohan, Kyle Lo, and Iz Beltagy. 2021. Flex: Unifying evaluation for few-shot nlp. arXiv preprint arXiv:2107.07170.  
+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. arXiv preprint arXiv:2005.14165.  
+Nicholas Carlini, Florian Tramér, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ülfar Erlingsson, Alina Oprea, and Colin Raffel. 2020.
+
+Extracting training data from large language models. arXiv preprint arXiv:2012.07805.  
+Ilias Chalkidis, Emmanouil Fergadiotis, Prodromos Malakasiotis, and Ion Androutsopoulos. 2019. Large-scale multi-label text classification on EU legislation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6314-6322, Florence, Italy. Association for Computational Linguistics.  
+Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. *BoolQ: Exploring the surprising difficulty of natural yes/no questions.* 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 2924–2936, Minneapolis, Minnesota. 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.  
+Tianyu Gao, Adam Fisch, and Danqi Chen. 2021. Making pre-trained language models better few-shot learners. 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 3816-3830, Online. Association for Computational Linguistics.  
+Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt. 2020. Aligning AI With Shared Human Values. arXiv e-prints, page arXiv:2008.02275.  
+Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt. 2021a. Aligning {ai} with shared human values. In International Conference on Learning Representations.  
+Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021b. Measuring massive multitask language understanding. In International Conference on Learning Representations.  
+John Hewitt and Percy Liang. 2019. Designing and interpreting probes with control tasks. 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 2733-2743, Hong Kong, China. Association for Computational Linguistics.
+
+Ari Holtzman, Peter West, Vered Shwartz, Yejin Choi, and Luke Zettlemoyer. 2021. Surface form competition: Why the highest probability answer isn't always right. CoRR, abs/2104.08315.  
+Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. 2020. UNIFIEDQA: Crossing format boundaries with a single QA system. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1896-1907, Online. Association for Computational Linguistics.  
+Tven Le Scao and Alexander Rush. 2021. How many data points is a prompt worth? In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2627-2636, Online. Association for Computational Linguistics.  
+Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691.  
+Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190.  
+Xin Li and Dan Roth. 2002. Learning question classifiers. In COLING 2002: The 19th International Conference on Computational Linguistics.  
+Hairong Liu, Mingbo Ma, Liang Huang, Hao Xiong, and Zhongjun He. 2019. Robust neural machine translation with joint textual and phonetic embedding. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3044-3049, Florence, Italy. Association for Computational Linguistics.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Ro{bert}a: A robustly optimized {bert} pretraining approach.  
+Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011a. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142-150, Portland, Oregon, USA. Association for Computational Linguistics.  
+Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011b. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142-150, Portland, Oregon, USA. Association for Computational Linguistics.
+
+Bryan McCann, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. 2018. The natural language decathlon: Multitask learning as question answering. arXiv preprint arXiv:1806.08730.  
+Tsvetomila Mihaylova, Georgi Karadzhov, Pepa Atanasova, Ramy Baly, Mitra Mohtarami, and Preslav Nakov. 2019. SemEval-2019 task 8: Fact checking in community question answering forums. In Proceedings of the 13th International Workshop on Semantic Evaluation, pages 860–869, Minneapolis, Minnesota, USA. Association for Computational Linguistics.  
+Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. 2021. Natural instructions: Benchmarking generalization to new tasks from natural language instructions. arXiv preprint arXiv:2104.08773.  
+Rishabh Misra. 2019. Imdb spoiler dataset.  
+Rishabh Misra, Mengting Wan, and Julian McAuley. 2018. Decomposing fit semantics for product size recommendation in metric spaces. In Proceedings of the 12th ACM Conference on Recommender Systems, pages 422-426. ACM.  
+Saif Mohammad, Felipe Bravo-Marquez, Mohammad Salameh, and Svetlana Kiritchenko. 2018. SemEval-2018 task 1: Affect in tweets. In Proceedings of The 12th International Workshop on Semantic Evaluation, pages 1-17, New Orleans, Louisiana. Association for Computational Linguistics.  
+Saif Mohammad, Svetlana Kiritchenko, Parinaz Sobhani, Xiaodan Zhu, and Colin Cherry. 2016. SemEval-2016 task 6: Detecting stance in tweets. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval-2016), pages 31-41, San Diego, California. Association for Computational Linguistics.  
+Shikhar Murty, Tatsunori B Hashimoto, and Christopher D Manning. 2021. Dreca: A general task augmentation strategy for few-shot natural language inference.  
+Bo Pang and Lillian Lee. 2004. A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics (ACL-04), pages 271-278, Barcelona, Spain.  
+Pushpankar Kumar Pushp and Muktabh Mayank Srivastava. 2017. Train once, test anywhere: Zero-shot learning for text classification. arXiv preprint arXiv:1712.05972.  
+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.
+
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683.  
+Sachin Ravi and Hugo Larochelle. 2016. Optimization as a model for few-shot learning.  
+Sara Rosenthal, Noura Farra, and Preslav Nakov. 2017. SemEval-2017 task 4: Sentiment analysis in Twitter. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), pages 502-518, Vancouver, Canada. Association for Computational Linguistics.  
+Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. 2020. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR.  
+Timo Schick and Hinrich Schütze. 2020a. Exploiting cloze questions for few-shot text classification and natural language inference. arXiv preprint arXiv:2001.07676.  
+Timo Schick and Hinrich Schütze. 2020b. It's not just size that matters: Small language models are also few-shot learners. arXiv preprint arXiv:2009.07118.  
+Taylor Shin, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh. 2020. AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Prompts. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4222-4235, Online. Association for Computational Linguistics.  
+Sasha Spala, Nicholas Miller, Franck Dernoncourt, and Carl Dockhorn. 2020. SemEval-2020 task 6: Definition extraction from free text with the DEFT corpus. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 336-345, Barcelona (online). International Committee for Computational Linguistics.  
+Sasha Spala, Nicholas A. Miller, Yiming Yang, Franck Dernoncourt, and Carl Dockhorn. 2019. DEFT: A corpus for definition extraction in free- and semi-structured text. In Proceedings of the 13th Linguistic Annotation Workshop, pages 124–131, Florence, Italy. Association for Computational Linguistics.  
+Iulia Turc, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Well-read students learn better: On the importance of pre-training compact models. arXiv preprint arXiv:1908.08962.  
+Cynthia Van Hee, Els Lefever, and Véronique Hoste. 2018. SemEval-2018 task 3: Irony detection in English tweets. In Proceedings of The 12th International Workshop on Semantic Evaluation, pages 39–50, New Orleans, Louisiana. Association for Computational Linguistics.
+
+Eric Wallace, Tony Z Zhao, Shi Feng, and Sameer Singh. 2020. Customizing triggers with concealed data poisoning. arXiv preprint arXiv:2010.12563.  
+Alex Warstadt, Amanpreet Singh, and Samuel R Bowman. 2018. Neural network acceptability judgments. arXiv preprint arXiv:1805.12471.  
+Orion Weller, Nicholas Lourie, Matt Gardner, and Matthew Peters. 2020. Learning from task descriptions. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1361-1375, Online. Association for Computational Linguistics.  
+Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112-1122, New Orleans, Louisiana. Association for Computational Linguistics.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, 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.  
+Qinyuan Ye, Bill Yuchen Lin, and Xiang Ren. 2021. Crossfit: A few-shot learning challenge for cross-task generalization in NLP. CoRR, abs/2104.08835.  
+Qinyuan Ye and Xiang Ren. 2021. Zero-shot learning by generating task-specific adapters. arXiv preprint arXiv:2101.00420.  
+Mingzhang Yin, George Tucker, Mingyuan Zhou, Sergey Levine, and Chelsea Finn. 2020. Metalearning without memorization. In International Conference on Learning Representations.  
+Wenpeng Yin. 2020. Meta-learning for few-shot natural language processing: A survey. arXiv preprint arXiv:2007.09604.  
+Wenpeng Yin, Jamaal Hay, and Dan Roth. 2019. Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach. 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 3914-3923, Hong Kong, China. Association for Computational Linguistics.
+
+Marcos Zampieri, Shervin Malmasi, Preslav Nakov, Sara Rosenthal, Noura Farra, and Ritesh Kumar. 2019. SemEval-2019 task 6: Identifying and categorizing offensive language in social media (OffensEval). In Proceedings of the 13th International Workshop on Semantic Evaluation, pages 75–86, Minneapolis, Minnesota, USA. Association for Computational Linguistics.  
+Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, NIPS'15, page 649-657, Cambridge, MA, USA. MIT Press.  
+Tony Z. Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. arXiv preprint arXiv:2102.09690.  
+Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, pages 19-27.
+
+# A Ethics
+
+Data and incentives In the existing prompting framework, end users send the natural language descriptions and a few training examples to the large language model inference API to perform few-shot learning (Brown et al., 2020). This becomes a natural source of training data for metatuning. Hence, the success of meta-tuning presented in this paper might incentivize for-profit organizations who provide language model inference APIs to collect prompts from the users, and train on these data.
+
+Privacy, security, and fairness If a model is meta-tuned on user-provided data, certain security, privacy and fairness concerns can potentially emerge. For example, Carlini et al. (2020) shows that it is possible to extract the training data from large language models, and hence meta-tuned systems might expose some users' prompts to other users. Wallace et al. (2020) shows that it is possible to poison the model through training data and trigger unwanted behaviors; the meta-tuning procedure might be susceptible to these data poisoning attacks as well. Finally, meta-tuning might perpetuate existing societal biases hidden in the users' prompts (Bolukbasi et al., 2016).
+
+If not addressed properly, these concerns might have a broader negative societal impact through meta-tuning. Compared to other domain-specific and task-specific machine learning applications, meta-tuned models might be applied to a much wider range of tasks, deployed at a larger scale, and serving a more diverse set of user population. Therefore, biased or poisoned training data for one task from one user population might compromise fairness and performance of another task and harm another user population; additionally, malicious or biased data might even tamper with the few-shot learning capability ("meta-poisoning").
+
+Potential abuse As shown in Figure 6, the AUC-ROC score for a lot of tasks are still well below 0.9, and hence our system is far from solving a significant fraction of tasks. Therefore, even though our system is flexible and has the potential to perform a wide range of tasks, it does not present an elixir to all classification tasks. Particularly, it should not be applied to higher stake scenarios (e.g. hate speech detection, fake news detection, etc), since its efficacy, robustness, and fairness properties remain unknown.
+
+# B Datasets
+
+IMDB movie review sentiment classification (Maas et al., 2011b). Classifies whether the user likes the movie.
+
+POSITIVE: "My favourite police series of all time turns to a TV-film. Does it work? Yes. ..."
+
+NEGATIVE: "Stupid! Stupid! Stupid! I can not stand Ben stiller anymore."
+
+Zero Shot Emotion Classification (Yin et al., 2019). This task classifies a textual input into 9 emotion types {"sadness", "joy", "anger", "disgust", "fear", "surprise", "shame", "guilt", "love"}, and none-type if not any of the above. For example,
+
+JOY: "Making new friends is always fun, especially when playing dress up"
+
+ANGER: "People that smoke cigarettes irritate my soul."
+
+Zero Shot topic Classification (Yin et al., 2019). This task classifies an articles into 10 topic classes, such as "Health", 'Sports", "Family & Relationships", "Science & Mathematics", etc.
+
+SCIENCE: "How the human species evolved? How the human species evolved? A tough question as it overlaps science and theology ..."
+
+SPORTS: "Number one formula one driver? Depends on your question. Kimi Raikkonen won more races ..."
+
+Zero Shot Situation Classification (Yin et al., 2019). This task classifies what type of situation the user is having and what aids it needs. The classes include "food", "terrorism", etc.
+
+FOOD: "livestock, crops, rural and urban housing, roads, river and canal embankments, and other infrastructure suffered extensive damage"
+
+TERRORISM: "Daish attacked Tahunmahu village with chemical gas."
+
+SemEval 2017 Task 4 (Barbieri et al., 2020) (Rosenthal et al., 2017). This task classifies the sentiment of a tweet as {"negative", "neutral", "positive)}. For example,
+
+NEGATIVE: "That sucks if you have to take the SATs tomorrow"
+
+NEUTRAL: "@user Well said on HMW. Can you now address why Texans fans file out of the stadium midway through the 4th qtr of every game?"
+
+POSITIVE: "This is the last Sunday w/o football ....NFL is back baby""  
+ETHICS Justice (Hendrycks et al., 2020). This task identifies whether the justification provided for a given scenario is deemed reasonable. For example,  
+REASONABLE: "I deserve to be paid by my boss because I keep her house clean daily."  
+UNREASONABLE: "I deserve to be given a brand new car by my boss because I keep her house clean daily."  
+ETHICS Deontology (Hendrycks et al., 2020). This task identifies whether the excuse provided for a given scenario is a reasonable excuse. For example,  
+SCENARIO: "Aren't you supposed to be preparing for the meeting?"  
+EXCUSE (REASONABLE): "No because I have everything already set up in the conference room."  
+EXCUSE (UNREASONABLE): "No because you don't have a meeting"  
+ETHICS Common Sense (Hendrycks et al., 2020). This task identifies identifies whether a scenario demonstrates common sense. For example,  
+COMMON SENSE: "I went to the principal's office to change my records before going to a different school."  
+NOT COMMON SENSE: "I secured the loan because I would make the payments."  
+EURLEX57K (Chalkidis et al., 2019). This task classifies European legislation.  
+NATIONAL CURRENCY: "Council Regulation (EC) No 2595/2000 of 27 November 2000 amending Regulation (EC) No 1103/97 on certain provisions relating to the introduction of the euro"  
+SOUTHERN AFRICA: "95/458/EC: Commission Regulation (EC) No 302/2006 of 20 February 2006 on import licences in respect of beef and veal products originating in Botswana, Kenya, Madagascar, Swaziland, Zimbabwe and Namibia"  
+SemEval 2019 Task 6 (Barbieri et al., 2020) (Zampieri et al., 2019). This task classifies the tweet as either offensive or not offensive. For example,  
+OFFENSIVE: "@user She has become a parody unto herself? She has certainly taken some heat for being such an...well idiot. Could be optic too
+
+Who know with Liberals They're all optics. No substance"  
+NOT OFFENSIVE: "@user @user She is great. Hi Fiona!"
+
+Click Bait Detection 10 This task detects whether a news title is a click bait.
+
+CLICK BAIT: "Can You Pass This Basic Trigonometry Quiz"
+
+NON CLICK BAIT: "NASCAR driver Kyle Busch wins 2011 Jeff Byrd 500".
+
+Abstract Domain Classification 11 This classifies the abstract into 4 domains: "Physcis", "Maths", "Computer Science", "Statistics". For example,
+
+PHYSICS: "a ever-growing datasets inside observational astronomy have challenged scientists inside many aspects, including an efficient and interactive data exploration and visualization. many tools have been developed to confront this challenge ..."
+
+MATHS: "a main result of this note was a existence of martingale solutions to a stochastic heat equation (she) inside the riemannian manifold ..."
+
+SemEval 2019 Task 5 (Barbieri et al., 2020) (Basile et al., 2019). This task identifies whether the tweet contains hate speech towards women and/or immigrants or not. For example,
+
+HATE SPEECH: "This account was temporarily inactive due to an irrational woman reporting us to Twitter. What a lack of judgement, shocking. #YesAllMen"
+
+NO HATE SPEECH: "@user nice new signage. Are you not concerned by Beatlemania -style hysterical crowds crongregating on you..."
+
+SemEval 2019 Task 8 (Mihaylova et al., 2019). This task identifies whether the text is an example of a question asking for factual information, an example of a question asking for an opinion, or an example of socializing. For example,
+
+FACTUAL: "is there any place i can find scented massage oils in qatar?"
+
+OPINION: "hi there; i can see a lot of massage center here; but i dont which one is better.
+
+can someone help me which massage center is good...and how much will it cost me? thanks"
+
+SOCIALIZING: "Hello people...let's play this game...you have to write something good about the person whose 'post' is above you on  $QL$ . You can write anything and you can write&#160; multiple times."
+
+SemEval 2018 Task 3 (Barbieri et al., 2020) (Van Hee et al., 2018). This task identifies whether the tweet contains irony or not. For example,
+
+IRONY: "seeing ppl walking w/ crutches makes me really excited for the next 3 weeks of my life"
+
+NO IRONY: "@user on stage at #flzjingleball at the @user in #Tampa #iheartradio"
+
+SemEval 2018 Task 1 (Barbieri et al., 2020; Mohammad et al., 2018) This task classifies a tweet as one of 4 emotion types {"sadness", "joy", "anger", "optimism)}. For example,
+
+SADNESS: "@user I so wish you could someday come to Spain with the play, I can't believe I'm not going to see it #sad"
+
+JOY: "#ThisIsUs has messed with my mind &amp; now I'm anticipating the next episode with #apprehension &amp; #delight! #isthereahelplineforthis"
+
+ANGER: "@user Haters!!! You are low in self worth. Self righteous in your delusions. You cower at the thought of change. Change is inevitable."
+
+OPTIMISM: "Don't be #afraid of the space between your #dreams and #reality. If you can #dream it, you can #make it so"
+
+SemEval 2016 Task 6 (Mohammad et al., 2016; Barbieri et al., 2020) This task classifies a tweet's stance as {"neutral", "against", "favor)}. Each tweet contains a stance on one of the five different target topics {"abortion", "atheism", "climate change", "feminism", "hillary)}. For example,
+
+NEUTRAL: "@user maybe that's what he wants #SemST"
+
+AGAINST: "Life is #precious & so are babies, mothers, & fathers. Please support the sanctity of Human Life. Think #SemST"
+
+FAVOUR: "@user @user Nothing to do with me. It's not my choice, nor is it yours, to dictate what another woman chooses. #feminism #SemST"
+
+SemEval 2020 Task 6 (Spala et al., 2020). This task classifies whether textbook sentence contains a definition. For example,
+
+CONTAINS DEFINITION: "Since 2005, automated sequencing techniques used by laboratories are under the umbrella of next-generation sequencing, which is a group of automated techniques used for rapid DNA sequencing"
+
+DOESN'T CONTAIN DEFINITION: "These automated low-cost sequencers can generate sequences of hundreds of thousands or millions of short fragments (25 to 500 base pairs) in the span of one day."
+
+TREC (Li and Roth, 2002). This task classifies a question into one of six question types: DESC (description), ABBR (abbreviation), ENTY (entity), HUM (people/individual), LOC (location), NUM (numeric information), each of which have specific fine-grained sub-categories. For example,
+
+DESC: "How did serfdom develop in and then leave Russia?"
+
+ABBR: "What is the full form of.com?"
+
+ENTY: "What films featured the character Pope eye Doyle?"
+
+HUM: "What contemptible scoundrel stole the cork from my lunch?"
+
+LOC: "What sprawling U.S. state boasts the most airports?"
+
+NUM: "How many Jews were executed in concentration camps during WWII?"
+
+SUBJ (Pang and Lee, 2004). This task classifies a sentence as being subjective or objective. For example,
+
+SUBJECTIVE: "smart and alert, thirteen conversations about one thing is a small gem."
+
+OBJECTIVE: "the movie begins in the past where a young boy named sam attempts to save celebi from a hunter."
+
+The Corpus of Linguistic Acceptability (Warstadt et al., 2018). This task detects if sentences are grammatically acceptable by their original authors. For example,
+
+GRAMMATICALLY ACCEPTABLE: "Her little sister will disagree with her."
+
+GRAMMATICALLY NOT ACCEPTABLE: "Has not Henri studied for his exam?"
+
+The Multi-Genre NLI Corpus (Williams et al., 2018). This task detects if a premise is a contradiction or entailment of a hypothesis, or if a hypothesis holds neutral view on the premise.. For example,
+
+NEUTRAL: "Premise: Exoatmospheric Kill Vehicles orbiting Earth would be programmed to collide with warheads. Hypothesis: Exoatmospheric Kill Vehicles would be very expensive and hard to make."
+
+ENTAILMENT: "Premise: so we have to run our clocks up forward an hour and i sure do hate to loose that hour of sleep in the morning. Hypothesis: I don't like the time change that results in losing an hour of sleeping time."
+
+CONTRADICTION: "Premise: The mayor originally hoped groundbreaking would take place six months ago, but it hasn't happened yet. Hypothesis: The mayor doesn't want groundbreaking to happen at all."
+
+Metaphor as a Medium for Emotion: An Empirical Study (?) This task detects if the application of a word is Literal or Metaphorical. For example,
+
+WORD: ABUSE
+
+LITERAL: "This boss abuses his workers."
+
+METAPHORICAL: "Her husband often abuses alcohol."
+
+Political Preference Classification (Allaway and McKeown, 2020). This task predicts a comment's stand point on a political topic. For example,
+
+# TOPIC: COMPANIES REGULATION
+
+CON: "Regulation of corporations has been subverted by corporations. States that incorporate corporations are not equipped to regulate corporations that are rich enough to influence elections, are rich enough to muster a legal team that can bankrupt the state. Money from corporations and their principals cannot be permitted in the political process if democracy is to survive."
+
+PRO: "Regulation is to a corporation what a conscience is to a living person. Without a conscience, we would all be sociopaths. Corporations do not have a conscience, thus they need regulation to make sure they are focused on benefiting society instead on merely benefiting themselves."
+
+NEUTRAL: "Without government to ensure their behavior, companies will attempt to make a profit even to the DETRIMENT of the society that supports the business. We have seen this in the environment, in finances, in their treatment of workers and customers. Enough."
+
+Airline Service Review 12 This task classifies if an airline review has a positive or negative sentiment. For example,
+
+POSITIVE: "This is such a great deal! Already thinking about my 2nd trip to Australia; I haven't even gone on my 1st trip yet!"
+
+NEGATIVE: "amazing to me that we can't get any cold air from the vents."
+
+Covid-19 Tweets Sentiment Analysis 13 This task classifies if a tweet has a positive or negative sentiment. For example,
+
+POSITIVE: "Taken by Henk Zwoferink on Saturday in Wargl, our black beauty hauled a train bringing the last tourists home. Our colleagues are #workinghard to keep supply chains running while respecting the measures to ensure everyone's #safety. A pleasure to work with such #Dedicated People!"
+
+NEGATIVE: "So far, the Minister does not seem to have made statement on the catastrophe that can develop if the issue of markets operation is not addressed. Food insecurity has potential to make current Covid-19 panic look like a kindergarten and could lead to riots. I submit."
+
+Hotel Review 14 This task predicts if a hotel review is a positive or negative review. For example,
+
+NEGATIVE: "The single rooms like hospital rooms single rooms hotel sparse intentional know ugly like trapped hospital white walls sink basin room small rectangle shape.the beds hard rocks blankets rough really noisy.this overrated hotel stayed fans type hotels"
+
+POSITIVE: "loved stay, stayed univ, inn 10 days april 2005 thoroughly enjoyed, free parking clean spacious room friendly staff great breakfast snack, loved location, definitely stay,"
+
+Stock Market Sentiment  $^{15}$  This task predicts if a comment holds a positive or negative view on the performance of the stock market. For example,
+
+NEGATIVE: "GPS wow that was a fast fast fade..."
+
+POSITIVE: "user Maykiljil posted that: I agree that MSFT is going higher & possibly north of 30"
+
+AG-News (Zhang et al., 2015). This task classifies the topic of news based on their contents. For example,
+
+WORLD NEWS: "Greek duo could miss drugs hearing"
+
+SPORTS NEWS: "AL Wrap: Olerud Cheers Yankees by Sinking Ex-Team"
+
+BUSINESS NEWS: "Lowe's Second-Quarter Profit Rises"
+
+TECH NEWS: "Satellite boosts Olympic security"
+
+Real and Fake News 16 This task classifies if a news is fake or real. For example,
+
+REAL: "WASHINGTON (Reuters) - Alabama Secretary of State John Merrill said he will certify Democratic Senator-elect Doug Jones as winner on Thursday despite opponent Roy Moorea x80
+
+$x99s$  challenge, in a phone call on CNN. Moore, a conservative who had faced allegations of groping teenage girls when he was in his 30s, filed a court challenge late on Wednesday to the outcome of a U.S. Senate election he unexpectedly lost."
+
+FAKE: "Ronald Reagan shut down the Berkeley protests many years ago THIS is how you do it!"
+
+Disaster Tweets  $^{17}$  This task detects if a tweet announces an emergency or a disaster. For example,
+
+CONTAINS DISASTER: "Our Deeds are the Reason of this #earthquake May ALLAH Forgive us all."
+
+DOES NOT CONTAIN DISASTER: "My dog attacked me for my food #pugprobs."
+
+Obama vs Trump Tweets  $^{18}$  This task detects if a tweet was send by Obama or Trump. For example,
+
+OBAMA: "Michelle and I are delighted to congratulate Prince Harry and Meghan Markle on their engagement. We wish you a lifetime of joy and happiness together."
+
+TRUMP: "Together, we dream of a Korea that is free, a peninsula that is safe, and families that are reunited once again!"
+
+Kaggle Sexually Explicit Tweets  $^{19}$  This dataset provides positive examples of profane comments. For example,
+
+EXPLICIT "What do guys say when you get naked in front of them for the first time?"
+
+Democratic vs Republican Tweets  $^{20}$  This task detects if a tweet was send by the Democratic or Republican Party. For example,
+
+DEMOCRATIC: "YuccaMountain would require moving tens of thousands of metric tons of radioactive waste across the country and through Southern Nevada."
+
+REPUBLICAN: "Stopped by One Hour Heating&amp; Air Conditioning to discuss the benefits tax reform will bring to their business."
+
+Women E-commerce Clothing Reviews 21
+
+This task predicts if the buyer likes or recommends a product base on its review. For example,
+
+LIKE: "After reading the previous reviews, i ordered a size larger. i am so glad i did it! it fits perfectly! i am 5'4"/115/32dd and went with the s regular. so beautiful! i can't wait to wear it!"
+
+DISLIKE: "The zipper broke on this piece the first time i wore it. very disappointing since i love the design. I'm actually going to try to replace the zipper myself with something stronger, but annoying that it's come to that."
+
+Quora Question Pairs 22 This task predicts if a pair of Quora question is asking for the same thing. For example,
+
+SAME: "Question 1: How many months does it take to gain knowledge in developing Android apps from scratch?; Question 2: How much time does it take to learn Android app development from scratch?"
+
+DIFFERENT: "Question 1: How would you review the site Waveclues? ; Question 2: Is there a good pay for reviews site out there?"
+
+Headline Sarcasm Detection This task detects if is a news headline contains sarcasm. For example,
+
+SARCASM: "guy who just wiped out immediately claims he's fine"
+
+NO SARCASM: "Donald trump effigies burn across Mexico in Easter ritual"
+
+Company Account Tweets  $^{23}$  This task detects whether the tweet is targeted towards a company account. For example,
+
+YES: "@VirginTrains Oh, that's nice. What are you doing about it? What are you targets next year?"
+
+No: "@115738 That's the best kind of trick-or-treating. All treats, my friend. -Becky"
+
+SMS Spam Detection (Almeida et al., 2013) This task detects whether the SMS is a spam message. For example,
+
+SPAM: "Thank you, winner notified by sms. Good Luck! No future marketing reply STOP to 84122 customer services 08450542832"
+
+HAM: "Lol great now I am getting hungry."
+
+Clothing Fitness (Misra et al., 2018) Checking whether the customer complains that the cloth is too small or too large.
+
+SMALL: "runs a bit small. wish it fit".
+
+LARGE: "too big".
+
+Water Problem Topic Classification 24 Classifying the topic of a report on water problems. The labels include "biological", "climatic indicator", "environmental technology", etc. For example,
+
+BIOLOGICAL: "Mineralization of organic phosphorus in bottom sediments reaches  $40 - 80\%$  and as we found out during the project implementation it intensified in autumn-winter period."
+
+CLIMATIC INDICATOR: "The average amount of precipitation in the lower part of the basin makes  $470~\mathrm{mm}$  to  $540~\mathrm{mm}$ . The relative average annual air humidity makes  $60 - 65\%$ .
+
+ENVIRONMENTAL TECHNOLOGY: "Most of wastewater treatment facilities require urgent modernization and reconstruction".
+
+Sexist Statement Detection 25 This task classifies whether the statement is sexist. For example,
+
+SEXIST: "It's impossible for a girl to be faithful."
+
+NON SEXIST: "Without strength, can we work to create wealth?"
+
+Movie Spoiler Detection (Misra, 2019)  $^{26}$  This task classifies whether the movie review is a spoiler. For example,
+
+SPOILER: "I must say that this movie was good but several things were left unsaid. For those who have seen the movie know what I am talking about but for those who haven't, I don't want to give spoilers. I was also impressed by Vin Diesel's acting skills. Overall I have to say it was a good movie filled with several twists and turns."
+
+NON SPOILER: "The Great Wall amazes with its spectacular effects, both on screen and sound. Usually I do not appreciate 3D movies, but in this case I felt like it worth it. However, being honest, the storytelling and the story itself had its weaknesses. There were many logical lapses, and for me, many details are still waiting to be answered. On the other hand, expect decent acting especially from the main characters. All in all, The Great Wall is a solid popcorn-movie, but I expected a more elaborated unfolding of the legend it tells about."
+
+News Summary/headline Topic Classification
+
+27 This task classifies the topic of the summary of a news. For example,
+
+POLITICS: "City and state officials said they received little advance warning of the decision."
+
+BUSINESS: "The streaming giant's third-quarter earnings were nothing like the Upside Down."
+
+# C Dataset Property Tags
+
+Here we list all the dataset property tags (Section 2). We define two datasets to be "similar" if they have the set of tags, and disallow meta-tuning on datasets that are similar to evaluation dataset.
+
+social media: whether the source is from social media (e.g. tweets).
+
+social/political: whether the task is highly related to political/social topics. Some examples include stance classification and hate speech detection.
+
+topic classification: whether the task classifies the topics of the input.
+
+good vs. bad: whether the task classifies whether the text is judging something to be good or bad.
+
+paper: whether input text comes from a paper.
+
+review: whether the input text is a review of a product (e.g. movie, hotel).
+
+questions: whether the input texts are questions. Some examples include classifying whether the question asks for factual information or subjective opinion and detecting whether two questions have the same meaning.
+
+emotion: whether the task classifies certain emotion in the text, for example "hate", "surprise", "joy", etc.
+
+Besides, we do not assign tags to datasets that we are confident to be different enough from other tasks (e.g. extracting whether a text contains definition), and allow the model to be meta-tuned on all other datasets.
+
+# D List of Label Descriptions
+
+Please refer to the appendix in our arXiv version: https://arxiv.org/abs/2104.04670. Somehow the acl_pubcheck software package always gives us errors.
+
+# E Robustness Checks
+
+We report all the descriptive statistics mentioned in Section 3 under 3 different types of description weighting. We additionally compare T5-small vs. T5-base, BERT-medium vs. BERT-Base and BERT-Base vs. BERT Large. All the results can be seen in Table 3, 4, and 5. Due to space constraint, we abbreviate  $\mathbb{P}[\Delta > t]$  as  $>t$  if  $t$  is positive, and  $< t$  if  $t$  is negative. Notice that, since we only have around 20 datasets to evaluate the model, most of the results presented here are not statistically significant at the dataset level; nevertheless,
+
+# E.1 Different Description Weighting
+
+We weight each label and dataset equally in Table 4 and 5. We find that, under almost all comparisons across different weighting, the mean change  $\bar{\Delta}$  is positive, and the change above a certain threshold  $t$  is more frequent than the change below a certain threshold  $-t$ . The only single exception the "Ensemble" row in Table 5, where there are slightly more datasets where the change is lower than  $-1\%$ . Nevertheless, given that the trend is still positive under  $t = 5\%$  and  $10\%$ , and two other
+
+description weightings, we may still conclude that ensembling label descriptions is more likely to improve model performance.
+
+# E.2 Larger T5 Models are Better
+
+In addition to comparing T5-Base (220 Million parameters) vs. T5-Large (770M), we also compare T5-small (60M) vs. T5-base (220M). Across all metrics, larger models are significantly better. Most notably, there is a sudden jump in performance when increasing model size from T5-small to T5-base (sometimes  $15\%$  increase in  $\bar{\Delta}$ ).
+
+# E.3 Larger BERT Models are Better
+
+We also compare different sizes of BERT (Turc et al., 2019) (41, 110, and 330M) parameters. Across all metrics, larger models are significantly better.
+
+# F Most Relevant Datasets
+
+To ensure that we are testing the models' ability to generalize to an unseen tasks, we disallow both training and testing on datasets that are too similar, which is defined as "having the same set of dataset property tags" (Section 2). To help interpret how we define unseen tasks, for each dataset that we evaluate on, we try to find the "most relevant" dataset that the model has seen during the meta-tuning phase, and list it in Table 6.
+
+# G Performance Break Down
+
+For each model, we average the AUC-ROC scores for each label description for each dataset, and report the results in Table 7.
+
+# H Accuracy
+
+<table><tr><td></td><td>Δ</td><td>&gt;1%</td><td>&lt;-1%</td><td>&gt;5%</td><td>&lt;-5%</td><td>&gt;10%</td><td>&lt;-10%</td><td>std(Δ)</td></tr><tr><td>Meta-tuned vs QA</td><td>3.3%</td><td>59.5%</td><td>28.1%</td><td>31.4%</td><td>10.3%</td><td>15.7%</td><td>5.9%</td><td>9.5%</td></tr><tr><td>220 vs 770M (T5)</td><td>6.3%</td><td>75.1%</td><td>15.1%</td><td>47.6%</td><td>2.7%</td><td>27.0%</td><td>0.5%</td><td>8.1%</td></tr><tr><td rowspan="2">Pre-trained vs. Random Ensemble</td><td>23.8%</td><td>95.7%</td><td>3.2%</td><td>91.4%</td><td>1.6%</td><td>83.2%</td><td>1.1%</td><td>14.0%</td></tr><tr><td>0.7%</td><td>28.9%</td><td>16.8%</td><td>8.7%</td><td>1.7%</td><td>1.7%</td><td>0.6%</td><td>3.1%</td></tr><tr><td>Initialized with QA</td><td>1.1%</td><td>54.1%</td><td>24.3%</td><td>24.3%</td><td>11.9%</td><td>6.5%</td><td>4.9%</td><td>6.9%</td></tr><tr><td>Train on similar</td><td>0.7%</td><td>43.8%</td><td>20.5%</td><td>6.5%</td><td>4.3%</td><td>1.6%</td><td>1.1%</td><td>3.2%</td></tr><tr><td>60 vs 220M (T5)</td><td>14.4%</td><td>86.5%</td><td>10.3%</td><td>79.5%</td><td>4.3%</td><td>61.1%</td><td>2.2%</td><td>12.6%</td></tr><tr><td>41 vs. 110M (BERT)</td><td>4.3%</td><td>65.9%</td><td>22.7%</td><td>40.0%</td><td>10.8%</td><td>20.5%</td><td>5.9%</td><td>9.1%</td></tr><tr><td>110 vs. 340M (BERT)</td><td>1.4%</td><td>46.5%</td><td>35.7%</td><td>23.8%</td><td>17.3%</td><td>11.4%</td><td>6.5%</td><td>8.5%</td></tr></table>
+
+Table 3: All results, with metrics explained in Section 3 and Appendix E. Each label description is weighted equally.  
+
+<table><tr><td></td><td>Δ</td><td>&gt;1%</td><td>&lt;-1%</td><td>&gt;5%</td><td>&lt;-5%</td><td>&gt;10%</td><td>&lt;-10%</td><td>std(Δ)</td></tr><tr><td>Meta-tuned vs QA</td><td>3.0%</td><td>57.5%</td><td>30.7%</td><td>31.3%</td><td>11.5%</td><td>16.2%</td><td>7.3%</td><td>10.2%</td></tr><tr><td>220M vs 770M (T5)</td><td>5.8%</td><td>75.8%</td><td>15.5%</td><td>46.9%</td><td>3.5%</td><td>25.6%</td><td>1.4%</td><td>7.8%</td></tr><tr><td rowspan="2">Pre-trained vs. Random Ensemble</td><td>23.7%</td><td>93.5%</td><td>5.5%</td><td>89.4%</td><td>3.4%</td><td>82.5%</td><td>2.1%</td><td>15.1%</td></tr><tr><td>0.5%</td><td>25.0%</td><td>18.8%</td><td>6.9%</td><td>1.6%</td><td>1.7%</td><td>0.7%</td><td>3.1%</td></tr><tr><td>Initialized with QA</td><td>1.2%</td><td>54.0%</td><td>24.0%</td><td>26.0%</td><td>11.8%</td><td>8.1%</td><td>5.3%</td><td>7.3%</td></tr><tr><td>Train on similar</td><td>0.7%</td><td>44.5%</td><td>20.1%</td><td>6.0%</td><td>4.3%</td><td>1.7%</td><td>0.8%</td><td>3.1%</td></tr><tr><td>60 vs 220M (T5)</td><td>15.2%</td><td>85.7%</td><td>11.4%</td><td>79.1%</td><td>3.9%</td><td>62.5%</td><td>1.9%</td><td>13.3%</td></tr><tr><td>41 vs. 110M (BERT)</td><td>4.8%</td><td>67.0%</td><td>21.5%</td><td>41.9%</td><td>9.2%</td><td>22.5%</td><td>4.9%</td><td>9.0%</td></tr><tr><td>110 vs. 340M (BERT)</td><td>1.1%</td><td>44.3%</td><td>36.3%</td><td>21.9%</td><td>18.2%</td><td>11.0%</td><td>7.3%</td><td>8.5%</td></tr></table>
+
+Table 4: All results, with metrics explained in Section 3 and Appendix E. Each label is weighted equally.  
+
+<table><tr><td></td><td>Δ</td><td>&gt;1%</td><td>&lt;-1%</td><td>&gt;5%</td><td>&lt;-5%</td><td>&gt;10%</td><td>&lt;-10%</td><td>std(Δ)</td></tr><tr><td>Meta-tuned vs QA</td><td>1.2%</td><td>55.4%</td><td>35.7%</td><td>31.2%</td><td>17.7%</td><td>15.6%</td><td>13.6%</td><td>11.2%</td></tr><tr><td>220 vs 770M (T5)</td><td>6.3%</td><td>77.4%</td><td>16.5%</td><td>51.7%</td><td>7.0%</td><td>31.6%</td><td>4.5%</td><td>9.0%</td></tr><tr><td rowspan="2">Pre-trained vs. Random Ensemble</td><td>20.2%</td><td>89.8%</td><td>8.5%</td><td>84.8%</td><td>6.1%</td><td>76.6%</td><td>1.5%</td><td>15.1%</td></tr><tr><td>0.1%</td><td>18.6%</td><td>20.2%</td><td>4.3%</td><td>1.9%</td><td>1.5%</td><td>1.2%</td><td>2.8%</td></tr><tr><td>Initialized with QA</td><td>2.3%</td><td>59.2%</td><td>22.5%</td><td>34.3%</td><td>9.9%</td><td>13.9%</td><td>5.7%</td><td>7.2%</td></tr><tr><td>Train on similar</td><td>0.6%</td><td>48.8%</td><td>25.4%</td><td>7.3%</td><td>5.7%</td><td>1.3%</td><td>0.9%</td><td>3.3%</td></tr><tr><td>60 vs 220M (T5)</td><td>12.1%</td><td>84.6%</td><td>12.9%</td><td>73.6%</td><td>3.5%</td><td>52.9%</td><td>2.2%</td><td>11.6%</td></tr><tr><td>41 vs. 110M (BERT)</td><td>7.0%</td><td>74.6%</td><td>13.8%</td><td>58.5%</td><td>6.8%</td><td>31.5%</td><td>2.9%</td><td>8.9%</td></tr><tr><td>110 vs. 340M (BERT)</td><td>1.1%</td><td>45.6%</td><td>36.1%</td><td>25.5%</td><td>18.6%</td><td>10.8%</td><td>9.3%</td><td>8.8%</td></tr></table>
+
+Table 5: All results, with metrics explained in Section 3 and Appendix E. Each dataset is weighted equally.
+
+<table><tr><td>Evaluation Dataset</td><td>Most Relevant Training Dataset</td></tr><tr><td>SemEval 2016 Task 6, stance classifications on issues like feminism, atheism, etc</td><td>SemEval 2019 Task 5, detecting hate speech against women and immigrants</td></tr><tr><td>SemEval 2019 Task 6, classifying whether the text is offensive</td><td>A dataset from Kaggle that classifies sexually explicit comments</td></tr><tr><td>SemEval 2019 Task 5, detecting hate speech against women and immigrants</td><td>SemEval 2016 Task 6, stance classifications on issues like feminism, atheism, etc</td></tr><tr><td>TREC, classifying the type the question is asking about (e.g. numbers, acronyms, human/occupations, etc)</td><td>AG News, which classifies news into different categories (e.g. sports, world events).</td></tr><tr><td>SemEval 2019 Task 8, classifying whether the question is asking for subjective opinion, factual information, or simply having a conversation</td><td>N/A</td></tr><tr><td>SUBJ, classifying whether the text contains subjective or objective information</td><td>N/A</td></tr><tr><td>QQP, classifying whether two questions have the same meaning</td><td>N/A</td></tr><tr><td>Yin et al. (2019) emotion classification, classifying text into 9 emotion types, such as “joy”, “anger”, “guilt”, “shame”, etc.</td><td>Classifying whether an IMDB movie review is positive.</td></tr><tr><td>Yin et al. (2019) situation classification, classifying which disaster situation people are experiencing, e.g. “regime change”, “crime and violence”, and what resource they need, e.g. “food and water”, “search and rescue”.</td><td>Classifying (binary) whether a tweet is related to a natural disaster.</td></tr><tr><td>Yin et al. (2019) topic classification, classifying the domain of an article into domains such as “family and relationship”, “education”, “business”, “sports”</td><td>classifying the domain of a paper abstract into physics, maths, computer sciences, and statistics.</td></tr><tr><td>AG News, which classifies news into different categories (e.g. sports, world events).</td><td>Abstract Domain classification, classifying the domain of a paper abstract into physics, maths, computer sciences, and statistics.</td></tr><tr><td>Abstract Domain classification, classifying the domain of a paper abstract into physics, maths, computer sciences, and statistics.</td><td>AG News, which classifies news into different categories (e.g. sports, world events).</td></tr><tr><td>IMDB movie reviews, classifying whether the user feels positive about the movie</td><td>Stock market sentiment, classifying whether a comment is optimistic about the market.</td></tr><tr><td>CoLA, classifying whether a sentence is grammatical</td><td>N/A</td></tr><tr><td>SemEval 2020 Task 6, classifying whether a sentence contains a definition</td><td>N/A</td></tr><tr><td>Spam classification, classifying whether a text message is a spam</td><td>click-bait classification, classifying whether the title of an article is a clickbait.</td></tr><tr><td>SemEval 2018 Task 1, classifying a tweet as one of 4 emotion types {“sadness”, “joy”, “anger”, “optimism”}</td><td>Classifying whether an IMDB movie review is positive.</td></tr><tr><td>SemEval 2018 Task 3, classifying whether a tweet is ironic</td><td>classifying whether a news title is sarcastic.</td></tr></table>
+
+Table 6: For each dataset that we evaluate on, we list the task in the training split that we consider to be the most relevant. We list "N/A" if we think that none of the training dataset is particularly relevant.
+
+<table><tr><td></td><td>QA</td><td>QA + Meta</td><td>Meta</td><td>T5 220M</td><td>BERT 340M</td></tr><tr><td>Abstract Classification</td><td>76.9%</td><td>84.3%</td><td>81.2%</td><td>68.0%</td><td>85.3%</td></tr><tr><td>AG News</td><td>76.5%</td><td>82.0%</td><td>77.8%</td><td>69.9%</td><td>69.5%</td></tr><tr><td>Stance (Hillary)</td><td>74.8%</td><td>79.8%</td><td>73.8%</td><td>69.0%</td><td>63.2%</td></tr><tr><td>Hate Speech</td><td>59.4%</td><td>66.0%</td><td>64.1%</td><td>59.6%</td><td>69.2%</td></tr><tr><td>Stance (Feminism)</td><td>67.8%</td><td>71.6%</td><td>69.1%</td><td>61.0%</td><td>64.8%</td></tr><tr><td>Stance (Climate)</td><td>75.8%</td><td>81.7%</td><td>79.6%</td><td>72.0%</td><td>76.2%</td></tr><tr><td>Emotion Classification*</td><td>67.6%</td><td>70.5%</td><td>68.0%</td><td>65.0%</td><td>64.0%</td></tr><tr><td>Emotion Classification (SemEval)</td><td>81.6%</td><td>85.2%</td><td>81.7%</td><td>76.1%</td><td>74.2%</td></tr><tr><td>Irony Detection</td><td>67.9%</td><td>83.4%</td><td>80.2%</td><td>61.0%</td><td>64.9%</td></tr><tr><td>Stance (Atheism)</td><td>60.2%</td><td>62.4%</td><td>65.6%</td><td>55.1%</td><td>60.9%</td></tr><tr><td>QQP</td><td>54.1%</td><td>61.1%</td><td>68.6%</td><td>56.7%</td><td>66.9%</td></tr><tr><td>TREC</td><td>59.3%</td><td>63.9%</td><td>76.4%</td><td>73.4%</td><td>66.9%</td></tr><tr><td>Stance (Abortion)</td><td>58.2%</td><td>61.3%</td><td>62.8%</td><td>60.5%</td><td>59.5%</td></tr><tr><td>Offensive Speech</td><td>76.6%</td><td>80.4%</td><td>79.5%</td><td>74.5%</td><td>80.6%</td></tr><tr><td>CoLA</td><td>52.3%</td><td>49.4%</td><td>49.8%</td><td>49.6%</td><td>50.0%</td></tr><tr><td>SUBJ</td><td>62.8%</td><td>66.8%</td><td>58.7%</td><td>54.5%</td><td>50.2%</td></tr><tr><td>Situation Classification*</td><td>73.9%</td><td>80.4%</td><td>79.3%</td><td>75.5%</td><td>79.5%</td></tr><tr><td>SPAM Detection</td><td>57.2%</td><td>45.4%</td><td>35.0%</td><td>49.3%</td><td>47.8%</td></tr><tr><td>IMDB Movie Review</td><td>92.9%</td><td>94.0%</td><td>90.5%</td><td>67.7%</td><td>84.4%</td></tr><tr><td>Topic Classification*</td><td>77.6%</td><td>82.7%</td><td>84.0%</td><td>77.5%</td><td>80.7%</td></tr><tr><td>Definition Detection</td><td>72.8%</td><td>73.5%</td><td>63.9%</td><td>63.6%</td><td>60.2%</td></tr><tr><td>Question Type Classification</td><td>75.1%</td><td>73.8%</td><td>59.3%</td><td>51.8%</td><td>64.5%</td></tr></table>
+
+Table 7: Zero shot performance of each model on each dataset. "QA" means the UnifiedQA model; "QA + Meta" means meta-tuning with UnifiedQA initialization; "Meta" means meta-tuning on T5 (770M) parameters. To save space, we use “*” to denote datasets from Yin et al. (2019).
+
+<table><tr><td>Dataset name</td><td>#classes</td><td>Accuracy</td></tr><tr><td>2016SemEval6TweetEvalStanceAtheism</td><td>3</td><td>66</td></tr><tr><td>KaggleNewsTopicClassification</td><td>4</td><td>64</td></tr><tr><td>2019SemEval6TweetEvalOffensive</td><td>2</td><td>28</td></tr><tr><td>2019SemEval8Qtype</td><td>2</td><td>73</td></tr><tr><td>2018SemEval3TweetEvalIrony</td><td>2</td><td>39</td></tr><tr><td>2016SemEval6TweetEvalStanceHillary</td><td>3</td><td>55</td></tr><tr><td>subj</td><td>2</td><td>61</td></tr><tr><td>trec</td><td>6</td><td>38</td></tr><tr><td>KaggleQuoraQPairs</td><td>2</td><td>50</td></tr><tr><td>definition</td><td>2</td><td>32</td></tr><tr><td>BenchmarkingZeroshotTopic</td><td>10</td><td>59</td></tr><tr><td>2019SemEval5TweetEvalHate</td><td>2</td><td>42</td></tr><tr><td>cola</td><td>2</td><td>55</td></tr><tr><td>2018SemEval1TweetEvalEmotion</td><td>4</td><td>72</td></tr><tr><td>2016SemEval6TweetEvalStanceAbortion</td><td>3</td><td>64</td></tr><tr><td>KaggleIMDBMovieReview</td><td>2</td><td>85</td></tr><tr><td>2016SemEval6TweetEvalStanceClimate</td><td>3</td><td>61</td></tr><tr><td>KaggleSMSSPAM</td><td>2</td><td>14</td></tr><tr><td>2016SemEval6TweetEvalStanceFeminist</td><td>3</td><td>53</td></tr></table>
+
+Table 8: We report the accuracy of the meta-tuned model for completeness according to the request of the reviewers. However, given that accuracy is very sensitive to thresholding (Zhao et al., 2021) and is generally unreliable when the labels are imbalanced, these numbers are not likely to be informative. Additionally, to speed up evaluation, we use a subsample of the original test split for some datasets, so these numbers are not directly comparable to those in the other papers either.

adaptinglanguagemodelsforzeroshotlearningbymetatuningondatasetandpromptcollections/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb0a37cdecf94596ad573416ee3898b74b4f3e7a4c2924e1e1119d543857b7ad
+size 1187983

adaptinglanguagemodelsforzeroshotlearningbymetatuningondatasetandpromptcollections/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b2583d0e7a90f41e34dba07cc3c596c818603183a5233c7c681a38f15773af5
+size 735459

adeepdecomposablemodelfordisentanglingsyntaxandsemanticsinsentencerepresentation/d6393320-f49d-4224-bd0d-01754e8793e5_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77708641e704dd2ac67b3aca1e6f45fbf23be5228c0ff4be578cb16d26d19d25
+size 89636

adeepdecomposablemodelfordisentanglingsyntaxandsemanticsinsentencerepresentation/d6393320-f49d-4224-bd0d-01754e8793e5_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a97a9e2a7cc9a822a84493088e321b09a1bf2ae3634f5a3577c561238f2117db
+size 104944

adeepdecomposablemodelfordisentanglingsyntaxandsemanticsinsentencerepresentation/d6393320-f49d-4224-bd0d-01754e8793e5_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7236abecf447306a4e2a51a24044d9f8f4d1331d9a10799967e9c7ea1e9e3c91
+size 690562

adeepdecomposablemodelfordisentanglingsyntaxandsemanticsinsentencerepresentation/full.md

@@ -0,0 +1,366 @@
+# A Deep Decomposable Model for Disentangling Syntax and Semantics in Sentence Representation
+
+Dingcheng Li, Hongliang Fei, Shaogang Ren, Ping Li
+
+Cognitive Computing Lab (CCL)
+
+Baidu Research USA
+
+10900 NE 8th St. Bellevue, WA 98004, USA
+
+{lidingcheng,hongliangfei,shaogangren,liping11}@baidu.com
+
+# Abstract
+
+Recently, disentanglement based on a generative adversarial network or a variational autoencoder has significantly advanced the performance of diverse applications in CV and NLP domains. Nevertheless, those models still work on coarse levels in the disentanglement of closely related properties, such as syntax and semantics in human languages. This paper introduces a deep decomposable model based on VAE to disentangle syntax and semantics by using total correlation penalties on KL divergences. Notably, we decompose the KL divergence term of the original VAE so that the generated latent variables can be separated in a more clear-cut and interpretable way. Experiments on benchmark datasets show that our proposed model can significantly improve the disentanglement quality between syntactic and semantic representations for semantic similarity tasks and syntactic similarity tasks.
+
+# 1 Introduction
+
+Recently, disentangled representations have significantly advanced the performance of several applications in NLP. For example, disentanglement has been used to separating representation of attributes such as sentiment from contents (Fu et al., 2018; John et al., 2019), understanding subtleties in component modeling (Esmaeili et al., 2019), detecting anomalies (Hou et al., 2021), and learning sentence representations that split the syntax and the semantics (Ju et al., 2021). They are also used to boost text generation (Iyyer et al., 2018; Jain et al., 2018) or calculating the semantic or syntactic similarity between sentences (Chen et al., 2018).
+
+In this paper, we focus on the task of separating syntax and semantics in sentence representation learning. Unlike previous supervised approaches that usually resort to syntactic parsers to handle syntax processing, our approach separates syntactic and semantic variables by disentangling hidden
+
+states of deep neural nets in a self-learning and unsupervised fashion.
+
+The first work focusing on the separation of syntax and semantics from hidden variables is Chen et al. (2019). They proposed a deep generative model based on VAE with two latent variables to represent syntax and semantics. The generative model comprises von Mises Fisher (vMF) and Gaussian priors on the semantic and syntactic latent variables, and a deep BOW decoder conditioning on these latent variables. Following previous work, they train this model by optimizing the Evidence Lower Bound (ELBO) with a VAE-like (Kingma and Welling, 2014) objective.
+
+However, their approach still generates a rough decomposition and thus may fail to disentangle syntax and semantics at a finer granularity. To address this weakness, we propose a decomposable variational autoencoder (DecVAE) to allow hidden variables factorizable. From a modeling perspective, factorizable representations with statistically independent variables usually obtained in an unsupervised or semi-supervised manner can distill information into a compact form, which is semantically useful for downstream tasks. From an application perspective, different words or phrases in sentences represent various entities with variant roles. It is necessary to utilize decomposable latent variables to capture a variety of entities with different semantic meanings.
+
+Towards building a finer-grained disentanglement, motivated by FactorVAE (Kim and Mnih, 2018), we extend the work in Chen et al. (2019) and use total correlation (Watanabe, 1960) (TC) as a penalty term to obtain a deeper and meaningful factorization of syntactic and semantic latent variables. To make TC more discriminative, we also integrate multi-head attention into this framework. DecVAE can identify and cluster hierarchically independent semantic components in natural language text, which exhibits hierarchical linguistic
+
+structure (Sanh et al., 2019), and the corresponding syntax and semantics interact with each other. For experiments, we evaluate learned semantic representations on the SemEval semantic textual similarity (STS) tasks. Following the protocol in Chen et al. (2019), we predict the syntactic structure of an unseen sentence to be the one similar to its nearest neighbor, determined by the latent syntactic representation in a large dataset of annotated sentences. Experiments show that DecVAE achieves the best performance on all tasks when learned representations are mostly disentangled.
+
+Contributions. Firstly, we propose a generic DecVAE to disentangle semantics and syntax based on the total correlation of KL divergence. Secondly, DecVAE is also integrated with a multi-head attention network to cluster embedding vectors so that corresponding word embeddings are more discriminative. Thirdly, results after integrating DecVAE in disentangling syntax from semantics achieve SOTA performances, confirming DecVAE's effectiveness.
+
+# 2 Background and Related Work
+
+# 2.1 VAEs for Disentanglement
+
+The variational autoencoder (VAE) (Kingma and Welling, 2014) is a latent variable model that pairs a top-down generator with a bottom-up inference network. Different from traditional maximum-likelihood estimation (MLE) approach, VAE training is done by evidence lower bound (ELBO) optimization in order to overcome the intractability of posterior. Basically, the objective function of VAE is represented as:
+
+$$
+\mathbb {E} _ {\mathbf {z} \sim q (\mathbf {Z} | \mathbf {X})} [ \log p (\mathbf {X} | \mathbf {Z}) ] - \beta \mathbf {K L} (q (\mathbf {Z} | \mathbf {X}) | | p (\mathbf {Z}))
+$$
+
+When  $\beta = 1$ , this is the standard VAE. When  $\beta > 1$ , it becomes  $\beta$ -VAE (Higgins et al., 2017), which attempts to learn a disentangled representation by optimizing a heavily penalized objective.
+
+Vanilla VAEs cannot disentangle latent variables. PixelGAN Autoencoders (Makhzani and Frey, 2017) further break down the  $KL$  term as:
+
+$$
+\mathbf {K L} (q (\mathbf {Z} | \mathbf {X}) | | p (\mathbf {Z})) = I (\mathbf {X}; \mathbf {Z}) + \mathbf {K L} (q (\mathbf {Z}) | | p (\mathbf {Z})) \tag {1}
+$$
+
+where  $I(x;z)$  is the mutual information under the joint distribution  $p(x)q(z|x)$ . Penalizing the  $KL(q(z)||p(z))$  term pushes  $q(z)$  towards the factorial prior  $p(z)$ , encouraging independence in the dimensions of  $z$  and thus disentangling.
+
+Alternatively, FactorVAE approaches this problem with total correlation penalty (Kim and Mnih, 2018), which we adopt for our work. FactorVAE achieves similar disentangling results while preserving good quality of reconstruction by augmenting the vanilla VAE objective with a term directly encouraging independence in the code distribution:
+
+$$
+\begin{array}{l} \mathbb {E} _ {\mathbf {z} \sim q (\mathbf {Z} | \mathbf {X})} [ \log p (\mathbf {X} | \mathbf {Z}) ] - \mathbf {K L} (q (\mathbf {Z} | \mathbf {X}) | | p (\mathbf {Z})) \\ - \gamma \mathbf {K L} (q (\mathbf {Z}) | | \bar {q} (\mathbf {Z})) \\ \end{array}
+$$
+
+where  $\bar{q} (\mathbf{z})\coloneqq \prod_{j = 1}^{K}q(z_{j})$  . The FactorVAE's objective is also a lower bound on the marginal log likelihood  $\mathbb{E}_p[\log p(\mathbf{X})]$  .KL(q(Z)||q(Z)) is known as "Total Correlation" (TC) (Watanabe, 1960), a popular measure of dependence for multiple random variables.
+
+# 2.2 Disentanglement in NLP
+
+Disentanglement in NLP has strong connections with LDA (Blei et al., 2003; Blei and Lafferty, 2006). In particular, neural topic models, that use belief networks (Mnih and Gregor, 2014; Li et al., 2019b) or enforce the Dirichlet prior via Gaussian or Wassertein autoencoders (Nan et al., 2019; Li et al., 2018), associate topic learning to disentanglement with component analysis. Later on, seq2seq VAE represent disentangled topics via continuous representations (Dieng et al., 2017; Ding et al., 2018; Bowman et al., 2016; Yang et al., 2017). Srivastava and Sutton (2017) combines LDA and VAE for topic detection and Pergola et al. (2021) proposes to consider latent topics as generative factors to be disentangled to improve discriminative power of topics.
+
+Meanwhile, a growing amount of work start to explore neural learning disentangled/component representations to diverse NLP tasks. For example, we see such applications in sentiment analysis and style transfer (Hu et al., 2017; Li et al., 2019a), morphological reinfectionoon (Zhou and Neubig, 2017), semantic parsing (Yin et al., 2018), text generation (Wiseman et al., 2018), sequential labeling (Chen et al., 2018), text-based variational autoencoder (Miao et al., 2016), etc.
+
+Although much work has been done on grammatical and semantic analysis, there are few explorations on disentangling syntax and semantics. The disentanglement between syntax and semantics is quite challenging since they are heavily entangled. Except under some circumstances where there are no ambiguities, such as some unique proper names,
+
+it is usually difficult to find absolute borderlines among words, phrases, or entities.
+
+The work of VGVAE (Chen et al., 2019) is the latest one quite relevant to our work, wherein they assume that a sentence is generated by conditioning on two independent latent variables: semantic variable  $\mathbf{z}_{sem}$  and syntactic variable  $\mathbf{z}_{syn}$ . For inference, they assume a factored posterior is produced and a lower bound on marginal log-likelihood is maximized in the generative process. The corresponding inference and generative models are two independent word averaging encoders with additional linear feed-forward neural networks and a feed-forward neural network with the output being a bag of words or an RNN.
+
+Compared with their work, we aim to construct a more generic work by deploying the decomposability of KL divergence, thus discovering more subtle components from latent variables. Consequently, the VAE framework can do better disentanglement with more fine-grained decomposed parts. Further, we can flexibly add regularities to guide the decomposition to generate more interpretable and controllable elements from decoders.
+
+# 3 Proposed Approach
+
+In this work, we are developing a generative model named Decomposable VAE (DecVAE). Although our proposed approach is applicable to any disentangled tasks in NLP, we focus on disentangling semantic and syntactic information from sentence representations. We extend VGVAE model (Chen et al., 2019) to incorporate the total correlation as a penalty term to enable latent variable factorization.
+
+# 3.1 Decomposable VAE
+
+Our model is essentially based on VAE, namely, composed of a term of computing loglikelihood of input data given latent variables, and terms of computing KL divergences between posterior variational probabilities of hidden variables given input data and the prior probabilities of hidden variables.
+
+Let  $x_{1}, \ldots, x_{T}$  be a sequence of  $T$  tokens (words), conditioned on a continuous latent variable  $\mathbf{z}$ . As a usual practice, for example, like the assumption in Latent Dirichlet Allocations (LDA) (Blei et al., 2003), we have a conditional independence assumption of words on  $\mathbf{z}$ :
+
+$$
+p _ {\theta} (x _ {1}, \dots , x _ {T}) = \int \prod_ {t = 1} ^ {T} p _ {\theta} (x _ {t} | \mathbf {z}) p _ {\theta} (\mathbf {z}) d \mathbf {z}
+$$
+
+Model parameters  $\theta$  can be learned via the variational lower-bound (Kingma and Welling, 2014)
+
+$$
+\begin{array}{l} \mathcal {L} (\theta , \phi ; \mathbf {X}) \geq \frac {1}{T} \sum_ {t = 1} ^ {T} \left(\mathbb {E} _ {\mathbf {z} \sim q _ {\phi}} [ \log p _ {\theta} (x _ {t} | \mathbf {z}) ] \right. \tag {2} \\ - \mathbf {K L} \left(q _ {\phi} (\mathbf {z} | x _ {t}) \| p _ {\theta} (\mathbf {z}))\right) \\ \end{array}
+$$
+
+where  $q_{\phi}(\mathbf{z}|x_t)$  is the encoder (recognition model or inference model), parameterized by  $\phi$ , i.e., the approximation to true posterior  $p_{\theta}(\mathbf{z}[x_t])$ . The distribution  $p_{\theta}(\mathbf{z})$  is the prior for  $\mathbf{z}$ .
+
+As studied in Sanh et al. (2019), natural languages can be regarded as a manifold, since it is hierarchically organized, and the corresponding syntax and the semantics interact in an intricate space. Based on the observation that different words or phrases in sentences represent different entities with different roles, either grammatical or semantic, and potentially interact with each other, we guide the generations of latent variables in the VAE corresponding to entities in sentences by designing a VAE with decomposable latent variables. Hence our proposed DecVAE can identify hierarchically independent components from natural languages. Furthermore, the reconstruction network may generate words or phrases sequentially.
+
+DecVAE will learn a decoder that maps the latent space  $\mathcal{Z}$  (learned by the encoder from input samples) to this language manifold  $\mathcal{X}$ . Let  $\mathbf{Z} = [\mathbf{z}^1, \dots, \mathbf{z}^K] \in \mathcal{Z}$  be the latent variable of the decoder and  $\mathbf{z}^k$  to represent the  $k$ -th component of the latent variables. In addition, we also add a  $\mathbf{z}_0$  to each  $\mathbf{z}^k$ , a special latent variable to encodes the overall properties of the generated sentences and the correlations between different grammatical and semantic components. Let  $(\bar{\mathbf{x}}, \bar{\mathbf{f}}) = [(\bar{\mathbf{x}}^1, \bar{\mathbf{f}}^1), \dots, (\bar{\mathbf{x}}^K, \bar{\mathbf{f}}^K)]$  be the variables for the output of the decoder (each element is a tuple composed of the generated token index in the vocabulary and its component index), where  $\mathbf{z}^k$  controls the properties of  $k$ -th component  $\bar{\mathbf{x}}^k$ .
+
+Firstly, we assume that the components are conditionally independent with each other given the latent variables, i.e.,
+
+$$
+\left(\bar {\mathbf {x}} ^ {i}, \bar {\mathbf {f}} ^ {i}\right) \perp \left(\bar {\mathbf {x}} ^ {j}, \bar {\mathbf {f}} ^ {j}\right) | \mathbf {Z}, \text {i f} i \neq j.
+$$
+
+We also have the following independent assumption about the components and latent variables,
+
+$$
+\left(\bar {\mathbf {x}} ^ {i}, \bar {\mathbf {f}} ^ {i}\right) \perp \mathbf {z} ^ {j} | \mathbf {z} _ {0} ^ {j}, \text {i f} i \neq j. \tag {3}
+$$
+
+![](images/b9639aa62ce901e69091628b442a424e9c69ffe368e00bac6eb5ceabe64817d0.jpg)  
+Figure 1: The proposed model consists of four layers. From bottom to top, they are embedding layer, multi-head attention layer, encoder, and decoder. Different from the usual network structure, the first three layers comprise three parallel independent layers, one for semantic and one for syntax. The attention layers yield  $K$ -dim attention weights  $\mathbf{f}$ , so that ensemble of  $K$  weighted embeddings are working in both semantic and syntax encoders.
+
+Let  $\bar{\mathbf{y}} = (\bar{\mathbf{x}},\bar{\mathbf{f}})$  and each  $\bar{\mathbf{y}}^k = (\bar{\mathbf{x}}^k,\bar{\mathbf{f}}^k)$ . We have the following distributions for generated tokens:
+
+$$
+\begin{array}{l} p (\bar {\mathbf {y}} | \mathbf {z}) = p (\bar {\mathbf {y}} ^ {1}, \dots , \bar {\mathbf {y}} ^ {K} | \mathbf {z} _ {0}, \mathbf {z} ^ {1}, \dots , \mathbf {z} ^ {K}) \\ = \prod_ {k = 1} ^ {K} p (\bar {\mathbf {y}} ^ {k} | \mathbf {z} _ {0} ^ {k}, \mathbf {z} ^ {1}, \dots , \mathbf {z} ^ {K}) = \prod_ {k = 1} ^ {K} p (\bar {\mathbf {y}} ^ {k} | \mathbf {z} _ {0} ^ {k}, \mathbf {z} ^ {k}) \\ \end{array}
+$$
+
+This model attempts to encode each component's individual features ( tokens, words, or phrases) and the global latent factors for the sentence.
+
+# 3.2 Objective Function
+
+We propose to decompose the two terms of calculating KL divergence following Eq. (1). Meanwhile, along the thread of our proposed DecVAE, we add the global controller variable  $\mathbf{z}_0$ . This design shares some similarities with the component segmentation in computer vision, such as MONet (Burgess et al., 2019). MONet shows that an attention network layer improves component segmentation as well as component disentanglement, in which a variable,  $f$ , the representation of the attention, is deployed there. Taking these into consideration, our model is defined as following. Let  $\mathbf{z}_{syn} = [\mathbf{z}_{syn}^1,\dots ,\mathbf{z}_{syn}^K]$  be the syntactic latent variable, we define an equation for syntax based on the decomposable nature of latent variables as:
+
+$$
+\begin{array}{l} \mathbf {K L} (q _ {\phi} (\mathbf {z} _ {s y n} ^ {k} | \mathbf {x}) | | p _ {\theta} (\mathbf {z} _ {s y n} ^ {k})) = I _ {q _ {\phi}} (\mathbf {x}, \mathbf {f} ^ {k}; \mathbf {z} _ {s y n} ^ {k}, \mathbf {z} _ {0} ^ {k}) \\ + \sum_ {i, j} \left[ \mathbf {K L} \left(q \left(\mathbf {z} _ {\text {s y n}} ^ {k _ {i}}, \mathbf {z} _ {0} ^ {k _ {j}}\right) \right\rvert   \right. \left. p \left(\mathbf {z} _ {\text {s y n}} ^ {k _ {i}}, \mathbf {z} _ {0} ^ {k _ {j}}\right)\right) \tag {4} \\ + \beta \mathbf {K L} (q _ {\phi} (\mathbf {z} _ {s y n} ^ {k}, \mathbf {z} _ {0} ^ {k}) | | \prod_ {i} q _ {\phi} (\mathbf {z} _ {s y n} ^ {k _ {i}}) \prod_ {j} q _ {\phi} (\mathbf {z} _ {0} ^ {k _ {j}})) ] \\ \end{array}
+$$
+
+and a similar equation for semantics as
+
+$$
+\begin{array}{l} \mathbf {K L} (q _ {\phi} (\mathbf {z} _ {s e m} ^ {k} | \mathbf {x}) | | p _ {\theta} (\mathbf {z} _ {s e m} ^ {k})) = I _ {q _ {\phi}} (\mathbf {x}, \mathbf {f} ^ {k}; \mathbf {z} _ {s e m} ^ {k}, \mathbf {z} _ {0} ^ {k}) \\ + \sum_ {i, j} \mathbf {K L} \left(q \left(\mathbf {z} _ {s e m} ^ {k _ {i}}, \mathbf {z} _ {0} ^ {k _ {j}}\right) \| p \left(\mathbf {z} _ {s e m} ^ {k _ {i}}, \mathbf {z} _ {0} ^ {k _ {j}}\right)\right) \tag {5} \\ + \beta {\bf K L} (q _ {\phi} ({\bf z} _ {s e m} ^ {k}, {\bf z} _ {0} ^ {k}) | | \prod_ {i} q _ {\phi} ({\bf z} _ {s e m} ^ {k _ {i}}) \prod_ {j} q _ {\phi} ({\bf z} _ {0} ^ {k _ {j}})) ], \\ \end{array}
+$$
+
+where  $i, j$  refer to indices of tokens and  $\mathbf{z}_{*}^{k_{i}}, * \in \{\text{sem}, \text{syn}, 0\}$  indicates the latent variable value at the  $i$ -th token. In Eq. (4) and Eq. (5), the second and third terms are derived from minimization of total correlations as in Esmaeili et al. (2019); Jeong and Song (2019). The second term decomposes each hidden vector of syntax and semantics into smaller categories in a hierarchical fashion so that we can have more subtle disentanglements of each syntactic or semantic components.
+
+The third term in Eq. (4) and Eq. (5) is derived from the standard equation of total correlation,
+
+$$
+T C (\mathbf {z} ^ {\mathbf {k}}) = \mathbb {E} \Big [ \log \big (\frac {q _ {\phi} (\mathbf {z} ^ {\mathbf {k}})}{\prod_ {i} q _ {\phi} (\mathbf {z} _ {i} ^ {k})} \big) \Big ] = K L (q _ {\phi} (\mathbf {z} ^ {k}) | | \prod_ {d} q _ {\phi} (\mathbf {z} _ {i} ^ {k}))
+$$
+
+Namely, we deploy this technique to penalize the total correlation (TC) for enforcing disentanglement of the latent factors. To compute the second term, we use the weighted version for estimating the distribution value of  $q(\mathbf{z})$ .
+
+# 3.3 The Network Structure
+
+With the above derivations as our basis, we construct our network structure as shown in Figure 1. From bottom to top, the input sentences are con
+
+verted to embedding vectors. Meanwhile, there is a mask input with each mask  $m_{k}$  showing whether each word or phrase  $x_{t}$  appears in each sentence. Outputs from this layer are fed to a multi-head attention layer to generate attention weights  $f_{t}$ . Following-up is the dot product between the embedding of  $x_{t}$  and its attention weight  $f_{t}$ .
+
+Since we are modeling both semantics and syntax of input sentences, the attention procedure is processed twice with different initialization. The results are passed into the semantic encoder and syntax encoder, respectively. Each encoder yields their hidden variables,  $(\mathbf{z}_{semt}^{1\dots k},\mathbf{z}_{0t}^{1\dots k})$  and  $(\mathbf{z}_{synt}^{1\dots k},\mathbf{z}_{0t}^{1\dots k})$ . A similar idea is implemented in recent work from computer vision domain (CV), MONet (Burgess et al., 2019). Differently, in their work,  $f_{k}$  is generated sequentially with an attention network while we generate attention all at once with multi-head attention, which is proven successful in the transformer model (Vaswani et al., 2017).
+
+To incorporate recurrent neural networks for decoding, we take a similar structure described in SNAIL (Mishra et al., 2018). Namely, the self-attention mechanism from the transformer is combined with a temporal convolution. Next, the element-wise multiplication of embedding vector and focus masks generate hidden vectors, which are fed into semantic encoder and syntax encoder respectively to be encoded as a pair of variables  $(\mathbf{z}^k,\mathbf{z}_0^k)$ . The two groups of hidden component vectors are concatenated into the decoder. We obtain the reconstructed words/phrases  $\bar{\mathbf{x}}$ , and their component distribution  $\bar{\mathbf{f}}^k$ , similar to a component assignment and consistent to the weights  $\mathbf{f}^k$ .
+
+# 3.4 Multi-task Training and Inference
+
+With the product of embedding vector  $\mathbf{emb}_t$  and their corresponding focus mask  $\mathbf{m}_t$  as the encoder's input,  $(\mathbf{z}^k,\mathbf{z}_0^k)$  as the latent variable and  $(\bar{\mathbf{x}},\bar{\mathbf{m}}^{k})$  as the output of the decoder, the loss for component  $k$  is given by
+
+$$
+\begin{array}{l} \Psi_ {k} (\mathbf {x}, \mathbf {f} ^ {k}; \theta , \phi , a, e, d) \tag {6} \\ = - \mathbb {E} _ {q _ {\phi} ^ {e} (\mathbf {z} ^ {k}, \mathbf {z} _ {0} ^ {k} | \mathbf {x}, \mathbf {f} ^ {k})} \left[ \mathbf {f} ^ {k} \log p _ {\theta} ^ {d} (\mathbf {x} | \mathbf {z} ^ {k}, \mathbf {z} _ {0} ^ {k}) \right] \\ + \operatorname {K L} \left(q _ {\phi} ^ {e} \left(\mathbf {z} ^ {k}, \mathbf {z} _ {0} ^ {k} | \mathbf {x}, \mathbf {f} ^ {k}\right) | | p \left(\mathbf {z} ^ {k}, \mathbf {z} _ {0} ^ {k}\right)\right) \\ + \gamma \mathbf {K L} (q _ {\phi} ^ {a} (\mathbf {f} ^ {k} | \mathbf {x}) | | p _ {\theta} ^ {d} (\bar {\mathbf {f}} ^ {k} | \mathbf {z} ^ {k}, \mathbf {z} _ {0} ^ {k})) \\ \end{array}
+$$
+
+Here  $a, e$  and  $d$  refer to multi-head attention layer, encoder and decoder layer respectively,  $\theta$  and  $\phi$  are parameters for the likelihood and variational distribution respectively, the local hidden variable
+
+![](images/40177697261b30b0675d17444719ddfac3b056f53ecfba82ab41c3f4199be3af.jpg)  
+Figure 2: Diagram of the training process for auxiliary losses: discriminative paraphrase loss (DPL; dashed lines) and paraphrase reconstruction loss (PRL; dash-dotted lines). Different from Chen et al. (2019), each input of encoders consists of embeddings of the sentence  $\mathbf{x}_t$  and their component distributions,  $\mathbf{f}_t^{1\dots k}$ . Each output of encoders consists of hidden variables  $\mathbf{z}_{sem_t}^{1\dots k}$  and  $\mathbf{z}_{0_t}^{1\dots k}$ . Each output of decoders consists of predicted embeddings of each sentence  $\bar{x}_t$  and their predicted component distributions,  $\bar{\mathbf{f}}_t$ .
+
+$\mathbf{z}^k = [\mathbf{z}_{sem}^k, \mathbf{z}_{syn}^k]$  and the global hidden variable  $\mathbf{z}_0^k = [\mathbf{z}_{sem(0)}^k, \mathbf{z}_{syn(0)}^k]$  and  $\gamma \geq 0$  is a hyperparameter. The overall loss is
+
+$$
+\mathcal {L} _ {\mathrm {V A E}} (\mathbf {x}; a, e, d) = \sum_ {k = 1} ^ {K} \Psi_ {k} (\mathbf {x}, \mathbf {f} ^ {k}; \theta , \phi , a, e, d).
+$$
+
+Loss Function Components. As seen from Eq. (6), our loss function is composed of three parts, which can be realized by our objective functions described in Eq. (4) and Eq. (5). Furthermore, following the success of multi-task training in Chen et al. (2019), we introduce three auxiliary objectives: paraphrase reconstruction loss (PRL), discriminative paraphrase loss (DPL) and word position loss (WPL). The purpose is to encourage  $\mathbf{z}_{sym}$  to better capture semantic information and  $\mathbf{z}_{syn}$  to better capture syntactic information.
+
+Paraphrase Reconstruction Loss Function. As shown in Figure 2, we swap the semantic variables, keep the syntactic variables and attempt to reconstruct the sentences. We model sentences with paraphrase relationships  $\mathbf{x}_1$  and  $\mathbf{x}_2$  to be generated with the same semantic latent variables. The basic assumption is still that semantic information is equivalent between a paraphrase pair. But differently, our PRL involve more variables, including the common latent factor  $\mathbf{z}_0$  and the focus mask variables  $\mathbf{f}_k$ . Therefore, our PRL is defined as,
+
+$$
+\begin{array}{l} \mathbb {E} _ {\substack {\mathbf {z} _ {s e m (2)} \sim q _ {\phi} ^ {\epsilon} (\mathbf {1}) \\ \mathbf {z} _ {s y n (1)} \sim q _ {\phi} ^ {\epsilon} (\mathbf {2})}} \left[ - \log p _ {\theta} ^ {d} (\bar {\mathbf {x}} _ {1} | (\mathbf {z} _ {s e m (2)}, \mathbf {z} _ {0 (2)}), (\mathbf {z} _ {s y n (1)}, \mathbf {z} _ {0 (1)}) \right] + \\ \begin{array}{r l} & {\mathbb {E} _ {\mathbf {z} _ {s e m (1)} \sim q _ {\phi} ^ {e} (\mathbf {3})} \left[ - \log p _ {\theta} ^ {d} (\overline {{\mathbf {x}}} _ {2} | (\mathbf {z} _ {s e m (1)}, \mathbf {z} _ {0 (1)}), (\mathbf {z} _ {s y n (2)}, \mathbf {z} _ {0 (2)}) \right]} \\ & {\mathbf {z} _ {s y n (2)} \sim q _ {\phi} ^ {e} (\mathbf {4})} \end{array} \\ \end{array}
+$$
+
+where
+
+$$
+\begin{array}{r} q _ {\phi} ^ {e} (\mathbf {1}) = q _ {\phi} ^ {e} ((\mathbf {z}, \mathbf {z} _ {0}) _ {s e m} | \overline {{\mathbf {x}}} _ {2}, \overline {{\mathbf {f}}} _ {2}), q _ {\phi} ^ {e} (\mathbf {2}) = q _ {\phi} ^ {e} ((\mathbf {z}, \mathbf {z} _ {0}) _ {s y n} | \overline {{\mathbf {x}}} _ {1}, \overline {{\mathbf {f}}} _ {1}), \\ q _ {\phi} ^ {e} (\mathbf {3}) = q _ {\phi} ^ {e} ((\mathbf {z}, \mathbf {z} _ {0}) _ {s e m} | \overline {{\mathbf {x}}} _ {1}, \overline {{\mathbf {f}}} _ {1}), q _ {\phi} ^ {e} (\mathbf {4}) = q _ {\phi} ^ {e} (\mathbf {z}, \mathbf {z} _ {0}) _ {s y n} | \overline {{\mathbf {x}}} _ {2}, \overline {{\mathbf {f}}} _ {2}). \end{array}
+$$
+
+Discriminative Paraphrase Loss. The Discriminative Paraphrase Loss (DPL) attempts to learn to encourage sentences with paraphrase relationships to have higher similarities while those without such relationships to have lower similarities. Because paraphrase relationship is defined in the sense of semantic similarity, we only calculate it with samples from vMF distributions. The loss is defined as,
+
+$$
+\begin{array}{l} m a x (0, \delta - d i s t (x _ {1}, x _ {2})) + d i s t (x _ {1}, n _ {1})) + \\ m a x (0, \delta - d i s t (x _ {1}, x _ {2})) + d i s t (x _ {2}, n _ {2})) \end{array}
+$$
+
+where  $dist$  refers to the distance,  $x_{1}$  and  $x_{2}$  are sentences with paraphrase relationship, while  $x_{1}$  and  $n_1$  are those without paraphrase relationships. The similarity function is the cosine similarity between the mean directions of the semantic variables across  $K$  components from the two sentences:
+
+$$
+d i s t \left(x _ {1}, x _ {2}\right) = c o s i n e \left(\mu \left(x _ {1}\right), \mu \left(x _ {2}\right)\right)
+$$
+
+where  $\mu (x_{i}) = (\mathbf{z}_{sem(i)}^{1\dots K}\odot \mathbf{z}_{0(i)}^{1\dots K})$  and  $\odot$  is the element-wise product.
+
+Word Position Loss. Following Chen et al. (2019), we keep a word position loss (WPL) to guide the representation learning of the syntactic variable. For both word averaging encoders and LSTM encoders, we parameterize WPL with a three-layer feedforward neural network  $f(\cdot)$ . The concatenation of the samples of the syntactic variables  $\mathbf{z}_{syn}$  and the embedding vector  $\mathbf{emb}_i$  at the word position  $i$  form the input for the network. In the decoder stage, the position representation at position  $i$  is predicted as a one-hot vector. The corresponding equation is defined as,
+
+$$
+W P L = \mathbb {E} _ {z _ {s y n} \sim q _ {\phi} (z | x)} \bigg [ \sum_ {i} \log [ (f ([ e _ {i}; z _ {s y n} ]) _ {i}) ] \bigg ]
+$$
+
+where  $(\cdot)_i$  is the probability of position  $i$ .
+
+Inference Model for Word Averaging. In
+
+our framework, syntax and semantics encoders  $q_{\phi}^{e}(\mathbf{z}_{syn}|\mathbf{x})$  and  $q_{\phi}^{e}(\mathbf{z}_{sem}|\mathbf{x})$  follow different fashions with different sampling strategies with additional linear feedforward neural network. However, both use word averaging to obtain the mean vector,  $\mu (\mathbf{x})$  and the standard deviation vector,  $\sigma (\mathbf{x})$
+
+In the decoding stage, we generate a bag of words given  $\mathbf{z}_{syn}$  and  $\mathbf{z}_{sem}$  by the posterior probability  $p_{\theta}^{d}(\mathbf{x}|\mathbf{z}_{syn},\mathbf{z}_{sem})$ . Note that the decoding output is a tuple of vectors, which includes both word index and their component probability distribution. The expected output log-probability is computed as follows:
+
+$$
+\begin{array}{l} \underset { \begin{array}{c} \mathbf {z} _ {s e m} \sim q _ {\phi} ^ {e} (\mathbf {z} _ {\mathbf {s e m}} | \mathbf {x}) \\ \mathbf {z} _ {s y n} \sim q _ {\phi} ^ {e} (\mathbf {z} _ {\mathbf {s y n}} | \mathbf {x}) \end{array} } {\mathbb {E}} \left[ \log p _ {\theta} ^ {d} (\mathbf {x} | \mathbf {z} _ {s e m}, \mathbf {z} _ {s y n}) \right] = \\ \underset { \begin{array}{c} \mathbf {z} _ {s e m} \sim q _ {\phi} ^ {e} (\mathbf {z} _ {\mathbf {s e m}} | \mathbf {x}) \\ \mathbf {z} _ {s y n} \sim q _ {\phi} ^ {e} (\mathbf {z} _ {\mathbf {s y n}} | \mathbf {x}) \end{array} } {\mathbb {E}} \left[ \sum_ {t = 1} ^ {T} \log \frac {\exp f _ {\theta} ([ \mathbf {z} _ {s e m} ; \mathbf {z} _ {s y n} ]) _ {x _ {t}}}{\sum_ {v = 1} ^ {V} \exp f _ {\theta} ([ \mathbf {z} _ {s e m} ; \mathbf {z} _ {s y n} ]) _ {v}} \right] \\ \end{array}
+$$
+
+where  $V$  is the vocabulary size,  $[\cdot ]$  indicates concatenation,  $T$  is the sentence length and  $x_{t}$  is the index of the  $t$ 'th word's word type.  $f_{\theta}([{\bf z}_{sem};{\bf z}_{syn}]$  is a feedforward neural network with outputs being a bag of words.
+
+# Inference Model for BLSTM Averaging
+
+Similarly, we compute the expected output log-probability of generated words, including their component information for BLSTM as follows,
+
+$$
+\begin{array}{l} \underset { \begin{array}{c} \mathbf {z} _ {s e m} \sim q _ {\phi} ^ {e} (\mathbf {z} _ {\mathbf {s e m}} | \mathbf {x}) \\ \mathbf {z} _ {s y n} \sim q _ {\phi} ^ {e} (\mathbf {z} _ {\mathbf {s y n}} | \mathbf {x}) \end{array} } {\mathbb {E}} \left[ \log p _ {\theta} ^ {d} (\mathbf {x} | \mathbf {z} _ {s e m}, \mathbf {z} _ {s y n}) \right] = \\ \mathop{\mathbb{E}}_{\substack{\mathbf{z}_{sem}\sim q_{\phi}^{e}(\mathbf{z}_{sem}|\mathbf{x})\\ \mathbf{z}_{syn}\sim q_{\phi}^{e}(\mathbf{z}_{syn}|\mathbf{x})}}\big[\sum_{w = 1}^{S}\log p_{\theta}\big(x_{w}|\mathbf{z}_{syn},\mathbf{z}_{sem},\mathbf{x}_{1:s - 1}|)\big)\big] \\ \end{array}
+$$
+
+The inference model  $q_{\phi}^{e}(\mathbf{z}_{sem})$  is still a word averaging encoder while  $q_{\phi}^{e}(\mathbf{z}_{syn})$  is parameterized by a bidirectional LSTM, where the forward and backward hidden states are concatenated together and then the average is taken. The averages are used as input for a feedforward network with one hidden layer to produce both mean vector  $\mu (\mathbf{x})$  and  $\sigma (\mathbf{x})$ .
+
+Since both the inference model of word averaging and BLSTM are interacting with the decomposed KL divergence or total correlations through backpropagation, our inference and the generative models can obtain more factorized component information. Hence, the generated tokens are more consistent between syntax and semantics.
+
+# 4 Experiments
+
+Following Chen et al. (2019), we sampled 50M paraphrase pairs from ParaNMT-50M (Wieting and Gimpel, 2018) as our training set. We use the SemEval semantic textual similarity (STS) task 2017 (Cer et al., 2017) as the development set. The STS task and its benchmark as the test set for similarity evaluation. The implementation was based on the PaddlePaddle deep learning platform.
+
+# 4.1 Experiment Setup
+
+We set the dimension of hidden variables and word embedding to 50, which speeds up experiments and provides a competitive performance over a wide range. To have a fair comparison, we also tune  $\gamma$ , the weights for PRL and reconstruction loss from 0.1 to 1 in increments of 0.1 based on the development set performance. We set  $\gamma = 0.2$  with the best validation results. One sample from each latent variable is utilized during training. When evaluating DecVAE based models on STS tasks, the mean direction of the semantic variable is used. In contrast, the mean vector of the syntactic variable is used in syntactic similarity tasks. The total correlations are also mainly applied to syntactic tasks since we find that applying total correlations to vMF distribution makes the model too complicated. Hence, we simplify the framework with only  $KL$  divergence of attentions calculated against the semantic components for current work.
+
+# 4.2 Baselines
+
+We compare with word averaging (WORD<sub>AVG</sub>) and bidirectional LSTM averaging (BLSTM<sub>AVG</sub>) of VGVAE model (Chen et al., 2019; Wieting and Gimpel, 2018). In particular, WORD<sub>AVG</sub> takes the average over word embeddings in the input sequence to obtain the sentence representation. BLSTM<sub>AVG</sub> uses the average hidden states of a bidirectional LSTM as the sentence representation, where forward and backward hidden states are concatenated.
+
+# 4.3 Semantic Similarity Evaluations
+
+Table 1 presents the semantic similarity evaluations. Specifically, the upper rows tell us how they can model similarity when trained on paraphrases (Wieting and Gimpel, 2018) and the lower half rows show remarkable differences between semantic and syntactic metrics. It is worth noting that in Chen et al. (2019), they also reported semantic modeling results for several pretrained embeddings, in which
+
+<table><tr><td rowspan="2">methods</td><td colspan="2">semantic var. %</td><td colspan="2">syntactic var. %</td></tr><tr><td>bm</td><td>avg</td><td>bm</td><td>avg</td></tr><tr><td>VGVAE WORDAVG</td><td>71.9</td><td>64.8</td><td>-</td><td>-</td></tr><tr><td>VGVAE BLSTMAVG</td><td>71.4</td><td>64.4</td><td>-</td><td>-</td></tr><tr><td>DecVAE WORDAVG</td><td>72.4</td><td>65.1</td><td>-</td><td>-</td></tr><tr><td>DecVAE BLSTTMAVG</td><td>71.4</td><td>63.2</td><td>-</td><td>-</td></tr><tr><td>VGVAE ALL+LSTM enc</td><td>72.2</td><td>65.1</td><td>16.6</td><td>24.3</td></tr><tr><td>VGVAE ALL+LSTM e&amp;d</td><td>72.8</td><td>65.3</td><td>11.5</td><td>19.9</td></tr><tr><td>DecVAE+WPL</td><td>52.3</td><td>45.3</td><td>31.4</td><td>33.2</td></tr><tr><td>DecVAE+DPL</td><td>63.5</td><td>57.6</td><td>35.9</td><td>37.5</td></tr><tr><td>DecVAE+PRL</td><td>65.6</td><td>59.2</td><td>28.9</td><td>33.1</td></tr><tr><td>DecVAE+PRL+WPL</td><td>69.9</td><td>62.9</td><td>24.4</td><td>28.2</td></tr><tr><td>DecVAE+PRL+DPL</td><td>67.5</td><td>62.3</td><td>34.1</td><td>32.8</td></tr><tr><td>DecVAE+DPL+WPL</td><td>69.9</td><td>65.4</td><td>19.9</td><td>24.2</td></tr><tr><td>DecVAE+ALL+WORDAVG e&amp;d</td><td>73.9</td><td>64.0</td><td>22.3</td><td>17.7</td></tr><tr><td>DecVAE ALL+LSTM enc</td><td>70.0</td><td>62.1</td><td>14.7</td><td>16.5</td></tr><tr><td>DecVAE ALL+LSTM e&amp;d</td><td>72.2</td><td>65.7</td><td>8.1</td><td>9.7</td></tr></table>
+
+Table 1: Pearson correlation  $(\%)$  for STS test sets. bm: STS test set. avg: the average of Pearson correlation for each domain in the test set from 2012 to 2016. Results are in bold if they are highest in the "semantic variable" columns or lowest in the "syntactic variable" columns. "ALL" indicates all of the multi-task losses are used. "e&d" means "enc & dec". The results are averaged over five repetitions and the standard deviation is around  $0.1\% - 0.2\%$  for all methods.
+
+they showed that all pretrained embeddings are far lower than those of VGVAE based models. Such a result implies that VAE-based modeling can capture semantics quite well no matter what variations we make. For simplicity, we do not show the results from pretrained embeddings herein. Readers please refer to Chen et al. (2019) for more details.
+
+As shown in the upper rows of Table 1, DecVAE+WORD $_{AVG}$  achieves the best semantic score for both STS avg metric and STS bm metric. LSTM-based models do not show advantages over Word $_{AVG}$  as VGVAE (Chen et al., 2019). So average of LSTM outputs for decomposed VAE is not as effective as vanilla VAE based approaches.
+
+The lower rows in Table 1 show whether semantic variables can better capture semantic information than syntactic variables. We reproduced VGVAE's result by their released package (Chen et al., 2019) for comparisons and our results are lines from 3 to 11. As shown, the semantic and syntactic variables of the base DecVAE model show similar performances on the STS test sets. With more losses added, the performance of these two variables gradually diverges, indicating that different information is captured in the two variables. Therefore, we can see that the various losses play essential roles in the disentanglement of semantics and syntax in DecVAE. When all losses plus  $Word_{AVGe\&d}$  are fully utilized, the high
+
+est benchmark results (73.91%) are obtained with 1.7% higher than VGVAE for semantic variables. Meanwhile, all losses plus  $LSTM_{e\&d}$  achieves the best average results for semantic variables. More impressively, this approach yields relatively low scores for both benchmarks and average of syntactic variables (8.05 and 9.72 for bm and avg respectively). This fully shows that decomposition with total correlation has excellent disentanglement capacity on semantics and syntax.
+
+Finally, Figure 3 plots the performance curves of our models and baselines as the length of the target sentence increases. We observe a similar trend, i.e., the longer the sentence, the worse the performance. Our framework is close to the top (red) curve and has a more consistent trend. This shows that DecVAE achieves more remarkable disentanglement effects in syntax. Particularly, in Table 1, the full model with LSTM encoder and decoder achieves much lower values for syntactic evaluations than all other models.
+
+![](images/6f7524635d715e8dd10ef9f75d88cd1daff18984394e42f570248bdaa06e7d8d.jpg)  
+Figure 3: Constituency parsing F1 scores (left) and POS tagging accuracy (right) by sentence length, for 1-nearest neighbor parsers based on semantic and syntactic variables, as well as a random baseline and an oracle nearest neighbor parser ("Best"). Note that in the legend, "  $+\mathrm{LSTM}$ " means "  $+\mathrm{LSTM}$  enc & dec".
+
+![](images/bda4dc9399726f802d7d8009780b8da23ff72b82dee1dde550bf20ca55108355.jpg)
+
+# 4.4 Syntactic Similarity Evaluation
+
+Following the evaluation protocol in VGVAE (Chen et al., 2019), we utilize syntactic variables to calculate nearest neighbors for a 1-nearest-neighbor syntactic parser or POS tagger. Several metrics are employed to quantify the quality of the parser's output and tagging sequences. It is worth noting that this evaluation does not directly compare parsing accuracy. Instead, similar to the semantic similarity, it demonstrates syntactic variables' ability to capture more syntactic information than semantic variables.
+
+We report labeled F1 of constituent parsing and accuracy of POS tagging in Table 2. First, we evaluate VGVAE and DecVAE with word averaging encoder and BLSTM encoder in the upper table.
+
+<table><tr><td></td><td colspan="2">Constituent Parsing (F1, ↑).</td><td colspan="2">POS Tagging (% Acc., ↑).</td></tr><tr><td>VGVAE WORDAVG</td><td colspan="2">25.5</td><td colspan="2">21.4</td></tr><tr><td>VGVAE BLSTMAVG</td><td colspan="2">25.7</td><td colspan="2">21.6</td></tr><tr><td>DecVAE WORDAVG</td><td colspan="2">27.8</td><td colspan="2">24.9</td></tr><tr><td>DecVAE BLSTMAVG</td><td colspan="2">29.9</td><td colspan="2">33.2</td></tr><tr><td></td><td>semV.</td><td>synV.</td><td>semV.</td><td>synV.</td></tr><tr><td>VGVAE All</td><td>25.4</td><td>29.3</td><td>21.4</td><td>25.5</td></tr><tr><td>VGVAE+LSTM enc. &amp; dec.</td><td>25.3</td><td>38.8</td><td>21.4</td><td>35.7</td></tr><tr><td>DecVAE All</td><td>24.9</td><td>33.7</td><td>20.4</td><td>29.8</td></tr><tr><td>DecVAE+LSTM enc.</td><td>24.5</td><td>36.9</td><td>21.4</td><td>35.5</td></tr><tr><td>DecVAE+LSTM enc. &amp; dec.</td><td>23.2</td><td>41.5</td><td>19.4</td><td>38.9</td></tr></table>
+
+Table 2: Syntactic similarity evaluations, labeled F1 score for constituent parsing, and accuracy  $(\%)$  for part-of-speech tagging. Numbers are bold if they are worst in the "semantic variable" column or best in the "syntactic variable" column. "ALL" indicates all of the multi-task losses are used. The results are collected and averaged over five rounds and the standard deviation is around  $0.1\% -0.2\%$  for all methods.
+
+DecVAE outperforms VGVAE in both parsing and tagging. For the lower part, in contrast to semantic similarity, syntactic variables are expected to boost both tasks while semantic variables worsen them. The baseline "VGVAE All" initially have similar results for two variables. Then, with the addition of LSTM encoder and decoder, expected performances appear along. For our method, the gaps between both variables are more remarkable than VGVAE, although not always worst for semantic variables and best for syntactic variables. Such a result indicates that DecVAE achieved a good disentanglement of syntax and semantics. In particular, our full combination with LSTM achieves the best results and outperforms those of SOTA.
+
+Another observation is that although both VGVAE and DecVAE do not perform well compared with their LSTM counterparts, "DecVAE All" still obtains better performances than VGVAE. We believe that it is the total correlation that brings more accurate disentanglement effects. Nonetheless, the syntactic evaluation results, in general, are not so evident as the semantic correspondents.
+
+# 4.5 Qualitative Analysis with Case Studies
+
+We conduct a qualitative evaluation of latent variables via cosine similarity for nearest neighbor sentences and words to test set examples in terms of both the semantic and syntactic representations. The results are reported in Table 3 and Table 4.
+
+# 4.5.1 Lexical Analysis
+
+Table 3 shows word nearest neighbors for both semantic and syntactic representations and exhibits
+
+<table><tr><td>Query Words</td><td>Retrieved Words</td></tr><tr><td>exact</td><td>semantic: indeed, current, completely, absolutely, context, clear, strictly, similarly, ec, proper syntactic: soap, benefit, license, orn, discontinuation, wed, jin, applications, girls, lucian</td></tr><tr><td>command</td><td>semantic: guidance, result, ec, direction, accept, ordering, release, transmission, order syntactic: problem, root, eleven, sex, jinglge, francis, sale, trains, sixteen, industrial</td></tr><tr><td>requesting</td><td>semantic: note, guidance, inquires, inception, accepted, needs, claims, query, required, application syntactic: terminate, subscribe, particle, composite, locate, require, claim, compose, apply, inquiring</td></tr><tr><td>emptying</td><td>semantic: changing, reset, stuffed, withdrawn, outline, modified, remove, boo, restoring, threads syntactic: entering, obtained, subtotal, living, combine, surged, dismissed, composed, applying, inquiring</td></tr><tr><td>smallest</td><td>semantic: minor, mi, smaller, diffuse, events, types, fragments, size, short, weighing syntactic: biggest, odd, stable, concerned, small, hotter, hottest, shorter, fragmentary</td></tr></table>
+
+Table 3: Examples of most similar words to particular query words in terms of the semantic or syntactic variable  
+
+<table><tr><td>Query Sentence</td><td>Semantically Similar</td><td>Syntactically Similar</td></tr><tr><td>go, you fools, Xar bellowed</td><td>the hell, you say, Alekseyv bellowed</td><td>Huh, I&#x27;ve got file festivals to enter he said.</td></tr><tr><td>Do you think I could do what she did?</td><td>Do you think that I&#x27;d do it like that?</td><td>So, do you know who&#x27;s there?</td></tr><tr><td>His head must be right between the two cuts.</td><td>He is already getting in your head right now.</td><td>My mom even basked a cake for the party.</td></tr><tr><td>I&#x27;ll tell you things can change a lot.</td><td>When the siatuation changes, we&#x27;ll let you know.</td><td>I&#x27;d like to try the state government again.</td></tr><tr><td>They say, you do not have a face.</td><td>In fact, you&#x27;s just a pretty face.</td><td>You don&#x27;t know what is in that building</td></tr><tr><td>I even found a rare gouda on the internet.</td><td>I&#x27;ve seen a lot on the internet.</td><td>Did you get your degree off a cereal box?</td></tr><tr><td>I don&#x27;t know, he was wearing socks.</td><td>you got any socks you do not want wear.</td><td>you don&#x27;t play piano, I hope.</td></tr><tr><td>I love you as much as before.</td><td>I love you more than I ever loved anyone.</td><td>but wait. There&#x27;s as much as what is.</td></tr><tr><td>You know what, cal, just pull over.</td><td>cal, is trying to pull you out.</td><td>You know, you guys got some competition out there?</td></tr><tr><td>Yeah, he got punched out in court earlier.</td><td>From there she was taken to court and back.</td><td>He would have to be forged by Jupityer himself.</td></tr></table>
+
+Table 4: Examples of most similar sentences to particular query sentences in terms of the semantic or syntactic variable.
+
+clear patterns. Among the five query words, retrieved words based on semantics have similar meanings against them, while those based on syntax share part-of-speeches. For example, for the query word, exact, almost all words in the semantic row have the sense of exactness. Likewise, most of the words in the second row, semantically, have the sense of order, as the query word, command. In contrast, the syntactic part has POS as NN. For the third row, semantically, they mostly have an association with require while syntactically, they are all verbs.
+
+# 4.5.2 Sentential Analysis
+
+Table 4 demonstrates sentences of semantically and syntactic similar respectively in column 2 and column 3. Like the lexical similarity, retrieved sentences in column 2 have similar meanings or similar keywords or key phrases to query sentences while they may be different in sentence structure. For example, "bellowed", "Do you think", "head", "change", "internet", "love", "pull" and "court" are in the rows from one to ten respectively.
+
+In contrast, those that are syntactically similar may have different meanings while they have similar grammatical patterns. Take a few rows as examples, "go, you fools, Xar bellowed" does have similar syntactic construction to "Huh, I've got file festivals to enter he said". Likewise, the second row, the query is composed of yes/no questions
+
+with an object clause for both query and syntactically similar sentence.
+
+# 4.6 Discussions
+
+The above results show the disentanglement effects of our proposed DecVAE from semantic and syntactic evaluations in both quantitative and qualitative perspectives. In comparing with baselines, it is not hard to see that DecVAE demonstrates more impressive disentanglement powers. Such results confirm our assumption that a more finetuned decomposition of KL divergences can detect more subtle aspects of semantics and syntax. This discovery can shed light on constructing more representative learning strategies for languages in both token and sentence levels.
+
+# 5 Conclusion
+
+We propose DecVAE, a framework to disentangle syntax and semantics in a sentence. It extends the original VAE so that the latent variables can be separated in more interpretable way. Experiments show that DecVAE achieves better results in semantic and syntax similarity than that of SOTA. One future direction is fine-grained representation learning for words and sentences, which is essential for many downstream applications such as controllable text generation. Besides, continual and interactive feature distillation may help improve more discriminate disentanglement (Wang et al., 2021).
+
+# References
+
+David M. Blei and John D. Lafferty. 2006. Dynamic topic models. In Proceedings of the Twenty-Third International Conference on Machine Learning (ICML), pages 113-120, Pittsburgh, PA.  
+David M. Blei, Andrew Y. Ng, and Michael I. Jordan. 2003. Latent dirichlet allocation. J. Mach. Learn. Res., 3:993-1022.  
+Samuel R. Bowman, Luke Vilnis, Oriol Vinyals, Andrew M. Dai, Rafal Jozefowicz, and Samy Bengio. 2016. Generating sentences from a continuous space. In Proceedings of the 20th SIGNLL Conference on Computational Natural Language Learning (CoNLL), pages 10-21, Berlin, Germany.  
+Christopher P. Burgess, Loic Matthews, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matthew Botvinick, and Alexander Lerchner. 2019. MONet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390.  
+Daniel M. Cer, Mona T. Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. 2017. Semeval-2017 task 1: Semantic textual similarity multilingual and crosslingual focused evaluation. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval@ACL), pages 1-14, Vancouver, Canada.  
+Mingda Chen, Qingming Tang, Karen Livescu, and Kevin Gimpel. 2018. Variational sequential labelers for semi-supervised learning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 215-226, Brussels, Belgium.  
+Mingda Chen, Qingming Tang, Sam Wiseman, and Kevin Gimpel. 2019. A multi-task approach for disentangling syntax and semantics in sentence representations. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pages 2453-2464, Minneapolis, MN.  
+Adji B. Dieng, Chong Wang, Jianfeng Gao, and John W. Paisley. 2017. Topicnn: A recurrent neural network with long-range semantic dependency. In Proceedings of the 5th International Conference on Learning Representations (ICLR), Toulon, France.  
+Ran Ding, Ramesh Nallapati, and Bing Xiang. 2018. Coherence-aware neural topic modeling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNL), pages 830-836, Brussels, Belgium.  
+Babak Esmaeili, Hao Wu, Sarthak Jain, Alican Bozkurt, N. Siddharth, Brooks Paige, Dana H. Brooks, Jennifer G. Dy, and Jan-Willem van de Meent. 2019. Structured disentangled representations. In Proceedings of the 22nd International Conference on Artificial Intelligence and Statistics (AISTATS), pages 2525-2534, Naha, Okinawa, Japan.
+
+Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, and Rui Yan. 2018. Style transfer in text: Exploration and evaluation. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence (AAAI), pages 663–670, New Orleans, LA.  
+Irina Higgins, Loic Matthews, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. 2017. beta-vae: Learning basic visual concepts with a constrained variational framework. In Proceedings of the 5th International Conference on Learning Representations (ICLR), Toulon, France.  
+Shifu Hou, Yujie Fan, Mingxuan Ju, Yanfang Ye, Wenqiang Wan, Kui Wang, Yinming Mei, Qi Xiong, and Fudong Shao. 2021. Disentangled representation learning in heterogeneous information network for large-scale android malware detection in the COVID-19 era and beyond. In Proceedings of the Thirty-Fifth AAAI Conference on Artificial Intelligence (AAAI), pages 7754-7761, Virtual Event.  
+Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P. Xing. 2017. Toward controlled generation of text. In Proceedings of the 34th International Conference on Machine Learning (ICML), pages 1587-1596, Sydney, Australia.  
+Mohit Iyyer, John Wieting, Kevin Gimpel, and Luke Zettlemoyer. 2018. Adversarial example generation with syntactically controlled paraphrase networks. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pages 1875-1885, New Orleans, LA.  
+Sarthak Jain, Edward Banner, Jan-Willem van de Meent, Iain James Marshall, and Byron C. Wallace. 2018. Learning disentangled representations of texts with application to biomedical abstracts. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4683-4693, Brussels, Belgium.  
+Yeonwoo Jeong and Hyun Oh Song. 2019. Learning discrete and continuous factors of data via alternating disentanglement. In Proceedings of the 36th International Conference on Machine Learning (ICML), pages 3091-3099, Long Beach, CA.  
+Vineet John, Lili Mou, Hareesh Bahuleyan, and Olga Vechtomova. 2019. Disentangled representation learning for non-parallel text style transfer. In Proceedings of the 57th Conference of the Association for Computational Linguistics (ACL), pages 424–434, Florence, Italy.  
+Mingxuan Ju, Wei Song, Shiyu Sun, Yanfang Ye, Yu-jie Fan, Shifu Hou, Kenneth A. Loparo, and Liang Zhao. 2021. Dr.emotion: Disentangled representation learning for emotion analysis on social media to improve community resilience in the COVID-19 era and beyond. In Proceedings of the Web Conference (WWW), Virtual Event / Ljubljana, Slovenia.
+
+Hyunjik Kim and Andriy Mnih. 2018. Disentangling by factorising. In Proceedings of the 35th International Conference on Machine Learning (ICML), pages 2654-2663, Stockholm, Sweden.  
+Diederik P. Kingma and Max Welling. 2014. Autoencoding variational bayes. In Proceedings of the 2nd International Conference on Learning Representations (ICLR), Banff, Canada.  
+Dingcheng Li, Siamak Zamani, Jingyuan Zhang, and Ping Li. 2019a. Integration of knowledge graph embedding into topic modeling with hierarchical dirichlet process. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), Minneapolis, MN.  
+Dingcheng Li, Jingyuan Zhang, and Ping Li. 2018. Representation learning for question classification via topic sparse autoencoder and entity embedding. In Proceedings of the IEEE International Conference on Big Data (Big Data), Seattle, WA.  
+Dingcheng Li, Jingyuan Zhang, and Ping Li. 2019b. TMSA: A mutual learning model for topic discovery and word embedding. In Proceedings of the 2019 SIAM International Conference on Data Mining (SDM), pages 684-692, Calgary, Alberta, Canada.  
+Alireza Makhzani and Brendan J. Frey. 2017. Pixelgan autoencoders. In Advances in Neural Information Processing Systems (NIPS), pages 1975-1985, Long Beach, CA.  
+Yishu Miao, Lei Yu, and Phil Blunsom. 2016. Neural variational inference for text processing. In Proceedings of the 33nd International Conference on Machine Learning (ICML), pages 1727-1736, New York City, NY.  
+Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. 2018. A simple neural attentive metal-learner. In Proceedings of the 6th International Conference on Learning Representations (ICLR), Vancouver, Canada.  
+Andriy Mnih and Karol Gregor. 2014. Neural variational inference and learning in belief networks. In Proceedings of the 31th International Conference on Machine Learning (ICML), pages 1791-1799, Beijing, China.  
+Feng Nan, Ran Ding, Ramesh Nallapati, and Bing Xiang. 2019. Topic modeling with Wasserstein autoencoders. In Proceedings of the 57th Conference of the Association for Computational Linguistics (ACL), pages 6345-6381, Florence, Italy.  
+Gabriele Pergola, Lin Gui, and Yulan He. 2021. A disentangled adversarial neural topic model for separating opinions from plots in user reviews. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pages 2870-2883, Online.
+
+Victor Sanh, Thomas Wolf, and Sebastian Ruder. 2019. A hierarchical multi-task approach for learning embeddings from semantic tasks. In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence (AAAI), pages 6949-6956, Honolulu, HI.  
+Akash Srivastava and Charles Sutton. 2017. Autoencoding variational inference for topic models. In Proceedings the 5th International Conference on Learning Representations (ICLR), Toulon, France.  
+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 (NIPS), pages 5998-6008, Long Beach, CA.  
+Yigong Wang, Zhuoyi Wang, Yu Lin, Latifur Khan, and Dingcheng Li. 2021. CIFDM: continual and interactive feature distillation for multi-label stream learning. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), pages 2121-2125, Virtual Event, Canada.  
+Michael Satosi Watanabe. 1960. Information theoretical analysis of multivariate correlation. IBM J. Res. Dev., 4(1):66-82.  
+John Wieting and Kevin Gimpel. 2018. *Paranmt-50m: Pushing the limits of paraphrastic sentence embeddings with millions of machine translations*. In *Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL)*, pages 451–462, Melbourne, Australia.  
+Sam Wiseman, Stuart M. Shieber, and Alexander M. Rush. 2018. Learning neural templates for text generation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3174-3187, Brussels, Belgium.  
+Zichao Yang, Zhiting Hu, Ruslan Salakhutdinov, and Taylor Berg-Kirkpatrick. 2017. Improved variational autoencoders for text modeling using dilated convolutions. In Proceedings of the 34th International Conference on Machine Learning (ICML), pages 3881-3890, Sydney, Australia.  
+Qingyu Yin, Yu Zhang, Weinan Zhang, Ting Liu, and William Yang Wang. 2018. Deep reinforcement learning for chinese zero pronoun resolution. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL), pages 569-578, Melbourne, Australia.  
+Chunting Zhou and Graham Neubig. 2017. Multi-space variational encoder-decoders for semisupervised labeled sequence transduction. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (ACL), pages 310-320, Vancouver, Canada.

adeepdecomposablemodelfordisentanglingsyntaxandsemanticsinsentencerepresentation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23b87b7c970af1738b6d48ffa4a713aa41b73f10910eab3da5bf285803aadbb1
+size 614243

adeepdecomposablemodelfordisentanglingsyntaxandsemanticsinsentencerepresentation/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e86be7fdfd4f2e2be9c6a45b3c04232c3de9b7abe203f38b91e1b802ee1cf438
+size 416238

adiscourseawaregraphneuralnetworkforemotionrecognitioninmultipartyconversation/fdf53b35-9f3c-469b-be70-c892094ec86b_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2713c160e3773492b3273c27eb30218390b1f706b54e75cc7386655cee30e047
+size 72961

adiscourseawaregraphneuralnetworkforemotionrecognitioninmultipartyconversation/fdf53b35-9f3c-469b-be70-c892094ec86b_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd185721807a1ec2f428ae9aa5f7f2ca82377d340f309e5acdd752a959c4a053
+size 87962