Add Batch 0d964477-6c06-48ba-974a-2d50656fed51

doc2dialagoalorienteddocumentgroundeddialoguedataset/79df7f58-acea-4529-bbed-016a158f27ae_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:771af63023ed4219c544cffcfb5bbf49cdacef51070e7421ed43e6844eaf2cd0
+size 74315

doc2dialagoalorienteddocumentgroundeddialoguedataset/79df7f58-acea-4529-bbed-016a158f27ae_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0651ee406e00ccefb5756ca66498bafbc7ac0e74f96affcfc37858ecdc9a11db
+size 87204

doc2dialagoalorienteddocumentgroundeddialoguedataset/79df7f58-acea-4529-bbed-016a158f27ae_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58bbeedf8163006ec4d0917dc2901ddca5d0541e72338b3707a5c1c73177f7eb
+size 698853

doc2dialagoalorienteddocumentgroundeddialoguedataset/full.md

@@ -0,0 +1,273 @@
+# doc2dial: A Goal-Oriented Document-Grounded Dialogue Dataset
+
+# Song Feng Hui Wan Chulaka Gunasekara Siva Sankalp Patel Sachindra Joshi Luis A. Lastras
+
+IBM Research AI
+
+{sfeng@us, hwan@us, chulaka.gunasekara@}.ibm.com {siva.sankalp.patel@, jsachind@in, lastrasl@us}.ibm.com
+
+# Abstract
+
+We introduce doc2dial, a new dataset of goal-oriented dialogues that are grounded in the associated documents. Inspired by how the authors compose documents for guiding end users, we first construct dialogue flows based on the content elements that corresponds to higher-level relations across text sections as well as lower-level relations between discourse units within a section. Then we present these dialogue flows to crowd contributors to create conversational utterances. The dataset includes over 4500 annotated conversations with an average of 14 turns that are grounded in over 450 documents from four domains. Compared to the prior document-grounded dialogue datasets, this dataset covers a variety of dialogue scenes in information-seeking conversations. For evaluating the versatility of the dataset, we introduce multiple dialogue modeling tasks and present baseline approaches.
+
+# 1 Introduction
+
+The task of reading documents and responding to queries has been the trigger of many recent research advances. On top of the development of contextual question answering QuAC (Choi et al., 2018) and CoQA (Reddy et al., 2019), more recent work MANtIS (Penha et al., 2019) and DoQA (Campos et al., 2020) included more kinds of user intents for querying over documents; while ShARC (Saeidi et al., 2018) added follow-up questions from agents and binary answers from users for the inference over documents. These exciting works confirm the importance of modeling document-grounded dialogue. Yet, it involves more complex scenes in practice, which requires better understanding of the inter-relations between conversations and documents. Thus, we aim to investigate how to create the training instances to further approach real-world applications of document-grounded dialogue for information seeking tasks.
+
+In this work, we propose a new dataset of goal-oriented document-grounded dialogue. Figure 1 shows sample utterances from dialogues D1, D2 and D3 between an assisting agent and a user, and an example document in the middle. D1 and D2 are grounded in the given document, while D3 is irrelevant to the document. It illustrates two different types of contexts that we aim to capture: (1) dialogue-based context, where a query could be formed by a single or multiple turns, and (2) document-based context, which corresponds to varied forms of knowledge represented in the document. More specifically, dialogue-based context of a query could be initiated by a user (e.g., U1 in D1) or an agent (e.g., A3 in D1), and carried out through multiple turns by both roles (e.g., all turns in D2). Document-based context could involve structural elements in documents, such as the headers T1 and T2 or list items of m1 and m2, as well as textual discourse units, such as clauses (e.g., "If your clothing has been damaged").
+
+For creating such dataset, we consider the document contents for social welfare websites, such as ssa.gov and va.gov, which guide users to access various forms of information. We develop a pipeline approach for dialogue data construction. Inspired by how human authors compose user-facing web content, we utilize both the high-level hierarchical relations between document components, as well as the low-level semantic relations between discourse units (Stede et al., 2019) to dynamically create outlines of dialogues, or we call dialogue flows. A dialogue flow is a sequence of interactions between an assisting agent and a user. Each turn contains a dialogue scene that is defined by a dialogue act, a role (user or agent) and a piece of grounding content from a document. Then we present these dialogue flows to crowd contributors to create conversational utterances. Such approach helps to avoid additional noise from the post-hoc
+
+![](images/098291ec61f752aa9d5bd28cba7eacc4df69f009d7f61c8b813bbe20e1da20be.jpg)  
+Figure 1: Sample segments of conversations (D1, D2 and D3) with various dialogue scenes that are grounded in a webpage (middle) from va.gov. The relevant content elements, such as hierarchical headers, list-items and spans, are highlighted. A / U indicates Agent / User role.
+
+human annotations of dialogue data, which is a known challenge (Geertzen and Bunt, 2009).
+
+The dataset contains about 4500 annotated conversations with an average of 14 turns per dialogue. The utterances are grounded in over 450 documents from four domains. Unlike the previous work on document-grounded question answering or dialogues (Choi et al., 2018; Reddy et al., 2019; Saeidi et al., 2018) that are based on a short text snippet, our dialogues are grounded in a much wider span of context in the associated documents.
+
+For evaluation, we propose three tasks that are related to identifying and generating responses with grounding content in documents: (1) user utterance understanding; (2) agent response generation; and (3) relevant document identification. For each task, we present baseline approaches and evaluation results. Our goal is to elicit further research efforts on building document-grounded dialogue models that can incorporate deeper contexts for tackling goal-oriented information-seeking tasks. We summarize our main contributions as follows:
+
+- We introduce a novel dataset for modeling dialogues that are grounded in documents from multiple domains. The dataset is available at http://doc2dial.github.io/.  
+- We develop a pipeline approach for dialogue data collection, which has been adapted and evaluated for varied domains.  
+- We propose multiple dialogue modeling tasks that are supported by our dataset, and present the baseline approaches.
+
+# 2 Doc2Dial
+
+We introduce doc2dial, a new dataset that includes (1) a set of documents; and (2) conversations between an assisting agent and an end user, which are grounded in the associated documents. Figure 1 presents sample utterances from different dialogues along with a sample document from va.gov in the middle. It illustrates some prominent features in our dataset, such as the cases where a conversation involves multiple interconnected sub-tasks under a general inquiry (e.g., D1); or the cases where a conversation involves multiple interactions to verify the conditional contexts for one query (e.g., D2).
+
+Recent work, such as Saeidi et al. (2018), has started to address the challenge of modeling complex contexts by allowing follow-up questions from agents based on natural language inference rules extracted from the relevant documents. However, it also simplified the task by using only restricted forms of questions and binary answers. In our work, we not only encourage free-form utterances, but also aim to include various dialogue scenes that provoke inquires with different document-based and dialog-based contexts. A user query can be formed in single-turn or multiple-turn manners: (1) the user explicitly states a context that is associated with a text-span that contains a solution to the query, e.g., U5 on T4; (2) the user describes an implicitly stated context associated with a solution, e.g., U7; (3) the user accepts or rejects a piece of agent-stated context that is associated with a solution, e.g., U4 (rejection), and U12 & U14 (acceptance). An agent response, on the other hand, either
+
+![](images/9159025b87272d3d9245e122bab0550b8a38dad73deb75b0c6a372608431c4a8.jpg)  
+Figure 2: The overview of the process for constructing and annotating doc2dial dataset.
+
+provides a solution or poses a query depending on the context of a given user query: (1) whether the query is irrelevant to the grounding document, e.g., A17; (2) whether the query is under-specified, if so, the agent will suggest associated context, e.g., A11 and A13; (3) whether a relevant answer is identified in the grounding document, e.g, A6, A8 and A15.
+
+# 2.1 Data Collection
+
+For collecting document-grounded dialogue data, we propose a pipeline approach derived from the framework proposed by Feng et al. (2020). As shown in Figure 2, it includes the components for: (1) processing the document contents; (2) generating dynamic dialogue flows; (3) crowdsourcing the dialogue utterances.
+
+# 2.1.1 Data Construction Approach
+
+Processing document contents We first select documents that contain the context-indicative elements, such as hierarchical headers and explicit discourse relations (Prasad et al., 2008, 2019), since those document contents could provoke more diversified dialogue flows. Then we extract text-spans to create a graph with the spans as nodes and semantic relations as edges. Some spans in the graph correspond to a piece of information for solving user problems, while some correspond to the conditional context of those solutions, such as SP2 and SP1 in Figure 1 respectively. The semantic relations are largely determined by the heuristics derived from the document structures (Mukherjee et al., 2003) and semantic connectives (Das et al., 2018) between discourse units or clauses. Both spans and semantic relations are labeled automatically via our tool. The labels can be reviewed and annotated via crowdsourcing platforms, which is also supported by our tool.
+
+Generating dynamic dialogue flows Each flow consists of a sequence of dialogue scenes. A dialogue scene is described with (1) role, either a user or an agent; (2) a selected span as the grounding content from the given document; (3) a dialogue
+
+act that determines how to describe the selected span in the given role. Thus, each turn is inherently annotated with the dialogue act and a reference to the document contents. The dynamics of the dialogue flows are introduced by varying the three factors that are constrained by the relations from the semantic graph and dialogue history. In principle, we randomly select content from a candidate pool of spans of conditional contexts and solutions. The pool is updated after every turn is generated based on the status of the previously selected span. The general rule for updating the candidate pool is to avoid re-selecting any spans with an established status. In addition, the dialogue flow is principally aligned with common practice of dialogue management, for instance, after an agent asks a user a question, we expect the next turn would be the user answering the question.
+
+Collecting human utterances Finally, we present the sequences of dialogue scenes to crowdsourced contributors to convert them into conversational utterances.
+
+# 2.1.2 Crowdsourcing Setup
+
+Our data collection task asks the crowd contributors to focus on one turn at a time so that they can carefully review the given dialogue scene and the dialogue history. Since the crowd generally prefers to work on tasks in batches, we try different settings to combine the tasks: (1) each writer plays the same role but for different dialogues per batch; or (2) each writer plays both agent and user role and completes entire dialogue in order, as inspired by Byrne et al. (2019). We also find that the conversations by the second setting tend to be more coherent and less time consuming. Many writers would make efforts to differentiate their writing styles for different roles. Therefore, our tasks were completed based on the second setting by about 70 qualified contributors from appen.com. We pay $1.5-$2 per conversation.
+
+# 2.2 Document Data
+
+For document contents, we consider the public government service websites that are designated to provide information to a vast group of users. We collect web contents from four domains and select about 450 documents for creating dialogue flows as shown in Table 1. Our dataset provides document contents in plain text and HTML, along with the meta information of titles and URLs. Each docu
+
+<table><tr><td rowspan="2">Domain</td><td rowspan="2">#Dials</td><td rowspan="2">#Docs</td><td colspan="4"># per doc</td></tr><tr><td>tk</td><td>sp</td><td>p</td><td>sec</td></tr><tr><td>ssa.gov</td><td>860</td><td>86</td><td>758</td><td>66</td><td>16</td><td>5</td></tr><tr><td>va.gov</td><td>1340</td><td>138</td><td>823</td><td>70</td><td>20</td><td>9</td></tr><tr><td>dmv.gov</td><td>1420</td><td>149</td><td>955</td><td>77</td><td>18</td><td>10</td></tr><tr><td>cdc.gov</td><td>850</td><td>85</td><td>1251</td><td>94</td><td>16</td><td>9</td></tr><tr><td>all</td><td>4470</td><td>458</td><td>947</td><td>77</td><td>18</td><td>8</td></tr></table>
+
+Table 1: The breakdown count of the dialogues, documents and average number of content elements per document by domain.  
+
+<table><tr><td>Role</td><td>DA</td><td>#Turns</td><td>#Tokens/Turn</td></tr><tr><td>user</td><td>request/query</td><td>25719</td><td>13</td></tr><tr><td>agent</td><td>request/query</td><td>8574</td><td>13</td></tr><tr><td>user</td><td>respond/yesOrNo</td><td>9254</td><td>7</td></tr><tr><td>agent</td><td>respond-reply</td><td>26273</td><td>24</td></tr><tr><td>total</td><td>all</td><td>69820</td><td>14</td></tr></table>
+
+Table 2: The total # of turns and the average # of tokens per turn, aggregated on dialogue act category.
+
+ment is also represented as a sequence of spans, for which we provide indexes to the plain text and the HTML respectively.
+
+Content elements To characterize the document contents, we examine the HTML source to extract the content elements with different scopes such as, tokens (tk), spans (sp), paragraphs (p) and titled sections (sec). Some of the spans within one sentence, such as SP1 and SP2 in Figure 1, are extracted via constituency parsers (Joshi et al., 2018). The paragraphs and sections are determined using HTML Markup. The average counts of these elements per document in Table 1 show the rich structures that are employed across domains. While this work starts to explore the simpler semi-structured information such as D2 in Figure 1; we are yet to explore various semantics from complex list structures, tables and other multi-modal contents in the webpages for future work.
+
+# 2.3 Dialogue Data
+
+Given a grounding document, we create about 10 unique dialogue flows with an average of 14 turns for this dataset. All dialogues are created based on a unique dialogue flow. In total, there are about 4500 conversations with close to 70,000 turns from four domains as shown in Table 1. Each dialogue utterance is annotated with a dialogue scene, i.e., role, dialogue act and the grounding span. As it is a known challenge to annotate conversation turns for the dialogue scenes (Geertzen and Bunt, 2009), our pipeline approach for data collection helps avoid
+
+![](images/83cf0f2280d755db35c098cc0855fdbda214632ce909d441082fdede866ab35a.jpg)  
+Figure 3: An illustration of the indexes of the relevant grounding contents in the documents.
+
+the cost and the noise from the additional human annotations. Next we further describe it from different perspectives regarding the dialogue scene.
+
+Dialogue acts We adopt the hierarchical dialogue act scheme by Pareti and Lando (2018) with a focus on the ones most essential to the information-seeking tasks. We describe those dialogue acts to the crowdsourced contributors pertaining to the selected grounding content and the assigned role (detailed descriptions in Appendix A). For future work, we plan to extend current dialogue scenes with other actions such as elucidations (Azzopardi et al., 2018) and social acts (Klüwer, 2011). To examine the dialogue distributions, we aggregate the hierarchical dialogue acts and list the total of turns, and the average length per turn under each category in Table 2. For example, "agent — request/query" corresponds to the queries based on document-guided dialogue management turns via an agent role; "user — respond/yesOrNo" corresponds to the scene where a user responds to an agent's query. Since we encourage the crowd to express "yes" or "no" in natural and creative writings, such as U10 in D2 in Figure 1, the average length of "respond/yesOrNo" is 7 tokens.
+
+Grounding content We aim to include the contents that are associated with varied conditional contexts based on the aforementioned span graph without introducing strong bias on certain index position in the document as discussed in Geva and Berant (2018). Therefore, we examine the coverage of the document contents from the generated dialogue flows. As illustrated in Figure 3, we create index of all the selected grounding contents to different document segments such as tokens, spans, paragraphs and titled sections (y-axis). The x-axis (numbered 1-10) indicates the position where 1 is closest to the beginning and 10 is closest to the end of a document. The numbers in the cells indicate the percentage distribution among all the ground
+
+<table><tr><td>feedback on rejected dialogue scene</td><td>%</td></tr><tr><td>The selected-text is not a contextual condition.</td><td>74.3</td></tr><tr><td>The selected-text is not a solution to the query.</td><td>10.5</td></tr><tr><td>Cannot write a turn to be coherent with the chat history.</td><td>10.1</td></tr><tr><td>There is not enough information in the selected (or adjacent) text.</td><td>2.4</td></tr><tr><td>The selected-text is not Comprehensible.</td><td>1.8</td></tr><tr><td>Other.</td><td>0.9</td></tr></table>
+
+Table 3: Feedback on the reasons for rejecting a dialogue scene by crowdsourced annotators.
+
+ing contents. The heatmap shows some degree of coverage on all parts of the documents, with a higher density at the beginning as we do include the scenarios of under-specified queries that typically correspond to the intro of a document.
+
+Dialogue flows For assessing the quality of the dialogue flows, we also ask the contributors to reject a dialogue turn when it is considered as infeasible to write a coherent utterance. We also solicit feedback via multiple choices on the reason as shown in Table 3. Out of 700 sampled dialogue flows, annotators reject about  $4\%$  of the turns. Among the rejected turns,  $70\%$  is due to not being able to interpret the selected span as applicable conditional context for user requests. In this dataset, we exclude the (sub)dialogues with rejected turns accordingly. However, we also observe certain "false positive" cases, where the crowd would rather try to adjust their writing for a less desirable dialogue scene rather than rejecting the turn, for which they get paid the same either way.
+
+# 2.4 Data Decomposition
+
+One benefit of constructing the dialogue data via our pipeline approach is that it provides a convenient and cost-effective way to reshape the existing dialogue data based on their dialogue flows. For instance, to ensure the quality, we can recollect or remove certain turns from the dialogues if they are rejected by the crowd contributors or affected by the changes in the grounding documents. In addition, for obtaining the training instances to identify the irrelevant queries, we modify an existing dialogue by inserting sub-dialogues created for another document or domain, for instance, adding D3 to D1 as irrelevant for va.org in Figure 1. Similarly, for creating dialogues that are grounded in multiple documents, we select sub-dialogues based on different documents and combine them into one.
+
+# 3 Tasks and Baselines
+
+For evaluation, we propose three tasks related to identifying the grounding content for a given dialogue: (1) user utterance understanding; (2) agent response prediction; (3) relevant document identification. In our tasks, we also aim to detect the cases that are irrelevant to the associated documents, for which we modify dialogues to include irrelevant (Irr) queries via data re-composition as described in Section 2.4. We split the dialogues into train/dev/test sets as  $70\%$ ,  $15\%$ ,  $15\%$  with half of the dev/test set grounded in "unseen" documents (not in training set). Experiment results are on test set unless otherwise stated. Numbers in the form of "mean  $\pm$  stdev" are computed out of 3 random seeds.
+
+# 3.1 User Utterance Understanding
+
+One of our main goals for creating this dataset is to broaden the coverage of different user queries for various task goals. Thus, our first task is interpreting a user utterance based on the dialogue history and the grounding document content. It aims to identify the associated dialogue scene, i.e., (1) grounding span in the document and (2) dialogue act, as described in the following two sections respectively.
+
+# 3.1.1 User Utterance Grounding
+
+In our dataset, all turns are associated with a dialogue scene that includes the grounding span. Interpreting the user utterance could be quite challenging, because in some cases, it would completely depend on the dialogue history such as U12 and U14; while some cases, such as U1 and U16, depend more on the user utterance itself. For the input of this task, it takes a user utterance along with (1) the dialogue history and 2) the document content with simplified document structure. The output is a span in the document as the text reference of the given user utterance. Each grounded user turn is considered a training instance, so a dialogue with  $n$  grounded user turns is considered as  $n$  instances, with overlapping dialogue context.
+
+Baseline Approach We formulate the problem as span selection, inspired by extractive question answering tasks such as SQuAD task (Rajpurkar et al., 2016, 2018). As a baseline, we adopt the extractive question answering model with transformers encoder by (Devlin et al., 2019). More specifically, we follow the QA example from Hugging-
+
+Face Transformers (Wolf et al., 2019). Pretrained bert-large-uncased-whole-word-masking model is used as encoder, and is fine-tuned during training.
+
+The document content serves as the context input of the model. The query input is the dialogue context, for which we experiment different settings of utilizing the dialogue history: (1) "last two turns", i.e., the input user utterance for which we want to identify the dialogue scene, and the agent utterance before the given user utterance; (2) "all prev", i.e., the input user utterance and all the utterances before it; (3) "all prev w/DA", i.e., context in (2) along with the corresponding dialogue acts. The dialogue context is concatenated in reversed time order where the latest user utterance appears first.
+
+Often the grounding document is longer than the maximum sequence length of transformers. In such cases, we truncate the documents in sliding windows with a stride. The dialogue context and each document trunk form one instance to be fed in batch into the encoder. The sequence of the encoded embeddings is then sent to a linear layer, which maps each embedding in the sequence into two logits, representing the probability of the corresponding position being the start and end position of the span. During training, we apply the Cross Entropy loss function to compute the loss. If the ground truth span does not fall in the document trunk, the start and end positions are both considered to be the beginning of the sequence. During decoding, the start-position and end-position logits from all document trunks are considered together to find the span most favored by the model.
+
+Evaluation Metrics For evaluation we use Exact Match score and token-level F1 score, as in the evaluation script 2.0 of SQuAD. In addition, since our data comes with predefined spans in each document, we map predicted span to the closest predefined span start index and span end index, and evaluate the mapped span with Exact Match score, as "ts EM" in Table 4 and Table 7.
+
+Experiment Results The experiment results are summarized in Table 4. Generally, the model performance improves with more information added to the dialogue context. It indicates that the queries in our datasets are highly conversational contextual and our dataset could serve as a valuable source for evaluating dialogue models' capability of learning from deeper context. We also conduct an experiment using the w/ Irr data with the "all prev"
+
+<table><tr><td>dial-ctxt</td><td>text EM</td><td>text F1</td><td>ts EM</td></tr><tr><td>last two turns</td><td>52.6 ± 0.3</td><td>64.3 ± 0.3</td><td>52.5 ± 0.4</td></tr><tr><td>all prev</td><td>54.3 ± 0.5</td><td>66.2 ± 0.3</td><td>54.4 ± 0.2</td></tr><tr><td>all prev w/ DA</td><td>55.1 ± 0.4</td><td>66.3 ± 0.3</td><td>55.2 ± 0.4</td></tr></table>
+
+Table 4: Results for user utterance grounding.  
+
+<table><tr><td>all prev</td><td>text EM</td><td>text F1</td></tr><tr><td>wo/Irr</td><td>54.3 ± 0.5</td><td>66.2 ± 0.3</td></tr><tr><td>w/Irr</td><td>62.7 ± 0.4</td><td>70.1 ± 0.6</td></tr><tr><td>has_ans turns</td><td>53.3 ± 0.4</td><td>62.7 ± 0.8</td></tr><tr><td>Irr turns</td><td>99.1 ± 0.3</td><td>99.1 ± 0.3</td></tr></table>
+
+Table 5: Comparison of w/Irr and wO/Irr settings for user utterance grounding.
+
+dialogue context. Table 5 summarizes the results in comparison with  $\mathrm{wo} / \mathrm{Irr}$  data. Irr turns impose noise in understanding the context, reduce the model accuracy from 54.3 to 53.3 on the original turns that are grounded to the document. However, the Irr turns themselves are easy to identify and achieve a high score of 99.1. As a result, the overall score including the Irr turns is increased to 62.7.
+
+# 3.1.2 User Dialogue Act Identification
+
+Dialogue act prediction using dialogue context as input is an important task in dialogue systems modeling (Liu et al., 2017; Tran et al., 2017). We identify the dialogue act of each user turn considering three different cases of dialogue context as input: (1) U: only the input user utterance, (2) U + A: the input user utterance and previous agent utterance, and (3) U + A(w. da): inputs in (2) along with agent turn's dialogue act. We use the hidden state of the tokens as the representation of the dialogue context, and further process it by a linear layer to identify the probability distribution over the total number of user dialogue acts. There are 7 dialogue acts for w/ Irr, 6 dialogue acts for wo/ Irr. We use the common metrics of accuracy (Acc), recall (R) and precision (P) for evaluation.
+
+Baselines and Experiment Results As a baseline we adopted BertForSequenceClassification model, a multi-class sequence classifier popular for GLUE tasks (Wang et al., 2018). We use pretrained bert-base-uncased model as the encoder and fine-tune during training.
+
+The results in Table 6 indicate much room for improvement, e.g., by adding document context, or building a joint model with the user utterance grounding task. The macro-averaged P and R are much lower than the micro-averaged Acc because
+
+<table><tr><td rowspan="2">dial-ctxt</td><td colspan="3">w/Irr</td><td colspan="3">wo/Irr</td></tr><tr><td>Acc.</td><td>R.</td><td>P.</td><td>Acc.</td><td>R.</td><td>P.</td></tr><tr><td>U</td><td>60.2</td><td>34.5</td><td>35.6</td><td>77.2</td><td>45.2</td><td>47.2</td></tr><tr><td>U+A</td><td>72.7</td><td>51.8</td><td>53.7</td><td>79.3</td><td>50.8</td><td>48.6</td></tr><tr><td>U+A(w. da)</td><td>76.4</td><td>53.8</td><td>55.7</td><td>80.6</td><td>50.9</td><td>55.2</td></tr><tr><td>-</td><td>53.3</td><td>14.3</td><td>7.6</td><td>67.4</td><td>16.7</td><td>11.2</td></tr></table>
+
+of the imbalanced DA distribution as shown in Table 2. The results also reflect the challenges effectively posed by the introduction of Irr turns, which was intended by our task design.
+
+# 3.2 Agent Response Prediction
+
+For this task, we aim at predicting the next agent turn with a focus on identifying the reference to the grounding document for the response. Such task can be a very important step towards building explainable conversational systems with practicality.
+
+# 3.2.1 Agent Response Grounding Prediction
+
+This task takes as input 1) the dialogue context; and 2) the document content with simplified document structure, and predicts a span in the document that grounds the next agent response. This task looks very similar to the user-turn grounding text prediction task in Section 3.1.1 in that they both take dialogue context and document context as input and perform a span selection inside the document. However, they are essentially different: the user-turn grounding text prediction is to understand what the user has already said, whereas this task is to predict what the agent would want to respond.
+
+# Baseline Approach and Evaluation Metrics
+
+As opposed to investigating this task from the aspect of dialogue management and planning, as a first attempt, we continue with our focus on identifying the associated grounding content in the document. Thus, we treat this as a span selection task, and adopt the same evaluation metrics and baseline approach as in Section 3.1.1. Note that with the same input dialogue context and text context, the model output in Section 3.1.1 is the dialogue scene corresponding to the given user utterance, while the model output of this task is the dialogue scene predicted for the next agent response.
+
+Experiment Results The experiment results are summarized in Table 7. The scores are much lower than the ones from our previous task in Table 4 due
+
+Table 6: Results for user dialogue act identification by BERT. The last row is by majority vote. Acc. is micro-averaged, while R. and P. are macro-averaged.  
+
+<table><tr><td>dial-ctxt</td><td>text EM</td><td>text F1</td><td>ts EM</td></tr><tr><td>last two turns</td><td>33.4 ± 0.4</td><td>49.6 ± 0.8</td><td>34.7 ± 0.5</td></tr><tr><td>all prev</td><td>34.3 ± 0.2</td><td>50.0 ± 0.8</td><td>35.9 ± 0.2</td></tr><tr><td>all prev w/ DA</td><td>36.2 ± 0.4</td><td>52.6 ± 1.0</td><td>37.6 ± 0.7</td></tr></table>
+
+Table 7: Results for agent response grounding prediction.  
+
+<table><tr><td>all prev</td><td>text EM</td><td>text F1</td></tr><tr><td>wo/Irr</td><td>34.3 ± 0.2</td><td>50.0 ± 0.8</td></tr><tr><td>w/Irr</td><td>47.3 ± 0.2</td><td>57.6 ± 0.6</td></tr><tr><td>has_ans turns</td><td>33.8 ± 0.3</td><td>46.7 ± 0.7</td></tr><tr><td>Irr turns</td><td>98.8 ± 0.3</td><td>98.8 ± 0.3</td></tr></table>
+
+Table 8: Comparison of w/Irr and wo/Irr settings for agent response grounding prediction.
+
+to the challenging nature of the task. However, we do see a significant improvement after including dialogue act information, which directs our further work on dialogue management to further improve the performance. Table 8 compares the experiment result in the w/Irr and wo/Irr settings, where we see a similar trend as in Table 5 unsurprisingly.
+
+# 3.2.2 Agent Response Generation
+
+Next we evaluate the dataset via the task of generating agent response. One primary goal of our task is to enable document-guided agent response, which overlaps with the primary goal of ShARC (Saeidi et al., 2018). However, our dataset includes more types of dialogue scenes and sets no restriction on the natural language forms of queries and responses. Thus, we investigate how one of the best performing end-to-end approaches to-date for ShARC works on our dataset. Compared to ShARC, our user queries do not come with the scenario description but are annotated with the grounding span, and the grounding documents is much longer. Therefore, we truncate the document into sub-documents with a size of 200 tokens. We try different ways to truncate the text: (1) only at the end of a span (ts); (2) only at the end of a paragraph (p).
+
+Baseline Approach and Experiment Results We adopt the model from Zhong and Zettlemoyer (2019). The input is the user query with dialogue history of up to last 4 turns as well as their grounding spans and the document content; the output is the agent utterance. The model learns to extract the relevant spans implicitly that are entailed by the dialogue-based and document-based contexts, and then edit them to generate the agent response.
+
+The BLEU scores are reported in Table 9. We ob
+
+<table><tr><td>doc-ctxt</td><td>BLEU-1</td><td>BLEU-2</td><td>BLEU-3</td><td>BLEU-4</td></tr><tr><td>ours (ts)</td><td>40.45</td><td>34.65</td><td>31.84</td><td>29.98</td></tr><tr><td>ours (p)</td><td>58.12</td><td>54.26</td><td>52.53</td><td>51.51</td></tr><tr><td>(ShARC)</td><td>(67.14)</td><td>(60.59)</td><td>(56.46)</td><td>(53.67)</td></tr></table>
+
+serve better results with the preprocessing method that maintains the original document structure at larger scale. Compared to the results by the same model reported on ShARC dev dataset, our BLEU scores are significantly lower. This is related to the more dynamic forms of agent responses in our dataset; another factor is the length of relevant document context, even when truncated, ours is 4 times longer.
+
+# 3.3 Relevant Document Identification
+
+Given that the goal-oriented dialogues could correspond to different tasks from a same document or multiple documents, in order to facilitate the understanding of such challenge, we experiment with two settings for the task on retrieving the grounding document(s): (1) the dialogues that are grounded in a single document; (2) the dialogues that are grounded in multiple documents.
+
+# 3.3.1 Single-Document Retrieval
+
+This task is to identify the relevant grounding document given limited dialogue history information. Thus, the input is certain dialogue context and a pool of 594 documents from all four domains.
+
+Baselines and Experiment Results We consider two different baselines for this task: (1) BM25 (Robertson and Zaragoza, 2009) based Information Retrieval method, and (2) A multi-class sequence classifier based on BertForMultipleChoice, using pretrained bert-base-uncased model as the encoder (Zellers et al., 2018).
+
+BM-25 method takes the full document into account to create the index and match them against the provided dialogue contexts. BERT model takes the dialogue context  $d$  and a document  $y$  together as a sequence. We use 512 tokens and feed BERT with the 256 tokens each from  $d$  and  $y$ . For each dialogue context, we create a set of triples: one triple containing the correct document (labeled with 1), and  $m$  triples containing incorrect documents sampled randomly from the set of all documents (labeled with 0). Table 10 corresponds to the setting  $m = 4$ . During evaluation, we evaluate a given dialogue context against the set of all documents. The
+
+Table 9: Results for agent turn generation (dev set).  
+
+<table><tr><td rowspan="2">n</td><td colspan="3">BM-25</td><td colspan="3">BERT</td></tr><tr><td>R@1</td><td>R@5</td><td>R@10</td><td>R@1</td><td>R@5</td><td>R@10</td></tr><tr><td>1</td><td>26.1</td><td>44.8</td><td>53.5</td><td>32.4</td><td>59.6</td><td>67.3</td></tr><tr><td>2</td><td>49.3</td><td>74.2</td><td>78.8</td><td>50.5</td><td>77.8</td><td>85.1</td></tr><tr><td>3</td><td>49.4</td><td>73.9</td><td>79.0</td><td>51.7</td><td>83.7</td><td>88.8</td></tr><tr><td>4</td><td>56.0</td><td>80.4</td><td>84.9</td><td>57.6</td><td>84.3</td><td>89.4</td></tr><tr><td>5</td><td>59.3</td><td>80.7</td><td>86.0</td><td>60.2</td><td>85.6</td><td>90.7</td></tr></table>
+
+Table 10: Results for single-document retrieval with  $n$  previous turns as input.  
+
+<table><tr><td>Domain</td><td>R@1</td><td>R@5</td><td>R@10</td></tr><tr><td>va.org</td><td>52.3</td><td>78.3</td><td>86.4</td></tr><tr><td>dmv.org</td><td>50.5</td><td>76.4</td><td>86.4</td></tr><tr><td>ssa.org</td><td>33.6</td><td>74.2</td><td>86.1</td></tr><tr><td>cdc.org</td><td>46.8</td><td>74.1</td><td>83.2</td></tr><tr><td>Weighted Average</td><td>47.5</td><td>76.1</td><td>85.9</td></tr></table>
+
+Table 11: Results for multi-document retrieval in single domain.
+
+task is evaluated with the commonly used recall  $(R@k)$  metric in retrieval tasks, which measures the fraction of times the correct document is found in the top-  $k$  predictions.
+
+As shown in Table 10, DL-based approach shows better performance consistently. From the perspective of examining the quality of our dataset, we also see the numbers confirms that as more turns are included, the better the dialogue is grounded to the relevant document.
+
+# 3.3.2 Multi-Document Retrieval
+
+We construct the dialogues that are grounded in multiple documents as described in Section 2.4. To make the tasks more challenging and closer to real-life applications, the segments of a dialogue are all grounded in the documents from the same domain. This dataset contains 2051 conversations, out of which 1640, 206 and 205 conversations were used in the train, dev and test sets respectively.
+
+Baselines and Experiment Results A baseline similar to Section 3.3.1 was constructed for this task using BertForMultipleChoice. At each user turn, we predict which document it should be grounded to, given the user utterance, previous agent utterance and the domain.
+
+This task is essentially related to conversational search task Penha et al. (2019), which predicts a link to the relevant document given a dialogue. Even though our document pool is not large compared to IR tasks, each document is quite long. The results in Table 11 show much room for improvement, and our dataset could be valuable resource for further deep document modeling.
+
+# 4 Related Work
+
+Our work is mainly focused on modeling dialogues that are grounded in documents. It is generally inspired by the recent substantial interests on the challenges of machine reading comprehension and conversational QA, such as CoQA (Reddy et al., 2019), QuAC (Choi et al., 2018) and DoQA (Campos et al., 2020). Those tasks aim to support conversational question answering, which involves understanding a text passage and answering a series of interconnected questions that appear in a conversation. These tasks add the complexity of coreference resolution and contextual reasoning to the reading comprehension challenges such as SQuAD (Rajpurkar et al., 2016, 2018), yet aim at identifying a solution from a given list of candidates by reasoning over spans from a document. Our task shares those challenges and additionally introduces the dialogue scenes where the agent asks questions when the user query is identified as under-specified or additional verification required for a resolute solution.
+
+Another recent work Kim et al. (2020) extends MultiWOZ (Budzianowski et al., 2018) by adding turns that are grounded in the FAQ knowledge for certain entity and domain. The document-based knowledge used in our work is beyond FAQs with entity as context but whole documents with more complex contexts. In addition, ours is also largely related to conversational search tasks, such as MANtIS (Penha et al., 2019). Similarly, it also provides multi-turn conversations with varied user intents that are grounded in documents from Stack Exchange website. In addition to the domain difference, one major distinction is that the grounding in MANtIS is determined by the hyperlinks to a document. Our grounding is defined at at a much finer level in addition to the link to a document.
+
+To the best of our knowledge, the closest related work to ours is ShARC (Saeidi et al., 2018) with dialogues that are grounded to a span of a given text snippet. It also proposes to address under-specified questions by requiring follow-up questions that are answerable with yes/no answers in similar domains. Our dataset goes beyond ShARC in several aspects nonetheless: we exploit not only paragraph-level structure but also higher-level document structure, we create conversations over much longer span of document content, where utterances are free-formed, as opposed to yes/no answers.
+
+# 5 Conclusion
+
+We have introduced doc2dial, a new dialogue dataset for goal-oriented tasks that are grounded in documents from multiple domains. Compared to previous work, our dialogues cover a greater variety of dialogue scenes that correspond to a much wider span of document content. For evaluation, we investigated three types of dialogue tasks and proposed baseline approaches. We hope this work will inspire and assist both dialogue and document modeling for tackling more real-life dialogue tasks.
+
+# Acknowledgments
+
+We thank the anonymous reviewers for their insightful comments. We thank Vera Liao and Kshitij Fadnis for their advice during the early stage of this project. We also thank crowd contributors for their valuable inputs for building our tool and dataset.
+
+# References
+
+Leif Azzopardi, Mateusz Dubiel, Martin Halvey, and Jeffery Dalton. 2018. Conceptualizing agent-human interactions during the conversational search process. In The Second International Workshop on Conversational Approaches to Information Retrieval.  
+Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, Inigo Casanueva, Stefan Ultes, Osman Ramadan, and Milica Gasic. 2018. Multiwoz-a large-scale multi-domain wizard-of-oz dataset for task-oriented dialogue modelling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5016-5026.  
+Bill Byrne, Karthik Krishnamoorthi, Chinnadhurai Sankar, Arvind Neelakantan, Ben Goodrich, Daniel Duckworth, Semih Yavuz, Amit Dubey, Kyu-Young Kim, and Andy Cedilnik. 2019. Taskmaster-1: Toward a realistic and diverse dialog dataset. 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 4516-4525, Hong Kong, China. Association for Computational Linguistics.  
+Jon Ander Campos, Arantxa Otegi, Aitor Soroa, Jan Deriu, Mark Cieliebak, and Eneko Agirre. 2020. DoQA - accessing domain-specific FAQs via conversational QA. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7302-7314, Online. Association for Computational Linguistics.  
+Eunsol Choi, He He, Mohit Iyyer, Mark Yatskar, Wentau Yih, Yejin Choi, Percy Liang, and Luke Zettle-moyer. 2018. QuAC: Question answering in context. In Proceedings of the 2018 Conference on
+
+Empirical Methods in Natural Language Processing, pages 2174-2184, Brussels, Belgium. Association for Computational Linguistics.  
+Debopam Das, Tatjana Scheffler, Peter Bourgonje, and Manfred Stede. 2018. Constructing a lexicon of english discourse connectives. In Proceedings of the 19th Annual SIGdial Meeting on Discourse and Dialogue, pages 360-365.  
+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.  
+Song Feng, Kshitij Fadnis, Q Vera Liao, and Luis A Lastras. 2020. Doc2dial: a framework for dialogue composition grounded in documents. In Thirty-Fourth AAAI Conference on Artificial Intelligence.  
+Jeroen Geertzen and Harry Bunt. 2009. Measuring annotator agreement in a complex hierarchical dialogue act annotation scheme. In Proceedings of the 7th SIGdial Workshop on Discourse and Dialogue, pages 126-133. Association for Computational Linguistics.  
+Mor Geva and Jonathan Berant. 2018. Learning to search in long documents using document structure. In Proceedings of the 27th International Conference on Computational Linguistics, pages 161-176, Santa Fe, New Mexico, USA. Association for Computational Linguistics.  
+Vidur Joshi, Matthew Peters, and Mark Hopkins. 2018. Extending a parser to distant domains using a few dozen partially annotated examples. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1190-1199, Melbourne, Australia. Association for Computational Linguistics.  
+Seokhwan Kim, Mihail Eric, Karthik Gopalakrishnan, Behnam Hedayatnia, Yang Liu, and Dilek Hakkani-Tur. 2020. Beyond domain apis: Task-oriented conversational modeling with unstructured knowledge access. arXiv preprint arXiv:2006.03533.  
+Tina Kluwer. 2011. "i like your shirt"-dialogue acts for enabling social talk in conversational agents. In International Workshop on Intelligent Virtual Agents, pages 14-27. Springer.  
+Yang Liu, Kun Han, Zhao Tan, and Yun Lei. 2017. Using context information for dialog act classification in dnn framework. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2170-2178.
+
+Saikat Mukherjee, Guizhen Yang, Wenfang Tan, and IV Ramakrishnan. 2003. Automatic discovery of semantic structures in html documents. In *Seventh International Conference on Document Analysis and Recognition*, 2003. Proceedings., pages 245-249. IEEE.  
+Silvia Pareti and Tatiana Lando. 2018. Dialog intent structure: A hierarchical schema of linked dialog acts. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC-2018).  
+Gustavo Penha, Alexandru Balan, and Claudia Hauff. 2019. Introducing mantis: a novel multi-domain information seeking dialogues dataset. arXiv preprint arXiv:1912.04639.  
+Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Miltsakaki, Livio Robaldo, Aravind Joshi, and Bonnie Webber. 2008. The Penn discourse TreeBank 2.0. In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC'08), Marrakech, Morocco. European Language Resources Association (ELRA).  
+Rashmi Prasad, Bonnie Webber, Alan Lee, and Aravind Joshi. 2019. Penn Discourse Treebank Version 3.0. In LDC2019T05. Philadelphia: Linguistic Data Consortium.  
+Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don't know: Unanswerable questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784-789, Melbourne, Australia. Association for Computational Linguistics.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics.  
+Siva Reddy, Danqi Chen, and Christopher D. Manning. 2019. CoQA: A conversational question answering challenge. Transactions of the Association for Computational Linguistics, 7:249-266.  
+Stephen Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: Bm25 and beyond. Found. Trends Inf. Retr., 3(4):333-389.  
+Marzieh Saeidi, Max Bartolo, Patrick Lewis, Sameer Singh, Tim Roktaschel, Mike Sheldon, Guillaume Bouchard, and Sebastian Riedel. 2018. Interpretation of natural language rules in conversational machine reading. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2087-2097, Brussels, Belgium. Association for Computational Linguistics.
+
+M Stede, T Scheffler, and A Mendes. 2019. Connective-lex: A web-based multilingual lexical resource for connectives. discours. Revue de linguistique, psycholinguistique et informatique. A journal of linguistics, psycholinguistics and computational linguistics,(24).  
+Quan Hung Tran, Gholamreza Haffari, and Ingrid Zukerman. 2017. A generative attentional neural network model for dialogue act classification. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 524-529.  
+Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. Glue: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R'emi Louf, Morgan Funtowicz, and Jamie Brew. 2019. Huggingface's transformers: State-of-the-art natural language processing. ArXiv, abs/1910.03771.  
+Rowan Zellers, Yonatan Bisk, Roy Schwartz, and Yejin Choi. 2018. Swag: A large-scale adversarial dataset for grounded commonsense inference. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 93-104.  
+Victor Zhong and Luke Zettlemoyer. 2019. E3: Entailment-driven extracting and editing for conversational machine reading. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2310-2320, Florence, Italy. Association for Computational Linguistics.

doc2dialagoalorienteddocumentgroundeddialoguedataset/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da6c75cea03fcee411ab8e5325895fa33e71499934530ace1feb0fc63c288524
+size 376029

doc2dialagoalorienteddocumentgroundeddialoguedataset/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bbb0d25399d8d9930e0a6a57182d0183ee6a19c45e605c8a207f4759152e1876
+size 276258

idratherjustgotobedunderstandingindirectanswers/2a73c933-c83f-46be-9755-7e06ae53dc89_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff35792e22b8b2ca3cd67fafe4edcaf2ec4df1322e560f465775051146d9a2f6
+size 118257

idratherjustgotobedunderstandingindirectanswers/2a73c933-c83f-46be-9755-7e06ae53dc89_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7759c809287ea94d68c20fc1a2e41a4c139b9fb22c48c689b19bcd6e9590abdd
+size 154484

idratherjustgotobedunderstandingindirectanswers/2a73c933-c83f-46be-9755-7e06ae53dc89_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:297c22ee2052df37ce0e42318a48675104ba9e390790b51b16c8a3869ccbbf96
+size 1378009

idratherjustgotobedunderstandingindirectanswers/full.md

@@ -0,0 +1,643 @@
+# "I'd rather just go to bed": Understanding Indirect Answers
+
+Annie Louis
+
+Google Research, UK
+
+annielouis@google.com
+
+Dan Roth*
+
+University of Pennsylvania
+
+danroth@seas.upenn.edu
+
+Filip Radlinski
+
+Google Research, UK
+
+filiprad@google.com
+
+# Abstract
+
+We revisit a pragmatic inference problem in dialog: Understanding indirect responses to questions. Humans can interpret 'I'm starving.' in response to 'Hungry?', even without direct cue words such as 'yes' and 'no'. In dialog systems, allowing natural responses rather than closed vocabularies would be similarly beneficial. However, today's systems are only as sensitive to these pragmatic moves as their language model allows. We create and release<sup>1</sup> the first large-scale English language corpus 'Circa' with 34,268 (polar question, indirect answer) pairs to enable progress on this task. The data was collected via elaborate crowdsourcing, and contains utterances with yes/no meaning, as well as uncertain, middle-ground, and conditional responses. We also present BERT-based neural models to predict such categories for a question-answer pair. We find that while transfer learning from entailment works reasonably, performance is not yet sufficient for robust dialog. Our models reach  $82 - 88\%$  accuracy for a 4-class distinction, and  $74 - 85\%$  for 6 classes.
+
+# 1 Introduction
+
+Humans produce and interpret complex utterances even in simple scenarios. For example, for the polar (yes/no) question 'Want to get dinner?', there are many perfectly natural responses in addition to 'yes' and 'no', as in Table 1. How should a dialog system interpret these INDIRECT answers? Many can be understood based on the answer text alone, e.g. 'I would like that' or 'I'd rather just go to bed'. For others, the question is crucial, e.g. 'Dinner would be lovely.' is a positive reply here, but a negative answer to 'Want to get lunch?' In this
+
+# "Want to get some dinner together?"
+
+"I know a restaurant we could get a reservation at."  
+"I have already eaten recently."  
+"I hope to make it home by supper but I'm not sure I can."  
+"Dinner would be lovely."  
+"I'd rather just go to bed."  
+"There's a few new restaurants we could go to."  
+"I would like that."  
+"We could do dinner this weekend."  
+"I would like to go somewhere casual."  
+"I'd like to try the new Italian place."
+
+Table 1: A polar question with 10 indirect responses, taken from our corpus.
+
+paper, we present the first large scale corpus and models for interpreting such indirect answers.
+
+Previous attempts to interpret indirect yes/no answers have been small scale and without data-driven techniques (Green and Carberry, 1999; de Marneffe et al., 2010). However, recent success on many language understanding problems (Wang et al., 2019), the impressive generation capabilities of modern dialog systems (Zhang et al., 2019; Adiwardana et al., 2020), as well as the huge interest in yes/no question-answering (Choi et al., 2018; Clark et al., 2019) have created a conducive environment for revisiting this hard task.
+
+We introduce Circa, a new dataset with 34K pairs of polar questions and indirect answers in the English language. This high quality corpus consists of natural responses collected via crowd workers, and goes beyond binary yes/no meaning to include conditionals, uncertain, and middle-ground answers. Circa contains many phenomena of interest, although the first step, which we address here, is how to robustly classify a question-answer pair into one of the above meaning categories. We find that BERT (Devlin et al., 2019) fine-tuned on entailment data is an effective initial approach, mirroring the success in question-answering work involving yes/no questions (Clark et al., 2019). It reaches
+
+an accuracy of  $85 - 88\%$  for responses in the same situational context, and  $6 - 10\%$  lower accuracy on held-out scenarios. The answer text itself (as in 'I would like that') contains strong cues leading to  $78 - 82\%$  accuracy, however the best results come from jointly analyzing the question and the answer.
+
+# 2 Related work
+
+Indirect answers to polar questions are reasonably frequent and warrant deep study. Early work put the proportion at  $13\%$  in face-to-face and telephone conversations (Stenström, 1984), and at  $27\%$  in an instruction giving/following map task (Rossen-Knill et al., 1997; Hockey et al., 1997). For a more recent and larger analysis, consider the Cornell Movie Dialog Corpus (Danescu-Niculescu-Mizil and Lee, 2011). We heuristically mined yes/no questions and their following utterances, finding 6,327 pairs. Direct answers (i.e., answers with 'yes', 'no', 'maybe' and related terms such as 'okay', 'yup', etc.) only cover  $53\%$  of the pairs. This suggests that indirect responses could be even more frequent in natural open-domain dialogue.
+
+Even when a direct answer is possible, speakers use indirect answers to be cooperative and address anticipated follow-up questions (Stenström, 1984), to provide explanations in the case of a negative answer (Stenström, 1984), to block misleading interpretations that may arise from a curt 'yes' or 'no' reply (Hirschberg, 1985), and since indirect answers may appear more polite (Brown and Levinson, 1978). But we lack a large corpus of such answers to study these multiple pragmatic functions. Our work aims to fill this gap.
+
+On the computational side, there are impressive efforts towards planning, generation, and detection of indirect answers, albeit, on a small scale, and without data-driven approaches. Green and Carberry (1999)'s early work leverages discourse relations for generating indirect answers. For example, an 'elaboration' may be relevant for a 'yes' response, and a 'contrast' might help convey a 'no' answer. de Marneffe et al. (2009) reason about such answers using Markov Logic Networks. In subsequent work, de Marneffe et al. (2010) present one of the first data-driven studies into indirect answers containing scalar adjectives. They mine a set of 224 question-answer pairs from interview transcripts and dialog corpora. Using polarity information from review data, and manual coding of test samples, they achieve an accuracy of  $71\%$
+
+on three classes 'yes', 'no' and 'uncertain'. Our work aims to collect a much larger and more diverse natural corpus, and demonstrates the first automatic approach using recent advances in natural language inference (NLI). de Marneffe et al. (2010) also demonstrated the first crowd annotation of indirect answers, and we draw on many aspects of their formulation for the creation of our corpus.
+
+An unexpected limelight on yes/no questions has also arisen in recent question-answering (QA) work. Researchers have noticed that yes/no questions are complex and naturally arise (as high as  $20\%$ ) when questions are posed one after the other in a conversation (Reddy et al., 2019; Choi et al., 2018). Their goal is to produce direct 'yes' or 'no' answers, but obtaining them requires inference against a paragraph or excerpt, an analogous task to our yes/no inference from indirect answers. Very recent work (Clark et al., 2019) has specifically sought to improve this ability in QA systems, by building a corpus of 16K yes/no factual questions paired with Wikipedia passages from which the yes/no answer can be inferred. Departing from factual texts, our focus is on single-turn conversational responses in everyday situations. The latter are faithful, cooperative, and grounded in world knowledge. Still, transfer learning from factual corpora could prove useful and we explore this too.
+
+# 3 The Circa Corpus
+
+Circa (meaning approximately) is our crowdsourced corpus for research on indirect answers. It contains 34,268 question-answer pairs, comprising 3,431 unique questions with up to 10 answers to each. Each question-answer pair is also annotated with meaning based on a fixed set of categories. We explain all aspects of the collection process below, but in the interest of space, further details (complete annotator instructions, interfaces, and examples) are in the Appendix.
+
+Our data consists of very short dialogues, each containing a question and its indirect answer. The texts are varied in semantic and syntactic forms, and grounded in 10 different situational contexts. Table 2 presents some examples, showing that the answers go beyond binary yes/no distinctions. We created this data via an elaborate crowd-annotation exercise which comprised of 4 steps described next.
+
+<table><tr><td>S1: Talking to a friend about food preferences.
+Q: “Do you like pizza?” 
+A: “I like it when the toppings are meat, not vegetable.”</td></tr><tr><td>S2: Talking to a friend about music preferences.
+Q: “Do you like guitars?” 
+A: “I practice playing each weekend.”</td></tr><tr><td>S3: Talking to a friend about weekend activities.
+Q: “Are you available this Sunday evening?” 
+A: “What did you have in mind?”</td></tr><tr><td>S4: Talking to a friend about book preferences.
+Q: “Are you a fan of comic books?” 
+A: “I read an Archie every time I have lunch.”</td></tr><tr><td>S5. Your friend is visiting from out of town.
+Q: “Would you like to go out for dinner?” 
+A: “I could go for some Mexican.”</td></tr><tr><td>S6. Two colleagues leaving work on a Friday.
+Q: “Long week?” 
+A: “I’ve had worse weeks.”</td></tr><tr><td>S7. You friend is planning to buy a flat in New York.
+Q: “Does the flat’s price fit your long-term budget?” 
+A: “Well, if it doesn’t I will definitely refinance my mortgage.”</td></tr><tr><td>S8. Your friend is thinking of switching jobs.
+Q: “Do you have to travel far?” 
+A: “My commute is about 10 minutes.”</td></tr><tr><td>S9. Two childhood neighbours unexpectedly run into each other at a cafe.
+Q: “Are you going to the high school reunion in June?” 
+A: “I forgot all about that.”</td></tr><tr><td>S10. Meeting your new neighbour for the first time.
+Q: “Did you move from near-by?” 
+A: “I am from Canada.”</td></tr></table>
+
+Table 2: Examples of questions and answers in our 10 dialogue scenarios.
+
+# 3.1 Step 1: Dialog scenarios
+
+We designed 10 diverse prompts to serve as situational contexts. These everyday situations can lead to productive dialog, and simple yes/no questions (i.e., they do not depend on elaborate prior context for initiating a conversation). We designed them manually to intentionally cover a number of situations that encourage variety in questions and responses. As only a small number were needed, and the desiderata are not trivial, a crowd task was not suitable for prompt development. In Table 2, S1-S10 are the titles of the 10 dialog scenarios (each also consisting of a longer description).
+
+The rest of the data were collected using crowd workers. We ran pilots for each step of data collection, and perused their results manually to ensure clarity in guidelines, and quality of the data. We recruited native English speakers, mostly from the USA, and a few from the UK and Canada. We did not collect any further information about the crowd workers.
+
+# Annotator instructions
+
+In this task, we will ask you to provide yes/no questions in a social dialogue situation.
+
+Example: Suppose that you are trying to learn about a friend's movie preferences, but can only ask yes/no questions. Provide 5 useful questions that can be answered "Yes" or "No".
+
+Note: (1) We are looking for variety in the questions. For instance:
+
+'Have you watched Star Wars?'
+
+'Do you like movies with a complicated plot?'
+
+'Did you enjoy the last movie we saw together?'
+
+'Want to go watch a thriller this weekend?'
+
+'Are you into the Avengers series?'
+
+Note that the questions have different forms as well as different content.
+
+(2) Remember that the setting is a conversation with a friend (or neighbour or colleague). Please keep the questions casual, so they would be natural during an informal conversation.
+
+Table 3: Annotator instructions for Step 2.
+
+# 3.2 Step 2: Question collection
+
+In this phase, we ask annotators to write yes/no questions for a given dialog scenario.
+
+Table 3 shows the instructions displayed to annotators. (Also see Figure 1 and Table 11.) 100 annotators each provided 5 questions per scenario, resulting in 5,000 questions.
+
+Questions where annotators did not adhere to topic provided were removed. Of the remaining 4,710 questions,  $84\%$  were unique. Understandably, some scenarios had more repetitions than others: For food preferences,  $76\%$  of the questions were unique, as opposed to  $95\%$  when talking about a friend's job change. Below we show the most and least common questions in the food context.
+
+Most common food questions  
+
+<table><tr><td>25 times</td><td>Do you like spicy food?</td></tr><tr><td>11 times</td><td>Are you vegetarian?</td></tr><tr><td>10 times</td><td>Do you eat meat?</td></tr><tr><td colspan="2">Sample of least common food questions</td></tr><tr><td>1 time</td><td>Have you ever tried vegan cuisine?</td></tr><tr><td>1 time</td><td>Do you have a gluten allergy?</td></tr><tr><td>1 time</td><td>Are you familiar with Thai food?</td></tr></table>
+
+The most common one 'Do you like spicy food?' was suggested by 25 out of 100 annotators.
+
+One important aspect of our design is that each annotator was asked to provide five questions at the same time. Obvious questions often showed up as the first question, with later questions becoming more complex and more interesting. Question properties such as length, inverse document frequency, and type-token ratio confirm this difference.
+
+# 3.3 Step 3: Answer elicitation
+
+We sampled 350 non-redundant questions per scenario, with an equal number from the 5 question positions (see previous section). Using a different set of annotators than Step 2, we then elicited 10 indirect answers for each question. The annotator instructions are provided in Table 4. For faster annotation, and to encourage diverse answers, we displayed five questions (in same situational context) simultaneously. The five questions were chosen to be diverse (based on cosine similarity between nouns and adjectives in the questions), and annotators were instructed to treat them independently. See Figure 2 for an example display.
+
+Importantly, a key design consideration was that we do not instruct annotators to produce an answer with a specific meaning. Rather the annotator composes a natural response without reflecting upon a required meaning. We believe this flexibility is important to ensure that varied meanings (even ambiguous ones) are present in our data. This is verified in our analysis in Section 3.6.
+
+Table 1 provides example answers from this step. Note that the 10 answers have varied meanings departing from definite 'yes' and 'no'. The answers were high quality and diverse in form.  $83\%$  of answers appear only once in the corpus. At the same time, there are a few prototypical answers for the different meanings. Below we show the most repeated responses (and frequency) for 'yes' and 'no' meanings, and also for cases where the answer is conditional upon some situation, and middle ground responses.
+
+<table><tr><td colspan="2">‘yes’ answers</td><td colspan="2">‘no’ answers</td></tr><tr><td>59</td><td>I would love to</td><td>21</td><td>I don’t drink</td></tr><tr><td>40</td><td>Let’s do it</td><td>18</td><td>I prefer pop</td></tr><tr><td>40</td><td>That would be good</td><td>18</td><td>I wish!</td></tr><tr><td colspan="2">‘conditional yes’</td><td colspan="2">‘middle-ground’</td></tr><tr><td>9</td><td>If the weather is nice</td><td>14</td><td>I’m not sure yet</td></tr><tr><td>6</td><td>If I can afford it</td><td>10</td><td>Which one?</td></tr><tr><td>6</td><td>Depends where you want to go</td><td>9</td><td>It’s OK</td></tr></table>
+
+These responses indicate that strong lexical signals for meaning are present in the answer.
+
+# 3.4 Step 4: Marking interpretations
+
+Finally, we ask a third set of annotators to mark interpretations for all the QA (question, indirect
+
+# Annotator Instructions
+
+You will be given a social situation, for example, talking to your friend or neighbour.
+
+Task: You will be asked to respond to a question from your friend/neighbour but without using the words 'yes' or 'no' (or similar words like 'yeah', etc). Please provide a possible answer, it does not have to be your real opinion. Rather you should provide a possible answer from which your friend will be able to infer whether you mean 'yes', 'no', 'maybe' or 'sometimes'.
+
+Example: Here are three such answers to your friend's question about movies.
+
+Do you like movies with sad endings?
+
+(a) I often watch them. (Meaning=Yes)  
+(b) I prefer movies which make me laugh. (Meaning=No)  
+(c) When the plot is also good. (Meaning=Sometimes)
+
+Table 4: Annotator instructions for Step 3.
+
+# Annotator instructions
+
+You will be shown short dialogues between two friends/colleagues (X and Y) in a certain context. For example:  
+Context: X wants to know about Y's movie preferences.  
+X: "Do you like movies with sad endings?"  
+Y: "I often watch them."  
+In all the dialogues, X asks a simple 'Yes/No' question, and Y answers it with a short sentence or a phrase.  
+Task: We need your help to interpret Y's answer. Read the dialog and tell us how you think X will interpret Y's answer. Your options are: X will think that Y means (1) 'yes'  
+(2) 'probably yes' / 'sometimes yes'  
+(3) 'yes, subject to some conditions'  
+(4) 'no'  
+(5) 'probably no'  
+(6) 'in the middle, neither yes nor no'  
+(7) 'I am not sure how X will interpret Y's answer' If Y's response does not fit any of the above, please choose the option (8) 'Other', and leave a short comment.  
+For our example above, the likely interpretation is 'yes'.
+
+Table 5: Annotator instructions for Step 4.
+
+answer) pairs from Step 2. In particular, they are asked to judge how the question-seeker would interpret the answer provided. As in most NLP tasks, interpretations will vary, and so we obtain five annotations per pair.
+
+The correct label categories are not readily clear, but the variety of examples in our corpus made it certain that just 'yes' and 'no' will not suffice. Building on prior work by de Marneffe et al. (2010), and a pilot experiment, we identified categories that can be annotated reliably. These are shown in Table 5. The annotators were asked to assume the dialogs were co-operative and casual. They were advised to use 'probably yes/no' when they cannot infer a definite meaning. If X will not infer a 'yes' without some condition being met, then the class 'yes, subject to some conditions' was to be chosen. Detailed instructions with exact phrasing,
+
+<table><tr><td>Yes
+Q: Do you have any pets?
+A: My cat just turned one year old.</td><td>Probably yes / sometimes yes
+Q: Do you like mysteries?
+A: I have a few that I like.</td><td>Yes, subject to some conditions
+Q: Do you enjoy drum solos?
+A: When someone&#x27;s a master.</td></tr><tr><td>No
+Q: Do you have a house?
+A: We are in a 9th floor apartment.</td><td>Probably no
+Q: Are you interested in fishing this weekend? A: It&#x27;s supposed to rain.</td><td>In the middle
+Q: Did you find this week good?
+A: It was the same as always.</td></tr></table>
+
+and practice questions are in the appendix (Figure 3, Tables 13 and 14). Annotators took an average of 23 seconds per question-answer pair.
+
+Annotators were also given an option to flag improper content. We remove those QA pairs which were flagged by even one of the five annotators. The authors also read every question, and used a blacklist of words for additional filtering. The remaining 34,268 pairs comprise the final corpus.
+
+# 3.5 Gold standard labels
+
+Each (question, indirect answer) pair was marked by five annotators, so we use majority judgement as the gold standard, subject to at least three annotators making that judgement.
+
+We use two aggregation schemes. The STRICT scheme keeps all eight class distinctions from Table 5. A more RELAXED label is computed by collapsing the uncertain classes with the definite ones: 'probably yes / sometimes yes'  $\rightarrow$  'yes', 'probably no'  $\rightarrow$  'no', and 'I am not sure how X will interpret Y's answer'  $\rightarrow$  'In the middle, neither yes nor no'. These classes were commonly confused by the raters. The 'Other' class was used mostly when the question was not polar (e.g., disjunctive ones such as 'Do you like to dine-in or take-out?'). To illustrate the richness of the Circa corpus, Table 6 shows one QA pair from each class.
+
+# 3.6 Label distributions
+
+We now analyze the distribution of the gold standard labels. For STRICT labels (Table 7), only  $8\%$  of the examples (marked 'N/A') do not receive a majority vote. The most frequent class is 'yes'  $(42\%)$  of the data). 'No' is less frequent  $(32\%)$ . The third most frequent is 'conditional yes', indicating that conditional preferences may be common. The 'probably' classes are around  $3 - 4\%$ , each with over a thousand examples in the corpus. There is also a notable number of 'in the middle' examples. With the collapsed RELAXED labels (Table 8), the proportion of 'yes' and 'no' increase slightly, and 'N/A' examples drop to only  $2\%$ . These distributions reflect the rich patterns in our data.
+
+Table 6: Example question and answer pairs where all 5 annotators agreed on the label.  
+
+<table><tr><td>Label</td><td colspan="2">STRICT</td></tr><tr><td>Yes</td><td>14,504</td><td>(42.3%)</td></tr><tr><td>No</td><td>10,829</td><td>(31.6%)</td></tr><tr><td>Probably yes / sometimes yes</td><td>1,244</td><td>(3.6%)</td></tr><tr><td>Yes, subject to some conditions</td><td>2,583</td><td>(7.5%)</td></tr><tr><td>Probably no</td><td>1,160</td><td>(3.4%)</td></tr><tr><td>In the middle, neither yes nor no</td><td>638</td><td>(1.9%)</td></tr><tr><td>I am not sure</td><td>63</td><td>(0.2%)</td></tr><tr><td>Other</td><td>504</td><td>(1.5%)</td></tr><tr><td>N/A</td><td>2,743</td><td>(8.0%)</td></tr></table>
+
+Table 7: Distribution of STRICT gold standard labels. 'N/A' indicates lack of majority agreement.  
+
+<table><tr><td>Label</td><td colspan="2">RELAXED</td></tr><tr><td>Yes</td><td>16,628</td><td>(48.5%)</td></tr><tr><td>No</td><td>12,833</td><td>(37.5%)</td></tr><tr><td>Yes, subject to some conditions</td><td>2,583</td><td>(7.5%)</td></tr><tr><td>In the middle, neither yes nor no</td><td>949</td><td>(2.8%)</td></tr><tr><td>Other</td><td>504</td><td>(1.5%)</td></tr><tr><td>N/A</td><td>771</td><td>(2.2%)</td></tr></table>
+
+# 3.7 Annotator agreement
+
+The Fleiss kappa scores are 0.61 for STRICT and 0.76 for RELAXED labels (p-values  $< 0.0001$ ) indicating substantial agreement. In fact, full agreement, as in Table 6 where all five annotators agree on the STRICT class, occurs for  $49\%$  of pairs, a high proportion given the complexity of the task. When the labels are RELAXED, this reaches  $71\%$ . The full agreement distributions are:
+
+Table 8: Distribution of RELAXED gold standard labels. 'N/A' indicates lack of majority agreement.  
+
+<table><tr><td>Agreement</td><td>STRICT</td><td>RELAXED</td></tr><tr><td>5 annotators</td><td>49.1%</td><td>71.0%</td></tr><tr><td>4 annotators</td><td>23.8%</td><td>17.2%</td></tr><tr><td>3 annotators</td><td>19.1%</td><td>9.6%</td></tr></table>
+
+# 3.8 Dialog scenarios and question type
+
+As one would expect, different scenarios prompt different types of questions and answers, and hence different label distributions. For the 'friend switching jobs' scenario,  $54.5\%$  of elicited answers are have 'yes' meaning (RELAXED labels). For book preferences, this is only  $42\%$ .
+
+Perhaps more unexpected is that a few questions
+
+have labels predominantly of the same kind. For example, all 10 answers to 'Ready for the weekend?' receive a 'yes' label, and they are all 'no' for 'Are you offering the asking price?''. While the first question is largely rhetorical, the second involves common sense: Most people negotiate real-estate prices. We found that  $3\%$  of questions (95) have all answers with the same label, and for another  $20\%$ , 8-9 answers have the same label. A model would still need to identify the label as either 'yes' or 'no', but these skews indicate that there may be some (weak) signals for the label in the question itself. After all, who isn't ready for the weekend?
+
+# 3.9 Annotation Protocol
+
+While our annotation method is comprehensive, one might wonder how alternative approaches would have fared. We have not performed controlled tests of different approaches but we briefly document our choices to aid future work.
+
+We performed pilot annotations for each step of our process. Our goal was to start with less restrictive settings, where annotators are given minimal and simple instructions. If we did not receive quality responses, we intended to give more specific directions. For example, in an initial pilot, sometimes annotators gave long answers which may be unrealistic in conversations eg. 'I can tell you, without a shadow of a doubt, that there are very few things that I enjoy more than sitting in front of my computer.' So in later annotations for all tasks, we added instructions 'to keep the conversation casual' which at least reminds annotators about this need. We performed spot checks on the results, but did not perform controlled tests.
+
+For answer elicitation, our simple instruction pilot produced answers with diverse meanings. The distribution of meanings also varied according to the question (some questions had a mix of meanings, others were skewed towards a few meanings). So we decided against explicitly asking annotators to provide a certain meaning as that would create uniform meaning distributions, which may turn out unrealistic and miss dominant tendencies. We also considered that the explicit approach may reduce answer quality (for example, if a person who eats meat were to have to answer indirectly that they are vegetarian, they may be more likely to get their facts wrong, or try to rush through the task). In fixing our choice, we took care to ensure that quality was not affected, and answers were varied.
+
+# 4 Learning Task
+
+Obviously answers contain numerous cues for subsequent dialog flow, but in this first work we focus on meaning prediction. Specifically, given a question-answer pair, we classify it into one of the meaning categories in Tables 7 and 8.
+
+We consider two experimental settings: In the matched setup, we assume that the response scenario is seen during training (randomly dividing our corpus examples into  $60\%$  training,  $20\%$  each for development/test). The unmatched setting is aimed at understanding the performance on unseen scenarios, i.e., whether models can learn the semantics of yes/no answers beyond the patterns specific to individual scenarios. As our data contains 10 scenarios, we carry out 10 leave-one-out tasks, each time holding out one scenario (for example, 'buying a flat in New York') as the test data, and use the remaining nine for training and development.
+
+For both the matched and unmatched setting, we consider two variants of the classification problem: STRICT (with 6 different labels, namely all except 'other' and 'N/A' in Table 7) and RELAXED with 4 labels (Table 8). We ignore the examples without a majority label and also those marked 'unsure' or 'other'. Thus our experiment data sizes are:
+
+<table><tr><td>Experimental Setting</td><td>Train</td><td>Dev.</td><td>Test</td></tr><tr><td>STRICT-matched</td><td>18,574</td><td>6,192</td><td>6,192</td></tr><tr><td>RELAXED-matched</td><td>19,795</td><td>6,599</td><td>6,599</td></tr><tr><td>STRICT-umatched</td><td>24,746</td><td>3,115</td><td>3,095</td></tr><tr><td>RELAXED-unmatched</td><td>26,404</td><td>3,289</td><td>3,299</td></tr></table>
+
+For the unmatched setting, these sizes are the average across the 10 leave-one-out sets.
+
+# 5 Models
+
+Building upon recent NLI systems, our approach leverages representations from unsupervised pretraining, and finetunes a multiclass classifier over the BERT model (Devlin et al., 2019). However, we first consider other models for related tasks.
+
+# 5.1 Related baselines and corpora
+
+BOOLQ is a question-answering dataset focused on factual yes/no questions (Clark et al., 2019). Here yes/no questions from web search queries are paired with Wikipedia paragraphs that help answer the question. There are 9.4k train, 3.2k development, and 3.2k test set examples, with two target classes, namely 'yes' and 'no'. We train our own BOOLQ model with BERT pre-training. It reaches
+
+a development accuracy of  $74.1\%$ . This model only predicts two classes 'yes' and 'no'.
+
+MNLI. The MultiNLI corpus (Williams et al., 2018) is a large corpus for textual entailment. It consists of premise-hypothesis sentence pairs which are marked with three target classes 'entailment', 'contradiction' and 'neutral'. There are 392K train, 9K dev., and 9K test examples.
+
+Although not applicable to all indirect answers, semantic consequence can be leveraged for interpreting some of them. For the question (Q) 'Do you like Italian food?', consider two possible answers (A) 'I love Tuscan food.' and (B) 'I prefer Mexican cuisine'. Let  $\mathbf{Q}'$  be the declarative (positive) sentence derived from Q i.e.  $\mathbf{Q}' = \mathbf{I}$  like Italian food'.  $(\mathbf{Q}')$  can be obtained by inverting the subject and auxiliary, and changing pronoun to first person). The meaning of A and B above, can then be obtained from an entailment system:  $\mathrm{A} \Longrightarrow \mathrm{Q}'$ , hence 'yes', while B contradicts  $\mathbf{Q}'$ , hence 'no'.
+
+We thus obtain predictions from an MNLI system, and map the predicted three NLI classes in a post-processing step: 'contradiction'  $\rightarrow$  'no', 'entailment'  $\rightarrow$  'yes', and 'neutral'  $\rightarrow$  'in the middle'. Note that this approach cannot predict all the classes in the corpus. Before prediction, we rewrite our questions into declarative form using syntactic rules on a constituency parse. Performance is much worse without this rewriting. Our models for the MNLI task start from a BERT checkpoint and reach a development accuracy of  $84\%$ .
+
+This model improves with the syntactic rewriting of questions which we are able to do fairly accurately. We based the rewrite rules on 50 questions. On a different set of 50 questions, 38 were rewritten fully accurately (manual inspection). Some errors arose from incorrect parsing and some are due to deficient rules. For example, we do not handle verb re-inflection. So 'Did you enjoy the movie?' gets rewritten into 'I did enjoy the movie.' rather than 'I enjoyed the movie.' This rewriting technique helped only the MNLI model. For other finetuning based models, which involve training, the models are able to learn from the question form itself.
+
+Majority baseline. This method predicts the most frequent label, 'yes', for all examples.
+
+# 5.2 Training with Question or Answer only
+
+Answer only. In many cases, the answer to a question suffices for predicting the label (see Table
+
+1 and Section 3.3); for example "I would like that." or "I wish!". This Answer only model fine-tunes BERT to predict the class based only on the answer. Similar experiments are done on NLI datasets to test if the hypothesis alone is at times sufficient for entailment prediction (Poliak et al., 2018; Gururangan et al., 2018). Such results are problematic for entailment (since it is defined to depend on the truth of the premise). In contrast, our problem is primarily about the meaning of answers. This experiment will provide insight into the cues within indirect responses, an aspect not understood so far.
+
+Question only. Some questions commonly elicit certain answers (see Section 3.8). These models test how well the question predicts the label.
+
+# 5.3 Question-Answer Pair models
+
+These models take both the question and the answer. They all finetune BERT checkpoints, and the question-answer pair is passed with a separator.
+
+BERT-YN is BERT finetuned only on our Circas corpus (YN).
+
+We also explore how to transfer the strength of parameters learned for three related inference tasks.
+
+BERT-BOOLQ-YN finetunes a BOOLQ model checkpoint (see Section 5.1) on our corpus, with a new output layer. Since BOOLQ is a Yes/No question answering system, even if developed for a different domain, we expect to learn many semantics of yes/no answers from this data.
+
+BERT-MNLI-YN is first fine-tuned on the MNLI corpus, followed by our YN data. This configuration tests if the signals we hoped to capture with the out-of-the-box MNLI model (Section 5.1) can be strengthened by training on our target task.
+
+BERT-DIS-YN. As discussed, indirect answers also have discourse relations with the speaker's intent (Green and Carberry, 1999). We implement this idea via a discourse connective prediction task. Consider again the texts from Section 5.1: The likely connective between  $Q'$  and A is 'because' as in 'I like Italian food [because] I love Tuscan food.' For  $Q'$  and B, 'but' would be more reasonable: 'I like Italian food [but] I prefer Mexican cuisine.' We hypothesize that these discourse relations will help learn the yes/no meaning via transfer learning.
+
+We use 400K examples (matching the MNLI data size) of explicit connectives and their arguments (a subset of Nie et al. (2019)). We aim to predict the 5 connectives (because, but, if, when,
+
+and) based on their arguments. This task itself can be done with a development accuracy of  $82\%$ . The best checkpoint is then finetuned on YN data.
+
+# 6 Experiments
+
+We use pre-trained BERT models (with 12 layers, 768 hidden units, and 12 attention heads, 110M parameters) for all our experiments. The experiments were done on a single Cloud TPU, and finetuning on our corpus takes under 30 minutes.
+
+# 6.1 Setup and Hyperparameter Tuning
+
+For the base models (MNLI, BOOLQ, and DIS), we tuned the learning rate (values 5e-5, 3e-5, 2e-5), the number of epochs (2, 3, 4), and train batch size (16, 32) in an exhaustive combination. For finetuning on yes/no data, we tune the learning rate while setting the epochs to 3 and training batch size to 32. We also perform three random restarts in each configuration. Performance was stable across the restarts (accuracy variation  $\leq 1\%$ ). So we take the best model on the development set as our final model. The best hyperparameters are in the Appendix. For the unmatched setting, we do 10 leave-one-out tasks. Here we take the best parameters from the matched setting, and use that configuration to train all 10 experiments.
+
+# 6.2 Results
+
+We report the development accuracy and detailed test results in Table 9 (RELAXED setting) and Table 10 (STRICT setting). For the unmatched setting, we report the mean accuracy and standard deviation across the 10 folds, and the min and max values. We first discuss results for the RELAXED labels. The findings for STRICT are similar.
+
+The majority baseline ('yes' class) leads to an accuracy of  $49\%$ . The MNLI to yes/no label mapping (no finetuning) is reasonable in terms of F-score for the 'no' class, but is poor for 'yes'. BOOLQ is the best baseline with  $63\%$  accuracy. However, there is no recall for labels other than 'yes' and 'no'.
+
+The question-only and answer-only results are noteworthy. The question-only model outperforms the majority baseline. On the other hand, the answer text contains strong signals, with  $82\%$  accuracy, or about  $20\%$  better than the best baseline.
+
+But models using both question and answer reach  $5 - 6\%$  greater accuracy and significantly outperform the answer-only model (McNemar's test, p-values  $< 1\mathrm{e} - 6$ ). As expected, these joint models
+
+are necessary when a string is a possible answer to multiple questions. An answer-only model is easily misled in these cases, as the examples below show:
+
+(1) "Is there something you absolutely won't
+
+eat?" "I really dislike bananas."
+
+Answer-only prediction: 'No'
+
+Question+Answer prediction: 'Yes'
+
+(2) "Do you need a nap?"
+
+"I have plenty of energy."
+
+Answer-only prediction: 'Yes'
+
+Question+Answer prediction: 'No'
+
+The best F-scores are obtained by an MNLI transfer task, reaching  $88.2\%$  accuracy in the matched setting. But it is not significantly better than a no-transfer BERT-YN model (McNemar's test).
+
+The unmatched setting shows that the models are worse when a scenario has not been seen in the training data. While it may not be possible for every conversational system to generalize across scenarios, a semantic classification such as yes/no should ideally be robust to such changes. Instead we see a  $6 - 10\%$  accuracy gap between the in-scenario test accuracies, and minimum out-of-scenario accuracy. The best accuracies are reached by a MNLI model. The highest accuracy on a scenario is  $90\%$  ('music preferences', 'weekend activities'), and lowest is  $82\%$  ('buying a flat in New York' and 'switching jobs'). The latter scenarios are quite different than the rest, indicating scope for improving the models.
+
+The 'in the middle' class has much worse results compared to the rest. The class has low frequency but also comprises responses of different types. Uncertain responses such as 'I am not sure.' appear easy to classify. But responses which do not take a stance: 'Do you know if it's raining outside? I'm prepared regardless.' are harder. Sometimes, the interpretation is left to the judgement of the listener. Eg. 'travelling an hour away' could be interpreted as 'far away' or 'close by' depending on the context and perceptions of the listener. These cases need models to deeply connect the question and answer, and are missed by our technique.
+
+The general trends for STRICT labels is similar: The best accuracy is again reached with MNLI pretraining. It is  $85\%$  for the matched case, a small but significant gain over BERT-YN (McNemar's test, p-value  $< 0.02$ ). The accuracy is  $10\%$  lower  $(74\%)$  for hardest experiment in leave-one-out.
+
+For STRICT labels, the 'probably no' class is hardest to predict even though it is as frequent as 'probably yes' and close to double the size of the 'in the middle' class. We found that 'probably no'
+
+<table><tr><td rowspan="3">Model</td><td colspan="6">Matched setting</td><td colspan="4">Unmatched setting</td></tr><tr><td rowspan="2">Accuracy Dev.</td><td rowspan="2">Test</td><td colspan="3">Test F-Score</td><td rowspan="2">Mid</td><td colspan="4">Test Accuracy</td></tr><tr><td>Yes</td><td>No</td><td>C.Yes</td><td>Mean</td><td>Std.</td><td>Min.</td><td>Max.</td></tr><tr><td colspan="11">Baselines (no finetuning)</td></tr><tr><td>Majority class</td><td>50.2</td><td>49.3</td><td>66.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>50.4</td><td>4.3</td><td>43.6</td><td>56.8</td></tr><tr><td>MNLI</td><td>28.4</td><td>28.9</td><td>34.4</td><td>52.8</td><td>0.0</td><td>6.9</td><td>28.1</td><td>2.8</td><td>24.2</td><td>34.1</td></tr><tr><td>BOOLQ</td><td>64.2</td><td>62.7</td><td>71.1</td><td>59.6</td><td>0.0</td><td>0.0</td><td>63.3</td><td>2.7</td><td>58.3</td><td>66.5</td></tr><tr><td colspan="11">BERT finetuned on Question or on Answer</td></tr><tr><td>BERT-YN (Question only)</td><td>56.4</td><td>56.0</td><td>63.1</td><td>54.1</td><td>9.1</td><td>1.0</td><td>53.3</td><td>2.9</td><td>48.0</td><td>58.4</td></tr><tr><td>BERT-YN (Answer only)</td><td>83.0</td><td>81.7</td><td>83.9</td><td>80.3</td><td>88.9</td><td>18.6</td><td>80.1</td><td>5.8</td><td>71.4</td><td>87.8</td></tr><tr><td colspan="11">BERT finetuned on Question + Answer</td></tr><tr><td>BERT-YN</td><td>88.4</td><td>87.8</td><td>89.8</td><td>87.9</td><td>89.9</td><td>28.2</td><td>85.5</td><td>3.9</td><td>79.0</td><td>90.2</td></tr><tr><td>BERT-MNLI-YN</td><td>89.6</td><td>88.2</td><td>90.4</td><td>88.5</td><td>89.3</td><td>29.4</td><td>87.1</td><td>3.0</td><td>81.9</td><td>90.3</td></tr><tr><td>BERT-DIS-YN</td><td>88.0</td><td>87.4</td><td>89.4</td><td>87.4</td><td>90.0</td><td>35.2</td><td>85.5</td><td>3.5</td><td>78.9</td><td>89.4</td></tr><tr><td>BERT-BOOLQ-YN</td><td>87.7</td><td>87.1</td><td>89.0</td><td>86.9</td><td>89.6</td><td>30.9</td><td>85.3</td><td>3.7</td><td>78.8</td><td>89.4</td></tr></table>
+
+Table 9: Performance on the relaxed labels. The highest value in each column is in bold. For matched setting, we show dev. and test accuracies, and also F-scores for the 4 labels ('yes', 'no', 'conditional yes' and 'in the middle'). In unmatched setting, we report summaries of 10 leave-one-out experiments. BERT-YN is significantly better than 'Answer only' (McNemar's test,  $p$ -value  $< 1e-6$ ); BERT-MNLI-YN is not significantly better than BERT-YN.  
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">Accuracy</td><td colspan="6">Matched setting</td><td colspan="4">Unmatched setting</td></tr><tr><td>Dev.</td><td>Test</td><td>Yes</td><td>P.Yes</td><td colspan="2">Test F-Score</td><td>No</td><td>P.No</td><td>Mean</td><td>Std.</td><td>Min.</td><td>Max.</td></tr><tr><td colspan="13">Baselines (no finetuning)</td></tr><tr><td>Majority class</td><td>47.5</td><td>47.0</td><td>63.9</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>0.0</td><td>46.9</td><td>3.9</td><td>40.0</td><td>52.3</td></tr><tr><td>MNLI</td><td>26.3</td><td>27.4</td><td>36.6</td><td>0.0</td><td>0.0</td><td>53.0</td><td>0.0</td><td>4.9</td><td>26.4</td><td>3.2</td><td>21.7</td><td>32.7</td></tr><tr><td>BOOLQ</td><td>59.4</td><td>59.2</td><td>70.4</td><td>0.0</td><td>0.0</td><td>57.0</td><td>0.0</td><td>0.0</td><td>58.9</td><td>3.0</td><td>53.8</td><td>63.7</td></tr><tr><td colspan="13">BERT finetuned on Question or Answer</td></tr><tr><td>BERT-YN (Question)</td><td>53.7</td><td>52.8</td><td>62.3</td><td>3.2</td><td>19.7</td><td>51.1</td><td>0.0</td><td>4.7</td><td>49.4</td><td>4.0</td><td>41.9</td><td>56.7</td></tr><tr><td>BERT-YN (Answer)</td><td>77.3</td><td>77.8</td><td>82.5</td><td>49.5</td><td>90.2</td><td>77.3</td><td>16.2</td><td>26.9</td><td>75.8</td><td>5.8</td><td>65.4</td><td>82.8</td></tr><tr><td colspan="13">BERT finetuned on Question + Answer</td></tr><tr><td>BERT-YN</td><td>83.6</td><td>84.0</td><td>88.7</td><td>49.9</td><td>90.2</td><td>85.4</td><td>18.6</td><td>42.6</td><td>81.2</td><td>4.6</td><td>71.8</td><td>85.6</td></tr><tr><td>BERT-MNLI-YN</td><td>85.0</td><td>84.8</td><td>89.8</td><td>51.8</td><td>89.8</td><td>86.6</td><td>18.0</td><td>41.3</td><td>82.8</td><td>4.0</td><td>74.4</td><td>86.7</td></tr><tr><td>BERT-DIS-YN</td><td>83.8</td><td>83.3</td><td>87.9</td><td>50.2</td><td>90.5</td><td>84.1</td><td>21.2</td><td>50.8</td><td>81.5</td><td>4.5</td><td>73.1</td><td>86.3</td></tr><tr><td>BERT-BOOLQ-YN</td><td>83.1</td><td>83.4</td><td>88.2</td><td>51.2</td><td>89.1</td><td>84.5</td><td>22.1</td><td>43.7</td><td>81.1</td><td>4.3</td><td>73.3</td><td>85.8</td></tr></table>
+
+Table 10: Performance on the strict labels. The highest value in each column is in bold.
+
+BERT-YN is significantly better than 'Answer only' (McNemar's test,  $p$ -value  $< 1e-6$ ), and BERT-MNLI-YN is better than BERT-YN ( $p$ -value  $< 0.02$ ).
+
+examples are heavily  $(69\%)$  mis-predicted into the 'no' class. Utterances which explicitly convey uncertainty eg. ('I don't believe so') or comparison ('Is everything good? Not the greatest.') are somewhat easier to predict. On the other hand, in:
+
+"Have you ever bought a romance novel?" "I
+
+prefer to read horror books."
+
+Best model: "No"
+
+Gold standard: "Probably no"
+
+The speaker prefers horror genre, but it does not preclude ever buying a romance novel. These subtleties are understandably harder for systems.
+
+Overall, while MNLI based transfer learning has led to small improvements, incorporating the right information for the task remains a challenge.
+
+# 7 Conclusion
+
+We have presented a new dataset containing natural indirect yes/no answers, as well as other significant pragmatic moves in the form of conditionals and uncertain utterances. Our first approach towards automatic interpretation is promising, but there is a significant gap especially for examples outside training scenarios. Our model does not yet classify additional information in responses ('Dinner? Let's go for Italian.' indicates not only a 'yes' answer, but also a preference for Italian food.). Moreover, we have explored the phenomena in English. There are exciting avenues for multilingual work to account for language and cultural differences.
+
+# References
+
+Daniel Adiwardana, Minh-Thang Luong, David R So, Jamie Hall, Noah Fiedel, Romal Thoppilan, Zi Yang, Apoorv Kulshreshtha, Gaurav Nemade, Yifeng Lu, et al. 2020. Towards a Human-Like Open-Domain Chatbot. arXiv preprint arXiv:2001.09977.  
+Penelope Brown and Stephen C Levinson. 1978. Universals in Language Usage: Politeness Phenomena. In E.N. Goody, editor, Questions and Politeness: Strategies in Social Interaction, pages 56-311. Cambridge University Press.  
+Eunsol Choi, He He, Mohit Iyyer, Mark Yatskar, Wentau Yih, Yejin Choi, Percy Liang, and Luke Zettlemoyer. 2018. QuAC: Question Answering in Context. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2174-2184.  
+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.  
+Cristian Danescu-Niculescu-Mizil and Lillian Lee. 2011. Chameleons in Imagined Conversations: A New Approach to Understanding Coordination of Linguistic Style in Dialogs. In Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics, ACL 2011.  
+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.  
+Nancy Green and Sandra Carberry. 1999. Interpreting and Generating Indirect Answers. Computational Linguistics, 25(3):389-435.  
+Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel Bowman, and Noah A. Smith. 2018. Annotation Artifacts in Natural Language Inference Data. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 107-112.  
+Julia B. Hirschberg. 1985. A Theory of Scalar Implicature. Ph.D. thesis, University of Pennsylvania.  
+Beth Ann Hockey, Deborah Rossen-Knill, Beverly Spejewski, Matthew Stone, and Stephen Isard. 1997. Can you Predict Responses to Yes/No Questions? Yes, No, and Stuff. In Fifth European Conference on Speech Communication and Technology.
+
+Marie-Catherine de Marneffe, Scott Grimm, and Christopher Potts. 2009. Not a Simple Yes or No: Uncertainty in Indirect Answers. In Proceedings of the SIGDIAL 2009 Conference: The 10th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 136-143.  
+Marie-Catherine de Marneffe, Christopher D Manning, and Christopher Potts. 2010. Was it Good? It was Provocative. Learning the Meaning of Scalar Adjectives. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 167-176.  
+Allen Nie, Erin Bennett, and Noah D. Goodman. 2019. DisSent: Learning Sentence Representations from Explicit Discourse relations. In Proceedings of the 57th Conference of the Association for Computational Linguistics, Volume 1: Long Papers, pages 4497-4510.  
+Adam Poliak, Jason Naradowsky, Aparajita Haldar, Rachel Rudinger, and Benjamin Van Durme. 2018. Hypothesis-Only Baselines in Natural Language Inference. In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, pages 180-191.  
+Siva Reddy, Danqi Chen, and Christopher D Manning. 2019. CoQA: A Conversational Question Answering Challenge. Transactions of the Association for Computational Linguistics, 7:249-266.  
+Deborah Rossen-Knill, Beverly Spejewski, Beth Ann Hockey, Stephen Isard, and Matthew Stone. 1997. Yes/No Questions and Answers in the Map Task Corpus. Technical report, University of Pennsylvania.  
+Anna-Brita Stenström. 1984. Questions and Responses in English Conversation: Lund Studies in English. Claes Schaar and Jan Svartvik, editors, Lund Studies in English 68, CWK Gleerup, Malmo Sweden.  
+Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. 2019. GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding. the Proceedings of ICLR.  
+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.  
+Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2019. *DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation*. arXiv preprint arXiv:1911.00536.
+
+# A Hyperparameter settings
+
+For the base models (MNLI, BOOLQ, DIS), we tuned the learning rate (values 5e-5, 3e-5, 2e-5), the number of epochs (2, 3, 4), and train batch size (16, 32) in an exhaustive combination of these parameters. The best performance on development data was obtained with the following settings:
+
+<table><tr><td>Model</td><td>learning rate</td><td>no. epochs</td><td>batch size</td></tr><tr><td>MNLI</td><td>2e-5</td><td>3</td><td>16</td></tr><tr><td>BOOLQ</td><td>3e-5</td><td>4</td><td>16</td></tr><tr><td>DIS</td><td>2e-5</td><td>2</td><td>32</td></tr></table>
+
+For finetuning on yes/no data (matched setting), we tune the learning rate while setting the number of epochs to 3, and training batch size to 32. We also perform three random restarts in each configuration. Performance was stable across the restarts (accuracy variation  $\leq 1\%$ ). We take the best model on the development set as our final model. The chosen learning rates are in the table below:
+
+<table><tr><td>Model</td><td>STRICT</td><td>RELAXED</td></tr><tr><td>BERT-YN (Question only)</td><td>2e-5</td><td>3e-5</td></tr><tr><td>BERT-YN (Answer only)</td><td>2e-5</td><td>2e-5</td></tr><tr><td>BERT-YN</td><td>3e-5</td><td>3e-5</td></tr><tr><td>BERT-MNLI-YN</td><td>2e-5</td><td>5e-5</td></tr><tr><td>BERT-DIS-YN</td><td>3e-5</td><td>5e-5</td></tr><tr><td>BERT-BOOKY-YN</td><td>3e-5</td><td>5e-5</td></tr></table>
+
+For the unmatched setting, we do 10 leave-one-out experiments. Here we use the best parameters from the matched setting, and use the same configuration for training in all 10 experiments.
+
+# B Annotation Instructions
+
+We now detail the complete instructions to annotators, along with interface examples, prompt texts, and practice items.
+
+Question collection. In this step, an annotator is shown a scenario, and asked to provide five yes/no questions. The instructions, and interface for an example item are in Figure 1. The descriptions used for each of the 10 dialog scenarios are in Table 11.
+
+Answer elicitation. Similarly, the instructions and interface for collecting answers are in Figure 2. Note that we show 5 questions on each screen, to reduce annotation time. The 5 questions are taken from the same scenario but such that they are not too similar. We enforce non-redundancy by keeping the pairwise similarity between any two questions on the same screen to less than 0.35 (measured by cosine similarity between the adjectives and nouns in the questions).
+
+We have the same 10 scenarios but their descriptions are changed slightly to suit to the answer elicitation task. These prompts are in Table 12.
+
+Marking interpretations. Finally, the question-answer pairs are annotated with meaning categories. Our complete instructions and annotator interface are in Figure 3. Again, the scenario descriptions are modified for the task, and are given in Table 13.
+
+This step is fairly complex, so every annotator also worked through 8 practice questions before starting the annotation. After they answered them, the correct answers were shown along with an explanation. These examples are given in Table 14.
+
+![](images/ca773dd8078ca3004ee67e611116189e6795df21935815a8011a00404cafe88a.jpg)  
+(a) Instructions  
+Figure 1: Annotator interface for question collection.
+
+![](images/60a2b496032f68d95ffe133185e2f1e6048b0b5fb3e1b0414b0fc78c7991c5d7.jpg)  
+(b) Example item
+
+(1) Suppose that you are trying to learn about a friend's food preferences, so that you can recommend a local restaurant, but can only ask yes/no questions. Provide 5 useful questions that can be answered "Yes" or "No".  
+(2) Suppose that you are trying to learn what sorts of activities a friend likes to do during the weekend, so that you can recommend local activities that might interest them, but can only ask yes/no questions. Provide 5 useful questions that can be answered "Yes" or "No".  
+(3) Suppose that you are trying to learn what sorts of books someone likes to read, but are only allowed to ask yes/no questions. Provide 5 questions you might ask about books that can be answered "Yes" or "No".  
+(4) Suppose that you are meeting your new neighbour for the first time. You want to find out more about him/her, but you are only allowed to ask yes/no questions. Provide 5 questions you might ask him/her that can be answered "Yes" or "No".  
+(9) Suppose that your friend tells you he is thinking of buying a flat in New York. In this context, provide 5 questions you might ask him/her. You are only allowed to ask yes/no questions.
+
+(5) On a Friday evening, you are leaving work and see your friend (and colleague) also at the door ready to leave. Provide 5 questions you might ask him/her that can be answered "Yes" or "No".  
+(6) Suppose that you are trying to learn about a friend's interests related to music. For instance, you could ask about music tastes, instruments played, events they go to, etc. You are only allowed to ask yes/no questions. Provide 5 questions you might ask.  
+(7) Your friend has arrived from out of town to visit you. Provide 5 questions you might ask your friend when he/she arrives and during your time together. You are only allowed yes/no questions.  
+(8) Suppose that you are at a cafe, and you run into your childhood neighbour. You haven't seen each other or had any contact for many years. Ask a few questions of your childhood neighbour. You are only allowed yes/no questions. Provide 5 questions you might ask.  
+(10) Your friend (not a colleague) is considering switching his/her job. You do not know much about the aspects of his/her job. Ask a few yes/no questions in this context. Provide 5 yes/no questions.
+
+Table 11: Descriptions of the 10 scenarios in the question collection step.
+
+You will be given a social situation, for example, talking to your friend or neighbour.
+
+TASK: You will be asked to respond to a question from your friend/neighbour but without using the words 'yes' or 'no' (or similar words like 'yeah', etc). Please provide a possible answer, it does not have to be your real opinion.
+
+EXAMPLE: Here are three such answers to your friend's question about movies.
+
+"Do you like movies with sad endings?"  
+"I often watch them." (Meaning = Yes)  
+"I prefer movies which make me laugh." (Meaning = No)  
+"When the plot is also good." (Meaning = Sometimes)
+
+# IMPORTANT:
+
+(1) You will be shown multiple questions for the same situation. For each question, provide one answer.  
+(2) Answer each question independently. The set of questions are not part of the same conversation.  
+(3) We are looking for variety in your answers, using various words, and conveying different meanings (yes, no, maybe, sometimes, often)  
+(4) Remember that the setting is a conversation with a friend (or neighbour or colleague). Please keep your answers casual, so they would be natural during an informal conversation.
+
+(a) Instructions
+
+Imagine that you are talking to a friend. Your friend asks you a question to know more about your food preferences.
+
+TASK: Answer your friend's question without using the words 'yes' or 'no'.
+
+1) Do you enjoy foreign cuisine?  
+2) Do you like rice?
+
+3) Are you willing to spend some money?  
+4) Before we go any further I have to know, do you like lasagna?  
+5) Did you enjoy the all you can eat food variety we had last week?
+
+SUBMIT
+
+(b) Example item
+
+Figure 2: Annotator interface for answer elicitation.
+
+(1) Imagine that you are talking to a friend. Your friend asks you a question to know more about your food preferences. Answer your friend's question without using the words 'yes' or 'no'.  
+(2) Imagine that you are talking to a friend. Your friend asks you a question to know more about what activities you like to do during weekends. Answer your friend's question without using the words 'yes' or 'no'.  
+(3) Imagine that you are talking to a friend. Your friend asks you a question to know more about what sorts of books you like to read. Answer your friend's question without using the words 'yes' or 'no'.  
+(4) Imagine that you have just moved into a neighbourhood. One of your new neighbours is a friendly person, and asks you a question. Answer your neighbour's question without using the words 'yes' or 'no'.  
+(9) Imagine that you have just told your friend that you are thinking of buying a flat in New York. Your friend asks you a question to know more about your plans. Answer your friend's question without using the words 'yes' or 'no'.
+
+(5) Imagine that at the end of the week you are leaving work and see your friend (and colleague) also at the door ready to leave. Your friend asks you a question. Answer your friend's question without using the words 'yes' or 'no'.  
+(6) Imagine you are talking to a friend. Your friend asks you a question to know more about your interests related to music. Answer your friend's question without using the words 'yes' or 'no'.  
+(7) Imagine that you have just travelled from a different city to visit your friend. Upon your arrival, your friend asks you a question. Answer your friend's question without using the words 'yes' or 'no'.  
+(8) Imagine that you run into your childhood neighbour at a cafe. You haven't seen each other or had any contact for many years. Your childhood neighbour asks you a question. Answer your neighbour's question without using the words 'yes' or 'no'.  
+(10 Imagine that you have just told your friend that you are considering switching your job. Your friend asks you a question to know more about your plans. Answer your friend's question without using the words 'yes' or 'no'.
+
+Table 12: Descriptions of the 10 scenarios as used in the answer elicitation step.
+
+# Please read these instructions carefully:
+
+You will be shown short dialogues between two
+
+friends/colleagues (X and Y) in a certain context. For example:
+
+Context : X wants to know about Y's movie preferences.
+
+X : "Do you like movies with sad endings?"  
+Y : "I often watch them."
+
+In all the dialogues, X asks a simple 'Yes/No' question, and Y answers it with a short sentence or a phrase.
+
+# Task: We need your help to interpret Y's answer
+
+Read the dialog and tell us how you think X will interpret Y's answer. Your options are:
+
+X will think that Y means
+
+1. Yes  
+2. Probably yes' / 'sometimes yes  
+3. Yes, subject to some conditions  
+4. No  
+5. Probably no  
+6. In the middle, neither yes nor no  
+7. I am not sure how X will interpret Y's answer
+
+If Y's response does not fit any of the above, please choose the option 'Other' and leave a short comment.
+
+# Example annotation:
+
+Context : X wants to know about Y's movie preferences.
+
+X : “Do you like movies with sad endings?”  
+Y : "I often watch them."
+
+X will think that Y means (1) Yes
+
+Previous
+
+![](images/1e7253bfdd6416962a0818230d57a08f889e3df1aaa22d903d902906ed651a5c.jpg)  
+(a) Instructions
+
+2 Next
+
+# Note (please read carefully):
+
+X and Y are friends. All their conversations are in casual contexts.  
+- The `Context' is provided to tell you about the situation where the dialogue happened.  
+- You should assume that Y has good intentions and always provided a relevant answer to the question  
+Think only about how X will interpret Y's answer in the given context and to the given question. Do not make a choice based on what the answer means in general.  
+- Use 'probably yes/sometimes yes' when the answer is not a definite yes. Similarly for 'probably no'.  
+- Use 'yes, subject to some conditions' only if X will not interpret a 'yes' without that condition being satisfied.  
+- Use the options 'probably yes', 'sometimes yes' and 'yes, subject to some conditions' only when X will not interpret the answer as a definite 'yes'.
+
+Eg. X: "Are you up for a movie?"
+
+Y: "Only to a comedy."
+
+X will think that Y means
+
+(3) yes, subject to some conditions.
+
+In contrast, in the following, Y means 'yes' even if his/her answer also contains a condition.
+
+X: "Have you ever seen a Star Wars movie?"
+
+Y : "Only every time I'm persuaded by my Star Wars fan wife."
+
+X will think that Y means
+
+(1) yes (because Y has seen at least one Star Wars movie)
+
+If the dialog contains offensive, obscene or sensitive content, please mark the checkbox provided. Such examples will be rare if any.
+
+Previous
+
+![](images/5f4a6ab6965990c1f219866f25b9a4cfd29d9bb5d0ba384edca5438e92768c60.jpg)  
+(b) Additional notes
+
+![](images/62c187ec03bc5330d4a5bb2035ccc7147b3e76d2b7ecc6475bc8a7a6c44a4f77.jpg)
+
+![](images/0931b421fdb018e1226485b2791e6f5bad907ca08ad1e2c24dffe8ad2bf2fa32.jpg)
+
+![](images/a0a4128fd3b809b6dca1fb7bd1ecca10eab17743f2e46d731ee287a9f5640cc1.jpg)  
+(c) Example item  
+Figure 3: Annotator interface for marking interpretation.
+
+(1) X wants to know about Y's food preferences.  
+(2) X wants to know what activities Y likes to do during weekends.  
+(3) X wants to know what sorts of books Y likes to read.  
+(4) Y has just moved into a neighbourhood and meets his/her new neighbour X.  
+(9) Y has just told X that he/she is thinking of buying a flat in New York.
+
+(5) X and Y are colleagues who are leaving work on a Friday at the same time.  
+(6) X wants to know about Y's music preferences.
+
+(7) Y has just travelled from a different city to meet X.  
+(8) X and Y are childhood neighbours who unexpectedly run into each other at a cafe.  
+(10) Y has just told X that he/she is considering switching his/her job.
+
+Table 13: Descriptions of the 10 scenarios as used in the interpretation marking step.
+
+(1) Context: X wants to know about Y's food preferences.  
+X: "Do you eat red meat?"  
+Y: "I am a vegetarian."  
+Answer: No (Vegetarians do not eat meat. So X will interpret it as a 'no' answer.)  
+(2) Context: X wants to know about Y's food preferences.  
+X: "Have you had Bulgogi?"  
+Y: "I am not sure."  
+Answer: In the middle, neither yes nor no. (Here Y's response is non-committal. So X's best option is to interpret it as neither a 'yes' nor a 'no')  
+(3) Context: Y has just travelled from a different city to meet X.  
+X: “Did you stop anywhere on the way?”  
+Y: “I had to get gas a few times.”  
+Answer: Yes (X will infer a definite ‘yes’ because Y stopped for gas.)  
+(4) Context: Y has just travelled from a different city to meet X.  
+X: “Did you stop anywhere on the way?”  
+Y: “I try to incorporate my errands into every trip I make.”  
+Answer: Probably yes / sometimes yes (In response to the question, Y mentions that his general tendency is to do errands on the way. So it is likely that he did the same on this trip. Hence ‘probably yes’)  
+(5) Context: X wants to know about Y’s music preferences.  
+X: “Would you go to a punk rock show?”  
+Y: “It depends on who is playing.”  
+Answer: Yes, subject to some conditions (Y might go depending on the artist who is performing. Hence X will interpret that Y is posing a condition for going to the punk show.)  
+(6) Context: X wants to know about Y's music preferences.  
+X: "Would you go to a punk rock show?"  
+Y: "I think I'd enjoy something less chaotic a little more."  
+Answer: Probably no (Y's indicates that he'd prefer some other activity. But he does not completely rule out the possibility of going to a punk show. Hence X will interpret it as 'probably no').  
+(7) Context: X and Y are colleagues who are leaving work on a Friday at the same time.  
+X: "Is your department busy this time of year?"  
+Y: "We usually end up working overtime."  
+Answer: Yes (Since Y is regularly working overtime, X will infer that Y's department is busy.)  
+(8) Context: X and Y are colleagues who are leaving work on a Friday at the same time.  
+X: "Is your department busy this time of year?"  
+Y: "Just as busy as the rest of the year."  
+Answer: In the middle, neither yes nor no (Since it is not known how busy Y's department is in general, it is unclear if Y is busy at this time. In such cases pick 'in the middle, neither yes nor no'. Note that if X had background knowledge that Y's department is usually never busy, he could interpret the answer as a 'no' or vice versa. You do not have to assume such information is available when providing your answer.)
+
+Table 14: The 8 practice questions used to train annotators for marking interpretations.

idratherjustgotobedunderstandingindirectanswers/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:363c6e51059961e607483149baca5ed78b75c386b533159656033a69761de1c0
+size 637166

idratherjustgotobedunderstandingindirectanswers/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c141432299dffb2a0433af80b86bcffd4cd2a35d09bd8f37a3e989ca3c73a112
+size 597235

whatdoyoumeanbythataparserindependentinteractiveapproachforenhancingtexttosql/c25c90ea-aaa3-480a-b3c2-a9760557a486_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:18293a17d1e425327f46fe9cfd917f001861ea5177a49a8a2802ecb04e5c92d4
+size 69340

whatdoyoumeanbythataparserindependentinteractiveapproachforenhancingtexttosql/c25c90ea-aaa3-480a-b3c2-a9760557a486_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84635d0d55242b2e61bb2909f34de127c579e7250656f058a5a3060615274f61
+size 83376

whatdoyoumeanbythataparserindependentinteractiveapproachforenhancingtexttosql/c25c90ea-aaa3-480a-b3c2-a9760557a486_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a27b5a5c8b49d66cbd1108eea0552c51896e74a1cbf10565838d62f24ec2c26f
+size 943925

whatdoyoumeanbythataparserindependentinteractiveapproachforenhancingtexttosql/full.md

@@ -0,0 +1,269 @@
+# "What Do You Mean by That?" A Parser-Independent Interactive Approach for Enhancing Text-to-SQL
+
+Yuntao Li $^{1*}$ , Bei Chen $^{2}$ , Qian Liu $^{3*}$ , Yan Gao $^{2}$ , Jian-Guang Lou $^{2}$ , Yan Zhang $^{1}$ , Dongmei Zhang $^{2}$
+
+$^{1}$ Department of Machine Intelligence, Peking University, Beijing, China
+
+<sup>2</sup>Microsoft Research, Beijing, China; <sup>3</sup>Beihang University, Beijing, China
+
+$^{1}\{\mathrm{li.} \mathrm{yt}, \mathrm{zhyzhy001}\} @ \mathrm{pku.edu.cn}; ^{3} \mathrm{qian.liu} @ \mathrm{buaa.edu.cn}$
+
+2{beichen, yan.gao, jlou, dongmeiz} @microsoft.com
+
+# Abstract
+
+In Natural Language Interfaces to Databases systems, the text-to-SQL technique allows users to query databases by using natural language questions. Though significant progress in this area has been made recently, most parsers may fall short when they are deployed in real systems. One main reason stems from the difficulty of fully understanding the users' natural language questions. In this paper, we include human in the loop and present a novel parser-independent interactive approach (PIIA) that interacts with users using multichoice questions and can easily work with arbitrary parsers. Experiments were conducted on two cross-domain datasets, the WikiSQL and the more complex Spider, with five state-of-the-art parsers. These demonstrated that PIIA is capable of enhancing the text-to-SQL performance with limited interaction turns by using both simulation and human evaluation.
+
+# 1 Introduction
+
+The past few years have witnessed a burgeoning interest in the study of text-to-SQL, the essential technique for Natural Language Interfaces to Databases (NLIDB) systems (Guo et al., 2019; Hwang et al., 2019; He et al., 2019a; Bogin et al., 2019a,b). By converting natural language (NL) questions into executable forms (i.e., Structured Query Language or SQL), text-to-SQL parsers relieve users from the burden of learning about techniques behind the queries. Though significant progress has been made in this field, most parsers are still less than desirable when deployed in real NLIDB systems. As users are not experts in database querying, a central challenge for the parsers is to fully understand the users' NL questions.
+
+Since users are who know the questions best, interacting with them has been seen as a promising
+
+way to tackle the above challenge in real NLIDB systems. Early works tried to get users involved in checking SQL queries (Li and Jagadish, 2014; Iyer et al., 2017; Yaghmazadeh et al., 2017), which are impracticable in real systems, as they can only succeed if users have a very good knowledge of SQL. In another attempt to involves users, Gur et al. (2018) proposed to interact with non-expert users by multi-choice questions. However, this approach is designed for relatively simple scenarios and cannot be easily applied to more complex ones. More recently, Yao et al. (2019) took an important step forward by measuring uncertainty of neural-based parsers and altering the behavior of them. However, as far as we know, most parsers in real systems are equipped with elaborate rules instead of using fully neural methods (Dhamdhere et al., 2017; Gliozzo et al., 2013; Lai et al., 2014). Moreover, in some situations, parsers are supplied by third parties, making it impossible to alter them. Therefore, assuming parsers are a black box, it is indispensable to conduct research on an interactive approach for enhancing the text-to-SQL technique in complex scenarios.
+
+In this paper, we propose a Parser-Independent Interactive Approach (PIIA) to interact with human users and help parsers better understand NL questions. To achieve this goal, we devised three modules: (1) Error Locator employs an alignment method to help parsers locate uncertain tokens in the NL questions. (2) Question Generator designs multi-choice questions in natural language for users, which offers a pleasant interactive experience. (3) NL Modifier rewrites the NL questions according to the users' feedback and produces more legible questions to facilitate downstream parsing. Our major contributions are:
+
+- We propose a novel interactive approach, named PIIA, to enhance the text-to-SQL for complex
+
+![](images/b4f3cca5c3860db0082245d4980dda76da41a8de5c95ac8f5b1197e8c8069f77.jpg)  
+Figure 1: The schema of PIIA, consisting of Error Locator, Question Generator and NL Modifier.  
+Figure 2: The grammar of the intermediate language. In a specific database, column refers to distinct column names while table comprises several table names and value indicates the value tokens expressed by the user.
+
+SQL queries in a cross-domain scenario.
+
+- The interaction process in PIIA is user-friendly that asks multi-choice questions and reduces the number of questions as much as possible.  
+PIIA is designed as a parser-independent approach that can easily collaborate with arbitrary base parsers and be deployed in real systems.  
+- We conduct a series of experiments with five base parsers on two large cross-domain datasets that demonstrate the effectiveness of PIIA by using both simulation and human evaluation.
+
+# 2 Methodology Overview
+
+While querying databases in an NLIDB system, users pose a natural language question that is denoted as  $\mathbf{x}$ . The text-to-SQL parser takes  $\mathbf{x}$  as input and predicts a SQL query, which is denoted as  $\mathbf{y}$ . The system then executes the predicted SQL and returns the result. As mentioned, users are not experts in database querying, so they may pose natural language questions with inexplicit expressions. To better understand the difficulties caused by inexplicit expressions, we carefully analyzed 300 mistakes made by IRNet (Guo et al., 2019), one of the state-of-the-art parsers on the Spider dataset (Yu et al., 2018b). We found that in  $47.3\%$  of the cases the parser couldn't understand database-related information, such as table and column names as well as values. Thus, we build PIIA upon parsers that can interactively revise inexplicit expressions in  $\mathbf{x}$  with the help of users' feedback, thus enhancing the performance of text-to-SQL.
+
+Our proposed PIIA, shown schematically in Figure 1, works between the user and the text-to-SQL parser and it consists of three modules, Error Locator, Question Generator and NL Modifier. After receiving  $\mathbf{x}$  and  $\mathbf{y}$ , the Error Locator helps the
+
+```txt
+Z := intersect  $R$ $R$  | union  $R$ $R$  | except  $R$ $R$  |  $R$ $R \coloneqq$  Select Filter Order | Select Filter | Select Order | Select  
+Select := A | AA | AA A | AA AA | AA AA A A  
+Filter := Filter and Filter | Filter or Filter | = AV | > AV | < AV | ≥ AV | ≤ AV | ≠ AV | between AV V  
+Order := asc A | asc A limit number | dec A | dec A limit number  
+A := none CT | max CT | min CT | count CT | sum CT | avg CT  
+C := column T := table V := value | R
+```
+
+parser find a set of uncertain tokens in  $\mathbf{x}$ . For each uncertain token, the Question Generator creates a natural language multi-choice question. After interactively asking the user all the multi-choice questions, the PIIA agent collects all the answers (i.e., the user's selections). Then, the NL Modifier corrects  $\mathbf{x}$  based on the answers and obtains a more legible question  $\hat{\mathbf{x}}$ . Finally, by feeding modified question  $\hat{\mathbf{x}}$  into the text-to-SQL parser, we can get a new predicted SQL query  $\hat{\mathbf{y}}$ . Compared to  $\mathbf{x}$ ,  $\hat{\mathbf{x}}$  combines user's feedback and contains clearer semantics. Hence  $\hat{\mathbf{y}}$  is likely more accurate than  $\mathbf{y}$ . Note that, our PIIA agent will not read the contents of the databases (i.e., values) due to privacy concerns. Details of the three modules are presented in Sections 3, 4 and 5.
+
+# 3 Error Locator
+
+The goal of Error Locator is to detect the tokens in  $\mathbf{x}$  that are hardly understood by the parser. These are called Uncertain Tokens, and in most cases are related to database information, such as table and column names as well as values. When the parser fails to understand them, the uncertain tokens will be mistranslated or ignored. Since the parser is a black box, the only information we can get from it is the predicted SQL. Hence, we devise a method to compare the NL question  $\mathbf{x}$  with the predicted SQL  $\mathbf{y}$ . All the informative tokens in  $\mathbf{x}$  may align with the corresponding tokens in  $\mathbf{y}$ , while the unaligned tokens in  $\mathbf{x}$  are extracted as uncertain tokens. To this end, we firstly restate the predicted SQL  $\mathbf{y}$  to an NL question  $\mathbf{x}'$  through SQL-to-Text Restatement, and then perform Token-to-Token Alignment between  $\mathbf{x}$  and  $\mathbf{x}'$ .
+
+![](images/2d584082ee83e2d544220a80296a10555782d548f94d19a41dbdb889d9bc375e.jpg)  
+Figure 3: Illustration of Error Locator with a real case from IRNet on the Spider dataset. The NL question  $\mathbf{x}$  is parsed to SQL  $\mathbf{y}$ , and then converted to restated NL question  $\mathbf{x}'$  via SQL-to-text restatement. Then, a token-to-token alignment similarity matrix between  $\mathbf{x}$  and  $\mathbf{x}'$  is computed to detect uncertain tokens (i.e., cat and aged).
+
+# 3.1 SQL-to-Text Restatement
+
+Compared to aligning an NL question with a SQL query, the alignment of two NL questions is more reasonable because it utilizes a similar linguistic structure and can make better use of pre-trained models (e.g., BERT). Thus, before the alignment, we restate the predicted SQL y into a natural language question  $\mathbf{x}^{\prime}$ . Previous work has proposed sequence-to-sequence based methods (Guo et al., 2018) that cannot ensure the restatement correctness. In contrast, we carefully design a template-based SQL-to-text method that solves this problem because the restatement correctness and integrality are critical for the alignment process.
+
+Since SQL is execution-oriented, its clauses are about database operations. Some of the clauses may not be expressed in the users' NL questions, such as GROUP BY and JOIN. To bridge the gap between natural language and SQL, we design an intermediate language that is inspired by Guo et al. (2019) and whose grammar is shown in Figure 2. The predicted SQL can be easily converted into the intermediate language, which can be naturally represented by a hierarchical tree structure. Each tree node corresponds to a grammar rule. For each kind of grammar rule, we design a few natural language templates to describe it. Thus, we can recursively convert the tree into a natural language question. As for nested SQL queries, we utilize subordinate clauses with "that/which" to handle the subqueries. The SQL-to-text restatement is depicted on the left of Figure 3, along with a concrete example. We also give some examples of templates in the figure. Finally, the restated  $\mathbf{x}^{\prime}$  has the same semantics
+
+of SQL y and is independent of database internal operations.
+
+# 3.2 Token-to-Token Alignment
+
+Given the user NL question  $\mathbf{x}$  and the restated NL question  $\mathbf{x}'$ , we perform the token-to-token alignment between them and find out the uncertain tokens in  $\mathbf{x}$ . As shown on the right of Figure 3, we adopt BERT (Devlin et al., 2019) as the encoder. BERT is pre-trained on a large corpus and equipped with the ability to encode sentences on the basis of contextual information. The input of BERT is the concatenation of the two questions with "[CLS]" and "[SEP)". Each token obtains an output vector from BERT, which is fed into a trainable Multi-Layer Perceptron (MLP) layer to further distill useful information. Assuming that  $\mathbf{x}$  has  $N$  tokens and  $\mathbf{x}'$  has  $M$  tokens, we can denote that  $\mathbf{x} = (x_{1}, x_{2}, \ldots, x_{N})$  and  $\mathbf{x}' = (x_{1}', x_{2}', \ldots, x_{M}')$ . The output embeddings of tokens in  $\mathbf{x}$  and  $\mathbf{x}'$  can respectively be denoted by
+
+$$
+H = \left(\mathbf {h} _ {1}, \mathbf {h} _ {2}, \dots , \mathbf {h} _ {N}\right) \quad \in \mathbb {R} ^ {d \times N}, \tag {1}
+$$
+
+$$
+U = \left(\mathbf {u} _ {1}, \mathbf {u} _ {2}, \dots , \mathbf {u} _ {M}\right) \quad \in \mathbb {R} ^ {d \times M},
+$$
+
+where  $d$  is the output embedding size.
+
+Based on  $H$  and  $U$ , we employ the cosine similarity to derive a token-level similarity matrix  $A \in \mathbb{R}^{N \times M}$ , where each entry  $A_{nm}$  indicates the similarity between  $x_{n}$  and  $x_{m}'$ :
+
+$$
+A _ {n m} = \frac {\mathbf {h} _ {n} ^ {\top} \cdot \mathbf {u} _ {m}}{\| \mathbf {h} _ {n} \| \cdot \| \mathbf {u} _ {m} \|}. \tag {2}
+$$
+
+The alignment between  $\mathbf{x}$  and  $\mathbf{x}'$  can be obtained using the similarity scores in  $A$ . After removing
+
+stop words and words from SQL-to-text restatement templates, we regard the similarity scores as the weights of a bipartite graph and apply the Hungarian maximum matching algorithm to find an optimized token-level one-to-one alignment. Finally, the tokens in  $\mathbf{x}$  with alignment scores less than the threshold  $p$  are extracted as uncertain tokens. It is also important to note that a schema-aware post-processing is operated on these scores. Since the tokens that appear more than once in the database schema may confuse the alignment process, the post-processing aims to give addition bias and help Error Locator detect potential uncertain tokens<sup>1</sup>.
+
+# 3.3 Training Process
+
+Since the annotations of token-to-token alignment are not available, fully supervised learning is infeasible. Inspired by the work of Legrand et al. (2016), we solve this problem by leveraging negative sampling to generate training data and adopt a weakly supervised training strategy. Concretely, we collect several pairs  $(\mathbf{x},\mathbf{x}_{pos}^{\prime})$ .  $\mathbf{x}$  is the user NL question and  $\mathbf{x}_{pos}^{\prime}$  is the corresponding positive restated NL question restated from the ground truth SQL of  $\mathbf{x}$ . For each  $(\mathbf{x},\mathbf{x}_{pos}^{\prime})$  pair, we generate negative restated NL questions  $\mathbf{x}_{neg}^{\prime}$  in two ways: random sampling, where we randomly pick an  $\mathbf{x}^{\prime}$  from other pairs as  $\mathbf{x}_{neg}^{\prime}$ ; and perturbed sampling, where we generate an  $\mathbf{x}_{neg}^{\prime}$  by replacing column names or/and value tokens in  $\mathbf{x}_{pos}^{\prime}$ . The random samples are vastly different from  $\mathbf{x}_{pos}^{\prime}$ , and the common tokens in positive and negative samples are uninformative (e.g., stop words). This kind of  $\mathbf{x}_{neg}^{\prime}$  helps distinguish the informative and uninformative tokens. The perturbed samples have the same uninformative tokens with  $\mathbf{x}_{pos}^{\prime}$ , and help model to focus on the alignment of informative tokens. We generate 50 random samples and 50 perturbed samples for each  $(\mathbf{x},\mathbf{x}_{pos}^{\prime})$  pair.
+
+By generating negative samples, we obtain the training data composed of triples  $(\mathbf{x},\mathbf{x}_{pos}^{\prime},\mathbf{x}_{neg}^{\prime})$ . The intuition behind the weakly supervised training is that  $\mathbf{x}$  is more similar to  $\mathbf{x}_{pos}^{\prime}$  than  $\mathbf{x}_{neg}^{\prime}$ . We measure the sentence-level similarity by averaging the token-level similarities:
+
+$$
+s \left(\mathbf {x}, \mathbf {x} ^ {\prime}\right) = \frac {1}{N} \sum_ {n = 1} ^ {N} \max _ {m = 1} ^ {M} A _ {n m}. \tag {3}
+$$
+
+![](images/293dffc70147e15bec60adb5a80d974bb072f5712da61b4ade8ca63e78a4dfce.jpg)  
+Figure 4: Question Generator and NL Modifier: an example. Shaded options in the multi-choice questions are selected by the user.
+
+Then, our goal is to increase  $s(\mathbf{x},\mathbf{x}_{pos}^{\prime})$  and decrease  $s(\mathbf{x},\mathbf{x}_{neg}^{\prime})$ . We employ hinge loss to maximize the margin of the two scores, which is also accompanied by  $L^{1}$ -norm on two corresponding similarity matrices to make them sparse. The loss function is:
+
+$$
+\begin{array}{l} L = \max  \left(0, m - \left(s \left(\mathbf {x}, \mathbf {x} _ {\text {p o s}} ^ {\prime}\right) - s \left(\mathbf {x}, \mathbf {x} _ {\text {n e g}} ^ {\prime}\right)\right)\right) \tag {4} \\ + \lambda \left(\left| A _ {p o s} \right| _ {1} + \left| A _ {n e g} \right| _ {1}\right), \\ \end{array}
+$$
+
+Where  $m$  is the margin and  $\lambda$  balances the hinge loss and the  $L^1$ -norm.
+
+# 4 Question Generator
+
+For each uncertain token detected by the Error Locator, the PIIA agent interacts with the user to get a more explicit explanation. Instead of asking the user to explain the uncertain token directly, we provide an NL multi-choice question. As users are non-expert and unfamiliar with database operations, simply picking an option is more natural and friendly. Thus, the Question Generator is designed to generate a multi-choice question for each uncertain token $^2$ , i.e., "What do you mean by that?" as shown in Figure 4.
+
+To make a multi-choice question, a set of candidate options are generated. As analyzed in Section 2, most of the uncertain tokens are related to database information. Thus, for each uncertain token in an NL question, we find out the corresponding database and add all the column and table
+
+names into the candidate set. Additionally, we observe that some uncertain tokens are about aggregation operations, so we also add the aggregation operations into the candidate set, such as min, max and sum. As a result, the set is quite large, especially for complex databases. For example, the set size can be larger than 40 in the Spider dataset.
+
+Hence, we devise a ranking method to find out the options with the highest correlations to the uncertain token. Concretely, for each candidate option (denoted by  $\mathbf{w}$ ), we calculate its similarity score with the uncertain token (denoted by  $z$ ). As each candidate option is a span with one or more tokens, we adopt both lexical and semantic similarities.  $\mathbf{w}$  and  $z$  are pre-processed with lemmatization. Then the Jaccard distance between them is computed for lexical similarity, which is the number of common tokens divided by the number of unique tokens in them. For semantic similarity, we present each token as an embedding vector (i.e., GloVe (Pennington et al., 2014)) and employ the Euclidean distance between  $\mathbf{w}$  and  $z$ . The embedding vector of a span is the average embedding over tokens in the span. All the candidate options are ranked by the summation of these two similarity scores, and three of them are picked as options. Additionally, we add two more options, Value and None, to each multi-choice question. Value indicates that  $z$  is related to a value in the database. None means that either the token does not need modification or all the other options are not related. The None option is essential as it prevents uncertain tokens from being modified unexpectedly. This alleviates the need for Error Locator to make exact error detection and ensures a higher recall rate. Following the example in Figure 3, we show the question generation process in Figure 4. As we can observe, the multi-choice questions are easy to be understood by non-expert users, and the provided options are reasonable. After interacting with users, the PIIA agent gets the information that "cat" is a value, and "aged" indicates the column "pet_age".
+
+# 5 NL Modifier
+
+The last module of PIIA is the NL Modifier, which corrects NL questions with the users' feedback, i.e., how users answer the multi-choice questions. The most straightforward way is to directly replace the uncertain tokens with the selected options. However, it is not always reasonable. Since the uncertain tokens can be not only nouns but also verbs
+
+or adjectives, directly replacing verbs or adjectives may cause incoherence. To avoid this problem, we carefully design several modifier rules according to different POS tags, option types, and user NL question contexts. As mentioned in Section 4, there are four option types: column name, table name, aggregation, and value. A concrete example is shown in Figure 4, where we also list the modifier rules that are applied. Since the noun "cat" is selected as a value, the single quotes are added. It is easier for the parser to recognize it as a value because values are always equipped with single quotes in the dataset. Moreover, as there are multiple columns about age in the database, the adjective "aged" is modified to "whose pet_age is" to make the column name easier to identify. Finally, the corrected NL question  $\hat{\mathbf{x}}$  is fed into the text-to-SQL parser. More modifier rules and examples are shown in Table 3.
+
+PIIA is designed to modify the user NL questions instead of the predicted SQLs. Modifying the predicted SQLs is more straightforward but impracticable. We conducted several surveys and found non-expert users had difficulty giving high-quality responses to modify SQLs directly, even for simple SQLs. Note that, the uncertain tokens found by PIIA are the unaligned tokens in NL questions, so the users' feedback to uncertain tokens cannot be used to modify the corresponding SQLs.
+
+# 6 Experiments
+
+In this section, we firstly introduce the experimental setup. Then we assess PIIA by using both simulation and human evaluation, and finally we perform a closer analysis of PIIA.
+
+# 6.1 Experimental Setup
+
+We conduct experiments on two cross-domain text-to-SQL datasets with five base parsers.
+
+The WikiSQL dataset (Zhong et al., 2017) collects 24,241 cross-domain single-table databases from Wikipedia and contains 80,654 hand-annotated pairs of NL questions and SQL queries. The SQL queries are relatively simple with only SELECT and WHERE clauses. Two parsers are selected: (1) SQLova (Hwang et al., 2019), currently the best open-sourced parser on WikiSQL, uses table-aware and context-aware representations of questions to generate SQL queries. (2) SQLNet (Xu et al., 2017) applies sequence-to-set prediction and employs a sketch-based approach to predict SQL queries. We report our PIIA results on the test
+
+set, which contains 15,878 samples.
+
+The Spider dataset (Yu et al., 2018b) is a human-labeled text-to-SQL dataset that consists of 10,181 NL questions and 5,693 unique complex SQL queries on 200 databases with multiple tables. It covers 138 different domains and is much more complex than the WikiSQL dataset because it has a greater number of complex questions and nested SQL queries. Three parsers are selected: (1) IRNet (Guo et al., 2019), currently the state-of-the-art open-sourced parser on Spider, employs the coarse-t-o-fine framework (Dong and Lapata, 2018) and designs an intermediate language. (2) IRNet+BERT takes BERT as NL encoder to enhance the performance of basic IRNet. (3) SyntaxSQLNet (Yu et al., 2018a) employs a SQL specific syntax tree based decoder and table-aware column attention encoders. The test set is not publicly available, so we evaluate PIIA on the development set, which contains 1,034 samples.
+
+In Error Locator, the similarity threshold is set to be the average score of all the  $(\mathbf{x},\mathbf{x}^{\prime})$  pairs in the training triples  $\mathcal{X} = \{(\mathbf{x},\mathbf{x}_{pos}^{\prime},\mathbf{x}_{neg}^{\prime})\}$  as follows:
+
+$$
+p = \frac {1}{2 | \mathcal {X} |} \sum_ {(\mathbf {x}, \mathbf {x} _ {p o s} ^ {\prime}, \mathbf {x} _ {n e g} ^ {\prime}) \in \mathcal {X}} \big (s (\mathbf {x}, \mathbf {x} _ {p o s} ^ {\prime}) + s (\mathbf {x}, \mathbf {x} _ {n e g} ^ {\prime}) \big).
+$$
+
+This threshold score is generally lower than alignment scores of certain tokens and higher than that of uncertain tokens, which can help to distinguish certain and uncertain token alignment. For hyperparameters, we set  $m = 1$  and  $\lambda = 0.5$ . As for the NL Modifier, the NLTK pos tagging model (Loper and Bird, 2002) is employed for pre-processing.
+
+# 6.2 Simulation Evaluation
+
+We build a simulator to interact with the PIIA agent, which aims to give ideal selections for multi-choice questions. We report the results achieved by five base parsers.
+
+Simulator The simulator chooses options on behalf of the real user. Given an NL question with  $T$  uncertain tokens, PIIA asks a multi-choice question with  $K$  options for each token. The simulator enumerates all  $K^T$  possible combinations of options and feeds them into the NL Modifier and the parser to get SQL queries. If one of these SQL queries is the same as the ground truth SQL, the simulator obtains the ideal selection. Otherwise, the PIIA fails to correct the NL question. Since it is too time-consuming to enumerate all  $K^T$  combinations, we
+
+<table><tr><td></td><td>Models</td><td>SQLAcc</td><td>ExeAcc</td><td>Avg.#T</td></tr><tr><td rowspan="4">WikiSQL</td><td>SQlova</td><td>80.7</td><td>86.2</td><td>N/A</td></tr><tr><td>+PIIA</td><td>84.9</td><td>88.9</td><td>1.3</td></tr><tr><td>SQLNet</td><td>61.7</td><td>68.0</td><td>N/A</td></tr><tr><td>+PIIA</td><td>68.4</td><td>73.2</td><td>1.7</td></tr><tr><td rowspan="6">Spider</td><td>IRNet</td><td>53.2</td><td>N/A</td><td>N/A</td></tr><tr><td>+PIIA</td><td>59.3</td><td>N/A</td><td>2.9</td></tr><tr><td>IRNet+BERT</td><td>61.9</td><td>N/A</td><td>N/A</td></tr><tr><td>+PIIA</td><td>63.4</td><td>N/A</td><td>2.4</td></tr><tr><td>SyntaxSQLNet</td><td>27.2</td><td>N/A</td><td>N/A</td></tr><tr><td>+PIIA</td><td>34.2</td><td>N/A</td><td>3.4</td></tr></table>
+
+Table 1: Simulation results of PIIA on the WikiSQL test set and the Spider development set.
+
+only rank and simulate the top 100. Concretely, we firstly filter out options whose tokens do not appear in the ground truth SQL. Then we provide a score to each left option, as Question Generator does in Section 4. Finally, the overall score for a combination is computed by summing up all the option scores in the combination.
+
+Model Comparison We evaluate PIIA with the simulator on both WikiSQL and Spider datasets. The results are shown in Table 1. For the WikiSQL dataset, we report the accuracy of SQL exact matching (SQLAcc) and the accuracy of execution (ExeAcc). We can observe that PIIA boosts the performance for both base parsers with fewer than two average interaction turns (Avg.#T). The absolute SQLAcc improvements for SQLova and SQLNet are  $4.2\%$  and  $6.7\%$  respectively, while the absolute ExeAcc improvements are  $2.7\%$  and  $5.2\%$ . The results on SQLNet are competitive with those of DailSQL (Gur et al., 2018), a parser-independent method designed for simple SQL. However, on average, PIIA interacts with users by 1.7 turns, while DialSQL needs about 4.8. This indicates that PIIA is quite efficient and able to enhance text-to-SQL with only a few interaction turns.
+
+Similar performance boosts can be observed on the Spider dataset, which has more complex multitable SQL queries. As execution results are not available, we only report the results of SQLAcc. The improvements on IRNet, IRNet+BERT and SyntaxSQLNet again demonstrate the effectiveness of PIIA. PIIA can also enhance the parser integrated with BERT, which further proves the necessity of PIIA. After interacting with users, PIIA provides the revised NL questions in a form that
+
+<table><tr><td>Models</td><td>w/o PIIA</td><td>PIIA(H)</td><td>PIIA(S)</td><td>Avg.#T</td></tr><tr><td>IRNet</td><td>49.0</td><td>52.7</td><td>54.7</td><td>2.8</td></tr><tr><td>IRNet+BERT</td><td>60.7</td><td>62.2</td><td>62.7</td><td>2.4</td></tr></table>
+
+Table 2: SQL Accuracy of human evaluation (H) and simulation (S) on 300 samples.
+
+is easier for parsers to understand. The average interaction turns are about three, a number users find acceptable. Fewer interaction turns are required by better parsers. Additionally, a smaller improvement is obtained with a better parser, as better parsers know NL questions better. PIIA is effective because it utilizes the feedback provided by users, who know the questions best, thus leading to a further narrowing of the gap between parsers and users.
+
+# 6.3 Human Evaluation
+
+We carry out the evaluation of PIIA with real users. The human evaluation is performed on the more complex Spider dataset and with two state-of-the-art parsers, i.e., IRNet and IRNet+BERT. We randomly sample 300 NL questions from the Spider development set and invite 30 volunteers majoring in liberal arts to interact with the PIIA agent. Each NL question is evaluated by three volunteers, all of whom are non-expert without any background knowledge of SQL queries. We provide them with the NL questions and the corresponding databases, and they interact with PIIA by answering the multi-choice questions.
+
+The SQLAcc results of the human evaluation are shown in Table 2. By interacting with real users, PIIA boosts the overall SQLAcc of both IRNet and IRNet+BERT by an absolute improvement of  $3.7\%$  and  $1.5\%$ , respectively. This indicates that PIIA provides a friendly way to interact with non-expert users, who are therefore able to understand the multi-choice questions and give proper answers. The average numbers of interaction turns are 2.8 and 2.4, respectively, which the users find acceptable. We also analyze the gap between human evaluation and simulation and find that some of the NL questions are ambiguous, making it hard for real users to distinguish similar options.
+
+# 6.4 Closer Analysis
+
+We use the IRNet parser on the Spider dataset to provide a closer analysis of PIIA. Similar observations can be obtained with other parsers.
+
+![](images/ec1729cb16406f645f03cbf759eac828ff891d7f3ba8d3d33651153aa7b5dc5e.jpg)  
+(a)  
+Figure 5: (a) Turn distribution and (b) SQLAcc w.r.t number of options by IRNet+PIIA on the Spider development set. Orange line in (b) indicates the result without PIIA.
+
+![](images/2c42bfc9b0addd7eaed69de8055892b8f2322ae9ce61f57ce8655cc5af3d7b24.jpg)  
+(b)
+
+Number of Turns In interactive systems, the number of interaction turns mostly determines whether users have a positive experience as too many turns may tire them out. With that in mind, we figure out the distribution of the number of turns (i.e., the number of multi-choice questions) over the Spider development set. The distribution is shown in Figure 5(a) with the average number being 2.9. As observed, the interaction process finishes in four turns in nearly  $90\%$  of the cases. Only 10 out of 1,034 cases require an interaction process with more than five turns, which indicates PIIA is able to process such a complex dataset with high efficiency. We also analyze the cases correctly modified by the simulation and find that about  $40\%$  of the multi-choice questions get the None answer, which is an acceptable percentage. Additionally, the performance of our Error Locator is adequate. Though there are an average of 12.4 tokens in each NL question on the Spider development set, the Error Locator is able to find about three uncertain tokens out of them effectively.
+
+Number of Options Providing a greater number of options in multi-choice questions increases the chances of including the correct one, thus enhancing the performance of the system. With more options, however, users have to make more effort. Therefore, we conduct experiments to analyze the influence of the number of options on the SQLAcc under simulation, as shown in Figure 5(b). The curve tends to be smooth after five options, meaning that this number is reasonable and able to balance the number of options and the need for correct ones. It's worth mentioning that each question contains two necessary options, i.e., None and Value,
+
+![](images/761a46da9c505e2d5dac084950122de71450e783cf918e836e49bf9e0dfec4fe.jpg)  
+Figure 6: The similarity matrix of a real case.
+
+so that the least number of options is three.
+
+Similarity Matrix The Error Locator module is responsible for calculating a similarity matrix  $A$  between a user's NL question  $x$  and a restated NL question  $x'$ . Following the example in Figure 3, we show the similarity matrix in Figure 6. "last name" is evidently more similar to the column name "lname" than to the others, which meets our expectations. Two uncertain tokens are extracted, namely "cat" and "age". They are not stop words, and their scores don't reach the threshold. Since the token "cat" is not expressed in the restated NL question, it has a low similarity score. Although "age" appears in the restated NL question, it exists in two column names in the database, i.e., "age" from the table "student" and "pet_age" from the table "pet". Thus the score of "age" is reduced and falls below the threshold.
+
+Modifier Rules We carefully design the modifier rules for the NL Modifier based on the uncertain tokens and the selected options. Concretely, we take into consideration the types of options (column, table, value, or aggregation), the POS taggings of the uncertain tokens, and the contexts of the uncertain tokens. As shown in Table 1, PIIA with the NL Modifier improves the efficacy of IRNet SQLAcc from  $53.2\%$  to  $59.3\%$ . Instead of using the NL Modifier, we try a straightforward way to modify the NL questions that involves directly replacing the uncertain tokens with selected options. With this approach, the simulated SQLAcc stands at  $54.8\%$ , a result that is much worse than what PIIA can achieve with the NL Modifier, thus proving that our module is indispensable.
+
+Case Study Table 3 shows more real cases of users' NL questions and corrected NL questions
+
+along with the corresponding modifier rules. The words in bold are uncertain tokens and their corrections. Though IRNet wrongly parses these six cases, PIIA manages to solve them correctly. The first five cases are modified by rules for nouns, verbs, and adjectives that are related to the column names in the databases. Different rules are applied to add the column names into NL questions, making it more explicit for the parser to understand them. Case 6 shows an example of how to modify the aggregation operator and the value-related tokens. PIIA revises the inexplicit NL questions by interacting with users. Equipped with the PIIA agent, the performance of the base parser is improved.
+
+# 7 Related Works
+
+The works most related to ours are those investigating interactive semantic parsing. For instance, DailSQL, proposed by Gur et al. (2018), aims to detect error spans and their categories based on an encoder-decoder architecture. But it is designed for relatively simple scenarios. In this research area, another impressive work involves a model-based interaction system, which detects uncertain tokens and asks questions relying on inner parser states (Yao et al., 2019). Unlike these studies, however, we design a parser-independent interactive approach that can also perform cross-domain complex SQL queries. In the field of applied systems, Gao et al. (2015) focused on user interface designing and proposed an interactive semantic parsing system called Datatone. In contrast to them, our main contribution lies in the realm of technology. Another topic our method related to is query reformulation. The idea of query reformulation is explored by Ray et al. (2018) and Rastogi et al. (2019), while they apply this idea in other domains with different scenarios. Our work is also related to semantic parsing, the process of converting natural language utterances into logical forms. Sequence-to-sequence methods are widely applied to solve this task (Berant et al., 2013; Dong and Lapata, 2016; Finegan-Dollak et al., 2018; Su et al., 2018). To reduce search space for decoding, several works employed intermediate representations to generate abstract representations (Cheng et al., 2017; Goldman et al., 2018; Dong and Lapata, 2018). Although these methods have achieved an impressive performance in experimental studies, there is still a long way to go before they can be successfully
+
+<table><tr><td rowspan="3">1</td><td>User NL question:</td><td>List the creation year, name and budget of each department.</td></tr><tr><td>Corrected NL question:</td><td>List the creation year, name and budget in billions of each department.</td></tr><tr><td>Modifier Rule:</td><td>[Noun; C] → [C] (replace the token with the whole column name)</td></tr><tr><td rowspan="3">2</td><td>User NL question:</td><td>What are the first names of all the different drivers in alphabetical order?</td></tr><tr><td>Corrected NL question:</td><td>What are the forename of all the different drivers in alphabetical order?</td></tr><tr><td>Modifier Rule:</td><td>[Noun; C] → [C] (detect the phrase and replace it with the whole column name)</td></tr><tr><td rowspan="3">3</td><td>User NL question:</td><td>What is the type of the document named &quot;David CV&quot;?</td></tr><tr><td>Corrected NL question:</td><td>What is the document type code of the document whose name is &quot;David CV&quot;?</td></tr><tr><td>Modifier Rule:</td><td>[Verb; C] → whose [C] is (directly tell the column name of the verb token)</td></tr><tr><td rowspan="3">4</td><td>User NL question:</td><td>Where is the club &quot;Hopkins Student Enterprises&quot; located?</td></tr><tr><td>Corrected NL question:</td><td>Where is the club &quot;Hopkins Student Enterprises&quot; &#x27;s location?</td></tr><tr><td>Modifier Rule:</td><td>[Verb; C] → &#x27;s [C] (directly tell the column name of the verb token, when the token is in the end of the sentence)</td></tr><tr><td rowspan="3">5</td><td>User NL question:</td><td>Show the enrollment and primary_conference of the oldest college.</td></tr><tr><td>Corrected NL question:</td><td>Show the enrollment and primary_conference of the oldest founded college.</td></tr><tr><td>Modifier Rule:</td><td>[Adj; C] → [Adj] [C] (directly add the column name of the adjective token)</td></tr><tr><td rowspan="3">6</td><td>User NL question:</td><td>Give the mean GNP and total population of nations which are considered US territory.</td></tr><tr><td>Corrected NL question:</td><td>Give the average GNP and total population of nations which are considered &#x27;US territory&#x27;.</td></tr><tr><td>Modifier Rule:</td><td>[Agg] → [Agg] (replace the aggregation token with the most similar aggregation word) [Noun; V] → &#x27;[V]&#x27; (add single quotes to indicate it is a value)</td></tr></table>
+
+Table 3: Cases by IRNet+PIIA on Spider. Texts highlighted in gray indicate column names in the databases.
+
+applied in real systems.
+
+Works dealing with the task of weakly supervised word alignment are also related to our research because our Error Locator module performs the same task. Some examples include the work of Liu and Sun (2015), who proposed a latent-variable log-linear model for word alignment, the research of Legrand et al. (2016), who used pairwise training with negative sampling to train the alignment model, and a study that introduced a gradient-based alignment method for machine translation (He et al., 2019b).
+
+# 8 Conclusion and Future Work
+
+We propose a parser-independent interactive approach, PIIA, to enhance the text-to-SQL process in NLIDB systems. PIIA interacts with users via multi-choice questions and can be built on arbitrary parsers. Experimental results show this approach leads to significant performance boosts on two cross-domain datasets with five different base parsers. In the future, we are interested in distilling and reusing the common knowledge from users' selections.
+
+# Acknowledgments
+
+We thank all the anonymous reviewers for their valuable comments. This work was supported in part by NSFC under Grant No. 61532001, National Key Research and Development Program of China under Grant No. 2018AAA0101902, and MOE-ChinaMobile Program under Grant No. MCM20170503.
+
+# References
+
+Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. 2013. Semantic parsing on Freebase from question-answer pairs. In EMNLP.  
+Ben Bogin, Matt Gardner, and Jonathan Berant. 2019a. Global reasoning over database structures for text-to-SQL parsing. In EMNLP.  
+Ben Bogin, Matt Gardner, and Jonathan Berant. 2019b. Representing schema structure with graph neural networks for text-to-SQL parsing. In ACL.  
+Jianpeng Cheng, Siva Reddy, Vijay Saraswat, and Mirella Lapata. 2017. Learning structured natural language representations for semantic parsing. In ACL.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In *NAACL*.  
+Kedar Dhamdhere, Kevin S McCurley, Ralfi Nahmias, Mukund Sundararajan, and Qiqi Yan. 2017. Analyza: Exploring data with conversation. In IUI.  
+Li Dong and Mirella Lapata. 2016. Language to logical form with neural attention. In ACL.  
+Li Dong and Mirella Lapata. 2018. Coarse-to-fine decoding for neural semantic parsing. In ACL.  
+Catherine Finegan-Dollak, Jonathan K Kummerfeld, Li Zhang, Karthik Ramanathan, Sesh Sadasivam, Rui Zhang, and Dragomir Radev. 2018. Improving text-to-SQL evaluation methodology. In ACL.  
+Tong Gao, Mira Dontcheva, Eytan Adar, Zhicheng Liu, and Karrie G Karahalios. 2015. Datatone: Managing ambiguity in natural language interfaces for data visualization. In Proceedings of the 28th Annual ACM Symposium on User Interface Software & Technology, pages 489-500.
+
+Alfio Gliozzo, Or Biran, Siddharth Patwardhan, and Kathleen McKeown. 2013. Semantic technologies in IBM Watson. In Proceedings of the Fourth Workshop on Teaching NLP and CL, pages 85-92.  
+Omer Goldman, Veronica Latcinnik, Udi Naveh, Amir Globerson, and Jonathan Berant. 2018. Weakly-supervised semantic parsing with abstract examples. In ACL.  
+Daya Guo, Yibo Sun, Duyu Tang, Nan Duan, Jian Yin, Hong Chi, James Cao, Peng Chen, and Ming Zhou. 2018. Question generation from SQL queries improves neural semantic parsing. In EMNLP.  
+Jiaqi Guo, Zecheng Zhan, Yan Gao, Yan Xiao, Jian-Guang Lou, Ting Liu, and Dongmei Zhang. 2019. Towards complex text-to-SQL in cross-domain database with intermediate representation. In ACL.  
+Izzeddin Gur, Semih Yavuz, Yu Su, and Xifeng Yan. 2018. DialSQL: Dialogue based structured query generation. In ACL.  
+Pengcheng He, Yi Mao, Kaushik Chakrabarti, and Weizhu Chen. 2019a. X-SQL: reinforce schema representation with context. arXiv preprint arXiv:1908.08113.  
+Shilin He, Zhaopeng Tu, Xing Wang, Longyue Wang, Michael R Lyu, and Shuming Shi. 2019b. Towards understanding neural machine translation with word importance. In EMNLP.  
+Wonseok Hwang, Jinyeung Yim, Seunghyun Park, and Minjoon Seo. 2019. A comprehensive exploration on WikiSQL with table-aware word contextualization. In KR2ML Workshop at NeurIPS.  
+Srinivasan Iyer, Ioannis Konstas, Alvin Cheung, Jayant Krishnamurthy, and Luke Zettlemoyer. 2017. Learning a neural semantic parser from user feedback. In ACL.  
+Sanjaya Lai, Kedar Doshi, Yamuna Esaiarasan, and Chaitanya Bhatt. 2014. Systems and methods for performing record actions in a multi-tenant database and application system. US Patent 8,818,940.  
+Joel Legrand, Michael Auli, and Ronan Collobert. 2016. Neural network-based word alignment through score aggregation. In SIGMT.  
+Fei Li and HV Jagadish. 2014. Constructing an interactive natural language interface for relational databases. VLDB Endowment, 8(1):73-84.  
+Yang Liu and Maosong Sun. 2015. Contrastive unsupervised word alignment with non-local features. In AAAI.  
+Edward Loper and Steven Bird. 2002. NLTK: the natural language toolkit. arXiv preprint cs/0205028.
+
+Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. GloVe: Global vectors for word representation. In EMNLP.  
+Pushpendre Rastogi, Arpit Gupta, Tongfei Chen, and Lambert Mathias. 2019. Scaling multi-domain dialogue state tracking via query reformulation. arXiv preprint arXiv:1903.05164.  
+Avik Ray, Yilin Shen, and Hongxia Jin. 2018. Learning out-of-vocabulary words in intelligent personal agents. In *IJCAI*, pages 4309-4315.  
+Yu Su, Ahmed Hassan Awadallah, Miaosen Wang, and Ryen W White. 2018. Natural language interfaces with fine-grained user interaction: A case study on web APIs. In SIGIR, pages 855-864. ACM.  
+Xiaojun Xu, Chang Liu, and Dawn Song. 2017. SQL-Net: Generating structured queries from natural language without reinforcement learning. arXiv preprint arXiv:1711.04436.  
+Navid Yaghmazadeh, Yuepeng Wang, Isil Dillig, and Thomas Dillig. 2017. SQLizer: query synthesis from natural language. ACM on Programming Languages, 1(OOPSLA):63.  
+Ziyu Yao, Yu Su, Huan Sun, and Wen-tau Yih. 2019. Model-based interactive semantic parsing: A unified framework and a text-to-SQL case study. In EMNLP.  
+Tao Yu, Michihiro Yasunaga, Kai Yang, Rui Zhang, Dongxu Wang, Zifan Li, and Dragomir Radev. 2018a. SyntaxSQLnet: Syntax tree networks for complex and cross-domain text-to-SQL task. In EMNLP.  
+Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, et al. 2018b. Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-SQL task. In EMNLP.  
+Victor Zhong, Caiming Xiong, and Richard Socher. 2017. Seq2SQL: Generating structured queries from natural language using reinforcement learning. In CoRR.

whatdoyoumeanbythataparserindependentinteractiveapproachforenhancingtexttosql/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d1eefbfb9f2a3b49f8c504b973cd69a3322d647c589b6349843abbca4f40e7e
+size 410106

whatdoyoumeanbythataparserindependentinteractiveapproachforenhancingtexttosql/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:840f8c1d9875de2349e5cc626bd8d3537858279a0a494274bdd8b91b47ac8b55
+size 362592

xlamrenablingcrosslingualamrparsingwithtransferlearningtechniques/45d8433e-75e6-4125-9552-3814479d6d92_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0854150a91c66fab235a47b0e9318c3896c1c862b7ec763378f119cd4ac29acc
+size 96357

xlamrenablingcrosslingualamrparsingwithtransferlearningtechniques/45d8433e-75e6-4125-9552-3814479d6d92_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c873e99a25e6f9d1ac6ea2107ca06d85edc7bfcff04a9a5a10e31ac2e8cee146
+size 119753

xlamrenablingcrosslingualamrparsingwithtransferlearningtechniques/45d8433e-75e6-4125-9552-3814479d6d92_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f389c9920a4660fcb4a711004581ba9debf47506d9ceacfb5edecdfea5379e1
+size 925303

xlamrenablingcrosslingualamrparsingwithtransferlearningtechniques/full.md

@@ -0,0 +1,311 @@
+# XL-AMR: Enabling Cross-Linguual AMR Parsing with Transfer Learning Techniques
+
+Rexhina Blloshmi Rocco Tripodi Roberto Navigli
+
+Sapienza NLP Group
+
+Department of Computer Science, Sapienza University of Rome
+
+{blloshmi,tripodi,navigli}@di.uniroma1.it
+
+# Abstract
+
+Abstract Meaning Representation (AMR) is a popular formalism of natural language that represents the meaning of a sentence as a semantic graph. It is agnostic about how to derive meanings from strings and for this reason it lends itself well to the encoding of semantics across languages. However, cross-lingual AMR parsing is a hard task, because training data are scarce in languages other than English and the existing English AMR parsers are not directly suited to being used in a cross-lingual setting. In this work we tackle these two problems so as to enable cross-lingual AMR parsing: we explore different transfer learning techniques for producing automatic AMR annotations across languages and develop a cross-lingual AMR parser, XL-AMR. This can be trained on the produced data and does not rely on AMR aligners or source-copy mechanisms as is commonly the case in English AMR parsing. The results of XL-AMR significantly surpass those previously reported in Chinese, German, Italian and Spanish. Finally we provide a qualitative analysis which sheds light on the suitability of AMR across languages. We release XL-AMR at github.com/SapienzaNLP/xl-amr.
+
+# 1 Introduction
+
+Abstract Meaning Representation (AMR) is a popular formalism for natural language (Banarescu et al., 2013). It represents sentences as rooted, directed and acyclic graphs in which nodes are concepts and edges are semantic relations among them. AMR unifies, in a single structure, a rich set of information coming from different tasks, such as Named Entity Recognition (NER), Semantic Role Labeling (SRL), Word Sense Disambiguation (WSD) and coreference resolution. Such representations are actively integrated in several Natural Language Processing (NLP) applications, inter
+
+alia, information extraction (Rao et al., 2017), text summarization (Hardy and Vlachos, 2018; Liao et al., 2018), paraphrase detection (Issa et al., 2018), spoken language understanding (Damonte et al., 2019), machine translation (Song et al., 2019b) and human-robot interaction (Bonial et al., 2020). It is therefore desirable to extend AMR semantic representations across languages along the lines of cross-lingual representations for grammatical annotation (de Marneffte et al., 2014), concepts (Conia and Navigli, 2020) and semantic roles (Akbik et al., 2015; Di Fabio et al., 2019). Furthermore, it could be especially useful to integrate cross-lingual semantic structures in multilingual applications of natural language understanding.
+
+A peculiar feature of the AMR formalism is that it aims at abstracting away from word forms. AMR graphs are unanchored, i.e., the linkage between tokens in a sentence and nodes in the corresponding graph is not explicitly annotated. Hence, the feature of being agnostic about how to derive meanings from strings makes AMR particularly suitable for representing semantics cross-lingually. However, AMR was initially designed for encoding the meaning of English sentences. Owing to this, the available resources and modelling techniques focus mostly on English, while leaving cross-lingual AMR understudied. Some preliminary studies showed the limits of AMR as an interlingua, categorizing them as due to distinctions in the underlying ontologies or structural divergences among languages (Xue et al., 2014; Hajic et al., 2014). More recent studies, instead, have provided evidence that AMR or a simplified version of it can be used as a formalism for cross-lingual semantic representation, showing that it is possible to overcome some of the structural linguistic divergences (Damonte and Cohen, 2018; Zhu et al., 2019).
+
+The underlying idea of this paper is that AMR can be used to represent semantic information in
+
+different languages since there exist key linguistic features that are shared across languages, such as predicates, roles and conjunctions (Von Fintel and Matthewson, 2008). However, developing an AMR parser for multiple languages is hard because the existing annotated training resources that are sufficiently large are available in English only, and, moreover, acquiring semantic annotations for a large number of sentences is well-known to be a slow and expensive process in NLP (Zhang et al., 2018; Pasini, 2020). To this end, we aim at exploiting and developing the necessary tools and resources for enabling cross-lingual AMR parsing, i.e., the task of transducing a sentence in the source language into an AMR graph based on English (Damonte and Cohen, 2018).
+
+We present XL-AMR, a cross-lingual AMR parser, and study different transfer learning techniques to enable its training: i) model transfer which relies on language-independent features, ii) annotation projection relying on parallel corpora and available English AMR parsers, and iii) automatic translation of the training corpora which guarantees gold AMR structures. We make the following contributions:
+
+- We develop and release XL-AMR, a cross-lingual AMR parser which disposes of word aligners, i.e., word-to-word and word-to-node, and surpasses the previously reported results on Chinese, German, Italian and Spanish, by a large margin.  
+- Exploration of different techniques to create cross-lingual AMR training data, showing how it is possible to transfer semantic structure information across different languages.  
+- Creation and release of diverse quality silver data for cross-lingual AMR parsing.  
+- Qualitative analysis of the ability of XL-AMR to transfer semantic structures across languages and of AMR to represent the meaning of sentences cross-lingually.
+
+# 2 Related Work
+
+Our work lies between two areas, namely, semantic parsing and cross-lingual transfer learning.
+
+Semantic parsing Semantic parsing is a key task required to complete the puzzle of Natural Language Understanding (Navigli, 2018), and one which is receiving growing attention in the scientific community. Besides AMR, various different formalisms have been proposed over the years
+
+to encode semantic structures: Elementary Dependency Structures (Oepen and Lønning, 2006, EDS), Prague Tectogrammatical Graphs (Hajic et al., 2012, PTG), Universal Conceptual Cognitive Annotation (Abend and Rappoport, 2013, UCCA), Universal Decompositional Semantics (White et al., 2016, UDS), inter alia. While some frameworks, such as UCCA and UDS, have been exploited in a cross-linguistic setting (Lyu et al., 2019; Zhang et al., 2018), cross-lingual AMR has mainly been studied within the scope of annotation analysis works (Xue et al., 2014; Hajic et al., 2014). These works point out the limitations of AMR as an interlingua, and consider them partly due to the distinctions in the underlying ontologies and structural divergences among languages. Zhu et al. (2019) also evaluate the properties of AMR across languages and aim at simplifying this formalism in order to express only essential semantic features of a sentence, such as predicate roles and linguistic relations. Cross-lingual AMR parsing, instead, has received relatively less attention. This is largely attributable to the lack of training data and evaluation benchmarks in languages other than English. Damonte and Cohen (2018) propose, to the best of our knowledge, the only cross-lingual AMR parser to date and, moreover, their proposed cross-lingual AMR evaluation benchmark has been released only very recently (Damonte and Cohen, 2020). The authors adapt a transition-based English AMR parser (Damonte et al., 2017) for cross-lingual AMR parsing, which is trained on silver annotated data. However, the performances it has achieved are not satisfying in terms of Smatch score (Cai and Knight, 2013), mostly as a result of concept identification errors, which in turn are directly related to the usage of noisy word-to-node alignments projected from English. Throughout the literature English AMR parsers commonly rely on AMR alignments which are automatically created using heuristics (Flanigan et al., 2014), or on pretrained aligners (Pourdamghani et al., 2014; Liu et al., 2018), treated as latent variables of the model (Lyu and Titov, 2018) or implicitly modelled through source-copy mechanisms (Zhang et al., 2019). These alignments, however, take advantage of the fact that AMR nodes and English words are highly related. $^{1}$  This dependency is therefore not suitable for cross-lingual parsing since similarity between words in
+
+the sentences and concepts in the graph does not hold at large. Our parser, instead, disposes of explicit and implicit AMR alignments using a seq2seq model for concept identification and achieves significantly higher performance on all the tested languages. On the other hand, to account for data sparsity, XL-AMR employs several common techniques in English AMR parsing literature (Konstas et al., 2017; Zhang et al., 2019), such as anonymization and recategorization, expanding them across languages by relying on multilingual resources.
+
+Transfer learning The idea behind this method is to leverage annotations available in one language, commonly English, to enable learning models that generalize to languages where labelled resources are scarce (Ruder et al., 2019). Different techniques include annotation projection, machine translation and language-independent feature-based models. Extensive works in this direction exist, applied to different NLP tasks, i.e., WSD (Barba et al., 2020), SRL (Padó and Lapata, 2009; Kozhevnikov and Titov, 2013), Dependency Parsing (Tiedemann, 2015), concept representation (Conia and Navigli, 2020), etc. In cross-lingual AMR parsing, annotation projection is employed by Damonte and Cohen (2018), who produce cross-lingual silver AMR annotations by exploiting parallel sentences selected from the Europarl corpus (Koehn, 2005): English sentences are parsed using an English parser (Damonte et al., 2017, AMREAGER) and the resulting graphs are associated with the corresponding parallel sentences. However, the data on which AMREAGER was trained is very different from those used to produce the silver annotations, thus affecting the quality and reliability of the AMR graphs produced. Here we test two different techniques: we conduct experiments with annotation projection using Europarl for comparison, and, in addition, we use translation techniques to produce better quality training corpora. This leads to significant improvements and provides evidence that better quality data – and models – allow for using AMR as an interlingua.
+
+# 3 Cross-Lingual AMR
+
+In what follows we first formalize the task (Section 3.1) and then detail our cross-lingual AMR parser (Section 3.2) and our proposed silver data creation methods (Section 3.3). Finally, we list the pre- and postprocessing cross-lingual techniques and resources we employ (Section 3.4).
+
+![](images/b04e81f853ae588ba8d0b90326c694d2db5dc3fcb0f6beb6bc94bc62629e1938.jpg)  
+Figure 1: Cross-Lingual AMR Parsing: (A) Sentences written in different languages sharing the same meaning; (B) concepts representing the words in the sentences; (C) the final AMR graph.
+
+# 3.1 The Task
+
+Cross-lingual AMR parsing is defined as the task of transducing a sentence in any language to the AMR graph of its English translation whose nodes are either English words, PropBank framesets (Kingsbury and Palmer, 2002) or special AMR keywords.
+
+Breaking down this definition, given an English sentence and its translation  $T_{L}$  in a language  $L$ , their meaning representation is ideally formalized by the same AMR,  $G = (V,E)$ , where  $V$  is a list of concept nodes and  $E$  is the set of semantic relations between them. Figure 1-A shows an example of a sentence in English, with its translations into Chinese, German, Italian and Spanish which have the same meaning and therefore the same abstract representation (Figure 1-C). Following state-of-the-art models for English AMR parsing (Zhang et al., 2019), we tackle cross-lingual AMR parsing as a two-stage approach, i.e., concept and relation identification, which we briefly
+
+overview here and later detail in Section 3.2. For concept identification, given the sequence  $T_{L} = (t_{1}, t_{2}, \ldots, t_{j})$ ,  $t_{i}$  being a word in language  $L$  ( $i \in \{1, \ldots, j\}$ ),  $L \in \{\mathrm{EN}, \mathrm{DE}, \mathrm{ES}, \mathrm{IT}, \mathrm{ZH}\}$ ), we train a neural network to generate the list of nodes  $V = (v_{1}, v_{2}, \ldots, v_{n})$ ,  $v_{i} \in$  English words  $\cup$  PropBank framesets  $\cup$  AMR keywords. In Figure 1-B we show the list of concepts that represent the words in the sentences of Figure 1-A. The relation identification procedure, instead, is inspired by the arc-factored approaches employed in dependency parsing (Kiperwasser and Goldberg, 2016), i.e., searching for the maximum-scoring connected subgraph over the identified concepts in the previous step. Thus, given the list of predicted nodes  $V = (v_{1}, v_{2}, \ldots, v_{n})$  and a learned score for each candidate edge, we search for the highest-scoring spanning tree and then merge the duplicate nodes based on unique node indices (see Section 3.2) to restore the final AMR graph. Figure 1-C shows the AMR representing the shared semantics of the sentences in Figure 1-A.
+
+# 3.2 XL-AMR Model
+
+XL-AMR is composed of two modules which are learned jointly, i.e., concept identification, modeled as a seq2seq problem, and relation identification, based on a biaffine attention classifier (Dozat and Manning, 2017). We use a seq2seq model to dispose of the need for an AMR alignment module. Lyu and Titov (2018) argue that alignments are important for injecting a useful inductive bias for AMR parsing and maintain that alignment-based parsers might be better than seq2seq for AMR parsing, owing to the relatively small amount of data available for AMR. However, aligning words to AMR nodes in cross-lingual parsing is challenging. The widely used AMR aligners are usually based on heuristics (Flanigan et al., 2014), or on the fact that AMR and English are highly cognate (Pourdamghani et al., 2014). Hence, these approaches would not be valid for cross-lingual alignment and, moreover, projecting the alignments across languages through English has shown to be noisy and to affect the parsing performance (Damonte and Cohen, 2018).
+
+Concept identification At training time we obtain the list of nodes by first converting the graph into a tree, duplicating the nodes occurring in multiple relations, and then using a pre-order traversal over the tree. To account for reentrancies we assign a unique index to each node during traversal,
+
+similarly to Zhang et al. (2019). Following the attention-based encoder-decoder architecture proposed by Bahdanau et al. (2015), our concept identification module consists of a bidirectional RNN encoder and a decoder that attends to the source sentence at each concept decoding step.
+
+The encoder employs an  $L$ -layer bidirectional RNN (Schuster and Paliwal, 1997) with LSTM cells (Hochreiter and Schmidhuber, 1997), i.e., BiLSTM, which encodes the input token embeddings  $e_i$  into hidden states  $h_i$ . Each hidden state  $h_i^l = [h_i^l; h_i^l]$ , is a concatenation of the forward hidden state and the backward hidden state at timestep  $i$ . Similarly to Zhang et al. (2019), the input token embedding  $e_i$  is a concatenation of contextualized embeddings, word embeddings, Part-of-Speech (PoS) embeddings, token anonymization indicator<sup>2</sup> and character-level embeddings. The subsequent BiLSTM layer, instead, takes the hidden states of the previous layer as input.
+
+The decoder also consists of  $L$  recurrent neural network (unidirectional) layers with LSTM cells. The decoder embedding layer concatenates word embeddings, node index embeddings and character-level embeddings. The layer  $l$  of the decoder calculates  $d_{t}^{l} = \text{decoder}_{l}(d_{t}^{l-1}, d_{t-1}^{l})$ , where  $d_{t}^{l-1}$  is the concept hidden state of the previous layer at timestep  $t$  while  $d_{t-1}^{l}$  that of previous timestep.  $d_{0}^{l}$  is initialized with the concatenation of the encoder's last hidden states  $h^{l} = [h^{l}; h^{l}]$ . We follow the input feeding approach of Luong et al. (2015), which concatenates the output of the decoder's embedding layer and an attentional vector computed at the previous timestep. We first compute the source attention distribution  $a_{t}$  using additive attention (Bahdanau et al., 2015) as follows:
+
+$$
+e _ {t, i} = v ^ {\top} \mathrm {t a n h} (W _ {h} h _ {i} ^ {L} + W _ {s} d _ {t} ^ {L} + b _ {s})
+$$
+
+$$
+a _ {t} = \operatorname {s o f t m a x} \left(e _ {t}\right)
+$$
+
+$$
+c _ {t} = \sum_ {i} a _ {t, i} h _ {i}
+$$
+
+where  $v, W_h, W_s$  and  $b_s$  are model parameters, and  $c_t$  is the source context vector. Then, we compute the attentional vector,  $\tilde{d}_t = \tanh(W_c[c_t; d_t^L] + b_c)$ , where  $W_c$  and  $b_c$  are model parameters.
+
+Zhang et al. (2019) used the attentional vector to allow the decoder to copy nodes predicted in the previous steps (target-copy), rather than only
+
+generating a new node from the vocabulary. As they provide empirical evidence that this is crucial for handling reentrancies, we employ their target-copy approach and use the attentional vector  $\tilde{d}_t$  to i) feed in a dense layer and softmax to produce a probability distribution over the vocabulary  $P_{\text{vocab}} = \text{softmax}(W_{\text{vocab}} \tilde{d}_t + b_{\text{vocab}})$ , ii) to learn a target attention distribution  $\hat{a}_t$  (similar to the source attention distribution above), iii) to calculate  $p_{\text{copy}}$  and  $p_{\text{generate}}$  probabilities that decide either to copy one of the previously predicted nodes by sampling a node from the target attention distribution  $\hat{a}_t$ , or to generate a new node from the output vocabulary. Each newly generated node is assigned a unique index, or it is assigned the index of the node copied from the previously generated concepts. At prediction time, we employ a beam search to decode the list of nodes based on the probability distribution computed above.
+
+Relation identification For this module, we follow Zhang et al. (2019) and use a deep biaffine classifier inspired by Dozat and Manning (2017), which takes as input the decoder states and factorizes the edge prediction in two components predicting i) whether there is an edge between a pair of nodes, and ii) the edge label for each possible edge, respectively. We direct the reader to Zhang et al. (2019) and Dozat and Manning (2017) for technical details on the biaffine attention classifier. At prediction time, to ensure the validity of the tree, given the list of predicted nodes and the score for candidate edges, we search for the highest-scoring spanning tree using the Chu-Liu-Edmonds algorithm. We then merge the duplicate nodes based on the node indices to restore the final AMR graph. The model is trained to jointly minimize the loss of reference nodes and edges.
+
+# 3.3 Silver Training Data
+
+In order to train cross-lingual AMR parsers and to evaluate the cross-lingual properties of AMR as an interlingua, we project existing AMR annotations for English sentences to target language sentences following two different approaches.
+
+Parallel sentences - silver AMR graphs We follow Damonte and Cohen (2018) and project AMR graphs from English sentences to target language sentences through a parallel corpus. Differently from Damonte and Cohen (2018), we do not need word-to-word and word-to-node aligners for training the concept identification module. Instead we
+
+directly pair a sentence in the target language with the AMR graph corresponding to its English counterpart. In this case, while the sentences are parallel, the AMR graphs are of silver standard quality, i.e., the English sentences of the parallel corpus are parsed using an existing AMR parser. We refer to this method as PARSENTS-SILVERAMR.
+
+Gold AMR graphs - silver translations In addition to pivoting through parallel sentences, we investigate whether considering human-annotated AMR graphs could bring more benefits than system produced AMR graphs. To this end, we make use of the existing gold standard datasets for AMR parsing, i.e., English sentence-AMR graph pairs, and use machine translation systems to translate the training sentences into the target language. This choice is motivated by the existence of reliable machine translation systems for the languages of our interest. Moreover, we validate the silver translations through a back-translation step (Sennrich et al., 2016). That is, firstly, we translate the sentences from English to the target language and, secondly, using the same neural translation model, we translate the target language translations back to English. Then, to filter out less accurate translations we apply a 1-NN strategy based on the cosine similarity between translations and source sentence semantic embeddings, similarly to Artetxe and Schwenk (2019a). If the nearest neighbour of a translation corresponds to its source English sentence, we consider it a good translation, otherwise we discard it. We employ semantic similarity since we have a two-step automatic translation, due to which lexical differences are introduced into translations compared to the original sentence. Typical machine translation metrics, e.g., BLEU, METEOR, rely on lexical similarity, which could lead good translations being discarded. In fact, we do not need the translation to be word-to-word aligned, but rather to preserve the meaning of the sentence, thus considering valid also the cases when certain words are translated into synonyms or related words. We refer to this method as GOLDAMR-SILVERTRNS.
+
+# 3.4 Pre- and Postprocessing
+
+AMR parsers in the literature rely on several pre- and postprocessing rules. We extend these rules for the cross-lingual AMR parsing task based on several multilingual resources such as Wikipedia, BabelNet 4.0 (Navigli and Ponzetto, 2010), DBpedia Spotlight API (Daiber et al., 2013) for wikifi
+
+<table><tr><td>Dataset</td><td>Lang</td><td>Train Insts</td><td>Dev Insts</td><td>Source</td></tr><tr><td>Gold</td><td>EN</td><td>36521</td><td>1368</td><td>AMR 2.0</td></tr><tr><td rowspan="2">PARSENTS</td><td>DE</td><td>20000</td><td>2000</td><td>Europarl</td></tr><tr><td>EN</td><td>20000</td><td>2000</td><td>Europarl</td></tr><tr><td rowspan="2">SILVERAMR</td><td>ES</td><td>20000</td><td>2000</td><td>Europarl</td></tr><tr><td>IT</td><td>20000</td><td>2000</td><td>Europarl</td></tr><tr><td rowspan="2">GOLDAMR</td><td>DE</td><td>34415</td><td>1319</td><td>AMR 2.0</td></tr><tr><td>ES</td><td>34552</td><td>1325</td><td>AMR 2.0</td></tr><tr><td rowspan="2">SILVERTRNS</td><td>IT</td><td>34521</td><td>1322</td><td>AMR 2.0</td></tr><tr><td>ZH</td><td>32154</td><td>1276</td><td>AMR 2.0</td></tr></table>
+
+Table 1: Dataset quality standard, instances per language, and the source corpus of the sentences.
+
+cation in all languages but Chinese, for which we use Babelfy (Moro et al., 2014) instead, Stanford CoreNLP (Manning et al., 2014) for English preprocessing pipeline, the Stanza Toolkit (Qi et al., 2020) for Chinese, German and Spanish sentences, and  $\mathrm{Tint}^3$  (Aprosio and Moretti, 2016) for Italian.
+
+The preprocessing steps consist of: i) lemmatization, ii) PoS tagging, iii) NER, iv) re-categorization of entities and senses, v) removal of wiki links and polarity attributes. The postprocessing steps consist of restoring i) anonymized subgraphs, ii) wikification, iii) senses, iv) polarity attributes. We give full details on pre- and postprocessing in Appendix A.
+
+# 4 Experiments
+
+We now present a set of experiments for cross-lingual AMR parsing when using different training techniques and the silver data we created (see Section 3.3). We discuss the results of our multiple settings and compare with previous approaches performing cross-lingual AMR parsing.
+
+Test bed We evaluate on the Abstract Meaning Representation 2.0 - Four Translations (Damonte and Cohen, 2020), a corpus containing translations of the test split of 1371 sentences from the LDC2017T10 (AMR 2.0), in Chinese (ZH), German (DE), Italian (IT) and Spanish (ES). This data is designed for use in cross-lingual AMR parsing (available to all LDC subscribers).
+
+Dataset In Section 3.3, we explained the two projection approaches for obtaining cross-lingual AMR data, i.e., PARSENTS-SILVERAMR and GOLDAMR-SILVERTRNS.
+
+For the first approach, inspired by Damonte and Cohen (2018), and for comparison purposes, we
+
+choose Europarl as parallel corpus.4 We predict the silver AMR using the model of Zhang et al. (2019).
+
+For the second approach, instead, i.e., GOLDAMR-SILVERTRNS, we choose AMR 2.0 as gold dataset and translate the sentences into Chinese, German, Italian and Spanish. For German, Italian and Spanish, for both translating and back-translating the sentences we use the machine translation models made available by Tiedemann and Thottingal (2020, OPUS-MT). For Chinese, instead, since OPUS-MT does not provide translation models, we employ the released  $\mathrm{MASS^6}$  (Song et al., 2019a) supervised neural translation models. Then, to filter out less accurate translations, we compute the cosine similarity between dense semantic representations of the original English sentence and its back-translated counterpart. To embed the sentences we use LASER (Artetxe and Schwenk, 2019b), a state-of-the-art model for sentence embeddings. Details on the number of instances per language and for each silver data approach are shown in Table 1.
+
+Training configurations We conduct experiments following different training approaches:
+
+- Zero-shot - the model is trained on English sentences only, relying on multilingual features, and is evaluated on all the target languages (henceforth  $\emptyset$ -shot).  
+- Language-specific - the model is trained only on target language data, i.e., DE, ES, IT or ZH, and evaluated in the same language.  
+- Bilingual - the model is trained on English data and one of either DE, ES, IT or ZH, and evaluated in the target language.  
+- Multilingual - the model is trained on data from all available languages per setting and evaluated on the target languages.
+
+Systems We denote the variations of XL-AMR, based on the above training configurations, as XL-AMR data where, data  $\in$  {par, trans, amr}, par referring to the data produced with PARSENTS-SILVERAMR approach, trans to GOLDAMR-SILVERTRNS approach, amr to the AMR 2.0 English gold standard, and data+ refers to combining par or trans with amr. The only existing crosslingual AMR parser from the literature to date is
+
+<table><tr><td>Parser</td><td>Configuration</td><td>DE</td><td>ES</td><td>IT</td><td>ZH</td></tr><tr><td>AMREAGER</td><td>Lang-Spec.</td><td>39.0</td><td>42.0</td><td>43.0</td><td>35.0</td></tr><tr><td>XL-AMRamr∅</td><td>∅-shot</td><td>32.7</td><td>39.1</td><td>37.1</td><td>25.9</td></tr><tr><td>XL-AMRpar+∅</td><td>∅-shot</td><td>38.3</td><td>41.8</td><td>41.0</td><td>23.9</td></tr><tr><td rowspan="3">XL-AMRpar</td><td>Lang-Spec.</td><td>40.8</td><td>44.2</td><td>43.4</td><td>-</td></tr><tr><td>Multiling.</td><td>41.5</td><td>45.6</td><td>45.0</td><td>-</td></tr><tr><td>Biling.</td><td>42.7</td><td>47.9</td><td>46.7</td><td>-</td></tr><tr><td rowspan="2">XL-AMRpar+</td><td>Multiling.</td><td>46.3</td><td>51.2</td><td>50.9</td><td>-</td></tr><tr><td>Biling.</td><td>47.0</td><td>53.0</td><td>51.4</td><td>-</td></tr><tr><td rowspan="3">XL-AMRtrans</td><td>Lang-Spec.</td><td>51.6</td><td>56.1</td><td>56.7</td><td>43.1</td></tr><tr><td>Multiling.</td><td>49.9</td><td>53.0</td><td>54.0</td><td>40.0</td></tr><tr><td>Multiling. (-ZH)</td><td>51.5</td><td>55.5</td><td>55.9</td><td>-</td></tr><tr><td rowspan="3">XL-AMRtrans+</td><td>Multiling.</td><td>49.9</td><td>53.2</td><td>53.5</td><td>41.0</td></tr><tr><td>Multiling. (-ZH)</td><td>52.1</td><td>56.2</td><td>56.7</td><td>-</td></tr><tr><td>Biling.</td><td>53.0</td><td>58.0</td><td>58.1</td><td>41.5</td></tr></table>
+
+Table 2: Smatch F1 scores on DE, ES, IT and ZH. Best scores per language are denoted in bold.
+
+the one of Damonte and Cohen (2018, AMREAGER Multilingual), henceforth AMREAGER. We compare the results of the XL-AMR variants with the projection method of AMREAGER on the gold dataset, i.e., AMR 2.0 - Four Translations. We remark that we do not consider the results of their Machine Translation<sup>7</sup> method, since, as emphasised by the authors, it is not informative in terms of cross-lingual properties of AMR (Damonte and Cohen, 2018) because it performs English AMR parsing. We provide details of our model hyperparameters in Appendix C.
+
+Results In Table 2 we show the Smatch $^{8}$  score of the models. This metric computes the degree of overlap of two AMR graphs (Cai and Knight, 2013).
+
+We point out the low score of the  $\emptyset$ -shot models, i.e.,  $\mathrm{XL - AMR}_{\emptyset}^{amr}$  and  $\mathrm{XL - AMR}_{\emptyset}^{par + }$ , which perform lower than AMREAGER, especially in the Chinese language. However,  $\mathrm{XL - AMR}_{\emptyset}^{par + }$  noticeably improves over  $\mathrm{XL - AMR}_{\emptyset}^{amr}$ , which can be explained by the fact that seq2seq requires a large amount of data in order to generalize. This is confirmed by a fine-grained analysis showing lower accuracy of  $\mathrm{XL - AMR}_{\emptyset}^{amr}$  compared to  $\mathrm{XL - AMR}_{\emptyset}^{par + }$  in concept identification, which, we recall, is a seq2seq module.
+
+Interestingly, the language-specific XL-AMR $^{par}$ , even if trained on less instances, outperforms the  $\emptyset$ -shot models by a large margin. Moreover, it also surpasses AMREAGER, which is trained on the same sentences from Europarl. The results are
+
+further improved when jointly training in multiple languages, i.e., when using the multilingual and bilingual configurations. We attribute this improvement to the ability of a seq2seq model to learn better when provided with a larger training set. The domain of the Europarl data is very specific, which does not enable the model to generalize in sentences from other domains. In fact, the XL-AMR $^{par+}$  models significantly improve over the XL-AMR $^{par}$  bilingual and multilingual models. We attribute the higher performances of XL-AMR $^{par+}$  to i) larger training dataset, ii) training on different domains, and iii) better quality of the data (AMR 2.0 data is human annotated).
+
+The XL-AMR<sup>trans</sup> models perform best: we note that the performances of the language-specific variants outperform those of the multilingual XL-AMR<sup>trans</sup> models, in contrast to the behaviour of the XL-AMR<sup>par</sup> models, suggesting that the addition of silver data in other languages is not beneficial. This may be due to the fact that the AMR graphs of translated sentences are the same, thus as a consequence the model does not access extra information. Moreover, the inclusion of translated sentences in other languages slightly harms the performances. This is confirmed by the removal from the training set of the most distant language, in the multilingual (-ZH) model, which in turn achieves around 2 F1 points more compared to the multilingual version including Chinese. This can be further explained by the linguistic differences between Chinese and the other languages, which prevent them from benefiting from the inclusion of Chinese instances in the training set. However, when adding English gold AMR 2.0, i.e., XL-AMR<sup>trans+</sup>, the model benefits from the better quality of this dataset. In fact, the bilingual version of XL-AMR<sup>trans+</sup> is the best performing across the board in German, Spanish and Italian, surpassing AMREAGER by at least 14 F1 points and both XL-AMR<sup>par</sup> and XL-AMR<sup>par+</sup> by at least 5 F1 points in each language. Interestingly, the best results in Chinese are achieved by the language-specific XL-AMR<sup>trans</sup> surpassing AMREAGER by 8 F1 points and the  $\emptyset$ -shot models by more than 17 F1 points. This is once again explained by the linguistic differences of Chinese as compared to the other languages, which render the additional data non-beneficial. Table 3 shows the fine-grained evaluation of AMREAGER and our best performing models for each data creation approach, for which we
+
+<table><tr><td rowspan="2">Metric</td><td colspan="4">AMREAGER</td><td colspan="4">XL-AMRpar+</td><td colspan="4">XL-AMRtrans+</td></tr><tr><td>DE</td><td>ES</td><td>IT</td><td>ZH</td><td>DE</td><td>ES</td><td>IT</td><td>ZH</td><td>DE</td><td>ES</td><td>IT</td><td>ZH</td></tr><tr><td>SMATCH</td><td>39.1</td><td>42.1</td><td>43.2</td><td>34.6</td><td>47.0</td><td>53.0</td><td>51.4</td><td>-</td><td>53.0</td><td>58.0</td><td>58.1</td><td>43.1</td></tr><tr><td>Unlabeled</td><td>45.0</td><td>46.6</td><td>48.5</td><td>41.1</td><td>52.0</td><td>58.3</td><td>57.1</td><td>-</td><td>57.7</td><td>63.0</td><td>63.4</td><td>48.9</td></tr><tr><td>No WSD</td><td>39.2</td><td>42.2</td><td>42.5</td><td>34.7</td><td>47.1</td><td>53.2</td><td>51.5</td><td>-</td><td>53.2</td><td>58.4</td><td>58.4</td><td>43.2</td></tr><tr><td>Reentrancies</td><td>18.6</td><td>27.2</td><td>25.7</td><td>15.9</td><td>33.6</td><td>40.1</td><td>39.2</td><td>-</td><td>39.9</td><td>46.6</td><td>46.1</td><td>34.7</td></tr><tr><td>Concepts</td><td>44.9</td><td>53.3</td><td>52.3</td><td>39.9</td><td>48.7</td><td>58.0</td><td>55.6</td><td>-</td><td>58.0</td><td>65.9</td><td>64.7</td><td>48.0</td></tr><tr><td>Named Ent.</td><td>63.1</td><td>65.7</td><td>67.7</td><td>67.9</td><td>63.1</td><td>61.6</td><td>62.7</td><td>-</td><td>66.0</td><td>66.2</td><td>70.0</td><td>60.6</td></tr><tr><td>Wikification</td><td>49.9</td><td>44.5</td><td>50.6</td><td>46.8</td><td>61.4</td><td>63.8</td><td>66.1</td><td>-</td><td>60.9</td><td>63.1</td><td>67.0</td><td>54.5</td></tr><tr><td>Negation</td><td>18.6</td><td>19.8</td><td>22.3</td><td>6.8</td><td>8.1</td><td>21.5</td><td>25.7</td><td>-</td><td>11.7</td><td>23.4</td><td>29.2</td><td>12.8</td></tr><tr><td>SRL</td><td>29.4</td><td>35.9</td><td>34.3</td><td>27.2</td><td>40.8</td><td>48.7</td><td>46.7</td><td>-</td><td>47.9</td><td>55.2</td><td>54.7</td><td>41.3</td></tr></table>
+
+Table 3: Fine-grained F1 scores DE, ES, IT and ZH. Best scores per language are denoted in bold.
+
+use the evaluation tools of Damonte et al. (2017). The fine-grained results for the AMREAGER are not reported by Damonte and Cohen (2018), therefore we run the evaluation using their released models. Our best model outperforms AMREAGER in all subtasks except for Negations in German and Named Entities in Chinese, which are prone to heuristic string matching errors in the pre- and postprocessing procedure of our models. XL-AMR<sup>trans+</sup> achieves significantly higher performance in Reentrancies, Concepts, SRL, in all the tested languages, compared to AMREAGER, thus demonstrating the effectiveness of our parser and data creation approaches.
+
+In summary, translating the gold standard training data, i.e., GOLDAMR-SILVERTRNS, leads XL-AMR to achieve higher performances than when trained on parallel sentences associated with silver AMR graphs, i.e., PARSENTS-SILVERAMR.
+
+# 5 Qualitative Analysis
+
+We manually check the predictions of XL-AMR in order to establish the nature of the mistakes based on the Smatch score between the gold and predicted AMR graphs and determine their severity. Then, we observe how XL-AMR handles the translation divergences, i.e., linguistic distinctions that make transfer across languages difficult (Dorr, 1994).
+
+Smatch errors The parser has difficulties with some compounded words in German, e.g., Uranproduktionsfähigkeit (uranium production capability), Kernkraftstoffkreislauf (nuclear fuel cycle), for which it fails to break their meaning down to the correct subgraph, e.g., (c / cycle-02 :ARG1 (f / fuel :mod (n / nucleus))), thus predicting a generic node,
+
+i.e., (t / thing). This issue can be alleviated using a better preprocessing to split the compounds.
+
+Several cases with low Smatch score are due to inconsistent translations of test set sentences into the target language, even though, we recall, the test set has been manually translated. This could be due to translator choices, but can lead to divergent meaning structures, e.g., Ich kann verstehen, wie Du Dich fuhlst (DE) (I can understand how you are feeling) whose original English sentence from which the AMR graph is projected is I know what you're feeling. The gold AMR graph is thus not appropriate for the German sentence, due to the sentence's different meaning. Thus these mistakes are not due to the parser, but to the translations.
+
+An interesting cause of drop in the Smatch arises from the prediction of concepts that are synonyms of the corresponding concepts in the gold graph, e.g., say-01  $\rightarrow$  state-01, stop-01  $\rightarrow$  halt-01, best friend  $\rightarrow$  best mate, demand-01  $\rightarrow$  urge-01, etc. We notice that the predicted concepts (to the left of the arrow) are less specific than the gold concepts, yet somehow preserve the meaning. These examples show that the parser captures a close meaning even when failing to predict the exact concept.
+
+Translation divergences We investigate how XL-AMR deals with the cases where there exist translation divergences, i.e., cases in which source and target language have different syntactic ordering properties (Dorr, 1990), as classified by Dorr (1994) using the following 7 categories: i) thematic, ii) promotional, iii) demotional, iv) structural, v) conflational, vii) categorial, vii) lexical.[11]
+
+A thematic divergence happens when the argument-predicate structure is different across lan
+
+guages, e.g., I like travelling where  $I$  is the subject, in Italian becomes  $M_i$  piace viaggiare, and  $M_i$  is now the object. XL-AMR overcomes this divergence and predicts the correct AMR, (1 / like-01 :ARG0 (i / I) :ARG1 (t / travel :ARG0 i)).
+
+Promotional and demotional divergences can be merged into the head switching macro-category. They arise when a modifier in one language is promoted to a main verb in the other, or vice versa, e.g., John usually goes home is Juan suele ir a casa (John is accustomed to go home) in Spanish. XL-AMR correctly parses the sentence into (g / go-01 :ARG0 (p / person :name (n / Juan)):ARG4 (h / home):mod (u / usual)).
+
+A structural divergence exists when a verbal object is realized as a noun phrase (NP) in one language and as prepositional phrase (PP) in the other, e.g., I saw John where John is NP, is translated as Vi a Juan (I saw to John) in Spanish where a Juan is PP. This also is not a problem for our parser, which predicts the correct graph, (s / see-01 :ARG0 (i / I) :ARG1 (p / person :name (n / Juan))).
+
+A conflational divergence refers to the translation of two or more words in one language into one word in the other. The above errors in German compounded words fall into this category and our model does not handle them properly. However, regarding other languages this problem is not common, e.g., I fear translates into Io ho paura (I have fear) in Italian and the parser correctly predicts the AMR graph, (f / fear-01 :ARG0 (i / I)).
+
+A categorical divergence arises when the same meaning is expressed by different syntactic categories across languages, e.g., I agree, where agree is a verb, is expressed by a noun in Italian and Spanish, Sono d'accordo and Estoy de acuerdo. The parser correctly predicts the same AMR for both languages, (a / agree-01 :ARG0 (i / I)).
+
+A lexical divergence arises when a verb in the source language is translated with a different lexical verb, e.g., John broke into the room, Juan forzo la entrada al cuarto, in which the verb break in English is translated with the verb forzar (force) in Spanish. XL-AMR predicts (f / force-01 :ARG0 (p / person :name (n / Juan)):ARG2 (e / enter-01 :ARG0 p :ARG1 (r / room))) for the Spanish sentence, which, even though it is correctly parsed, does not overcome the lexical difference of the action, which results in different AMR graphs for the same meaning. This is partially due to the fact that AMR is bounded to lexical forms in English.
+
+In summary, XL-AMR overcomes most of the foregoing structural divergences with the exception of two cases: i) the conflational divergence in German, that is caused by the language's compound words vocabulary, for the resolution of which a better preprocessing can be beneficial; ii) the lexical divergence that persists despite the parser predicting a valid graph. The latter divergence results in non-parallel structures for parallel meanings, and we believe this might be tackled by integrating a unified ontology for synonyms or related meanings within the AMR formalism, along the line of disjunctive  $\mathrm{AMR}^{12}$  (Banarescu et al., 2013). We leave exploration of this approach open for future work.
+
+# 6 Conclusion
+
+We explored transfer learning techniques to enable high performance cross-lingual AMR parsing. We created silver data based on annotation projection through parallel sentences and machine translation, on which we trained XL-AMR, a cross-lingual AMR parser that achieves the highest results reported to date on Chinese, German, Italian and Spanish. A qualitative evaluation showed that XL-AMR is able to handle most of the structural divergences among languages. The performance of XL-AMR together with the qualitative analysis suggests that carefully modeling cross-lingual AMR parsing leads to the production of suitable AMR structures across languages. It would therefore be promising to extend this line of our research to exploit larger multilingual semantic resources, in order to further improve the parsing quality. These AMR representations could then be integrated into downstream cross-lingual tasks to investigate their added value.
+
+# Acknowledgments
+
+![](images/488feb7fc60ff0d3a6e14b660373b7226caf9976b9921161b42e7070b63a742f.jpg)
+
+The authors gratefully acknowledge the support of the ERC Consolidator Grant MOUSSE No. 726487 and the ELEXIS project No. 731015 under the European Union's Horizon 2020 research and innovation programme.
+
+![](images/908216f05b7c9855314d8e605e53774f9b71dcf25f5935a63f89dbf640b569f4.jpg)
+
+This work was partially supported by the MIUR under the grant "Dipartimenti di eccellenza 2018-2022" of the Department of Computer Science of Sapienza University.
+
+The authors would like to thank Luigi Procopio for the valuable discussions during this work.
+
+# References
+
+Omri Abend and Ari Rappoport. 2013. Universal conceptual cognitive annotation (UCCA). In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 228-238, Sofia, Bulgaria.  
+Alan Akbik, Laura Chiticariu, Marina Danilevsky, Yun-yao Li, Shivakumar Vaithyanathan, and Huaiyu Zhu. 2015. Generating high quality proposition Banks for multilingual semantic role labeling. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 397-407, Beijing, China.  
+Alessio Palmero Aprosio and Giovanni Moretti. 2016. Italy goes to stanford: a collection of corenlp modules for italian. CoRR, abs/1609.06204.  
+Mikel Artetxe and Holger Schwenk. 2019a. Margin-based parallel corpus mining with multilingual sentence embeddings. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3197-3203, Florence, Italy.  
+Mikel Artetxe and Holger Schwenk. 2019b. Massively multilingual sentence embeddings for zero-shot cross-lingual transfer and beyond. Transactions of the Association for Computational Linguistics, 7:597-610.  
+Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.  
+Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffith, Ulf Hermjakob, Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. 2013. Abstract meaning representation for sembanking. In Proceedings of the 7th Linguistic Annotation Workshop and Interoperability with Discourse, pages 178-186, Sofia, Bulgaria.  
+Edoardo Barba, Luigi Procopio, Niccolò Campolungo, Tommaso Pasini, and Roberto Navigli. 2020. MuLaN: Multilingual label propagation for word sense disambiguation. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI-20, pages 3837-3844.  
+Claire Bonial, Lucia Donatelli, Mitchell Abrams, Stephanie M. Lukin, Stephen Tratz, Matthew Marge, Ron Artstein, David Traum, and Clare Voss. 2020. Dialogue-AMR: Abstract Meaning Representation for dialogue. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 684-695, Marseille, France.
+
+Shu Cai and Kevin Knight. 2013. Smatch: an evaluation metric for semantic feature structures. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 748-752, Sofia, Bulgaria.  
+Simone Conia and Roberto Navigli. 2020. Conception: Multilingually-enhanced, human-readable concept vector representations. In Proceedings of the 28th International Conference on Computational Linguistics, COLING 2020.  
+Joachim Daiber, Max Jakob, Chris Hokamp, and Pablo N. Mendes. 2013. Improving efficiency and accuracy in multilingual entity extraction. In ISEMANTICS 2013 - 9th International Conference on Semantic Systems, ISEM '13, Graz, Austria, September 4-6, 2013, pages 121-124.  
+Marco Damonte and Shay Cohen. 2020. Abstract Meaning Representation 2.0 - Four Translations LDC2020T07. Web Download, Philadelphia: Linguistic Data Consortium.  
+Marco Damonte and Shay B. Cohen. 2018. Crosslingual abstract meaning representation parsing. 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 1146-1155, New Orleans, Louisiana.  
+Marco Damonte, Shay B. Cohen, and Giorgio Satta. 2017. An incremental parser for Abstract Meaning Representation. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 536-546, Valencia, Spain.  
+Marco Damonte, Rahul Goel, and Tagyoung Chung. 2019. Practical semantic parsing for spoken language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Industry Papers), pages 16-23, Minneapolis - Minnesota.  
+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.  
+Andrea Di Fabio, Simone Conia, and Roberto Navigli. 2019. VerbAtlas: a novel large-scale verbal semantic resource and its application to semantic role labeling. 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 627-637, Hong Kong, China.
+
+Bonnie Dorr. 1990. Solving thematic divergences in machine translation. In 28th Annual Meeting of the Association for Computational Linguistics, pages 127-134, Pittsburgh, Pennsylvania, USA.  
+Bonnie J. Dorr. 1994. Machine translation divergences: A formal description and proposed solution. Computational Linguistics, 20(4):597-633.  
+Timothy Dozat and Christopher D. Manning. 2017. Deep biaffine attention for neural dependency parsing. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings.  
+Jeffrey Flanigan, Sam Thomson, Jaime Carbonell, Chris Dyer, and Noah A. Smith. 2014. A discriminative graph-based parser for the abstract meaning representation. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1426-1436, Baltimore, Maryland.  
+DongLai Ge, Junhui Li, Muhua Zhu, and Shoushan Li. 2019. Modeling source syntax and semantics for neural AMR parsing. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019, pages 4975-4981.  
+Jan Hajic, Ondrej Bojar, and Zdenka Urešová. 2014. Comparing Czech and English AMRs. In Proceedings of Workshop on Lexical and Grammatical Resources for Language Processing, pages 55-64, Dublin, Ireland.  
+Jan Hajic, Eva Hajicova, Jarmila Paneová, Petr Sgall, Ondrej Bojar, Silvie Cinková, Eva Fučíková, Marie Mikulova, Petr Pajas, Jan Popelka, Jiří Semecký, Jana Sindlerová, Jan Štepanek, Josef Toman, Zdenka Urešová, and Zdeněk Žabokrtský. 2012. Announcing Prague Czech-English dependency treebank 2.0. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC-2012), pages 3153-3160, Istanbul, Turkey.  
+Hardy Hardy and Andreas Vlachos. 2018. Guided neural language generation for abstractive summarization using Abstract Meaning Representation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 768-773, Brussels, Belgium.  
+Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural Computation, 9(8):1735-1780.  
+Fuad Issa, Marco Damonte, Shay B. Cohen, Xiaohui Yan, and Yi Chang. 2018. Abstract meaning representation for paraphrase detection. 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 442-452, New Orleans, Louisiana.
+
+Yoon Kim, Yacine Jernite, David A. Sontag, and Alexander M. Rush. 2016. Character-aware neural language models. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, February 12-17, 2016, Phoenix, Arizona, USA, pages 2741-2749.  
+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.  
+Paul Kingsbury and Martha Palmer. 2002. From TreeBank to PropBank. In Proceedings of the Third International Conference on Language Resources and Evaluation (LREC'02), Las Palmas, Canary Islands - Spain.  
+Eliyahu Kiperwasser and Yoav Goldberg. 2016. Simple and accurate dependency parsing using bidirectional LSTM feature representations. Transactions of the Association for Computational Linguistics, 4:313-327.  
+Philipp Koehn. 2005. Europarl: A Parallel Corpus for Statistical Machine Translation. In Conference Proceedings: the tenth Machine Translation Summit, pages 79-86, Phuket, Thailand. AAMT.  
+Ioannis Konstas, Srinivasan Iyer, Mark Yatskar, Yejin Choi, and Luke Zettlemoyer. 2017. Neural AMR: Sequence-to-sequence models for parsing and generation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 146-157, Vancouver, Canada.  
+Mikhail Kozhevnikov and Ivan Titov. 2013. Crosslingual transfer of semantic role labeling models. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1190-1200, Sofia, Bulgaria.  
+Kexin Liao, Logan Lebanoff, and Fei Liu. 2018. Abstract Meaning Representation for multi-document summarization. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1178-1190, Santa Fe, New Mexico, USA. Association for Computational Linguistics.  
+Yijia Liu, Wanxiang Che, Bo Zheng, Bing Qin, and Ting Liu. 2018. An AMR aligner tuned by transition-based parser. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2422-2430, Brussels, Belgium.  
+Thang Luong, Hieu Pham, and Christopher D. Manning. 2015. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1412-1421, Lisbon, Portugal.
+
+Chunchuan Lyu and Ivan Titov. 2018. AMR parsing as graph prediction with latent alignment. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 397-407, Melbourne, Australia.  
+Weimin Lyu, Sheng Huang, Abdul Rafae Khan, Shengqiang Zhang, Weiwei Sun, and Jia Xu. 2019. CUNY-PKU parser at SemEval-2019 task 1: Crosslingual semantic parsing with UCCA. In Proceedings of the 13th International Workshop on Semantic Evaluation, pages 92–96, Minneapolis, Minnesota, USA.  
+Christopher Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven Bethard, and David McClosky. 2014. The Stanford CoreNLP natural language processing toolkit. In Proceedings of 52nd Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 55-60, Baltimore, Maryland.  
+Marie-Catherine de Marneffe, Timothy Dozat, Natalia Silveira, Katri Haverinen, Filip Ginter, Joakim Nivre, and Christopher D. Manning. 2014. Universal Stanford dependencies: A cross-linguistic typology. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC-2014), pages 4585-4592, Reykjavik, Iceland.  
+Andrea Moro, Alessandro Raganato, and Roberto Navigli. 2014. Entity linking meets word sense disambiguation: a unified approach. Transactions of the Association for Computational Linguistics, 2:231-244.  
+Roberto Navigli. 2018. Natural language understanding: Instructions for (present and future) use. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, pages 5697-5702.  
+Roberto Navigli and Simone Paolo Ponzetto. 2010. *BabelNet: Building a very large multilingual semantic network*. In *Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics*, pages 216–225, Uppsala, Sweden.  
+Rik van Noord and Johan Bos. 2017. Neural semantic parsing by character-based translation: Experiments with abstract meaning representations. CoRR, abs/1705.09980.  
+Stephan Oepen and Jan Tore Lønning. 2006. Discriminant-based MRS banking. In Proceedings of the Fifth International Conference on Language Resources and Evaluation (LREC'06), Genoa, Italy.  
+Sebastian Padó and Mirella Lapata. 2009. Cross-lingual annotation projection for semantic roles. Journal of Artificial Intelligence Research, 36:307-340.
+
+Tommaso Pasini. 2020. The knowledge acquisition bottleneck problem in multilingual word sense disambiguation. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-20, Yokohama, Japan.  
+Xiaochang Peng, Chuan Wang, Daniel Gildea, and Nianwen Xue. 2017. Addressing the data sparsity issue in neural AMR parsing. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 366-375, Valencia, Spain.  
+Slav Petrov, Dipanjan Das, and Ryan McDonald. 2012. A universal part-of-speech tagset. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC-2012), pages 2089-2096, Istanbul, Turkey.  
+Nima Pourdamghani, Yang Gao, Ulf Hermjakob, and Kevin Knight. 2014. Aligning English strings with abstract meaning representation graphs. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 425-429, Doha, Qatar.  
+Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. Stanza: A python natural language processing toolkit for many human languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 101-108, Online.  
+Sudha Rao, Daniel Marcu, Kevin Knight, and Hal Daumé III. 2017. Biomedical event extraction using Abstract Meaning Representation. In BioNLP 2017, pages 126-135, Vancouver, Canada..  
+Sebastian Ruder, Ivan Vulic, and Anders Søgaard. 2019. A survey of cross-lingual word embedding models. Journal of Artificial Intelligence Research, 65:569-631.  
+Mike Schuster and Kuldip K. Paliwal. 1997. Bidirectional recurrent neural networks. IEEE Trans. Signal Process., 45(11):2673-2681.  
+Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 86-96, Berlin, Germany.  
+Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2019a. MASS: masked sequence to sequence pre-training for language generation. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, volume 97 of Proceedings of Machine Learning Research, pages 5926-5936.
+
+Linfeng Song, Daniel Gildea, Yue Zhang, Zhiguo Wang, and Jinsong Su. 2019b. Semantic neural machine translation using AMR. Transactions of the Association for Computational Linguistics, 7:19-31.  
+Robyn Speer, Joshua Chin, and Catherine Havasi. 2017. Conceptnet 5.5: An open multilingual graph of general knowledge. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, AAAI'17, page 4444-4451.  
+Jörg Tiedemann. 2015. Improving the cross-lingual projection of syntactic dependencies. In Proceedings of the 20th Nordic Conference of Computational Linguistics (NODALIDA 2015), pages 191–199, Vilnius, Lithuania.  
+Jörg Tiedemann and Santhosh Thottingal. 2020. OPUS-MT - building open translation services for the world. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation, pages 479-480, Lisboa, Portugal.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, 4-9 December 2017, Long Beach, CA, USA, pages 5998-6008.  
+Kai Von Fintel and Lisa Matthewson. 2008. Universals in semantics. The linguistic review, 25(1-2):139-201.  
+Aaron Steven White, Drew Reisinger, Keisuke Sakaguchi, Tim Vieira, Sheng Zhang, Rachel Rudinger, Kyle Rawlins, and Benjamin Van Durme. 2016. Universal decompositional semantics on universal dependencies. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1713-1723, Austin, Texas.  
+Nianwen Xue, Ondrej Bojar, Jan Hajic, Martha Palmer, Zdenka Urešová, and Xiuhong Zhang. 2014. Not an interlingua, but close: Comparison of English AMRs to Chinese and Czech. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC-2014), pages 1765–1772, Reykjavik, Iceland.  
+Sheng Zhang, Xutai Ma, Kevin Duh, and Benjamin Van Durme. 2019. AMR parsing as sequence-tograph transduction. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 80-94, Florence, Italy.  
+Sheng Zhang, Xutai Ma, Rachel Rudinger, Kevin Duh, and Benjamin Van Durme. 2018. Cross-lingual decompositional semantic parsing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1664–1675, Brussels, Belgium.
+
+Huaiyu Zhu, Yunyao Li, and Laura Chiticariu. 2019. Towards universal semantic representation. In Proceedings of the First International Workshop on Designing Meaning Representations, pages 177-181, Florence, Italy.
+
+# A Cross-Lingual AMR Pre- and Postprocessing
+
+English AMR parsers throughout the literature rely on several pre- and postprocessing rules. We extend these rules for the cross-lingual AMR parsing task based on several multilingual resources.
+
+Preprocessing This step consists of: i) lemmatization, ii) PoS-tagging, iii) NER, iv) re-categorization of entities and senses and v) removal of wiki links and polarity attributes. As NLP pipelines (steps i-iii) we use Stanford CoreNLP (Manning et al., 2014) for English sentences, the Stanza Toolkit (Qi et al., 2020) for Chinese, German and Spanish sentences, and Tint<sup>13</sup> (Aprosio and Moretti, 2016) for Italian. Re-categorization and anonymization of entities is often used in English AMR parsing to reduce data sparsity (Zhang et al., 2019; Lyu and Titov, 2018; Peng et al., 2017; Konstas et al., 2017). Here we follow Konstas et al. (2017); Zhang et al. (2019) and anonymize entity subgraphs, which are identified by an AMR entity type and the :name role. First, the entity subgraphs are mapped with the corresponding text span in the sentence and then the text span is replaced with the anonymized token, i.e., EN-TITY_TYPE_i. To match the entities in the AMR graphs, which are tied to English, with the corresponding text span in non-English sentences, we first collect all the possible lexicalizations of the entity in the target language using BabelNet 4.0 (Navigli and Ponzetto, 2010), a multilingual semantic network which brings together different resources such as WordNet, Wikipedia, etc., each node of which clusters together the lexicalizations that express the same concept in different languages. Then we search for the possible text spans in the sentence written in the target language. At test time, we anonymize the text spans which have been identified during the training data preprocessing and which are tagged by the NER tagger as entities.
+
+Postprocessing This step consists of restoring i) anonymized subgraphs, ii) wiki links, iii) senses and iv) polarity attributes. The anonymized subgraphs are restored using the anonymized text
+
+spans created during preprocessing. Then wiki links are restored using the DBpedia Spotlight  $\mathrm{API^{14}}$  (Daiber et al., 2013), commonly used in English AMR parsing (van Noord and Bos, 2017; Zhang et al., 2019; Ge et al., 2019). It provides models for multiple languages, except Chinese, for which we use Babelfy (Moro et al., 2014). Since the wiki links identified by DBpedia Spotlight API are language-specific to the text, we further use Wikipedia inter-language links to retrieve the corresponding wiki links for the English entities. We restore senses as the most frequent sense of the predicate in the training data (using -01 if unseen) similar to (Lyu and Titov, 2018; Zhang et al., 2019) and finally restore polarity attributes based on heuristic rules observed on the training data and linguistic rules specific to each language (included in the released code).
+
+# B OpusMT Translation Models
+
+For the translation and back-translation steps of GOLDAMR-SILVERTRNS data creation approach, we use the pretrained models $^{15}$  from the hugging-face transformers library $^{16}$  listed in Table 4.
+
+<table><tr><td>Source</td><td>Target</td><td>Model</td></tr><tr><td>German</td><td>English</td><td>Helsinki-NLP/opus-mt-de-en</td></tr><tr><td>Italian</td><td>English</td><td>Helsinki-NLP/opus-mt-it-en</td></tr><tr><td>Spanish</td><td>English</td><td>Helsinki-NLP/opus-mt-ROMANCE-en</td></tr><tr><td>English</td><td>German</td><td>Helsinki-NLP/opus-mt-en-de</td></tr><tr><td>English</td><td>Italian</td><td>Helsinki-NLP/opus-mt-en-it</td></tr><tr><td>English</td><td>Spanish</td><td>Helsinki-NLP/opus-mt-en-ROMANCE</td></tr></table>
+
+Table 4: OpusMT translation models.
+
+# C Model Hyperparameters
+
+The input features for all the models include: i) fixed mBERT $^{17}$  (Devlin et al., 2019) as contextual embeddings (dim = 768), ii) ConceptNet Numberbatch  $9.08^{18}$  (Speer et al., 2017) multilingual static word embeddings (dim = 300) which we set as trainable except in  $\emptyset$ -shot models, iii) trainable PoS embeddings (dim = 100) where we use the universal PoS-tags set by Petrov et al. (2012), iv) trainable anonymization indicator embeddings
+
+$(\dim = 50)$  , v) trainable character-level embeddings (dim  $= 100$  ), i.e., CharCNN (Kim et al., 2016).
+
+The encoder and decoder of the node prediction module are composed of 2 layers of 512 and 1024 LSTM units each, respectively. All the models are trained using Adam optimizer (Kingma and Ba, 2015) with learning rate 0.001, for 120 epochs and the best model hyperparameters are chosen on the basis of development set accuracy. The models are trained using 1 GeForce GTX TITAN X GPU, full training takes around 48 hours for models trained in the largest dataset XL-AMR $^{trans+}$  ( $\sim$ 84M trainable parameters) and XL-AMR $^{par+}$  ( $\sim$ 86M trainable parameters). At prediction time we set the size of beam search to 5.

xlamrenablingcrosslingualamrparsingwithtransferlearningtechniques/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ecc4005c652da56175cdc9e223bda5379e58b0a74b6a852a1d95c3cb3ec76ec
+size 242912

xlamrenablingcrosslingualamrparsingwithtransferlearningtechniques/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:15fa268844898d3cd6aa6adcf7c966b0f49754d61a98ae2b33f6e885a4a2763a
+size 420907

xlwicamultilingualbenchmarkforevaluatingsemanticcontextualization/bae702ad-b6cb-4875-9268-e1164d319f2c_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:065768fbd6cfae058fc1cc91a409a5542c092a8007e0c6442a0224616969e38e
+size 99742

xlwicamultilingualbenchmarkforevaluatingsemanticcontextualization/bae702ad-b6cb-4875-9268-e1164d319f2c_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d44bbf8759d08fc6e7792c662fdd1b9a522f7c4ed53335585b93bffc55414c28
+size 122319

xlwicamultilingualbenchmarkforevaluatingsemanticcontextualization/bae702ad-b6cb-4875-9268-e1164d319f2c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc5331e77661c708f1b4ccae51ab3b35a2aef65c5e85e3e4ad60cb60637690cb
+size 594716

xlwicamultilingualbenchmarkforevaluatingsemanticcontextualization/full.md

@@ -0,0 +1,347 @@
+# XL-WiC: A Multilingual Benchmark for Evaluating Semantic Contextualization
+
+Alessandro Raganato\* Tommaso Pasini\* Jose Camacho-Collados Mohammad Taher Pilehvar
+
+$\checkmark$  Department of Digital Humanities, University of Helsinki, Finland
+
+SapienzaNLP Group, Computer Science Department, Sapienza University of Rome, Italy
+
+\* School of Computer Science and Informatics, Cardiff University, United Kingdom
+
+$\spadesuit$  Tehran Institute for Advanced Studies, Iran
+
+$^{\diamond}$ alessandro.raganato@helsinki.fi,  $^{\diamond}$ pasini@di.uniromal.it,
+
+$\clubsuit$ camachocolladosj@cardiff.ac.uk,  $\clubsuit$ mp792@cam.ac.uk
+
+# Abstract
+
+The ability to correctly model distinct meanings of a word is crucial for the effectiveness of semantic representation techniques. However, most existing evaluation benchmarks for assessing this criterion are tied to sense inventories (usually WordNet), restricting their usage to a small subset of knowledge-based representation techniques. The Word-in-Context dataset (WiC) addresses the dependence on sense inventories by reformulating the standard disambiguation task as a binary classification problem; but, it is limited to the English language. We put forward a large multilingual benchmark, XL-WiC, featuring gold standards in 12 new languages from varied language families and with different degrees of resource availability, opening room for evaluation scenarios such as zero-shot cross-lingual transfer. We perform a series of experiments to determine the reliability of the datasets and to set performance baselines for several recent contextualized multilingual models. Experimental results show that even when no tagged instances are available for a target language, models trained solely on the English data can attain competitive performance in the task of distinguishing different meanings of a word, even for distant languages. XL-WiC is available at https://pilehvar.github.io/xlwic/.
+
+# 1 Introduction
+
+One of the desirable properties of contextualized models, such as BERT (Devlin et al., 2019) and its derivatives, lies in their ability to associate dynamic representations to words, i.e., embeddings that can change depending on the context. This provides the basis for the model to distinguish different meanings (senses) of words without the need to resort to an explicit sense disambiguation step. The conventional evaluation framework for this property has
+
+been Word Sense Disambiguation (Navigli, 2009, WSD). However, evaluation benchmarks for WSD are usually tied to external sense inventories (often WordNet (Fellbaum, 1998)), making it extremely difficult to evaluate systems that do not explicitly model sense distinctions in the inventory, effectively restricting the benchmark to inventory-based sense representation techniques and WSD systems. This prevents a direct evaluation of lexical semantic capacity for a wide range of inventory-free models, such as the dominating language model-based contextualized representations.
+
+Pilehvar and Camacho-Collados (2019) addressed this dependence on sense inventories by reformulating the WSD task as a simple binary classification problem: given a target word  $w$  in two different contexts,  $c_{1}$  and  $c_{2}$ , the task is to identify if the same meaning (sense) of  $w$  was intended in both  $c_{1}$  and  $c_{2}$ , or not. The task was framed as a dataset, called Word-in-Context (WiC), which is also a part of the widely-used SuperGLUE benchmark (Wang et al., 2019). Despite allowing a significantly wider range of models for direct WSD evaluations, WiC is limited to the English language only, preventing the evaluation of models in other languages and in cross-lingual settings.
+
+In this paper, we present a new evaluation benchmark, called XL-WiC, that extends the WiC dataset to 12 new languages from different families and with different degrees of resource availability: Bulgarian (BG), Chinese (ZH), Croatian (HR), Danish (DA), Dutch (NL), Estonian (ET), Farsi (FA), French (FR), German (DE), Italian (IT), Japanese (JA) and Korean (KO). With over 80K instances, our benchmark can serve as a reliable evaluation framework for contextualized models in a wide range of heterogeneous languages. XL-WiC can also serve as a suitable testbed for cross-lingual experimentation in settings such as zero-shot or
+
+few-shot transfer across languages. As an additional contribution, we tested several pretrained multilingual models on XL-WiC, showing that they are generally effective in transferring sense distinction knowledge from English to other languages in the zero-shot setting. However, with more training data at hand for target languages, monolingual approaches gain ground, outperforming their multilingual counterparts by a large margin.
+
+# 2 Related Work
+
+XL-WiC is a benchmark for inventory-independent evaluation of WSD models (Section 2.1), while the multilingual nature of the dataset makes it an interesting resource for experimenting with cross-lingual transfer (Section 2.2).
+
+# 2.1 Word Sense Disambiguation
+
+The ability to identify the intended sense of a polysemous word in a given context is one of the fundamental problems in lexical semantics. It is usually addressed with two different kinds of approaches relying on either sense-annotated corpora (Bevilacqua and Navigli, 2020; Scarlini et al., 2020; Blevins and Zettlemoyer, 2020) or knowledge bases (Moro et al., 2014; Agirre et al., 2014; Scozzafava et al., 2020). Both are usually evaluated on dedicated benchmarks, including at least five WSD tasks in Senseval and SemEval series, from 2001 (Edmonds and Cotton, 2001) to 2015 (Moro and Navigli, 2015a) that are included in the Raganato et al. (2017)'s test suite. All these tasks are framed as classification problems, where disambiguation of a word is defined as selecting one of the predefined senses of the word listed by a sense inventory. This brings about different limitations such as restricting senses only to those defined by the inventory, or forcing the WSD system to explicitly model sense distinctions at the granularity level defined by the inventory.
+
+Stanford Contextual Word Similarity (Huang et al., 2012) is one of the first datasets that focuses on ambiguity but outside the boundaries of sense inventories, and as a similarity measurement between two words in their contexts. Pilehvar and Camacho-Collados (2019) highlighted some of the limitations of the dataset that prevent a reliable evaluation, and proposed the Word-in-Context (WiC) dataset. WiC is the closest dataset to ours, which provides around 10K instances (1400 instances for 1184 unique target nouns and verbs in the test set),
+
+but for the English language only.
+
+# 2.2 Cross-lingual NLP
+
+A prerequisite for research on a language is the availability of relevant evaluation benchmarks. Given its importance, construction of multilingual datasets has always been considered as a key contribution in NLP research and numerous benchmarks exist for a wide range of tasks, such as semantic parsing (Hershcovich et al., 2019), word similarity (Camacho-Collados et al., 2017; Barzegar et al., 2018), sentence similarity (Cer et al., 2017), or WSD (Navigli et al., 2013; Moro and Navigli, 2015b). A more recent example is XTREME (Hu et al., 2020), a benchmark that covers around 40 languages in nine syntactic and semantic tasks.
+
+On the other hand, pre-trained language models have recently proven very effective in transferring knowledge in cross-lingual NLP tasks (Devlin et al., 2019; Conneau et al., 2020). This has further magnified the requirement for rigorous multilingual benchmarks that can be used as basis for this direction of research (Artetxe et al., 2020b).
+
+# 3 XL-WiC: The Benchmark
+
+In this section, we describe the procedure we followed to construct the XL-WiC benchmark. Our framework is based on the original WiC dataset, which we extend to multiple languages.
+
+# 3.1 English WiC
+
+Each instance of the original WiC dataset (Pilehvar and Camacho-Collados, 2019) is composed of a target word (e.g., justify) and two sentences where the target word occurs (e.g., "Justify the margins" and "The end justifies the means"). The task is a binary classification: to decide whether the same sense of the target word (justify) was intended in the two contexts or not. The dataset was built using example sentences from resources such as Wiktionary, WordNet (Miller, 1995) and VerbNet (Schuler et al., 2009).
+
+# 3.2 XL-WiC
+
+We followed Pilehvar and Camacho-Collados (2019) and constructed XL-WiC based on example usages of words in sense inventories. Example usages are curated in a way to be self contained and clearly distinguishable across different senses of a word; hence, they provide a reliable basis for the binary classification task. Specifically, for a word
+
+<table><tr><td>Lang.</td><td>Target Word</td><td>Sentence 1</td><td>Sentence 2</td><td>Label</td></tr><tr><td>EN</td><td>Beat</td><td>We beat the competition.</td><td>Agassi beat Becker in the tennis championship.</td><td>True</td></tr><tr><td>DA</td><td>Tro</td><td>Jeg tror på det, min mor fortalte.</td><td>Maria trode ekke sine egne øjne.</td><td>True</td></tr><tr><td>ET</td><td>Ruum</td><td>Ühel hetkel olin väljaspool aega ja ruumi.</td><td>Ümberringi oli lõputu tühi ruum.</td><td>True</td></tr><tr><td>FR</td><td>Causticité</td><td>Sa causticité lui a fait bien des ennemis.</td><td>La causticité des acides.</td><td>False</td></tr><tr><td>KO</td><td>TRLIMP</td><td>TRLIMP是什么呢지 doesn&#x27;t be any more.</td><td>그 어이 하는 절에 퍻리어였다면 Many do not be any more could be it.</td><td>False</td></tr><tr><td>ZH</td><td>發</td><td>建築師希望發大火燒掉城市的三分之一。</td><td>如果南美洲氣壓偏低,則印度可能發乾早</td><td>True</td></tr><tr><td>FA</td><td>şω</td><td>şωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşomega.</td><td>şωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωşωSz</td><td>False</td></tr></table>
+
+Table 1: Sample instances from XL-WiC for different languages.
+
+$w$  and for all its senses  $\{s_1^w,\dots,s_n^w\}$ , we extract from the inventory all the example usages. We then pair those examples that correspond to the same sense  $s_i^w$  to form a positive instance (True label) while examples from different senses (i.e.,  $s_i^w$  and  $s_j^w$  where  $i\neq j$ ) are paired as a negative instance (False label).
+
+We leveraged two main sense inventories for this extension: Multilingual WordNet (Section 3.2.1) and Wiktionary (Section 3.2.2).
+
+# 3.2.1 Multilingual WordNet
+
+WordNet (Miller, 1995) is the de facto sense inventory for English WSD. The resource was originally built as an English lexical database in 1995, but since then there have been many efforts to extend it to other languages (Bond and Paik, 2012). We took advantage of these extensions to construct XL-WiC. In particular, we processed the WordNet versions of Bulgarian (Simov and Osenova, 2010), Chinese (Huang et al., 2010), Croatian (Raffaelli et al., 2008), Danish (Pedersen et al., 2009), Dutch (Postma et al., 2016), Estonian (Vider and Orav, 2002), Japanese (Isahara et al., 2008), Korean (Yoon et al., 2009) and Farsi (Shamsfard et al., 2010).
+
+Farsi: Semi-automatic extraction. FarsNet v3.0 (Shamsfard et al., 2010) comprises 30K synsets with over 100K word entries. Many of these synsets are mapped to the English database; however, each synset provides just one example usage for a target word. This prevents us from applying the automatic extraction of positive examples. Therefore, we utilized a semi-automatic procedure for the construction of the Farsi set. To this end, for each word, we extracted all example usages from
+
+FarsNet, and asked an annotator to group them into positive and negative pairs. The emphasis was to make a challenging dataset with sense distinctions that are easily interpretable by humans. This can also be viewed as a case study to understand the real gap between human and machine performance in settings where manual curation of instances is feasible.
+
+Filtering. WordNet is often considered to be a fine-grained resource, especially for verbs (Duffield et al., 2007). In some cases, the exact meaning of a word can be hard to assess, even for humans. For example, WordNet lists 29 distinct meanings for the noun line, two of which correspond to the horizontally and the vertically organized line formations. Therefore, to cope with this issue, we followed Pilehvar and Camacho-Collados (2019) and filtered out all pairs whose target senses were connected by an edge (including sister-sense relations) in WordNet's semantic network or if they belonged to the same supersense, i.e., one of the 44 lexicographer files $^{2}$  in WordNet which cluster concepts into semantic categories, e.g., Animal, Cognition, Food, etc. For example, the Japanese instance "成長 中の企業は大な指導者的いなかをらむ" ("Growing companies must have bold leaders"), "彼は安定た大的な企業に投資するだけ") ("He just invested in big stable companies") for the target word "企業" ("company") is discarded as its corresponding synsets, i.e., "An organization created for business ventures" and "An institution created to conduct business", are grouped under the same supersense in WordNet, i.e., Group.
+
+Finally, all datasets are split into development<sup>3</sup> and test. At the end of this step, we ensure that both test and development sets have the same number of positive and negative instances. An excerpt of examples included in some of our datasets are
+
+<table><tr><td rowspan="2">Split</td><td rowspan="2">Stat</td><td>WiC</td><td colspan="9">Multilingual WordNet</td><td colspan="3">Wiktionary</td></tr><tr><td>EN</td><td>BG</td><td>DA</td><td>ET</td><td>FA</td><td>HR</td><td>JA</td><td>KO</td><td>NL</td><td>ZH</td><td>DE</td><td>FR</td><td>IT</td></tr><tr><td rowspan="3">Train</td><td>Instances</td><td>5428</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>48042</td><td>39428</td><td>1144</td></tr><tr><td>Unique Words</td><td>1265</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>23213</td><td>20221</td><td>721</td></tr><tr><td>Avg. Context Len</td><td>16.8</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>32.7</td><td>32.3</td><td>23.2</td></tr><tr><td rowspan="3">Dev</td><td>Instances</td><td>638</td><td>998</td><td>852</td><td>98</td><td>200</td><td>104</td><td>208</td><td>404</td><td>250</td><td>3046</td><td>8870</td><td>8588</td><td>198</td></tr><tr><td>Unique Words</td><td>599</td><td>354</td><td>542</td><td>63</td><td>174</td><td>82</td><td>137</td><td>183</td><td>150</td><td>867</td><td>4383</td><td>3517</td><td>136</td></tr><tr><td>Avg. Context Len</td><td>17.1</td><td>8.4</td><td>32.6</td><td>19.8</td><td>24.8</td><td>17.9</td><td>20.8</td><td>5.7</td><td>20.1</td><td>45.7</td><td>32.5</td><td>34.3</td><td>23.2</td></tr><tr><td rowspan="3">Test</td><td>Instances</td><td>1400</td><td>1220</td><td>3406</td><td>390</td><td>800</td><td>408</td><td>824</td><td>1014</td><td>1004</td><td>5538</td><td>24268</td><td>22232</td><td>592</td></tr><tr><td>Unique Words</td><td>1184</td><td>567</td><td>2088</td><td>276</td><td>533</td><td>305</td><td>476</td><td>475</td><td>600</td><td>1888</td><td>11734</td><td>3517</td><td>394</td></tr><tr><td>Avg. Context Len</td><td>17.2</td><td>8.5</td><td>32.6</td><td>19.4</td><td>23.5</td><td>18.1</td><td>20.6</td><td>6.0</td><td>19.9</td><td>46.0</td><td>32.9</td><td>36.4</td><td>23.4</td></tr></table>
+
+Table 2: Statistics for WordNet and Wiktionary datasets for different languages.
+
+shown in Table 1.
+
+# 3.2.2 Wiktionary
+
+Wiktionary is one of the richest free collaborative lexical databases, available for dozens of languages. In this online resource, each word is provided with definitions for its various potential meanings, some of which are paired with example usages. However, each language has a specific format, and therefore the compilation of these examples requires a careful language-specific parsing. We extracted examples for three European languages for which we did not have WordNet-based data, namely French, German, and Italian.4 Once these examples were compiled, the process to build the final dataset was analogous to that for the WordNet-based datasets (see Section 3.2.1), except for the filtering step, which was not feasible as Wiktionary entries are not connected through paradigmatic relations as in WordNet.
+
+For the case of Wiktionary, the number of examples was considerably higher; therefore, we also compiled language-specific training sets, which enabled a comparison between cross-lingual and monolingual models (see Section 5.2). All Wiktionary datasets are split into balanced training, development and test splits, in each of which there are equal number of positive and negative instances.
+
+# 3.3 Statistics
+
+Table 2 shows the statistics of all datasets, including the total number of instances, unique words, and the context length average.5 Wiktionary-based
+
+datasets are substantially larger than the WordNet-based ones, and also provide training sets. The Chinese datasets feature longer contexts on average and contain the largest number of development and testing instances among WordNet-based datasets. Korean, on the other hand, is the one with the shortest contexts, which is expected given its agglutinative nature. As for the training corpora, German and French datasets contain almost ten times the number of instances in the English training set. This allows us to perform a large-scale comparison between cross-lingual and monolingual settings (see Section 5.2) as well as a few-shot analysis (Section 6.2).
+
+# 3.4 Validation and human performance
+
+To verify the reliability of the datasets, we carried out manual evaluation for those languages for which we had access to annotators. To this end, we presented a set of 100 randomly sampled instances from each dataset to the corresponding annotator in the target language. Annotators were all native speakers of the target language with high-level education. They were provided with a minimal guideline: a brief explanation of their task and a few tagged examples. We did not provide any lexical resource (or any other detailed instructions) to the annotators with the emphasis to make a challenging dataset with sense distinctions that are easily interpretable to the layman. Given an instance, i.e., a pair of sentences containing the same target word, their task consisted of tagging it with a True or False label, depending on the intended meanings of the word in the two contexts.
+
+Table 3 reports human performance for eight
+
+<table><tr><td>WiC</td><td colspan="6">WordNet</td><td colspan="2">Wiktionary</td></tr><tr><td>EN</td><td>DA</td><td>FA</td><td>IT</td><td>JA</td><td>KO</td><td>ZH</td><td>DE</td><td>IT</td></tr><tr><td>80.0*</td><td>87.0</td><td>97.0</td><td>82.0</td><td>75.0</td><td>76.0</td><td>85.0</td><td>74.0</td><td>78.0</td></tr></table>
+
+Table 3: Human performance (in terms of accuracy) for different languages in XL-WiC. *From the original English WiC dataset.
+
+datasets in XL-WiC. All accuracy figures are around  $80\%$ , i.e., in the same ballpark as the original WiC English dataset, which attests the reliability of underlying resources and the construction procedure. The only exception is for Farsi, for which the checker annotators agree with the gold labels in  $97\%$  of the instances (by average). This corroborates our emphasis on the annotation procedure for this manually-created dataset to have sense distinctions that are easily interpretable by humans. As for Wiktionary, the human agreements are lower than those for the WordNet counterparts. This was partly expected given that the semantic network-based filtering step (see Section 3.2.1) was not feasible for the case of Wiktionary datasets due to the nature of the underlying resource.
+
+# 4 Experimental Setup
+
+For our experiments, we implemented a simple, yet effective, baseline based on a Transformer-based text encoder (Vaswani et al., 2017) and a logistic regression classifier, following Wang et al. (2019). The model takes as input the two contexts and first tokenizes them, splitting the input words into sub-tokens. The encoded representations of the target words are concatenated and fed to the logistic classifier. For those cases where the target word was split by the tokenizer into multiple sub-tokens, we followed Devlin et al. (2019) and considered the representation of its first sub-token.
+
+As regards the text encoder, we carried out the experiments with three different multilingual models, i.e., the multilingual version of BERT (Devlin et al., 2019) (mBERT) and the base and large versions of XLM-RoBERTa (Conneau et al., 2020) (XLMR-base and XLMR-large, respectively). In the monolingual setting, we used the following language-specific models: BERT-de<sup>8</sup>,
+
+CamemBERT-large (Martin et al., 2020) $^{9}$ , BERT $^{10}$ , and ParsBERT $^{11}$  (Farahani et al., 2020), respectively, for German, French, Italian, and Farsi. As for all the other languages covered by the WordNet datasets, i.e., Bulgarian, Chinese, Croatian, Danish, Dutch, Estonian, Japanese and Korean, we used the pre-trained models made available by TurkuNLP. $^{12}$  We refer to each language-specific model as L-BERT.
+
+In all experiments we trained the baselines to minimize the binary cross-entropy loss between their prediction and the gold label with the Adam (Kingma and Ba, 2015) optimizer. Training is carried out for 10 epochs with the learning rate fixed to  $1e^{-5}$  and weight decay set to 0. As for tuning, results are reported for the best training checkpoint (among the 10 epochs) according to the performance on the development set.
+
+# 4.1 Evaluation settings
+
+We evaluated the baselines with different configuration setups, depending on the data used for training and tuning.
+
+Cross-Lingual Zero-shot. This setting aims at assessing the capabilities of multilingual models in transferring knowledge captured in the English language to other languages. As training set, we used the English training set of WiC. As for tuning, depending on the setting, we either used the English development set of WiC or language-specific development sets of XL-WiC (Section 3.2.1). We report results on all WordNet and Wiktionary datasets of XL-WiC, i.e., Bulgarian, Chinese, Croatian, Danish, Dutch, Estonian, Farsi, French, German, Italian, Japanese and Korean.
+
+Multilingual Fine-Tuning. In this setting, models are first trained on the WiC's English training set, and then further fine-tuned on the development sets of the target languages in XL-WiC. Depending on the training set used, we report results for two configurations: (i)  $EN + Target$  Language, combining with WiC's training data and the language-specific WordNet development sets for each language, and (ii)  $EN + All$  Languages, combining that with all WordNet development sets for all languages in XL-WiC, merged as one dataset.
+
+<table><tr><td>Model</td><td>BG</td><td>DA</td><td>ET</td><td>FA</td><td>HR</td><td>JA</td><td>KO</td><td>NL</td><td>ZH</td></tr><tr><td></td><td></td><td></td><td></td><td colspan="3">Zero-shot cross-lingual setting</td><td colspan="3">Train: EN - Dev: EN</td></tr><tr><td>mBERT</td><td>58.28</td><td>64.86</td><td>62.56</td><td>71.50</td><td>63.97</td><td>62.26</td><td>59.76</td><td>63.84</td><td>69.36</td></tr><tr><td>XLMR-base</td><td>60.73</td><td>64.79</td><td>62.82</td><td>69.88</td><td>62.01</td><td>60.44</td><td>66.96</td><td>65.73</td><td>65.78</td></tr><tr><td>XLMR-large</td><td>66.48</td><td>71.11</td><td>68.71</td><td>75.25</td><td>72.30</td><td>63.83</td><td>69.63</td><td>72.81</td><td>73.15</td></tr><tr><td></td><td></td><td></td><td></td><td colspan="3">Test instances translated to English</td><td colspan="3">Train: EN - Dev: EN</td></tr><tr><td>mBERT</td><td>63.52</td><td>62.71</td><td>68.46</td><td>-</td><td>60.54</td><td>63.95</td><td>-</td><td>66.53</td><td>-</td></tr><tr><td>XLMR-base</td><td>60.98</td><td>60.24</td><td>62.82</td><td>-</td><td>60.78</td><td>61.77</td><td>-</td><td>64.64</td><td>-</td></tr><tr><td>XLMR-large</td><td>64.43</td><td>66.64</td><td>63.84</td><td>-</td><td>69.85</td><td>64.44</td><td>-</td><td>72.11</td><td>-</td></tr><tr><td>L-BERT</td><td>64.02</td><td>65.38</td><td>64.62</td><td>-</td><td>69.61</td><td>65.90</td><td>-</td><td>68.43</td><td>-</td></tr><tr><td colspan="7">Train and Dev instances translated from English to target language</td><td colspan="3">Train: T-EN - Dev: T-EN</td></tr><tr><td>mBERT</td><td>56.97</td><td>60.25</td><td>59.48</td><td>-</td><td>66.91</td><td>58.13</td><td>-</td><td>60.06</td><td>-</td></tr><tr><td>XLMR-base</td><td>56.07</td><td>52.85</td><td>57.18</td><td>-</td><td>64.22</td><td>56.19</td><td>-</td><td>60.56</td><td>-</td></tr><tr><td>XLMR-large</td><td>62.13</td><td>63.39</td><td>64.87</td><td>-</td><td>66.18</td><td>59.47</td><td>-</td><td>66.73</td><td>-</td></tr><tr><td>L-BERT</td><td>54.26</td><td>60.57</td><td>59.49</td><td>-</td><td>61.52</td><td>58.98</td><td>-</td><td>60.46</td><td>-</td></tr></table>
+
+Table 4: Results on the WordNet test sets when using only English training data in WiC, either in zero-shot cross-lingual setting (top block) or translation-based settings (the lower two blocks). T-EN is a target language dataset, automatically constructed by translating English instances in WiC.
+
+Monolingual. In this setting, we trained each model on the corresponding training set of the target language only. For the case of WordNet datasets (where no training sets are available), we used the development sets for training. In this case we split each development set into two subsets with 9:1 ratio (for training and development). As for the Wiktionary datasets, we used the corresponding training and development sets for each language (Section 3.2.2).
+
+Translation. In this last setting we make use of existing neural machine translation (NMT) models to translate either the training or the test set, essentially reducing the cross-lingual problem to a monolingual one. In particular, we used the general-domain translation models from the Opus-MT project $^{13}$  (Tiedemann and Thottingal, 2020) available for the following language pairs: English-Bulgarian, English-Croatian, English-Danish, English-Dutch, English-Estonian, and English-Japanese. The models are trained on all OPUS parallel corpora collection (Tiedemann, 2012), using the state-of-the-art 6-layer Transformer-based architecture (Vaswani et al., 2017). $^{14}$  In this configuration, as the original target word may be lost during automatic translation, we view the task as context (sentence) similarity
+
+as proposed by Pilehvar and Camacho-Collados (2019).<sup>15</sup> Therefore, for each model, the context vector is given by the start sentence symbol. We note that while training custom optimized NMT models for each target language may result in better overall performance, this is beyond the scope of this work.
+
+Evaluation Metrics. Since all datasets are balanced, we only report accuracy, i.e., the ratio of correctly predicted instances (true positives or true negatives) to the total number of instances.
+
+# 5 Results
+
+In this section, we report the results for the configurations discussed in the previous section on the XL-WiC benchmark. We organize the experiments into two parts, based on the test dataset: WordNet (Section 5.1) and Wiktionary (Section 5.2).
+
+# 5.1 WordNet datasets
+
+Using English data only. Table 4 shows results on the XL-WiC WordNet test sets, when only WiC's English data was used for training and tuning purposes. Across the board, XLMR-large consistently achieves the best results, while mBERT and XLMR-base attain scores in the same ballpark. Indeed, the massive pretraining and the number of
+
+<table><tr><td>Model</td><td>EN</td><td>BG</td><td>DA</td><td>ET</td><td>FA</td><td>HR</td><td>JA</td><td>KO</td><td>NL</td><td>ZH</td></tr><tr><td colspan="11">Train: EN - Dev: Target Language</td></tr><tr><td>mBERT</td><td>-</td><td>59.18</td><td>64.59</td><td>63.08</td><td>70.38</td><td>64.95</td><td>59.95</td><td>63.31</td><td>64.04</td><td>70.48</td></tr><tr><td>XLMR-base</td><td>-</td><td>60.74</td><td>64.80</td><td>60.77</td><td>67.75</td><td>62.50</td><td>57.65</td><td>66.96</td><td>61.85</td><td>65.78</td></tr><tr><td>XLMR-large</td><td>-</td><td>66.48</td><td>71.10</td><td>68.72</td><td>73.63</td><td>72.30</td><td>60.92</td><td>69.63</td><td>69.62</td><td>73.15</td></tr><tr><td colspan="11">Train: EN+Target Language - Dev: EN</td></tr><tr><td>mBERT</td><td>-</td><td>71.72</td><td>62.62</td><td>63.08</td><td>69.38</td><td>72.30</td><td>60.92</td><td>70.91</td><td>62.95</td><td>76.72</td></tr><tr><td>XLMR-base</td><td>-</td><td>64.51</td><td>64.45</td><td>60.00</td><td>65.38</td><td>71.57</td><td>58.37</td><td>65.68</td><td>64.54</td><td>73.46</td></tr><tr><td>XLMR-large</td><td>-</td><td>75.41</td><td>70.52</td><td>68.97</td><td>73.75</td><td>69.61</td><td>63.11</td><td>73.47</td><td>74.50</td><td>77.52</td></tr><tr><td colspan="11">Train: EN+All Languages - Dev: EN</td></tr><tr><td>mBERT</td><td>-</td><td>73.03</td><td>65.09</td><td>62.31</td><td>73.63</td><td>72.30</td><td>65.53</td><td>71.01</td><td>67.73</td><td>76.53</td></tr><tr><td>XLMR-base</td><td>-</td><td>67.30</td><td>67.62</td><td>59.49</td><td>64.50</td><td>66.18</td><td>57.77</td><td>67.06</td><td>66.33</td><td>71.02</td></tr><tr><td>XLMR-large</td><td>-</td><td>78.44</td><td>71.49</td><td>72.05</td><td>78.25</td><td>76.96</td><td>66.38</td><td>76.53</td><td>77.49</td><td>78.95</td></tr><tr><td colspan="11">Train: Target Language - Dev: Target Language</td></tr><tr><td>mBERT</td><td>66.71</td><td>82.30</td><td>62.13</td><td>58.21</td><td>63.75</td><td>77.45</td><td>61.04</td><td>70.71</td><td>64.84</td><td>76.09</td></tr><tr><td>XLMR-base</td><td>64.36</td><td>79.75</td><td>64.00</td><td>64.36</td><td>66.25</td><td>79.17</td><td>58.86</td><td>70.61</td><td>66.33</td><td>78.11</td></tr><tr><td>XLMR-large</td><td>70.14</td><td>82.05</td><td>66.53</td><td>59.23</td><td>68.00</td><td>76.72</td><td>55.22</td><td>73.08</td><td>69.42</td><td>81.83</td></tr><tr><td>L-BERT</td><td>69.60</td><td>81.23</td><td>62.60</td><td>58.46</td><td>76.63</td><td>76.47</td><td>56.07</td><td>58.68</td><td>68.73</td><td>77.36</td></tr></table>
+
+Table 5: Results on the WordNet test sets when using language-specific data, either for training or for tuning.
+
+parameters of XLMR-large play key roles behind this lead in performance. As regards the translation-based settings (lower two blocks), the performance generally falls slightly behind the zero-shot crosslingual counterpart. This shows that the usage of good quality English data and multilingual models provide a stronger training signal than noisy automatically-translated data. This somehow contrasts with the observations made on other crosslingual tasks in XTREME (Hu et al., 2020), especially in question answering datasets (Artetxe et al., 2020a; Lewis et al., 2020; Clark et al., 2020), where translating data was generally better. This difference could perhaps be reduced with larger monolingual models or accurate alignment, but this would further increase the complexity, and extracting these alignments from NMT models is not trivial (Koehn and Knowles, 2017; Ghader and Monz, 2017; Li et al., 2019).
+
+Utilizing language-specific data. Table 5 shows results for settings where target language-specific data was used for training or tuning. Comparing the results in the top block (where target language data was used for tuning) with the middle two blocks (where target language data was instead used for training) reveals that it is more effective to leverage the target language data for training, rather
+
+than using it for tuning only. Overall, it is clear that adding multilingual data during training drastically improves the results in all languages. In this case, training (fine-tuning) is performed on a larger dataset which, despite having examples from different languages, provides a stronger signal to the models, enabling them to better generalize across languages. On the contrary, when only using target language data for training and tuning (last block in the table), results drop for most languages. This highlights the fact that having additional training data is beneficial, reinforcing the utility of multilingual models and cross-lingual transfer.
+
+# 5.2 Wiktionary Datasets
+
+In Table 6 we show results for the Wiktionary datasets. Differently from the results reported for the WordNet datasets (Table 4), models are less effective in the zero-shot setting, performing from 10 to almost 20 points lower than their counterparts trained on data in the target language. This can be attributed to the size of the available training data. Indeed, while in the WordNet datasets we only have a very small amount of data at our disposal for training (see statistics in Table 2), Wiktionary training sets are much larger, hence providing enough data to the models to better generalize. Once again, XLMR-large proves to be the best model in the
+
+<table><tr><td></td><td>Model</td><td>DE</td><td>FR</td><td>IT</td></tr><tr><td rowspan="3">Z-Shot</td><td>mBERT</td><td>58.27</td><td>56.00</td><td>58.61</td></tr><tr><td>XLMR-base</td><td>58.30</td><td>56.13</td><td>55.91</td></tr><tr><td>XLMR-large</td><td>65.83</td><td>62.50</td><td>64.86</td></tr><tr><td rowspan="4">Mono</td><td>mBERT</td><td>81.58</td><td>73.67</td><td>71.96</td></tr><tr><td>XLMR-base</td><td>80.84</td><td>73.06</td><td>68.58</td></tr><tr><td>XLMR-large</td><td>84.03</td><td>76.16</td><td>72.30</td></tr><tr><td>L-BERT</td><td>82.90</td><td>78.14</td><td>72.64</td></tr></table>
+
+Table 6: Results on the Wiktionary test sets in different training settings: zero-shot (Z-Shot) and monolingual training (Mono). L-BERT stands for language-specific models, i.e., BERT-de, CamemBERT-large and BERT-it for German, French and Italian, respectively.
+
+zero-shot setting and a competitive alternative to the language-specific models (L-BERT row) in the monolingual setting, performing 1.1 points higher in German and 2 and 0.3 lower in French and Italian, respectively.
+
+# 6 Analysis
+
+In this section, we delve into the performance of the models on XL-WiC and analyze relevant aspects about their behaviour.
+
+# 6.1 Seen and Unseen Words
+
+For this analysis, we aim at measuring the difference in performance when a given target word was seen (as a target word) at training time or not. To this end, we evaluate our baselines when trained on the German, French, and Italian Wiktionary training sets and tested on two different subsets of the larger language-specific Wiktionary test sets: In-Vocabulary (IV), containing only the examples whose target word was seen at training time; and Out-of-Vocabulary (OOV), containing only the examples whose target word was not seen during training. We report the results in Table 7.
+
+In general, multilingual models are less reliable when classifying unseen instances, lagging between 1 and 12 points behind in performance, depending on the language and on the model considered. This can be attributed to the fact that their vocabulary is shared among several languages, and therefore may have less knowledge stored about particular words that do not occur often. The performance drop of language-specific models (L-BERT) is less pronounced, with the French architecture (CamemBERT-large) attaining even higher performance (0.4 points more) on the OOV set.
+
+<table><tr><td></td><td>Model</td><td>DE</td><td>FR</td><td>IT</td></tr><tr><td rowspan="4">N</td><td>mBERT</td><td>81.86</td><td>72.92</td><td>73.15</td></tr><tr><td>XLMR-base</td><td>81.17</td><td>71.92</td><td>70.69</td></tr><tr><td>XLMR-large</td><td>84.24</td><td>75.61</td><td>75.12</td></tr><tr><td>L-BERT</td><td>83.23</td><td>77.62</td><td>73.89</td></tr><tr><td rowspan="4">OOV</td><td>mBERT</td><td>70.08</td><td>71.24</td><td>68.54</td></tr><tr><td>XLMR-base</td><td>71.31</td><td>71.14</td><td>62.36</td></tr><tr><td>XLMR-large</td><td>72.54</td><td>73.93</td><td>65.17</td></tr><tr><td>L-BERT</td><td>76.64</td><td>78.00</td><td>69.10</td></tr></table>
+
+Table 7: Results on the in-vocabulary (IV) and out-of-vocabulary (OOV) Wiktionary test sets. L-BERT stands for each language-specific model, i.e., BERT-base-de, camemBERT and BERT-base-xxl-it for German, French and Italian, respectively.
+
+# 6.2 Few-shot Monolingual
+
+As an additional experiment, we investigate the impact of training size on performance. To this end, we leveraged the Wiktionary datasets for German, French and Italian, which allow us to use varying-sized training sets, and created 7 training sets with 10, 25, 50, 100, 250, 500, 1000 instances.[16]
+
+The results of this experiment are displayed in Figure 1. When providing only 10 examples, most of the models perform similarly or even worse than random, i.e.,  $50\%$  accuracy. In this setting, language-specific models (L-BERT) attain better results than their multilingual counterparts, showing better generalization capabilities when fewer examples are provided. This also goes in line with what we found in the previous experiment on seen and unseen words. With less than  $5\%$  of the training data (1000 instances in French and German and 50 instances in Italian), all models attain roughly  $85\%$  of their performance with full training data, comparable to results reported for the zero-shot setting (Table 6).
+
+# 7 Conclusions
+
+In this paper we have introduced XL-WiC, a large benchmark for evaluating context-sensitive models. XL-WiC comprises datasets for a heterogeneous set of 13 languages, including the original English data in WiC (Pilehvar and Camacho-Collados, 2019), providing an evaluation framework not only for contextualized models in those languages, but also for experimentation in a cross-lingual transfer setting. Our evaluations show that, even though cur
+
+![](images/8ed3e2f575a01e7ad184fcc6921b4e447bd667c5588ee5cebe947a7c4cf3fc6b.jpg)  
+Figure 1: The impact of training set size (# of instances) on performance, for the Wiktionary datasets.
+
+rent language models are effective performers in the zero-shot cross-lingual setting (where no instances in the target language are provided), there is still room for improvement, especially for far languages such as Japanese or Korean.
+
+As for future work, we plan to investigate using languages other than English for training (e.g., our larger French and German training sets) in our cross-lingual transfer experiments, since English may not always be the optimal source language (Anastasopoulos and Neubig, 2020). Finally, while in our comparative analysis we have focused on a quantitative evaluation for all languages, an additional error analysis per language would be beneficial in revealing the weaknesses and limitations of cross-lingual models.
+
+# Acknowledgments
+
+We would like to thank Qincheng Zhang (Chinese), Janine Siewert (Danish), Houman Mehrafarin, Hossein Mohebbi, and Ali Modarresi (Farsi), Angela Collados Aís (German), Asahi Ushio (Japanese), and Yunseo Joung (Korean) for their help with the manual evaluation of the datasets.
+
+Alessandro gratefully acknowledges the support of the FoTran project, funded by the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme (grant agreement No 771113), and the CSC
+
+erc - IT Center for Science, Finland, for computational resources.
+
+Tommaso gratefully acknowledges the support of the ERC Consolidator Grant MOUSSE No. 726487 under the European Union's Horizon 2020 research and innovation programme.
+
+# References
+
+Eneko Agirre, Oier Lopez de Lacalle, and Aitor Soroa. 2014. Random walks for knowledge-based word sense disambiguation. Computational Linguistics, 40(1):57-84.
+
+Antonios Anastasopoulos and Graham Neubig. 2020. Should all cross-lingual embeddings speak English? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8658-8679, Online. Association for Computational Linguistics.
+
+Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020a. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4623–4637, Online. Association for Computational Linguistics.
+
+Mikel Artetxe, Sebastian Ruder, Dani Yogatama, Gorka Labaka, and Eneko Agirre. 2020b. A call for more rigor in unsupervised cross-lingual learning. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7375-7388, Online. Association for Computational Linguistics.
+
+Jordi Atserias, Luís Villarejo, and German Rigau. 2004. Spanish WordNet 1.6: Porting the Spanish Word-net across Princeton versions. In Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC'04), Lisbon, Portugal. European Language Resources Association (ELRA).
+
+Siamak Barzegar, Brian Davis, Manel Zarrouk, Siegfried Handschuh, and Andre Freitas. 2018. SemR-11: A multi-lingual gold-standard for semantic similarity and relatedness for eleven languages. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC-2018), Miyazaki, Japan. European Languages Resources Association (ELRA).
+
+Laura Benitez, Sergi Cervell, Gerard Escudero, Monica López, German Rigau, and Mariona Taule. 1998. Methods and tools for building the catalan wordnet. Proceedings of ELRA Workshop on Language Resources for European Minority Languages.
+
+Michele Bevilacqua and Roberto Navigli. 2020. Breaking through the  $80\%$  Glass Ceiling: Raising the State of the Art in Word Sense Disambiguation by Incorporating Knowledge Graph Information. In Proc. of ACL, pages 2854-2864.  
+Terra Blevins and Luke Zettlemoyer. 2020. Moving Down the Long Tail of Word Sense Disambiguation with Gloss Informed Bi-encoders. In Proc. of ACL, pages 1006-1017.  
+Francis Bond and Kyonghee Paik. 2012. A survey of wordnets and their licenses. In GWC 2012 6th International Global Wordnet Conference, volume 8, pages 64-71.  
+Jose Camacho-Collados, Mohammad Taher Pilehvar, Nigel Collier, and Roberto Navigli. 2017. SemEval2017 task 2: Multilingual and cross-lingual semantic word similarity. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval2017), pages 15–26, Vancouver, Canada. Association for Computational Linguistics.  
+Daniel Cer, Mona 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-2017), pages 1-14, Vancouver, Canada. Association for Computational Linguistics.  
+Christos Christodouloupoulos and Mark Steedman. 2015. A massively parallel corpus: the bible in 100 languages. Language resources and evaluation, 49(2):375-395.  
+Jonathan H Clark, Eunsol Choi, Michael Collins, Dan Garrette, Tom Kwiatkowski, Vitaly Nikolaev, and Jennimaria Palomaki. 2020. Tydi qa: A benchmark for information-seeking question answering in typologically diverse languages. Transactionsof the Association of Computational Linguistics.  
+Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440-8451, Online. Association for Computational Linguistics.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186.  
+Cecily Jill Duffield, Jena D Hwang, Susan Windisch Brown, Dmitriy Dligach, Sarah Vieweg, Jenny
+
+Davis, and Martha Palmer. 2007. Criteria for the manual grouping of verb senses. In Proceedings of the Linguistic Annotation Workshop, pages 49-52.  
+Philip Edmonds and Scott Cotton. 2001. SENSEVAL-2: Overview. In Proceedings of SENSEVAL-2 Second International Workshop on Evaluating Word Sense Disambiguation Systems, pages 1-5, Toulouse, France. Association for Computational Linguistics.  
+Mehrdad Farahani, Mohammad Gharachorloo, Marzieh Farahani, and Mohammad Manthouri. 2020. Parsbert: Transformer-based model for persian language understanding.  
+Christiane Fellbaum, editor. 1998. WordNet: An Electronic Database. MIT Press, Cambridge, MA.  
+Darja Fisér, Jernej Novak, and Tomaz Erjavec. 2012. slownet 3.0: development, extension and cleaning. In Proceedings of 6th International Global Wordnet Conference (GWC 2012), pages 113-117.  
+Hamidreza Ghader and Christof Monz. 2017. What does attention in neural machine translation pay attention to? In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 30-39, Taipei, Taiwan. Asian Federation of Natural Language Processing.  
+Xavier Gómez Guinovart. 2011. Galnet: Wordnet 3.0 do galego. Languamática, 3(1):61-67.  
+Birgit Hamp and Helmut Feldweg. 1997. Germanet-a lexical-semantic net for german. In Automatic information extraction and building of lexical semantic resources for NLP applications.  
+Daniel Hershcovich, Zohar Aizenbud, Leshem Choshen, Elior Sulem, Ari Rappoport, and Omri Abend. 2019. SemEval-2019 task 1: Cross-lingual semantic parsing with UCCA. In Proceedings of the 13th International Workshop on Semantic Evaluation, pages 1-10, Minneapolis, Minnesota, USA. Association for Computational Linguistics.  
+Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. XTREME: A massively multilingual multitask benchmark for evaluating cross-lingual generalization. Proceedings of the 37th International Conference on Machine Learning (ICML).  
+Chu-Ren Huang, Shu-Kai Hsieh, Jia-Fei Hong, Yun-Zhu Chen, I-Li Su, Yong-Xiang Chen, and Sheng-Wei Huang. 2010. Chinese Wordnet: Design, Implementation and Application of an Infrastructure for Cross-Linguual Knowledge Processing. Journal of Chinese Information Processing, 24(2):14.  
+Eric Huang, Richard Socher, Christopher Manning, and Andrew Ng. 2012. Improving word representations via global context and multiple word prototypes. In Proceedings of the 50th Annual Meeting of
+
+the Association for Computational Linguistics (Volume 1: Long Papers), pages 873-882, Jeju Island, Korea. Association for Computational Linguistics.  
+Hitoshi Isahara, Francis Bond, Kiyotaka Uchimoto, Masao Utiyama, and Kyoko Kanzaki. 2008. Development of the Japanese WordNet. In Sixth International conference on Language Resources and Evaluation.  
+Marcin Junczys-Dowmunt, Roman Grundkiewicz, Tomasz Dwojak, Hieu Hoang, Kenneth Heafield, Tom Neckermann, Frank Seide, Ulrich Germann, Alham Fikri Aji, Nikolay Bogoychev, Andre F. T. Martins, and Alexandra Birch. 2018. Marian: Fast neural machine translation in  $\mathbf{C} + +$ . In Proceedings of ACL 2018, System Demonstrations, pages 116-121, Melbourne, Australia. Association for Computational Linguistics.  
+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.  
+Philipp Koehn and Rebecca Knowles. 2017. Six challenges for neural machine translation. In Proceedings of the First Workshop on Neural Machine Translation, pages 28-39, Vancouver. Association for Computational Linguistics.  
+Patrick Lewis, Barlas Oguz, Rudy Rinott, Sebastian Riedel, and Holger Schwenk. 2020. MLQA: Evaluating cross-lingual extractive question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7315-7330, Online. Association for Computational Linguistics.  
+Xintong Li, Guanlin Li, Lemao Liu, Max Meng, and Shuming Shi. 2019. On the word alignment from neural machine translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1293-1303, Florence, Italy. Association for Computational Linguistics.  
+Louis Martin, Benjamin Muller, Pedro Javier Ortiz Suárez, Yoann Dupont, Laurent Romary, Éric de la Clergerie, Djamé Seddah, and Benoit Sagot. 2020. CamemBERT: a tasty French language model. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7203-7219, Online. Association for Computational Linguistics.  
+Márton Miháltz, Csaba Hatvani, Judit Kuti, György Szarvas, János Cserik, Gábor Prószéky, and Tamás Varadi. 2008. Methods and Results of the Hungarian WordNet Project. In Proceedings of The Fourth Global WordNet Conference, pages 311-321.  
+George A Miller. 1995. WordNet: a lexical database for english. Communications of the ACM, 38(11):39-41.
+
+Andrea Moro and Roberto Navigli. 2015a. SemEval 2015 task 13: Multilingual all-words sense disambiguation and entity linking. In Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 288-297, Denver, Colorado. Association for Computational Linguistics.  
+Andrea Moro and Roberto Navigli. 2015b. Semeval-2015 task 13: Multilingual all-words sense disambiguation and entity linking. Proceedings of SemEval-2015.  
+Andrea Moro, Alessandro Raganato, and Roberto Navigli. 2014. Entity Linking meets Word Sense Disambiguation: a Unified Approach. Transactions of the Association for Computational Linguistics (TACL), 2:231-244.  
+Roberto Navigli. 2009. Word Sense Disambiguation: A survey. ACM Computing Surveys, 41(2):1-69.  
+Roberto Navigli, David Jurgens, and Daniele Vannella. 2013. SemEval-2013 task 12: Multilingual word sense disambiguation. In Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Proceedings of the Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 222-231, Atlanta, Georgia, USA. Association for Computational Linguistics.  
+Roberto Navigli and Simone Paolo Ponzetto. 2012. BabelNet: The automatic construction, evaluation and application of a wide-coverage multilingual semantic network. Artificial Intelligence, 193:217-250.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.  
+Bolette S. Pedersen, Sanni Nimb, Jorg Asmussen, Nicolai Hartvig Sorensen, Lars Trap-Jensen, and Henrik Lorentzen. 2009. DanNet: the challenge of compiling a wordnet for Danish by reusing a monolingual dictionary. Language Resources and Evaluation, 43:269-299.  
+Emanuele Pianta, Luisa Bentivogli, and Christian Girardi. 2002. MultiWordNet: Developing an Aligned Multilingual Database. In Proceedings of the 1st International Conference on Global WordNet, pages 293-302.  
+Mohammad Taher Pilehvar and Jose Camacho-Collados. 2019. WiC: the Word-in-Context dataset for evaluating context-sensitive meaning representations. 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 1267-1273, Minneapolis, Minnesota.
+
+Eli Pociello, Antton Gurrutxaga, Eneko Agirre, Izaskun Aldezabal, and German Rigau. 2008. WNTERM: Enriching the MCR with a terminological dictionary. In LREC 2008.  
+Marten Postma, Emiel van Miltenburg, Roxane Segers, Anneleen Schoen, and Piek Vossen. 2016. Open Dutch WordNet. In Proceedings of the Eight Global Wordnet Conference.  
+Quentin Pradet, Gael de Chalendar, and Jeanne Desormeaux Baguenier. 2014. Wonef, an improved, expanded and evaluated automatic french translation of wordnet. In Proceedings of the Seventh Global Wordnet Conference (GWC2014), pages 32-39.  
+Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. Stanza: A Python Natural Language Processing Toolkit for Many Human Languages. In In Association for Computational Linguistics (ACL) System Demonstrations.  
+Ida Raffaelli, Marko Tadic, Božo Bekavac, and Željko Agić. 2008. Building croatian wordnet. In Fourth global wordnet conference (gwc 2008).  
+Alessandro Raganato, Jose Camacho-Collados, and Roberto Navigli. 2017. Word Sense Disambiguation: A Unified Evaluation Framework and Empirical Comparison. In Proc. of EACL, pages 99-110.  
+Ervin Ruci. 2008. On the current state of Albanet and related applications. Technical report, University of Vlora. (http://fjalnet.com/technicalreportalbanet.pdf).  
+Bianca Scarlini, Tommaso Pasini, and Roberto Navigli. 2020. With More Contexts Comes Better Performance: Contextualized Sense Embeddings for All-Round Word Sense Disambiguation. In Proc. of EMNLP.  
+Karin Kipper Schuler, Anna Korhonen, and Susan Brown. 2009. VerbNet overview, extensions, mappings and applications. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Tutorial Abstracts, NAACL-Tutorials '09, page 13-14, USA. Association for Computational Linguistics.  
+Federico Scozzafava, Marco Maru, Fabrizio Brignone, Giovanni Torrisi, and Roberto Navigli. 2020. Personalized pagerank with syntagmatic information for multilingual word sense disambiguation. In Proc. of ACL, pages 37-46.  
+Mehrnoush Shamsfard, Akbar Hesabi, Hakimeh Fadaei, Niloofar Mansoory, Payam Noor, Ali Reza Gholi Famian, Somayeh Bagherbeigi, Elham Fekri, and Maliheh Monshizadeh. 2010. Semi automatic development of FarsNet; the Persian WordNet. In Proceedings of 5th global WordNet conference.
+
+Kiril Simov and Petya Osenova. 2010. Constructing of an Ontology-based Lexicon for Bulgarian. In Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC'10), Valletta, Malta. European Language Resources Association (ELRA).  
+Jörg Tiedemann. 2012. Parallel data, tools and interfaces in OPUS. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC'12), pages 2214-2218, Istanbul, Turkey. European Language Resources Association (ELRA).  
+Jörg Tiedemann and Santhosh Thottingal. 2020. OPUS-MT — Building open translation services for the World. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation (EAMT), Lisbon, Portugal.  
+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, pages 5998-6008.  
+Kadri Vider and Heili Orav. 2002. Estonian WordNet and Lexicography. In Proceedings of the Eleventh International Symposium on Lexicography, pages 549-555.  
+Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. SuperGLUE: A stickier benchmark for general-purpose language understanding systems. In Proceedings of NeurIPS.  
+Ae-Sun Yoon, Soon-Hee Hwang, Eun-Ryoung Lee, and Hyuk-Chul Kwon. 2009. Construction of Korean WordNet. Journal of KIISE: Software and Applications, 36(1):92-108.
+
+# Appendix
+
+# A Comparison between Italian WordNet and Wiktionary datasets
+
+In Table 8 we provide a small comparison between the Italian WordNet and Wiktionary datasets, which is the only language that overlaps. While this comparison is quite limited, it provides a few hints on the qualitative differences between Wiktionary and WordNet datasets.
+
+# B Models' Parameters
+
+In Table 9 we report the parameters of the models used in our experiments.
+
+<table><tr><td></td><td></td><td>WordNet</td><td>Wiktionary</td></tr><tr><td rowspan="4">Stats</td><td>Instances</td><td>260</td><td>592</td></tr><tr><td>Unique Words</td><td>105</td><td>394</td></tr><tr><td>Avg. Ctx length</td><td>14.53</td><td>23.39</td></tr><tr><td>Human acc.</td><td>82.0</td><td>78.0</td></tr><tr><td rowspan="3">Z-shot</td><td>XLM-R base</td><td>66.15</td><td>55.91</td></tr><tr><td>XLM-R large</td><td>80.00</td><td>64.86</td></tr><tr><td>mBERT</td><td>70.00</td><td>58.61</td></tr></table>
+
+Table 8: Statistics and comparison between the Italian WordNet and the Italian Wiktionary WiC datasets. Zero-shot results are computed by using the original English WiC (Pilehvar and Camacho-Collados, 2019) for training and development.  
+
+<table><tr><td>Model</td><td>Trainable Parameters</td></tr><tr><td>mBERT</td><td>110M</td></tr><tr><td>BERT-large</td><td>335M</td></tr><tr><td>XLMR-base</td><td>270M</td></tr><tr><td>XLMR-large</td><td>550M</td></tr></table>
+
+Table 9: Number of parameters for our comparison systems.
+
+# C Additional experimental results
+
+WordNet datasets. Table 10 includes details on the variability of the results, in particular the average results from three runs, including the standard deviation, for the zero-shot cross-lingual setting - this is the setting producing a higher variability in the results.
+
+Wiktionary Datasets. In Table 11 we show the development and test results in the monolingual settings of the multilingual language models trained, tuned and tested on the XL-WiC language-specific datasets from Wiktionary.
+
+Translation setting + Dictionary alignment. We include a setting where, after translating the English training set to each target language, we also retrieve the corresponding translation of the English target word through a multilingual dictionary. We use BabelNet (Navigli and Ponzetto, 2012) as multilingual dictionary for all languages, discarding the sentences where the translated target word could not be found. Table 12 shows the results.
+
+# D Translation models
+
+Translation models are trained using the MarianNMT framework (Junczys-Dowmunt et al., 2018)
+
+on a filtered version of all OPUS parallel corpora collection using a language identifier (CLD2). As hyper-parameters, each model is based on the base version of the Transformer architecture (Vaswani et al., 2017). All models and training details are available at https://github.com/Helsinki-NLP/Opus-MT. To give an idea of the translation quality, Table 13 reports the BLEU scores (Papineni et al., 2002) for each model. We report the performance, as described within the Opus-MT project, on the latest available test sets from the series of WMT news translation shared tasks, or on 5K sentences taken from either the Tatoeba corpus (Tiedemann, 2012), or the Bible corpus (Christodouloupoulos and Steedman, 2015):
+
+Bulgarian (BG): Tatoeba, model checkpoint EN  $\leftrightarrow$  BG opus-2019-12-18  
+- Danish (DA): Tatoeba, model checkpoint EN↔DA opus-2019-12-18  
+- Estonian (ET): newstest2018, model checkpoint EN  $\leftrightarrow$  ET opus-2019-12-18  
+- Croatian (HR): Tatoeba, model checkpoint EN→HR opus-2019-12-04, HR→EN opus-2019-12-05  
+- Japanese (JA): bible-uedin, model checkpoint EN  $\rightarrow$  JA opus-2020-01-08, JA  $\rightarrow$  EN opus-2019-12-18  
+- Dutch (NL): Tatoeba, model checkpoint EN  $\rightarrow$  NL opus-2019-12-04, NL  $\rightarrow$  EN opus-2019-12-05
+
+<table><tr><td>Language</td><td>mBERT</td><td>XLMR-base</td><td>XLMR-large</td><td>Size</td></tr><tr><td>BG</td><td>58.72 ± 1.60</td><td>58.42 ± 3.74</td><td>64.70 ± 2.05</td><td>1220</td></tr><tr><td>DA</td><td>63.26 ± 2.25</td><td>62.29 ± 2.34</td><td>68.91 ± 2.45</td><td>3406</td></tr><tr><td>ET</td><td>59.40 ± 2.74</td><td>62.74 ± 2.43</td><td>68.29 ± 2.97</td><td>390</td></tr><tr><td>FA</td><td>67.71 ± 3.36</td><td>64.67 ± 5.07</td><td>73.58 ± 1.49</td><td>800</td></tr><tr><td>HR</td><td>65.93 ± 2.25</td><td>63.07 ± 2.06</td><td>68.63 ± 3.20</td><td>408</td></tr><tr><td>JA</td><td>62.58 ± 0.37</td><td>59.63 ± 3.11</td><td>63.23 ± 1.72</td><td>824</td></tr><tr><td>KO</td><td>61.21 ± 2.34</td><td>64.04 ± 3.93</td><td>69.40 ± 3.61</td><td>1014</td></tr><tr><td>NL</td><td>63.55 ± 1.37</td><td>64.41 ± 1.40</td><td>72.14 ± 1.88</td><td>1004</td></tr><tr><td>ZH</td><td>68.85 ± 0.50</td><td>61.69 ± 3.78</td><td>70.68 ± 2.94</td><td>5538</td></tr></table>
+
+Table 10: Zero-shot results on mBERT, XLMR-base and XLMR-large on the WordNet-based datasets when using the English WiC training and development sets.  
+
+<table><tr><td rowspan="2">Language</td><td colspan="2">mBERT</td><td colspan="2">XLMR-base</td><td colspan="2">XLMR-large</td><td rowspan="2">Train</td><td rowspan="2">Size Dev</td><td rowspan="2">Test</td></tr><tr><td>Dev</td><td>Test</td><td>Dev</td><td>Test</td><td>Dev</td><td>Test</td></tr><tr><td>DE</td><td>79.73</td><td>81.58</td><td>78.93</td><td>80.84</td><td>83.03</td><td>84.03</td><td>48042</td><td>8870</td><td>24268</td></tr><tr><td>FR</td><td>71.62</td><td>73.67</td><td>71.43</td><td>73.06</td><td>75.00</td><td>76.16</td><td>39428</td><td>8588</td><td>22232</td></tr><tr><td>IT</td><td>73.23</td><td>71.96</td><td>75.25</td><td>68.58</td><td>74.24</td><td>72.30</td><td>1144</td><td>198</td><td>592</td></tr></table>
+
+Table 11: Results on mBERT, XLMR-base and XLMR-large on the Wiktionary-based datasets when using the language-specific training and development data.  
+
+<table><tr><td>Model</td><td>EN</td><td>BG</td><td>DA</td><td>ET</td><td>FA</td><td>HR</td><td>JA</td><td>KO</td><td>NL</td><td>ZH</td></tr><tr><td colspan="11">All instances translated to Target Language + Dictionary Alignment (Train: T-EN - Dev: T-EN)</td></tr><tr><td>mBERT</td><td>-</td><td>60.66</td><td>60.16</td><td>61.79</td><td>-</td><td>68.87</td><td>52.79</td><td>-</td><td>57.57</td><td>-</td></tr><tr><td>XLMR-base</td><td>-</td><td>57.30</td><td>57.34</td><td>51.79</td><td>-</td><td>59.80</td><td>51.70</td><td>-</td><td>60.26</td><td>-</td></tr><tr><td>XLMR-large</td><td>-</td><td>63.36</td><td>66.27</td><td>61.54</td><td>-</td><td>66.42</td><td>53.88</td><td>-</td><td>69.42</td><td>-</td></tr><tr><td>L-BERT</td><td>-</td><td>56.31</td><td>58.07</td><td>56.67</td><td>-</td><td>59.31</td><td>53.40</td><td>-</td><td>58.47</td><td>-</td></tr></table>
+
+Table 12: Results on the WordNet test sets when using automatically-translated data with a multilingual dictionary-alignment technique.  
+
+<table><tr><td>Opus-MT</td><td>BG</td><td>DA</td><td>ET</td><td>HR</td><td>JA</td><td>NL</td></tr><tr><td>EN→XX</td><td>50.0</td><td>60.4</td><td>23.3</td><td>48.3</td><td>42.1</td><td>57.1</td></tr><tr><td>XX→EN</td><td>59.4</td><td>63.6</td><td>30.3</td><td>58.7</td><td>41.7</td><td>60.9</td></tr></table>
+
+Table 13: BLEU score of the translation models.

xlwicamultilingualbenchmarkforevaluatingsemanticcontextualization/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da63873c3aa22b2181d69aae502e21b287cc17b9ce347c363dfb7e2eb59631ff
+size 705586

xlwicamultilingualbenchmarkforevaluatingsemanticcontextualization/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:787e0607a89ef286d34f7ea3aac391e9ce4de3211fb2832d4e60a2c4728773ad
+size 404903

youaregroundedlatentnameartifactsinpretrainedlanguagemodels/a2e3a095-4138-43c5-b379-03545f0322cc_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04c567e1d342aa12bbfc76acdc9be46bcd4ade91a2b18c490dccbec3c11c0016
+size 126609

youaregroundedlatentnameartifactsinpretrainedlanguagemodels/a2e3a095-4138-43c5-b379-03545f0322cc_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c55db3ba85a3f06a12a3f9941091fbd964b58ccd5523240633f1e34a5165ab39
+size 138319

youaregroundedlatentnameartifactsinpretrainedlanguagemodels/a2e3a095-4138-43c5-b379-03545f0322cc_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74e2a4703c7d769c8f1927b256c0799bcc113849403fa7a7cd4f4e8587d49836
+size 768211

youaregroundedlatentnameartifactsinpretrainedlanguagemodels/full.md

@@ -0,0 +1,366 @@
+# "You are grounded!": Latent Name Artifacts in Pre-trained Language Models
+
+Vered Shwartz $^{1,2}$ , Rachel Rudinger $^{1,2,3}$ , and Oyvind Tafjord $^{1}$
+
+$^{1}$ Allen Institute for Artificial Intelligence
+
+$^{2}$ Paul G. Allen School of Computer Science & Engineering, University of Washington
+
+<sup>3</sup>University of Maryland, College Park, MD
+
+{vereds,oyvindt}@allenai.org,rudinger@umd.edu
+
+# Abstract
+
+Pre-trained language models (LMs) may perpetuate biases originating in their training corpus to downstream models. We focus on artifacts associated with the representation of given names (e.g., Donald), which, depending on the corpus, may be associated with specific entities, as indicated by next token prediction (e.g., Trump). While helpful in some contexts, grounding happens also in underspecified or inappropriate contexts. For example, endings generated for 'Donald is a' substantially differ from those of other names, and often have more-than-average negative sentiment. We demonstrate the potential effect on downstream tasks with reading comprehension probes where name perturbation changes the model answers. As a silver lining, our experiments suggest that additional pre-training on different corpora may mitigate this bias.
+
+# 1 Introduction
+
+Pre-trained language models (LMs) have transformed the NLP landscape. State-of-the-art performance across tasks is achieved by fine-tuning the latest LM on task-specific data. LMs provide an effective way to represent contextual information, including lexical and syntactic knowledge as well as world knowledge (Petroni et al., 2019).
+
+LMs conflate generic facts (e.g. "the US has a president") with grounded knowledge regarding specific entities and events (e.g. "the (current) president is a male"), occasionally leading to gender and racial biases (e.g. "women can't be presidents") (May et al., 2019; Sheng et al., 2019).
+
+In this work we focus on the representations of given names in pre-trained LMs (Table 1). Prior work showed that the representations of named entities incorporate sentiment (Prabhakaran et al., 2019), which is often transferable across entities via a shared given name (Field and Tsvetkov, 2019).
+
+<table><tr><td>Model</td><td>Main Corpus Type</td><td>Gen.</td><td>Cls.</td></tr><tr><td>BERT (Devlin et al., 2019)</td><td>Wikipedia</td><td>×</td><td>✓</td></tr><tr><td>RoBERTa (Liu et al., 2019)</td><td>Web</td><td>×</td><td>✓</td></tr><tr><td>GPT (Radford et al., 2018)</td><td>Fiction</td><td>✓</td><td>×</td></tr><tr><td>GPT2 (Radford et al., 2019)</td><td>Web</td><td>✓</td><td>×</td></tr><tr><td>XLNet (Yang et al., 2019)</td><td>Web</td><td>✓</td><td>✓</td></tr><tr><td>TransformerXL (Dai et al., 2019)</td><td>Wikipedia</td><td>✓</td><td>×</td></tr></table>
+
+Table 1: Pre-trained LMs and whether they are typically used for generation (Gen.) or classification (Cls.).
+
+In a series of experiments we show that, depending on the corpus, some names tend to be grounded to specific entities, even in generic contexts.
+
+The most striking effect is of politicians in GPT2. For example, the name Donald: 1) predicts Trump as the next token with high probability; 2) generated endings of "Donald is a" are easily distinguishable from any other given name; 3) their sentiment is substantially more negative; and 4) this bias can potentially perpetuate to downstream tasks.
+
+Although these results are expected, their extent is surprising. Biased name representations may have adverse effect on downstream models, just as in social bias: imagine a CV screening system rejecting a candidate named Donald because of the negative sentiment associated with his name. Our experiments may be used to evaluate the extent of name artifacts in future LMs.<sup>1</sup>
+
+# 2 Last Name Prediction
+
+As an initial demonstration of the tendency of pretrained LMs to ground given names to prominent named entities in the media, we examine the next-word probabilities assigned by the LM. If high probability is placed on a named entity's last name conditioned on observing their given name (e.g.,  $P(\text{Trump} | \text{Donald}) = 0.99$ ), we take this as evidence that the LM is, in effect, interpreting the first-name mention as a reference to the named entity. We note that this is a lower bound on evidence
+
+<table><tr><td rowspan="2">Model</td><td colspan="5">Named Entities from News</td><td colspan="5">Named Entities from History</td></tr><tr><td>Minimal</td><td>News</td><td>History</td><td>Infrml</td><td>Avg</td><td>Minimal</td><td>News</td><td>History</td><td>Infrml</td><td>Avg</td></tr><tr><td>GPT</td><td>0.0</td><td>7.0</td><td>12.7</td><td>1.4</td><td>5.3</td><td>0.0</td><td>21.9</td><td>39.1</td><td>7.8</td><td>17.2</td></tr><tr><td>GPT2-small</td><td>22.5</td><td>63.4</td><td>50.7</td><td>15.5</td><td>38.0</td><td>12.5</td><td>29.7</td><td>56.2</td><td>12.5</td><td>27.7</td></tr><tr><td>GPT2-medium</td><td>33.8</td><td>64.8</td><td>49.3</td><td>12.7</td><td>40.2</td><td>21.9</td><td>32.8</td><td>62.5</td><td>4.7</td><td>30.5</td></tr><tr><td>GPT2-large</td><td>43.7</td><td>66.2</td><td>47.9</td><td>16.9</td><td>43.7</td><td>29.7</td><td>29.7</td><td>56.2</td><td>12.5</td><td>32.0</td></tr><tr><td>GPT2-XL</td><td>50.7</td><td>62.0</td><td>45.1</td><td>21.1</td><td>44.7</td><td>28.1</td><td>31.2</td><td>60.9</td><td>14.1</td><td>33.6</td></tr><tr><td>TransformerXL</td><td>14.1</td><td>18.3</td><td>15.5</td><td>12.7</td><td>15.2</td><td>35.9</td><td>43.8</td><td>51.6</td><td>37.5</td><td>42.2</td></tr><tr><td>XLNet-base</td><td>4.2</td><td>33.8</td><td>12.7</td><td>4.2</td><td>13.7</td><td>0.0</td><td>34.4</td><td>23.4</td><td>3.1</td><td>15.2</td></tr><tr><td>XLNet-large</td><td>11.3</td><td>40.8</td><td>23.9</td><td>9.9</td><td>21.5</td><td>6.2</td><td>29.7</td><td>31.2</td><td>7.8</td><td>18.7</td></tr><tr><td>Average</td><td>22.5</td><td>44.5</td><td>32.2</td><td>11.8</td><td>27.7</td><td>16.8</td><td>31.7</td><td>47.6</td><td>12.5</td><td>27.1</td></tr></table>
+
+Table 2: Percentage of named entities such that each LM greedily generates their last name conditioned on a prompt ending with their given name. Named entities are (1) frequently mentioned people in the U.S. news, or (2) prominent people from history.  
+
+<table><tr><td rowspan="2">Named Entity</td><td rowspan="2">Media Freq.</td><td rowspan="2">Rank</td><td colspan="2">Minimal Prompt</td><td colspan="2">News Prompt</td><td colspan="2">History Prompt</td><td colspan="2">Informal Prompt</td></tr><tr><td>Next Word</td><td>%</td><td>Next Word</td><td>%</td><td>Next Word</td><td>%</td><td>Next Word</td><td>%</td></tr><tr><td>Donald Trump</td><td>2,844,894</td><td>15</td><td>Trump</td><td>70.8</td><td>Trump</td><td>99.0</td><td>Trump</td><td>93.2</td><td>Trump</td><td>34.1</td></tr><tr><td>Hillary Clinton</td><td>373,952</td><td>788</td><td>Clinton</td><td>80.9</td><td>Clinton</td><td>91.6</td><td>Clinton</td><td>82.9</td><td>Clinton</td><td>46.5</td></tr><tr><td>Robert Mueller</td><td>322,466</td><td>3</td><td>B.[. Reich]</td><td>2.1</td><td>Mueller</td><td>82.2</td><td>F[. Kennedy]</td><td>13.5</td><td>.</td><td>16.6</td></tr><tr><td>Bernie Sanders</td><td>97,104</td><td>757</td><td>Sanders</td><td>66.8</td><td>Sanders</td><td>95.9</td><td>Sanders</td><td>84.8</td><td>Sanders</td><td>24.9</td></tr><tr><td>Benjamin Netanyahu</td><td>65,863</td><td>66</td><td>Netanyahu</td><td>10.8</td><td>Netanyahu</td><td>78.9</td><td>Franklin</td><td>61.3</td><td>.</td><td>15.7</td></tr><tr><td>Elizabeth Warren</td><td>58,370</td><td>5</td><td>,</td><td>4.7</td><td>Warren</td><td>90.1</td><td>Taylor</td><td>17.1</td><td>.</td><td>21.4</td></tr><tr><td>Marco Rubio</td><td>56,224</td><td>363</td><td>Rubio</td><td>15.2</td><td>Rubio</td><td>98.1</td><td>Polo</td><td>68.4</td><td>.</td><td>2.3</td></tr><tr><td>Richard Nixon</td><td>55,911</td><td>7</td><td>B[. Spencer]</td><td>2.1</td><td>Nixon</td><td>17.3</td><td>Nixon</td><td>76.8</td><td>.</td><td>20.0</td></tr></table>
+
+Table 3: Maximum next-word probabilities from GPT2-XL conditioned on prompts with first names of select people frequently mentioned in the media. Brackets represent additional (greedily) decoded tokens for disambiguation. Rank: aggregate 1990 U.S. Census data of most common male and female names.
+
+for grounding: while it is reasonable to assume that nearly all mentions of, e.g., “Hillary Clinton” in text are references to (the entity) Hillary Clinton, other references may use different strings (“Hillary Rodham Clinton,” “H.R.C.,” or just “Hillary”). We also note that the LM is not constrained to generate a last name but may instead select one of many other linguistically plausible continuations.
+
+We examine greedy decoding of named entity last names systematically for each generative LM. To this end, we compile two sets of prominent named entities from the media and from history. We construct four prompt templates ending with a given name to feed to each LM: (1) Minimal: “[NAME]”, (2) News: “A new report from CNN says that [NAME]”, (3) History: “A newly published biography of [NAME]”, and (4) Informal: “I want to introduce you to my best friend, [NAME]”. Table 2 shows, for each LM, the percentage of named entities for which the LM greedily generated that entity’s last name<sup>3</sup> conditioned on one of the four prompt templates.
+
+Overall, the GPT2 models (in particular, GPT2-XL), which are trained on web text - including news but excluding Wikipedia - are vastly more likely than other models to predict named entities from the news, across all prompts. The GPT2 models are also very likely to predict named entities from history, but primarily when conditioned with the History prompt. By contrast, the TransformerXL model, trained on Wikipedia articles, is overall more likely to predict historical named entities than any other model, and is substantially more likely to predict historical entities than news entities. The GPT model, trained on fiction is the least likely of any model to generate named entities from the news. These results clearly demonstrate that (1) the variance of named entity grounding effects across different LMs is great, and (2) these differences are likely at least partially attributable to differences in training data genre.
+
+Table 3 focuses on GPT2-XL and shows the next word prediction for 8 given names of named entities frequently appearing in the U.S. news media, which are also common in the general population. Due to the contextual nature of LMs, the prompt type affects the last-name probabilities. Intuitively, generating the last name of an entity seems appro
+
+<table><tr><td colspan="2">GPT</td><td colspan="2">GPT2-small</td><td colspan="2">GPT2-medium</td><td colspan="2">GPT2-large</td><td colspan="2">GPT2-XL</td><td colspan="2">TransformerXL</td><td colspan="2">XLNet-base</td><td colspan="2">XLNet-large</td></tr><tr><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td></tr><tr><td>Philip</td><td>0.739</td><td>Bernie</td><td>0.853</td><td>Bernie</td><td>0.884</td><td>Bernie</td><td>0.815</td><td>Bernie</td><td>0.966</td><td>Virginia</td><td>0.761</td><td>Grace</td><td>0.793</td><td>Brittany</td><td>0.808</td></tr><tr><td>Bryan</td><td>0.683</td><td>Donald</td><td>0.800</td><td>Donald</td><td>0.845</td><td>Barack</td><td>0.800</td><td>Donald</td><td>0.922</td><td>Dylan</td><td>0.742</td><td>Rose</td><td>0.705</td><td>Matthew</td><td>0.803</td></tr><tr><td>Beverly</td><td>0.670</td><td>Victoria</td><td>0.772</td><td>Irma</td><td>0.834</td><td>Theresa</td><td>0.773</td><td>Hillary</td><td>0.869</td><td>Hillary</td><td>0.731</td><td>Martha</td><td>0.702</td><td>Amber</td><td>0.788</td></tr><tr><td>Louis</td><td>0.641</td><td>Virginia</td><td>0.771</td><td>Christian</td><td>0.822</td><td>Donald</td><td>0.759</td><td>Barack</td><td>0.832</td><td>Jeff</td><td>0.715</td><td>Victoria</td><td>0.700</td><td>Hillary</td><td>0.782</td></tr><tr><td>Danielle</td><td>0.639</td><td>Gloria</td><td>0.763</td><td>Hillary</td><td>0.782</td><td>Victoria</td><td>0.702</td><td>Virginia</td><td>0.767</td><td>Alice</td><td>0.693</td><td>Alice</td><td>0.692</td><td>Teresa</td><td>0.771</td></tr><tr><td>Kelly</td><td>0.631</td><td>Hillary</td><td>0.756</td><td>Barack</td><td>0.774</td><td>Matthew</td><td>0.688</td><td>Christian</td><td>0.749</td><td>Thomas</td><td>0.690</td><td>Hillary</td><td>0.661</td><td>Grace</td><td>0.764</td></tr><tr><td>Nicholas</td><td>0.631</td><td>Cheryl</td><td>0.755</td><td>Victoria</td><td>0.766</td><td>Jacob</td><td>0.688</td><td>Jose</td><td>0.746</td><td>Judy</td><td>0.681</td><td>Mary</td><td>0.657</td><td>Virginia</td><td>0.762</td></tr><tr><td>Brenda</td><td>0.630</td><td>Jeff</td><td>0.733</td><td>Virginia</td><td>0.760</td><td>Billy</td><td>0.677</td><td>Irma</td><td>0.739</td><td>Gregory</td><td>0.677</td><td>Kenneth</td><td>0.656</td><td>Jordan</td><td>0.755</td></tr><tr><td>Vincent</td><td>0.628</td><td>Ann</td><td>0.697</td><td>Joyce</td><td>0.757</td><td>Virginia</td><td>0.676</td><td>Joseph</td><td>0.732</td><td>Samantha</td><td>0.676</td><td>Bobby</td><td>0.653</td><td>Madison</td><td>0.754</td></tr><tr><td>Russell</td><td>0.625</td><td>Christina</td><td>0.693</td><td>Alice</td><td>0.753</td><td>Paul</td><td>0.668</td><td>Sophia</td><td>0.717</td><td>Amber</td><td>0.675</td><td>Virginia</td><td>0.651</td><td>Barack</td><td>0.751</td></tr><tr><td colspan="2">0.526 ± 0.157</td><td colspan="2">0.568 ± 0.173</td><td colspan="2">0.572 ± 0.182</td><td colspan="2">0.545 ± 0.166</td><td colspan="2">0.549 ± 0.181</td><td colspan="2">0.552 ± 0.169</td><td colspan="2">0.525 ± 0.162</td><td colspan="2">0.548 ± 0.175</td></tr></table>
+
+Table 4: Top 10 most predictable names from the "is a" endings for each model, using Nucleus sampling with  $p = 0.9$  and limiting the number of generated tokens to 150. Bold entries mark given names that appear frequently in the media. Bottom: mean and STD of scores.
+
+priate and expected in news-like contexts ("A new report from CNN says that [NAME]") but less so in more personal contexts ("I want to introduce you to my best friend, [NAME]"). Indeed, Table 3 demonstrates grounding effects are strongest in news-like contexts; however, these effects are still clearly present across all contexts—appropriate or not—for more prominent named entities in the U.S. media (Donald, Hillary, and Bernie). When prompted with given name only, GPT2-XL predicts the last name of a prominent named entity in all but one case (Elizabeth). In three cases, the corresponding probability is well over  $50\%$  (Clinton, Trump, Sanders), and in one case generates the full name of a white supremacist, Richard B. Spencer.
+
+# 3 Given Name Recovery
+
+Given a text discussing a certain person, can we recover their (masked) given name? Our hypothesis was that it would be more feasible for a given name prone to grounding, due to unique terms that appear across multiple texts discussing this person.
+
+To answer this question, we compiled a list of the 100 most frequent male and female names in the U.S.,<sup>4</sup> to which we added the first names of the most discussed people in the media (Section 2). Using the template "[NAME] is a" we generated 50 endings of 150 tokens for each name, with each of the generator LMs (Table 1), using Nucleus sampling (Holtzman et al., 2019) with  $p = 0.9$ . For each pair of same-gender given names,<sup>5</sup> we trained a binary SVM classifier using the Scikit-learn library (Pedregosa et al., 2011) to predict the given name from the TF-IDF representation of the endings, excluding the name. Finally, we computed the average of pairwise  $F_{1}$  scores as a single score
+
+![](images/1b67f8d43646f2873ede1a82b3bb35944de71ce2fc2b435ff81dbdd1761399e7.jpg)  
+Figure 1: t-SNE projection of BERT vectors of the GPT2-large "is a" endings for Helen, Ruth, and Hillary.
+
+![](images/8f40b815ca52360bbc106e441f1d4f102d28e6c8006f80c5b2d13144d0dc5cd4.jpg)
+
+![](images/53769e39a603e406ade8bd760d94df082b573de2843d45c74129be93bbb27a0d.jpg)
+
+per given name.
+
+Table 4 displays the top 10 names with the most distinguishable "is a" endings. Bold entries mark given names of media entities, most prominent in the GPT2 models, trained on web text. Apart from U.S. politicians, Virginia (name of a state) and Irma (a widely discussed hurricane) are also predictable, supposedly due to their other senses. The results are consistent for different generation lengths and sampling strategies (see Section B).
+
+Figure 1 illustrates the ease of distinguishing texts discussing Hillary from others (GPT2-large). We masked the name ("[MASK] is a..."), computed the BERT vectors, and projected them to 2d using t-SNE (Maaten and Hinton, 2008). Similar results were observed for texts generated by other GPT2 models, for different names (e.g., Donald, Bernie), and with other input representations (TF-IDF).
+
+# 4 Sentiment Analysis
+
+Following Prabhakaran et al. (2019), we can expect endings (§3) discussing specific named entities to be associated with sentiment more consistently than those discussing hypothetical people. We predict sentiment using the AllenNLP sentiment analyzer (Gardner et al., 2018) trained on the Stanford Sentiment Treebank (Socher et al., 2013).
+
+Table 5 displays the top 10 most negative given names for each LM, where per-name score is the average of negative sentiment scores for their endings.
+
+<table><tr><td colspan="2">GPT</td><td colspan="2">GPT2-small</td><td colspan="2">GPT2-medium</td><td colspan="2">GPT2-large</td><td colspan="2">GPT2-XL</td><td colspan="2">TransformerXL</td><td colspan="2">XLNet-base</td><td colspan="2">XLNet-large</td></tr><tr><td>Name</td><td>Score</td><td>Name</td><td>Score</td><td>Name</td><td>Score</td><td>Name</td><td>Score</td><td>Name</td><td>Score</td><td>Name</td><td>Score</td><td>Name</td><td>Score</td><td>Name</td><td>Score</td></tr><tr><td>Noah</td><td>0.808</td><td>Bernie</td><td>0.619</td><td>Donald</td><td>0.629</td><td>Bernie</td><td>0.556</td><td>Alice</td><td>0.620</td><td>Sean</td><td>0.526</td><td>Judy</td><td>0.382</td><td>Kyle</td><td>0.324</td></tr><tr><td>John</td><td>0.802</td><td>Donald</td><td>0.591</td><td>Bernie</td><td>0.565</td><td>Hillary</td><td>0.537</td><td>Donald</td><td>0.546</td><td>Mitch</td><td>0.525</td><td>Albert</td><td>0.375</td><td>Rudy</td><td>0.318</td></tr><tr><td>Keith</td><td>0.800</td><td>Ryan</td><td>0.560</td><td>Jerry</td><td>0.559</td><td>Johnny</td><td>0.505</td><td>Chuck</td><td>0.526</td><td>Jack</td><td>0.512</td><td>Johnny</td><td>0.370</td><td>Johnny</td><td>0.318</td></tr><tr><td>Kenneth</td><td>0.795</td><td>Hillary</td><td>0.547</td><td>Kevin</td><td>0.546</td><td>Alice</td><td>0.490</td><td>Ryan</td><td>0.524</td><td>Johnny</td><td>0.507</td><td>Hillary</td><td>0.357</td><td>Sean</td><td>0.304</td></tr><tr><td>Kevin</td><td>0.790</td><td>Lisa</td><td>0.519</td><td>Joe</td><td>0.544</td><td>Barack</td><td>0.469</td><td>Judy</td><td>0.520</td><td>Brian</td><td>0.505</td><td>Alice</td><td>0.347</td><td>Evelyn</td><td>0.277</td></tr><tr><td>Virginia</td><td>0.782</td><td>Johnny</td><td>0.492</td><td>Jose</td><td>0.539</td><td>Wayne</td><td>0.463</td><td>Paul</td><td>0.513</td><td>Jessica</td><td>0.492</td><td>Henry</td><td>0.343</td><td>Steve</td><td>0.276</td></tr><tr><td>Billy</td><td>0.782</td><td>Rick</td><td>0.490</td><td>Brandon</td><td>0.532</td><td>Rudy</td><td>0.453</td><td>Barack</td><td>0.509</td><td>Boris</td><td>0.492</td><td>Rachel</td><td>0.342</td><td>Jane</td><td>0.252</td></tr><tr><td>Bernie</td><td>0.782</td><td>Dorothy</td><td>0.484</td><td>Bill</td><td>0.528</td><td>Bill</td><td>0.449</td><td>Hillary</td><td>0.490</td><td>Patricia</td><td>0.489</td><td>Gary</td><td>0.332</td><td>Jonathan</td><td>0.251</td></tr><tr><td>Randy</td><td>0.781</td><td>Jose</td><td>0.479</td><td>Jack</td><td>0.528</td><td>Jordan</td><td>0.446</td><td>Betty</td><td>0.489</td><td>Jennifer</td><td>0.488</td><td>Barbara</td><td>0.331</td><td>Stephanie</td><td>0.246</td></tr><tr><td>Madison</td><td>0.779</td><td>Noah</td><td>0.478</td><td>Hillary</td><td>0.522</td><td>Marco</td><td>0.442</td><td>Jerry</td><td>0.484</td><td>Amy</td><td>0.486</td><td>Rick</td><td>0.329</td><td>Gerald</td><td>0.244</td></tr><tr><td colspan="2">0.687 ± 0.052</td><td colspan="2">0.339 ± 0.073</td><td colspan="2">0.350 ± 0.079</td><td colspan="2">0.328 ± 0.067</td><td colspan="2">0.331 ± 0.077</td><td colspan="2">0.385 ± 0.055</td><td colspan="2">0.236 ± 0.053</td><td colspan="2">0.149 ± 0.049</td></tr></table>
+
+Table 5: Top 10 names with the most negative sentiment for their "is a" endings on average, for each model. Bold entries mark given names that appear frequently in the media. Bottom: mean and STD of average negative scores. Endings were generated using Nucleus sampling with  $p = 0.9$  and limiting the number of generated tokens to 150.
+
+Again, many of the top names are given names of people discussed in the media, mainly U.S. politicians, and more so in the GPT2 models.<sup>6</sup> We found the variation among the most positive scores to be low. We conjecture that LMs typically default to generating neutral texts about hypothetical people.
+
+# 5 Effect on Downstream Tasks
+
+Pre-trained LMs are now used as a starting point for a vast array of downstream tasks (Raffel et al., 2019), raising concerns about unintended consequences in such models. To study an aspect of this, we construct a set of 26 question-answer probe templates with [NAME1] and [NAME2] slots.
+
+We populate the templates with pairs of same-gender names sampled from the list in §2. We evaluate the expanded templates on a set of LMs fine-tuned for either SQuAD (exemplified in Figure 2; Rajpurkar et al., 2016), or (slightly tweaked) Winogrande (Sakaguchi et al., 2020), with optional pre-fine-tuning on RACE (Lai et al., 2017; Sun et al., 2018). We calculate how often the model prediction changes when [NAME1] and [NAME2] are swapped in the template (flips).
+
+Table 6 and Table 7 present the top names contributing to the name swap fragility and the overall LM scores. SQuAD models exhibit a significant effect for all LMs, from weak to strong. Conversely, Winogrande models are mostly insulated from this effect. We speculate that the nature of the Winogrande training set, having seen many examples of names used in generic fashion, have helped remove the inherent artifacts associated with names.
+
+We also note that extra pre-fine-tuning on RACE, although not helping noticeably with the original task, seems to increase robustness for name swaps.
+
+![](images/50e30b0f31b7158183961f1643d4c1f47a9a76595494a247197d95e326a017a5.jpg)  
+Figure 2: Sample name swap template and the per-slot accuracy on certain given names. Large gaps between the two slots may indicate grounding.
+
+# 6 Related Work
+
+Social Bias. There is multiple evidence that word embeddings encode gender and racial bias (Bolukbasi et al., 2016; Caliskan et al., 2017; Manzini et al., 2019; Gonen and Goldberg, 2019), in particular in the representations of given names (Romanov et al., 2019). Bias can perpetuate to downstream tasks such as coreference resolution (Webster et al., 2018; Rudinger et al., 2018; Zhao et al., 2018), natural language inference (Rudinger et al., 2017), machine translation (Stanovsky et al., 2019), and sentiment analysis (Díaz et al., 2018). In natural language generation, prompts with mentions of demographic groups (e.g., "The gay person was") may generate stereotypical texts (Sheng et al., 2019).
+
+Named Entities. Field and Tsvetkov (2019) used pre-trained LMs to analyze power, sentiment, and agency aspects of entities, and found the representations were biased towards the LM training corpus. In particular, frequently discussed entities such as politicians biased the representations of their given names. Prabhakaran et al. (2019) showed that bias reflected in the language describing named entities is encoded into their representations, in particular
+
+<table><tr><td colspan="2">RoBERTa-base</td><td colspan="2">RoBERTa-large</td><td colspan="2">RoBERTa-large w/RACE</td><td colspan="2">XLNet-base</td><td colspan="2">XLNet-large</td><td colspan="2">RoBERTa-largeW</td><td colspan="2">RoBERTa-largeW w/RACE</td></tr><tr><td>Name</td><td>flips</td><td>Name</td><td>flips</td><td>Name</td><td>flips</td><td>Name</td><td>flips</td><td>Name</td><td>flips</td><td>Name</td><td>flips</td><td>Name</td><td>flips</td></tr><tr><td>Meghan</td><td>36.8</td><td>Hillary</td><td>34.6</td><td>Hillary</td><td>17.1</td><td>Dianne</td><td>20.7</td><td>Emily</td><td>23.2</td><td>Chuck</td><td>7.5</td><td>Hillary</td><td>2.4</td></tr><tr><td>Hillary</td><td>26.9</td><td>Emily</td><td>19.6</td><td>Meghan</td><td>16.3</td><td>Donald</td><td>16.5</td><td>Irma</td><td>21.9</td><td>Hillary</td><td>5.4</td><td>Barack</td><td>2.2</td></tr><tr><td>Mark</td><td>25.6</td><td>Meghan</td><td>18.4</td><td>Lindsey</td><td>15.2</td><td>Meghan</td><td>16.4</td><td>Thomas</td><td>21.5</td><td>Dianne</td><td>5.4</td><td>Barbara</td><td>1.1</td></tr><tr><td>Andrew</td><td>25.3</td><td>Christopher</td><td>18.2</td><td>Mary</td><td>15.0</td><td>Irma</td><td>15.9</td><td>Jennifer</td><td>19.2</td><td>Kimberly</td><td>4.7</td><td>Margaret</td><td>0.6</td></tr><tr><td>Michelle</td><td>24.0</td><td>Barack</td><td>17.9</td><td>Donald</td><td>14.2</td><td>Mary</td><td>15.5</td><td>Christine</td><td>19.0</td><td>Timothy</td><td>4.2</td><td>Meghan</td><td>0.6</td></tr></table>
+
+Table 6: Top flipping names (bold for media names) for name swap probes in SQuAD and Winogrande  $(^{W})$  models.  
+
+<table><tr><td>Model</td><td>Task</td><td>Probe</td><td>Flips</td><td>Flips top-5</td></tr><tr><td>RoBERTa-base</td><td>91.2</td><td>49.6</td><td>15.7</td><td>51.0</td></tr><tr><td>RoBERTa-large</td><td>94.4</td><td>82.2</td><td>9.8</td><td>31.2</td></tr><tr><td>RoBERTa-large w/RACE</td><td>94.4</td><td>87.9</td><td>7.7</td><td>33.8</td></tr><tr><td>XLNet-base</td><td>90.3</td><td>54.5</td><td>7.3</td><td>24.3</td></tr><tr><td>XLNet-large</td><td>93.4</td><td>82.9</td><td>14.8</td><td>54.4</td></tr><tr><td>RoBERTa-largeW</td><td>79.3</td><td>90.5</td><td>2.5</td><td>12.7</td></tr><tr><td>RoBERTa-largeW w/RACE</td><td>81.5</td><td>96.1</td><td>0.2</td><td>0.8</td></tr></table>
+
+Table 7: Performance (SQuAD: dev  $F_{1}$ , Winogrande  $^{W}$ ): dev accuracy) on the main task (Task) and the name swap probes (Probe). Flips measures how often name pairs change model output when swapped, with top-5 computed over the 5 most affected templates.
+
+associating politicians with toxicity. The potential effect on downstream applications is demonstrated with the sensitivity of sentiment and toxicity systems to name perturbation, which can be mitigated by name perturbation during training.
+
+Reporting Bias. People rarely state the obvious (Grice et al., 1975), thus uncommon events are reported disproportionately, and their frequency in corpora does not directly reflect real-world frequency (Gordon and Van Durme, 2013; Sorower et al., 2011). A private case of reporting bias is towards named entities: not all Donalds are discussed with equal probability. Web corpora specifically likely suffer from media bias, making some entities more visible than others (coverage bias; D'Alessio and Allen, 2006), sometimes due to "newsworthiness" (structural bias; van Dalen, 2012).
+
+# 7 Ethical Considerations and Conclusion
+
+We explored biases in pre-trained LMs with respect to given names and the named entities that share them. We discuss two types of ethical considerations pertaining to this work: (1) the limitations of this work, and (2) the implications of our findings.
+
+Our methodology relies on a number of limitations that should be considered in understanding the scope of our conclusions. First, we evaluated only English LMs, thus we cannot assume these results will extend to LMs in different languages. Second, the lists of names we use to analyze these models are not broadly representative of English-speaking populations. The list of most common given names
+
+in the U.S. are over-representative of stereotypically white and Western names. The list of most frequently named people in the media as well as A&E's (subjective) list of most influential people of the millennium both are male-skewed, owing to many sources of gender bias, both historical and contemporary. For our last name prediction experiment, we are forced to filter named entities whose given names don't precede the surname, which is a cultural assumption that precludes naming conventions from many languages, like Chinese and Korean. We used statistical resources that treat gender as a binary construct, which is a reductive view of gender. We hope future work may better address this limitation, as in the work of Cao and Daumé III (2019). Finally, there are many other important types of biases pertaining to given names that we do not focus on, including biases on the basis of perceived race or gender (e.g. Bertrand and Mullainathan, 2004; Moss-Racusin et al., 2012). While our experiments shed light on artifacts of certain common U.S. given names, an equally important question is how LMs treat very uncommon names, effects which would disproportionately impact members of minority groups.
+
+What this work does do, however, is shed light on a particular behavior of pre-trained LMs which has potential ethical implications. Pre-trained LMs do not treat given names as interchangeable or anonymous; this has not only implications for the quality and accuracy of systems that employ these LMs, but also for the fairness of those systems. Furthermore, as we observed with GPT2-XL's freeform production of a white supremacist's name conditioned only on a common given name (Richard), further inquiry into the source of training data of these models is warranted.
+
+# Acknowledgments
+
+This research was supported in part by NSF (IIS-1524371, IIS-1714566), DARPA under the CwC program through the ARO (W911NF-15-1-0543), and DARPA under the MCS program through NIWC Pacific (N66001-19-2-4031).
+
+# References
+
+Marianne Bertrand and Sendhil Mullainathan. 2004. Are emily and greg more employable than lakisha and jamal? a field experiment on labor market discrimination. American economic review, 94(4):991-1013.  
+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, pages 4349-4357.  
+Aylin Caliskan, Joanna J Bryson, and Arvind Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334):183-186.  
+Yang Trista Cao and Hal Daumé III. 2019. Toward gender-inclusive coreference resolution. arXiv preprint arXiv:1910.13913.  
+Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc Le, and Ruslan Salakhutdinov. 2019. Transformer-XL: Attentive language models beyond a fixed-length context. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2978-2988, Florence, Italy. Association for Computational Linguistics.  
+Arjen van Dalen. 2012. Structural bias in cross-national perspective: How political systems and journalism cultures influence government dominance in the news. The International Journal of Press/Politics, 17(1):32-55.  
+Dave D'Alessio and Mike Allen. 2006. Media Bias in Presidential Elections: A Meta-Analysis. Journal of Communication, 50(4):133-156.  
+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, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Mark Díaz, Isaac Johnson, Amanda Lazar, Anne Marie Piper, and Darren Gergle. 2018. Addressing age-related bias in sentiment analysis. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, page 412. ACM.  
+Anjalie Field and Yulia Tsvetkov. 2019. Entity-centric contextual affective analysis. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2550-2560, Florence, Italy. Association for Computational Linguistics.  
+Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson F. Liu, Matthew Peters, Michael Schmitz, and Luke Zettlemoyer. 2018.
+
+AllenNLP: A deep semantic natural language processing platform. In Proceedings of Workshop for NLP Open Source Software (NLP-OSS), pages 1-6, Melbourne, Australia. Association for Computational Linguistics.  
+Hila Gonen and Yoav Goldberg. 2019. Lipstick on a pig: Debiasing methods cover up systematic gender biases in word embeddings but do not remove them. 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 609-614, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Jonathan Gordon and Benjamin Van Durme. 2013. Reporting bias and knowledge acquisition. In Proceedings of the 2013 workshop on Automated knowledge base construction, pages 25-30. ACM.  
+H Paul Grice, Peter Cole, Jerry Morgan, et al. 1975. Logic and conversation. 1975, pages 41-58.  
+Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. In International Conference on Learning Representations.  
+Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard H. Hovy. 2017. Race: Large-scale reading comprehension dataset from examinations. In EMNLP.  
+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.  
+Laurens van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of machine learning research, 9(Nov):2579-2605.  
+Thomas Manzini, Lim Yao Chong, Alan W Black, and Yulia Tsvetkov. 2019. Black is to criminal as caucasian is to police: Detecting and removing multiclass bias in word embeddings. 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 615-621, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Chandler May, Alex Wang, Shikha Bordia, Samuel R. Bowman, and Rachel Rudinger. 2019. On measuring social biases in sentence encoders. 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 622-628, Minneapolis, Minnesota. Association for Computational Linguistics.
+
+Corinne A Moss-Racusin, John F Dovidio, Victoria L Brescoll, Mark J Graham, and Jo Handelsman. 2012. Science faculty's subtle gender biases favor male students. Proceedings of the National Academy of Sciences, 109(41):16474-16479.  
+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.  
+Fabio Petroni, Tim Rocttäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China. Association for Computational Linguistics.  
+Vinodkumar Prabhakaran, Ben Hutchinson, and Margaret Mitchell. 2019. Perturbation sensitivity analysis to detect unintended model biases. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5739-5744, Hong Kong, China. Association for Computational Linguistics.  
+Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training. --  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. -  
+Colin Raffel, Noam Shazeer, Adam Kaleo 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, abs/1910.10683.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad:  $100,000+$  questions for machine comprehension of text. In EMNLP.  
+Alexey Romanov, Maria De-Arteaga, Hanna Wallach, Jennifer Chayes, Christian Borgs, Alexandra Chouldechova, Sahin Geyik, Krishnamaram Kenthapadi, Anna Rumshisky, and Adam Kalai. 2019. What's in a name? Reducing bias in bios without access to protected attributes. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4187-4195, Minneapolis, Minnesota. Association for Computational Linguistics.
+
+Rachel Rudinger, Chandler May, and Benjamin Van Durme. 2017. Social bias in elicited natural language inferences. In Proceedings of the First ACL Workshop on Ethics in Natural Language Processing, pages 74-79, Valencia, Spain. Association for Computational Linguistics.  
+Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. 2018. Gender bias in coreference resolution. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 8-14, New Orleans, Louisiana. Association for Computational Linguistics.  
+Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2020. Winogrande: An adversarial winograd schema challenge at scale. In AAAI.  
+Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. 2019. The woman worked as a babysitter: On biases in language generation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3398-3403, Hong Kong, China. Association for Computational Linguistics.  
+Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, Seattle, Washington, USA. Association for Computational Linguistics.  
+Mohammad S Sorower, Janardhan R Doppa, Walker Orr, Prasad Tadepalli, Thomas G Dietterich, and Xiaoli Z Fern. 2011. Inverting grice's maxims to learn rules from natural language extractions. In Advances in neural information processing systems, pages 1053-1061.  
+Gabriel Stanovsky, Noah A. Smith, and Luke Zettle-moyer. 2019. Evaluating gender bias in machine translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1679-1684, Florence, Italy. Association for Computational Linguistics.  
+Kai Sun, Dian Yu, Dong Yu, and Claire Cardie. 2018. Improving machine reading comprehension with general reading strategies. In *NAACL-HLT*.  
+Kellie Webster, Marta Recasens, Vera Axelrod, and Jason Baldridge. 2018. Mind the gap: A balanced corpus of gendered ambiguous pronouns. Transactions of the Association for Computational Linguistics, 6:605-617.  
+Zhilin Yang, Zihang Dai, Yiming Yang, Jaime G. Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. 2019.
+
+Xlnet: Generalized autoregressive pretraining for language understanding. CoRR, abs/1906.08237.
+
+Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2018. Gender bias in coreference resolution: Evaluation and debiasing methods. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 15–20, New Orleans, Louisiana. Association for Computational Linguistics.
+
+# A Lists of Given Names
+
+Tables 8 and 9 specify the given names used in this paper for females and males, respectively, along with named entities with each given name, and the sections of the experiments in which they were included (2 - last name prediction, 3 - given name recovery, 4 - sentiment analysis, and 5 - effect on downstream tasks).
+
+<table><tr><td>Name</td><td>Media</td><td>History</td><td>2</td><td>3-4</td><td>5</td><td>Name</td><td>Media</td><td>History</td><td>2</td><td>3-4</td><td>5</td></tr><tr><td>Abigail</td><td></td><td></td><td></td><td>✘</td><td></td><td>Joyce</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Alexis</td><td></td><td></td><td></td><td>✘</td><td></td><td>Judith</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Alice</td><td></td><td></td><td></td><td>✘</td><td></td><td>Judy</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Amanda</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Julia</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Amber</td><td></td><td></td><td></td><td>✘</td><td></td><td>Julie</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Amy</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Karen</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Andrea</td><td></td><td></td><td></td><td>✘</td><td></td><td>Katherine</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Angela</td><td>Merkel</td><td></td><td></td><td>✘</td><td>✘</td><td>Kathleen</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Ann</td><td></td><td></td><td></td><td>✘</td><td></td><td>Kathryn</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Anna</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Kayla</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Ashley</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Kelly</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Barbara</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Kimberly</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Betty</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Kirstjen</td><td>Nielsen</td><td></td><td>✘</td><td></td><td></td></tr><tr><td>Beverly</td><td></td><td></td><td></td><td>✘</td><td></td><td>Laura</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Brenda</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Lauren</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Brittany</td><td></td><td></td><td></td><td>✘</td><td></td><td>Linda</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Carol</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Lindsey</td><td>Graham</td><td></td><td>✘</td><td>✘</td><td>✘</td></tr><tr><td>Carolyn</td><td></td><td></td><td></td><td>✘</td><td></td><td>Lisa</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Catherine</td><td></td><td></td><td></td><td>✘</td><td></td><td>Lori</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Cheryl</td><td></td><td></td><td></td><td>✘</td><td></td><td>Madison</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Christina</td><td></td><td></td><td></td><td>✘</td><td></td><td>Margaret</td><td></td><td>Sanger</td><td>✘</td><td>✘</td><td>✘</td></tr><tr><td>Christine</td><td>Blasey Ford</td><td></td><td></td><td>✘</td><td>✘</td><td>Maria</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Cynthia</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Marie</td><td></td><td>Curie</td><td>✘</td><td>✘</td><td></td></tr><tr><td>Danielle</td><td></td><td></td><td></td><td>✘</td><td></td><td>Marilyn</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Deborah</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Martha</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Debra</td><td></td><td></td><td></td><td>✘</td><td></td><td>Mary</td><td></td><td>Wollstonecraft</td><td>✘</td><td>✘</td><td>✘</td></tr><tr><td>Denise</td><td></td><td></td><td></td><td>✘</td><td></td><td>Megan</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Diana</td><td></td><td></td><td></td><td>✘</td><td></td><td>Meghan</td><td>Markle</td><td></td><td>✘</td><td>✘</td><td>✘</td></tr><tr><td>Diane</td><td></td><td></td><td></td><td>✘</td><td></td><td>Melania</td><td>Trump</td><td></td><td>✘</td><td></td><td></td></tr><tr><td>Dianne</td><td>Feinstein</td><td></td><td></td><td>✘</td><td>✘</td><td>Melissa</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Donna</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Michelle</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Doris</td><td></td><td></td><td></td><td>✘</td><td></td><td>Nancy</td><td>Pelosi</td><td></td><td>✘</td><td>✘</td><td>✘</td></tr><tr><td>Dorothy</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Natalie</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Eleanor</td><td></td><td></td><td>Roosevelt</td><td>✘</td><td></td><td>Nicole</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Elizabeth</td><td>Warren</td><td></td><td>Stanton</td><td>✘</td><td>✘</td><td>Nikki</td><td>Haley</td><td></td><td>✘</td><td></td><td></td></tr><tr><td>Emily</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Olivia</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Emma</td><td></td><td></td><td></td><td>✘</td><td></td><td>Oprah</td><td>Winfrey</td><td></td><td>✘</td><td></td><td></td></tr><tr><td>Evelyn</td><td></td><td></td><td></td><td>✘</td><td></td><td>Pamela</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Florence</td><td></td><td></td><td>Nightingale</td><td>✘</td><td></td><td>Patricia</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Frances</td><td></td><td></td><td></td><td>✘</td><td></td><td>Rachel</td><td></td><td>Carson</td><td>✘</td><td>✘</td><td></td></tr><tr><td>Gloria</td><td></td><td></td><td></td><td>✘</td><td></td><td>Rebecca</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Grace</td><td></td><td></td><td></td><td>✘</td><td></td><td>Rose</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Hannah</td><td></td><td></td><td></td><td>✘</td><td></td><td>Ruth</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Harriet</td><td></td><td></td><td>Tubman</td><td>✘</td><td></td><td>Samantha</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Heather</td><td></td><td></td><td></td><td>✘</td><td></td><td>Sandra</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Helen</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Sara</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Hillary</td><td>Clinton</td><td></td><td></td><td>✘</td><td>✘</td><td>Sarah</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Irma</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Sharon</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Ivana</td><td>Trump</td><td></td><td></td><td>✘</td><td></td><td>Shirley</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Jacqueline</td><td></td><td></td><td></td><td>✘</td><td></td><td>Sophia</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Jane</td><td></td><td></td><td>Austen</td><td>✘</td><td>✘</td><td>Stephanie</td><td></td><td></td><td></td><td>✘</td><td>✘</td></tr><tr><td>Janet</td><td></td><td></td><td></td><td>✘</td><td></td><td>Susan</td><td>Collins</td><td></td><td>✘</td><td>✘</td><td>✘</td></tr><tr><td>Janice</td><td></td><td></td><td></td><td>✘</td><td></td><td>Teresa</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Jean</td><td></td><td></td><td></td><td>✘</td><td></td><td>Theresa</td><td>May</td><td></td><td>✘</td><td>✘</td><td>✘</td></tr><tr><td>Jennifer</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Victoria</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Jessica</td><td></td><td></td><td></td><td>✘</td><td>✘</td><td>Virginia</td><td></td><td></td><td></td><td>✘</td><td></td></tr><tr><td>Joan</td><td></td><td></td><td></td><td>✘</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr></table>
+
+Table 8: Female given names used in this paper.
+
+Media entities source: Most discussed people in 2018 U.S. news media (https://public.tableau.com/view/2018Top100/1_Top100). History entities source: A&E's Biography: 100 Most Influential People of the Millennium (https://wmich.edu/mus-gened/mus150/biography100.html), after filtering out names that are not simple Given Name + Last Name (e.g. Suleiman I, "The Beatles").
+
+# B Given Name Prediction
+
+In Section 3 we have presented the most predictable given names from the generated texts using Nucleus sampling with  $p = 0.9$  and limiting the number of generated tokens to 150. Here we present the result with different hyper-parameters. Specifically, Tables 10 and 11 display the results for different lengths, 75 and 300 respectively, while Table 12 shows the results with length 150 and top k sampling with  $k = 25$ . The results are highly consistent for the different hyperparameter values. We omitted the results for beam search because it tends to generate very homogeneous texts for each name, making it trivial to classify all the names.
+
+# C Sentiment Analysis
+
+Table 13 shows the most negative "is a" ending generated by GPT2-small for some of the people with the most negative average sentiment.
+
+In Section 4 we have presented the most negative given names based on the generated texts using Nucleus sampling with  $p = 0.9$  and limiting the number of generated tokens to 150. Here we present the result with different hyper-parameters. Specifically, Tables 16 and 14 display the results for different lengths, 75 and 300 respectively, while Table 15 shows the results with length 150 and top k sampling with  $k = 25$ . The results are highly consistent for the different hyperparameter values.
+
+# D Effect on Downstream Tasks
+
+Figure 3 shows 6 (out of 26) example name swap probing templates, along with the most affected given names for each model.
+
+<table><tr><td>Name</td><td>Media</td><td>History</td><td>2 3-4 5</td><td>Name</td><td>Media</td><td>History</td><td>2 3-4 5</td></tr><tr><td>Aaron</td><td>Rodgers</td><td></td><td>× ×</td><td>Jon</td><td>Gruden</td><td></td><td>×</td></tr><tr><td>Abraham</td><td></td><td>Lincoln</td><td>×</td><td>Jonas</td><td></td><td>Salk</td><td>×</td></tr><tr><td>Adam</td><td></td><td>Smith</td><td>× ×</td><td>Jonathan</td><td></td><td></td><td>×</td></tr><tr><td>Adolf</td><td></td><td>Hitler</td><td>×</td><td>Jordan</td><td></td><td></td><td>×</td></tr><tr><td>Alan</td><td></td><td></td><td>×</td><td>Jose</td><td></td><td></td><td>×</td></tr><tr><td>Albert</td><td></td><td>Einstein</td><td>× ×</td><td>Joseph</td><td></td><td>Stalin</td><td>× × ×</td></tr><tr><td>Alex</td><td>Cora</td><td></td><td>×</td><td>Joshua</td><td></td><td></td><td>× ×</td></tr><tr><td>Alexander</td><td></td><td>Fleming</td><td>× ×</td><td>Juan</td><td></td><td></td><td>×</td></tr><tr><td>Andrew</td><td>Cuomo</td><td></td><td>× × ×</td><td>Justin</td><td>Trudeau</td><td></td><td>× ×</td></tr><tr><td>Anthony</td><td>Kennedy</td><td></td><td>× × ×</td><td>Karl</td><td></td><td>Marx</td><td>×</td></tr><tr><td>Arthur</td><td></td><td></td><td>×</td><td>Keith</td><td></td><td></td><td>×</td></tr><tr><td>Austin</td><td></td><td></td><td>×</td><td>Kenneth</td><td></td><td></td><td>× ×</td></tr><tr><td>Baker</td><td>Mayfield</td><td></td><td>×</td><td>Kevin</td><td>Durant</td><td></td><td>× × ×</td></tr><tr><td>Barack</td><td>Obama</td><td></td><td>× × ×</td><td>Klay</td><td>Thompson</td><td></td><td>×</td></tr><tr><td>Benjamin</td><td>Netanyahu</td><td>Franklin</td><td>× × ×</td><td>Kyle</td><td></td><td></td><td>×</td></tr><tr><td>Bernie</td><td>Sanders</td><td></td><td>× × ×</td><td>Larry</td><td>Nassar</td><td></td><td>× ×</td></tr><tr><td>Bill</td><td>Clinton</td><td>Gates</td><td>× × ×</td><td>Lawrence</td><td></td><td></td><td>×</td></tr><tr><td>Billy</td><td></td><td></td><td>×</td><td>LeBron</td><td>James</td><td></td><td>×</td></tr><tr><td>Bobby</td><td></td><td></td><td>×</td><td>Logan</td><td></td><td></td><td>×</td></tr><tr><td>Boris</td><td></td><td></td><td>× ×</td><td>Louis</td><td></td><td>Pasteur</td><td>× ×</td></tr><tr><td>Bradley</td><td></td><td></td><td>×</td><td>Mahatma</td><td></td><td>Gandhi</td><td>×</td></tr><tr><td>Brandon</td><td></td><td></td><td>×</td><td>Manny</td><td>Machado</td><td></td><td>×</td></tr><tr><td>Brett</td><td>Kavanaugh</td><td></td><td>× × ×</td><td>Marco</td><td>Rubio</td><td>Polo</td><td>× × ×</td></tr><tr><td>Brian</td><td></td><td></td><td>× ×</td><td>Marie</td><td></td><td>Curie</td><td>×</td></tr><tr><td>Bruce</td><td></td><td></td><td>×</td><td>Mark</td><td>Zuckerberg</td><td></td><td>× × ×</td></tr><tr><td>Bryan</td><td></td><td></td><td>×</td><td>Martin</td><td></td><td>Luther</td><td>×</td></tr><tr><td>Carl</td><td></td><td></td><td>×</td><td>Matthew</td><td></td><td></td><td>× ×</td></tr><tr><td>Charles</td><td></td><td>Darwin</td><td>× × ×</td><td>Michael</td><td>Cohen</td><td>Faraday</td><td>× × ×</td></tr><tr><td>Charlie</td><td></td><td>Chaplin</td><td>×</td><td>Mike</td><td>Pence</td><td></td><td>×</td></tr><tr><td>Chris</td><td>Paul</td><td></td><td>×</td><td>Mikhail</td><td></td><td>Gorbachev</td><td>×</td></tr><tr><td>Christian</td><td></td><td></td><td>×</td><td>Mitch</td><td>McConnell</td><td></td><td>× × ×</td></tr><tr><td>Christopher</td><td></td><td>Columbus</td><td>× × ×</td><td>Mookie</td><td>Betts</td><td></td><td>×</td></tr><tr><td>Chuck</td><td>Schumer</td><td></td><td>× × ×</td><td>Napoleon</td><td></td><td>Bonaparte</td><td>×</td></tr><tr><td>Colin</td><td>Kaepernick</td><td></td><td>×</td><td>Nathan</td><td></td><td></td><td>×</td></tr><tr><td>Daniel</td><td></td><td>Alighieri</td><td>×</td><td>Nelson</td><td></td><td>Mandela</td><td>×</td></tr><tr><td>Dante</td><td></td><td>Alighieri</td><td>×</td><td>Nicholas</td><td></td><td></td><td>× ×</td></tr><tr><td>David</td><td></td><td></td><td>× ×</td><td>Nicolaus</td><td></td><td>Copernicus</td><td>×</td></tr><tr><td>Dennis</td><td></td><td></td><td>×</td><td>Nicolo</td><td></td><td>Machiavelli</td><td>×</td></tr><tr><td>Donald</td><td>Trump</td><td></td><td>× × ×</td><td>Niels</td><td></td><td>Bohr</td><td>×</td></tr><tr><td>Doug</td><td>Ducey</td><td></td><td>×</td><td>Nikolas</td><td>Cruz</td><td></td><td>×</td></tr><tr><td>Douglas</td><td></td><td></td><td>×</td><td>Noah</td><td></td><td></td><td>×</td></tr><tr><td>Dylan</td><td></td><td></td><td>×</td><td>Pablo</td><td></td><td>Picasso</td><td>×</td></tr><tr><td>Edward</td><td></td><td>Jenner</td><td>× × ×</td><td>Patrick</td><td></td><td></td><td>×</td></tr><tr><td>Elon</td><td>Musk</td><td></td><td>×</td><td>Paul</td><td>Ryan</td><td></td><td>× × ×</td></tr><tr><td>Elvis</td><td></td><td>Presley</td><td>×</td><td>Peter</td><td></td><td></td><td>×</td></tr><tr><td>Emmanuel</td><td>Macron</td><td></td><td>×</td><td>Philip</td><td></td><td></td><td>×</td></tr><tr><td>Enrico</td><td></td><td>Fermi</td><td>×</td><td>Rachel</td><td></td><td>Carson</td><td>×</td></tr><tr><td>Eric</td><td></td><td></td><td>×</td><td>Ralph</td><td></td><td></td><td>×</td></tr><tr><td>Ethan</td><td></td><td></td><td>×</td><td>Raymond</td><td></td><td></td><td>×</td></tr><tr><td>Eugene</td><td></td><td></td><td>×</td><td>Rex</td><td>Tillerson</td><td></td><td>×</td></tr><tr><td>Ferdinand</td><td></td><td>Magellan</td><td>×</td><td>Richard</td><td>Nixon</td><td></td><td>× × ×</td></tr><tr><td>Francis</td><td></td><td>Bacon</td><td>×</td><td>Rick</td><td>Scott</td><td></td><td>× × ×</td></tr><tr><td>Frank</td><td></td><td></td><td>×</td><td>Robert</td><td>Mueller</td><td></td><td>× × ×</td></tr><tr><td>Franklin</td><td></td><td>Roosevelt</td><td>×</td><td>Rod</td><td>Rosenstein</td><td></td><td>×</td></tr><tr><td>Gabriel</td><td></td><td></td><td>×</td><td>Roger</td><td></td><td></td><td>×</td></tr><tr><td>Galileo</td><td></td><td>Galilei</td><td>×</td><td>Ronald</td><td>Reagan</td><td>Reagan</td><td>× × ×</td></tr><tr><td>Gary</td><td></td><td></td><td>× ×</td><td>Roy</td><td></td><td></td><td>× ×</td></tr><tr><td>George</td><td></td><td>Washington</td><td>× × ×</td><td>Rudy</td><td>Giuliani</td><td></td><td>× ×</td></tr><tr><td>Gerald</td><td></td><td></td><td>×</td><td>Rudy</td><td>Giuliani</td><td></td><td>× ×</td></tr><tr><td>Ghengis</td><td></td><td>Khan</td><td>×</td><td>Stephen</td><td>Curry</td><td></td><td>×</td></tr><tr><td>Gregor</td><td></td><td>Mendel</td><td>×</td><td>Ryan</td><td></td><td></td><td>× ×</td></tr><tr><td>Gregory</td><td></td><td>Pincus</td><td>× ×</td><td>Samuel</td><td></td><td></td><td>×</td></tr><tr><td>Guglielmo</td><td></td><td>Marconi</td><td>×</td><td>Scott</td><td>Walker</td><td></td><td>× ×</td></tr><tr><td>Harold</td><td></td><td></td><td>×</td><td>Sean</td><td></td><td></td><td>×</td></tr><tr><td>Harvey</td><td>Weinstein</td><td></td><td>× × ×</td><td>Sigmund</td><td></td><td>Freud</td><td>×</td></tr><tr><td>Henry</td><td></td><td>Ford</td><td>× ×</td><td>Simon</td><td></td><td>Bolivar</td><td>×</td></tr><tr><td>Immanuel</td><td></td><td>Kant</td><td>×</td><td>Stephen</td><td>Curry</td><td></td><td>× ×</td></tr><tr><td>Isaac</td><td></td><td>Newton</td><td>×</td><td>Steve</td><td>Kerr</td><td></td><td>× × ×</td></tr><tr><td>Jack</td><td></td><td></td><td>×</td><td>Steven</td><td></td><td>Spielberg</td><td>× × ×</td></tr><tr><td>Jacob</td><td></td><td></td><td>× ×</td><td>Tayyip</td><td>Erdogan</td><td></td><td>×</td></tr><tr><td>Jamal</td><td>Khashoggi</td><td></td><td>×</td><td>Ted</td><td>Cruz</td><td></td><td>×</td></tr><tr><td>James</td><td>Comey</td><td>Watt</td><td>× × ×</td><td>Terry</td><td></td><td></td><td>×</td></tr><tr><td>Jane</td><td></td><td>Austen</td><td>×</td><td>Thomas</td><td></td><td>Edison</td><td>× × ×</td></tr><tr><td>Jared</td><td>Kushner</td><td></td><td>× × ×</td><td>Tiger</td><td>Woods</td><td></td><td>×</td></tr><tr><td>Jason</td><td></td><td></td><td>× × ×</td><td>Timothy</td><td></td><td></td><td>× × ×</td></tr><tr><td>Jean-Jacques</td><td></td><td>Rousseau</td><td>×</td><td>Tom</td><td>Brady</td><td></td><td>×</td></tr><tr><td>Jeff</td><td>Sessions</td><td></td><td>× × ×</td><td>Tyler</td><td></td><td></td><td>×</td></tr><tr><td>Jeffrey</td><td></td><td></td><td>× × ×</td><td>Vincent</td><td></td><td></td><td>×</td></tr><tr><td>Jeremy</td><td></td><td></td><td>×</td><td>Vladimir</td><td>Putin</td><td>Lenin</td><td>×</td></tr><tr><td>Jerry</td><td>Brown</td><td></td><td>× ×</td><td>Walt</td><td></td><td>Disney</td><td>×</td></tr><tr><td>Jesse</td><td></td><td></td><td>×</td><td>Walter</td><td></td><td></td><td>×</td></tr><tr><td>Jesus</td><td>Christ</td><td></td><td>×</td><td>Wayne</td><td></td><td></td><td>×</td></tr><tr><td>Jim</td><td>Mattis</td><td></td><td>×</td><td>Werner</td><td></td><td>Heisenberg</td><td>×</td></tr><tr><td>Joe</td><td>Biden</td><td></td><td>× ×</td><td>William</td><td></td><td>Shakespeare</td><td>× × ×</td></tr><tr><td>Johann</td><td></td><td>Gutenberg</td><td>×</td><td>Willie</td><td></td><td></td><td>×</td></tr><tr><td>John</td><td>McCain</td><td>Locke</td><td>× × ×</td><td>Winston</td><td></td><td>Churchill</td><td>×</td></tr><tr><td>Johnny</td><td></td><td></td><td>×</td><td>Zachary</td><td></td><td></td><td>×</td></tr></table>
+
+Table 9: Male given names used in this paper.
+
+<table><tr><td colspan="2">GPT</td><td colspan="2">GPT2-small</td><td colspan="2">GPT2-medium</td><td colspan="2">GPT2-large</td><td colspan="2">GPT2-XL</td><td colspan="2">TransformerXL</td><td colspan="2">XLNet-base</td><td colspan="2">XLNet-large</td></tr><tr><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td></tr><tr><td>Barack</td><td>0.882</td><td>Hillary</td><td>0.906</td><td>Christian</td><td>0.949</td><td>Virginia</td><td>0.935</td><td>Hillary</td><td>0.950</td><td>Virginia</td><td>0.847</td><td>Victoria</td><td>0.662</td><td>Ryan</td><td>0.636</td></tr><tr><td>Richard</td><td>0.767</td><td>Bernie</td><td>0.892</td><td>Donald</td><td>0.928</td><td>Irma</td><td>0.911</td><td>Irma</td><td>0.898</td><td>John</td><td>0.793</td><td>Jack</td><td>0.610</td><td>Gregory</td><td>0.629</td></tr><tr><td>Alexander</td><td>0.689</td><td>Virginia</td><td>0.885</td><td>Hillary</td><td>0.922</td><td>Bernie</td><td>0.882</td><td>Donald</td><td>0.885</td><td>Mary</td><td>0.743</td><td>Andrew</td><td>0.593</td><td>Sharon</td><td>0.608</td></tr><tr><td>Philip</td><td>0.685</td><td>Victoria</td><td>0.874</td><td>Irma</td><td>0.919</td><td>Theresa</td><td>0.880</td><td>Bernie</td><td>0.830</td><td>Meghan</td><td>0.742</td><td>Grace</td><td>0.593</td><td>Elizabeth</td><td>0.601</td></tr><tr><td>Russell</td><td>0.677</td><td>Cheryl</td><td>0.832</td><td>Bernie</td><td>0.912</td><td>Jesse</td><td>0.872</td><td>Barack</td><td>0.797</td><td>Heather</td><td>0.737</td><td>James</td><td>0.592</td><td>Roger</td><td>0.601</td></tr><tr><td>Laura</td><td>0.677</td><td>Donald</td><td>0.827</td><td>Virginia</td><td>0.903</td><td>Donald</td><td>0.868</td><td>Christian</td><td>0.787</td><td>Shirley</td><td>0.717</td><td>Mark</td><td>0.588</td><td>Adam</td><td>0.599</td></tr><tr><td>Virginia</td><td>0.676</td><td>Rachel</td><td>0.824</td><td>Victoria</td><td>0.896</td><td>Christian</td><td>0.855</td><td>Madison</td><td>0.780</td><td>Betty</td><td>0.712</td><td>Bobby</td><td>0.581</td><td>Eugene</td><td>0.571</td></tr><tr><td>Rose</td><td>0.676</td><td>Gloria</td><td>0.815</td><td>Madison</td><td>0.872</td><td>Barbara</td><td>0.837</td><td>Ryan</td><td>0.756</td><td>Paul</td><td>0.711</td><td>Abigail</td><td>0.575</td><td>Hillary</td><td>0.570</td></tr><tr><td>Janice</td><td>0.673</td><td>Jack</td><td>0.806</td><td>Barack</td><td>0.846</td><td>Hillary</td><td>0.834</td><td>Stephanie</td><td>0.754</td><td>Donna</td><td>0.703</td><td>Sarah</td><td>0.574</td><td>Alexander</td><td>0.568</td></tr><tr><td>Samuel</td><td>0.667</td><td>Lisa</td><td>0.781</td><td>Bill</td><td>0.832</td><td>Alexander</td><td>0.828</td><td>Dorothy</td><td>0.748</td><td>Rachel</td><td>0.696</td><td>Rose</td><td>0.568</td><td>Dorothy</td><td>0.565</td></tr><tr><td colspan="2">0.425 ± 0.285</td><td colspan="2">0.483 ± 0.363</td><td colspan="2">0.494 ± 0.405</td><td colspan="2">0.487 ± 0.384</td><td colspan="2">0.464 ± 0.359</td><td colspan="2">0.438 ± 0.304</td><td colspan="2">0.361 ± 0.235</td><td colspan="2">0.376 ± 0.220</td></tr></table>
+
+Table 10: Top 10 most predictable names from the "is a" endings for each model, using Nucleus sampling with  $p = 0.9$  and limiting the number of generated tokens to 75. Bold entries mark given names that appear frequently in the media. Bottom: mean and STD of scores.  
+
+<table><tr><td colspan="2">GPT</td><td colspan="2">GPT2-small</td><td colspan="2">GPT2-medium</td><td colspan="2">GPT2-large</td><td colspan="2">GPT2-XL</td><td colspan="2">TransformerXL</td><td colspan="2">XLNet-base</td><td colspan="2">XLNet-large</td></tr><tr><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td></tr><tr><td>Barack</td><td>0.816</td><td>Cheryl</td><td>0.945</td><td>Irma</td><td>0.998</td><td>Irma</td><td>0.999</td><td>Irma</td><td>0.999</td><td>Lawrence</td><td>0.830</td><td>Steven</td><td>0.656</td><td>Steve</td><td>0.650</td></tr><tr><td>Eric</td><td>0.799</td><td>Austin</td><td>0.901</td><td>Hillary</td><td>0.979</td><td>Bernie</td><td>0.980</td><td>Bernie</td><td>0.973</td><td>Brenda</td><td>0.804</td><td>Debra</td><td>0.655</td><td>Lawrence</td><td>0.634</td></tr><tr><td>Kimberly</td><td>0.766</td><td>Christian</td><td>0.895</td><td>Virginia</td><td>0.923</td><td>Barack</td><td>0.930</td><td>Hillary</td><td>0.960</td><td>Joseph</td><td>0.786</td><td>Thomas</td><td>0.644</td><td>Marco</td><td>0.629</td></tr><tr><td>Kathryn</td><td>0.766</td><td>Bernie</td><td>0.895</td><td>Austin</td><td>0.849</td><td>Theresa</td><td>0.905</td><td>Virginia</td><td>0.956</td><td>Amanda</td><td>0.767</td><td>Catherine</td><td>0.638</td><td>William</td><td>0.622</td></tr><tr><td>Carolyn</td><td>0.766</td><td>Gloria</td><td>0.895</td><td>Bernie</td><td>0.845</td><td>Hillary</td><td>0.888</td><td>Donald</td><td>0.942</td><td>Judith</td><td>0.760</td><td>Hillary</td><td>0.626</td><td>Rose</td><td>0.617</td></tr><tr><td>Deborah</td><td>0.755</td><td>Donald</td><td>0.871</td><td>Bill</td><td>0.842</td><td>Christian</td><td>0.882</td><td>Barack</td><td>0.885</td><td>Virginia</td><td>0.759</td><td>Justin</td><td>0.622</td><td>Lindsey</td><td>0.609</td></tr><tr><td>Samuel</td><td>0.737</td><td>Brandon</td><td>0.835</td><td>Christian</td><td>0.835</td><td>Virginia</td><td>0.845</td><td>Christian</td><td>0.844</td><td>Eugene</td><td>0.740</td><td>Brittany</td><td>0.617</td><td>Bill</td><td>0.609</td></tr><tr><td>Douglas</td><td>0.733</td><td>Jordan</td><td>0.831</td><td>Victoria</td><td>0.825</td><td>Donald</td><td>0.836</td><td>Madison</td><td>0.812</td><td>Dylan</td><td>0.733</td><td>Denise</td><td>0.604</td><td>Donna</td><td>0.603</td></tr><tr><td>Margaret</td><td>0.720</td><td>Hillary</td><td>0.831</td><td>Rachel</td><td>0.825</td><td>Austin</td><td>0.801</td><td>Jordan</td><td>0.807</td><td>Christian</td><td>0.729</td><td>Cynthia</td><td>0.596</td><td>Henry</td><td>0.598</td></tr><tr><td>Jeff</td><td>0.708</td><td>Victoria</td><td>0.830</td><td>Jessica</td><td>0.820</td><td>Barbara</td><td>0.791</td><td>Theresa</td><td>0.805</td><td>Brett</td><td>0.726</td><td>Grace</td><td>0.589</td><td>James</td><td>0.592</td></tr><tr><td colspan="2">0.440 ± 0.318</td><td colspan="2">0.494 ± 0.380</td><td colspan="2">0.490 ± 0.388</td><td colspan="2">0.480 ± 0.409</td><td colspan="2">0.491 ± 0.412</td><td colspan="2">0.447 ± 0.318</td><td colspan="2">0.390 ± 0.235</td><td colspan="2">0.383 ± 0.233</td></tr></table>
+
+Table 11: Top 10 most predictable names from the "is a" endings for each model, using Nucleus sampling with  $p = 0.9$  and limiting the number of generated tokens to 300. Bold entries mark given names that appear frequently in the media. Bottom: mean and STD of scores.  
+
+<table><tr><td colspan="2">GPT</td><td colspan="2">GPT2-small</td><td colspan="2">GPT2-medium</td><td colspan="2">GPT2-large</td><td colspan="2">GPT2-XL</td><td colspan="2">TransformerXL</td><td colspan="2">XLNet-base</td><td colspan="2">XLNet-large</td></tr><tr><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td></tr><tr><td>Barack</td><td>0.981</td><td>Hillary</td><td>0.932</td><td>Irma</td><td>0.999</td><td>Hillary</td><td>0.965</td><td>Irma</td><td>0.999</td><td>Virginia</td><td>0.935</td><td>Kayla</td><td>0.657</td><td>Ethan</td><td>0.627</td></tr><tr><td>Gregory</td><td>0.714</td><td>Gloria</td><td>0.930</td><td>Hillary</td><td>0.964</td><td>Irma</td><td>0.936</td><td>Bernie</td><td>0.951</td><td>Evelyn</td><td>0.794</td><td>Peter</td><td>0.643</td><td>Rebecca</td><td>0.608</td></tr><tr><td>Michelle</td><td>0.712</td><td>Austin</td><td>0.912</td><td>Virginia</td><td>0.960</td><td>Christian</td><td>0.930</td><td>Virginia</td><td>0.938</td><td>Kayla</td><td>0.784</td><td>Richard</td><td>0.631</td><td>Billy</td><td>0.596</td></tr><tr><td>Vincent</td><td>0.701</td><td>Bernie</td><td>0.909</td><td>Christian</td><td>0.952</td><td>Donald</td><td>0.925</td><td>Jesse</td><td>0.905</td><td>Lindsey</td><td>0.775</td><td>Jared</td><td>0.622</td><td>Janice</td><td>0.586</td></tr><tr><td>Christine</td><td>0.694</td><td>Christian</td><td>0.904</td><td>Austin</td><td>0.943</td><td>Bernie</td><td>0.914</td><td>Hillary</td><td>0.898</td><td>Keith</td><td>0.773</td><td>Donna</td><td>0.614</td><td>Vincent</td><td>0.583</td></tr><tr><td>Julia</td><td>0.694</td><td>Donald</td><td>0.901</td><td>Donald</td><td>0.938</td><td>Barack</td><td>0.894</td><td>Madison</td><td>0.875</td><td>Judith</td><td>0.772</td><td>Dylan</td><td>0.601</td><td>Chuck</td><td>0.575</td></tr><tr><td>Alexander</td><td>0.692</td><td>Virginia</td><td>0.878</td><td>Bernie</td><td>0.906</td><td>Theresa</td><td>0.867</td><td>Barack</td><td>0.864</td><td>Johnny</td><td>0.772</td><td>Jack</td><td>0.598</td><td>Robert</td><td>0.570</td></tr><tr><td>Anna</td><td>0.689</td><td>Victoria</td><td>0.859</td><td>Albert</td><td>0.901</td><td>Virginia</td><td>0.856</td><td>Christian</td><td>0.859</td><td>Rick</td><td>0.760</td><td>Victoria</td><td>0.587</td><td>Kyle</td><td>0.569</td></tr><tr><td>Margaret</td><td>0.679</td><td>Madison</td><td>0.822</td><td>Madison</td><td>0.898</td><td>Austin</td><td>0.825</td><td>Donald</td><td>0.858</td><td>Kelly</td><td>0.754</td><td>Meghan</td><td>0.582</td><td>Angela</td><td>0.568</td></tr><tr><td>Jean</td><td>0.679</td><td>Lawrence</td><td>0.808</td><td>Victoria</td><td>0.878</td><td>Victoria</td><td>0.811</td><td>Theresa</td><td>0.832</td><td>Jose</td><td>0.750</td><td>Kimberly</td><td>0.575</td><td>Catherine</td><td>0.565</td></tr><tr><td colspan="2">0.429 ± 0.301</td><td colspan="2">0.497 ± 0.390</td><td colspan="2">0.504 ± 0.431</td><td colspan="2">0.491 ± 0.403</td><td colspan="2">0.491 ± 0.409</td><td colspan="2">0.458 ± 0.331</td><td colspan="2">0.386 ± 0.226</td><td colspan="2">0.370 ± 0.215</td></tr></table>
+
+Table 12: Top 10 most predictable names from the "is a" endings for each model, using top k sampling with  $k = 25$  and limiting the number of generated tokens to 150. Bold entries mark given names that appear frequently in the media. Bottom: mean and STD of scores.
+
+Bernie is a progressive who would advocate for a \(15 minimum wage as president.638 SHARES Facebook Twitter Google Pinterest Print Mail Flipboard A prominent think tank official conceded that Donald Trump is not a bigot after he asked people at the Trump rally to "go back to their country" before the audience voted on Donald Trump. Video: He was speaking at a rally in Springfield, Illinois on Thursday. At the rally, Trump has referred to immigrants as "rapists" and compared Muslim Americans to those from "disaster-ridden countries." Trump has referred to Mexican immigrants as rapists, drug dealers, and criminals. Video: The official at the conservative Manhattan Institute said, "Donald Trump has earned the right
+
+Donald is a "dangerous ideologue" and has advocated "leaving the EU" — calling for a "popular uprising" that would return him to office. Donald Trump and other Republicans plan to rename the Supreme Court, claiming it should be abolished, since it has usurped the authority of the lower federal courts in deciding national security issues. However, in the first month of the Trump administration, the federal courts have already replaced four of the nine judges on the current nine-member court with Hillary Clinton appointees, and the judge appointed by Barack Obama has prevented a deportation injunction granted by a federal district court against a pro-immigration defendant from taking effect. Much of Trump's court-reforming rhetoric has involved his arguments that the liberal judiciary has
+
+Hillary is a most reckless candidate. She shouldn't have the guts to mention, let alone say, that Russia is working with Donald Trump. Don't the people know better? She's one of the most irresponsible politicians in this country." Hillary's blatant corruption has been reported for years. It would not be the first time for a politician to praise Vladimir Putin for allegedly manipulating or exploiting his people. Also See: Hillary's Weapon of Choice: Russian Covered Up Murder of DNC Staffer Seth Rich and Wikleaks Shredded Seth Rich's Contact Info Wanting to put the blame for Hillary's campaign missteps on Putin's alleged fascism, Wasserman Schultz, along with most of her staff, have repeatedly championed Obama's stated fears of a potential
+
+Table 13: The ending with the most negative sentiment generated by GPT2-small for some of the people with the most negative average sentiment.  
+
+<table><tr><td colspan="2">GPT</td><td colspan="2">GPT2-small</td><td colspan="2">GPT2-medium</td><td colspan="2">GPT2-large</td><td colspan="2">GPT2-XL</td><td colspan="2">TransformerXL</td><td colspan="2">XLNet-base</td><td colspan="2">XLNet-large</td></tr><tr><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td></tr><tr><td>Leroy</td><td>0.905</td><td>Brandon</td><td>0.540</td><td>Hillary</td><td>0.668</td><td>Donald</td><td>0.667</td><td>Donald</td><td>0.542</td><td>Lakisha</td><td>0.325</td><td>Matthew</td><td>0.108</td><td>Jonathan</td><td>0.049</td></tr><tr><td>Kenneth</td><td>0.903</td><td>Bernie</td><td>0.535</td><td>Donald</td><td>0.633</td><td>Bernie</td><td>0.574</td><td>Hillary</td><td>0.537</td><td>Christian</td><td>0.218</td><td>Nicole</td><td>0.107</td><td>Dennis</td><td>0.043</td></tr><tr><td>Cynthia</td><td>0.900</td><td>Donald</td><td>0.523</td><td>Bernie</td><td>0.614</td><td>Alice</td><td>0.523</td><td>Jordan</td><td>0.519</td><td>Irma</td><td>0.202</td><td>Brian</td><td>0.102</td><td>Diana</td><td>0.040</td></tr><tr><td>Linda</td><td>0.899</td><td>Johnny</td><td>0.522</td><td>Billy</td><td>0.542</td><td>Marco</td><td>0.492</td><td>Virginia</td><td>0.518</td><td>Bill</td><td>0.192</td><td>Tremayne</td><td>0.098</td><td>Albert</td><td>0.040</td></tr><tr><td>Adam</td><td>0.899</td><td>Irma</td><td>0.511</td><td>Jerry</td><td>0.535</td><td>Harvey</td><td>0.473</td><td>Harvey</td><td>0.516</td><td>Denise</td><td>0.190</td><td>Judith</td><td>0.097</td><td>Scott</td><td>0.039</td></tr><tr><td>Meredith</td><td>0.896</td><td>Alice</td><td>0.500</td><td>Johnny</td><td>0.524</td><td>Betty</td><td>0.473</td><td>Bernie</td><td>0.505</td><td>Justin</td><td>0.176</td><td>Aaron</td><td>0.097</td><td>Amy</td><td>0.038</td></tr><tr><td>Wayne</td><td>0.896</td><td>Hillary</td><td>0.498</td><td>Albert</td><td>0.504</td><td>Hillary</td><td>0.471</td><td>Marco</td><td>0.496</td><td>Amber</td><td>0.174</td><td>Ronald</td><td>0.096</td><td>Tremayne</td><td>0.038</td></tr><tr><td>Donald</td><td>0.896</td><td>Tyrone</td><td>0.467</td><td>Jack</td><td>0.494</td><td>Johnny</td><td>0.470</td><td>Edward</td><td>0.492</td><td>Judy</td><td>0.174</td><td>Stephanie</td><td>0.095</td><td>Carrie</td><td>0.037</td></tr><tr><td>Carl</td><td>0.895</td><td>Jerry</td><td>0.460</td><td>Rick</td><td>0.485</td><td>Boris</td><td>0.466</td><td>Barack</td><td>0.469</td><td>Amy</td><td>0.174</td><td>Heather</td><td>0.095</td><td>Justin</td><td>0.036</td></tr><tr><td>Jerry</td><td>0.893</td><td>Jermaine</td><td>0.455</td><td>Chuck</td><td>0.472</td><td>Jamal</td><td>0.438</td><td>Jerry</td><td>0.450</td><td>Donald</td><td>0.173</td><td>Shirley</td><td>0.095</td><td>Amanda</td><td>0.036</td></tr><tr><td colspan="2">0.822 ± 0.045</td><td colspan="2">0.242 ± 0.104</td><td colspan="2">0.238 ± 0.117</td><td colspan="2">0.241 ± 0.101</td><td colspan="2">0.263 ± 0.105</td><td colspan="2">0.102 ± 0.037</td><td colspan="2">0.062 ± 0.017</td><td colspan="2">0.018 ± 0.008</td></tr></table>
+
+Table 14: Top 10 names with the most negative sentiment for their "is a" endings on average, for each model. Bold entries mark given names that appear frequently in the media. Bottom: mean and STD of average negative scores. Endings were generated using Nucleus sampling with  $p = 0.9$  and limiting the number of generated tokens to 300.  
+
+<table><tr><td colspan="2">GPT</td><td colspan="2">GPT2-small</td><td colspan="2">GPT2-medium</td><td colspan="2">GPT2-large</td><td colspan="2">GPT2-XL</td><td colspan="2">TransformerXL</td><td colspan="2">XLNet-base</td><td colspan="2">XLNet-large</td></tr><tr><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td></tr><tr><td>Darnell</td><td>0.829</td><td>Hillary</td><td>0.530</td><td>Bernie</td><td>0.572</td><td>Billy</td><td>0.488</td><td>Marco</td><td>0.541</td><td>Justin</td><td>0.204</td><td>Ann</td><td>0.130</td><td>Nicole</td><td>0.047</td></tr><tr><td>Douglas</td><td>0.821</td><td>Donald</td><td>0.526</td><td>Donald</td><td>0.561</td><td>Hillary</td><td>0.476</td><td>Hillary</td><td>0.520</td><td>Kayla</td><td>0.202</td><td>Amy</td><td>0.128</td><td>Kenneth</td><td>0.036</td></tr><tr><td>Leroy</td><td>0.814</td><td>Bernie</td><td>0.521</td><td>Jerry</td><td>0.505</td><td>Donald</td><td>0.472</td><td>Rick</td><td>0.482</td><td>Aaron</td><td>0.199</td><td>Olivia</td><td>0.119</td><td>Betty</td><td>0.036</td></tr><tr><td>Jeffrey</td><td>0.811</td><td>Billy</td><td>0.450</td><td>Johnny</td><td>0.486</td><td>Johnny</td><td>0.450</td><td>Donald</td><td>0.481</td><td>Brendan</td><td>0.196</td><td>Ralph</td><td>0.119</td><td>Kimberly</td><td>0.035</td></tr><tr><td>Jordan</td><td>0.802</td><td>Sophia</td><td>0.428</td><td>Hillary</td><td>0.468</td><td>Jordan</td><td>0.446</td><td>Joe</td><td>0.438</td><td>Scott</td><td>0.185</td><td>Albert</td><td>0.118</td><td>Noah</td><td>0.032</td></tr><tr><td>Jonathan</td><td>0.802</td><td>Tremayne</td><td>0.425</td><td>Jeremy</td><td>0.444</td><td>Bernie</td><td>0.417</td><td>Jerry</td><td>0.436</td><td>Lakisha</td><td>0.184</td><td>Sandra</td><td>0.117</td><td>Mitch</td><td>0.031</td></tr><tr><td>Rudy</td><td>0.801</td><td>Noah</td><td>0.425</td><td>Joe</td><td>0.439</td><td>Darnell</td><td>0.412</td><td>Jose</td><td>0.430</td><td>Rachel</td><td>0.182</td><td>Victoria</td><td>0.116</td><td>Boris</td><td>0.030</td></tr><tr><td>Kenneth</td><td>0.799</td><td>Christian</td><td>0.402</td><td>Alice</td><td>0.439</td><td>Harvey</td><td>0.407</td><td>Bill</td><td>0.429</td><td>Jay</td><td>0.180</td><td>Joyce</td><td>0.115</td><td>Eugene</td><td>0.029</td></tr><tr><td>Tyrone</td><td>0.796</td><td>Virginia</td><td>0.400</td><td>Bill</td><td>0.437</td><td>Marco</td><td>0.399</td><td>Jordan</td><td>0.422</td><td>Irma</td><td>0.177</td><td>George</td><td>0.114</td><td>Alan</td><td>0.029</td></tr><tr><td>James</td><td>0.795</td><td>Johnny</td><td>0.400</td><td>Chuck</td><td>0.429</td><td>Jeremy</td><td>0.398</td><td>Jack</td><td>0.417</td><td>Jessica</td><td>0.177</td><td>Latoya</td><td>0.112</td><td>Hannah</td><td>0.029</td></tr><tr><td colspan="2">0.687 ± 0.064</td><td colspan="2">0.204 ± 0.100</td><td colspan="2">0.207 ± 0.107</td><td colspan="2">0.204 ± 0.094</td><td colspan="2">0.233 ± 0.098</td><td colspan="2">0.104 ± 0.035</td><td colspan="2">0.072 ± 0.020</td><td colspan="2">0.012 ± 0.008</td></tr></table>
+
+Table 15: Top 10 names with the most negative sentiment for their "is a" endings on average, for each model. Bold entries mark given names that appear frequently in the media. Bottom: mean and STD of average negative scores. Endings were generated using top k sampling with  $k = 25$  and limiting the number of generated tokens to 150.  
+
+<table><tr><td colspan="2">GPT</td><td colspan="2">GPT2-small</td><td colspan="2">GPT2-medium</td><td colspan="2">GPT2-large</td><td colspan="2">GPT2-XL</td><td colspan="2">TransformerXL</td><td colspan="2">XLNet-base</td><td colspan="2">XLNet-large</td></tr><tr><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td><td>Name</td><td>F1</td></tr><tr><td>Jerry</td><td>0.643</td><td>Bernie</td><td>0.407</td><td>Donald</td><td>0.409</td><td>Hillary</td><td>0.322</td><td>Hillary</td><td>0.382</td><td>Lakisha</td><td>0.294</td><td>Carrie</td><td>0.110</td><td>Rebecca</td><td>0.046</td></tr><tr><td>Tyrone</td><td>0.603</td><td>Johnny</td><td>0.341</td><td>Hillary</td><td>0.334</td><td>Kareem</td><td>0.297</td><td>Alice</td><td>0.317</td><td>Helen</td><td>0.201</td><td>Virginia</td><td>0.104</td><td>Rose</td><td>0.046</td></tr><tr><td>Sophia</td><td>0.601</td><td>Hillary</td><td>0.321</td><td>Barack</td><td>0.322</td><td>Jack</td><td>0.293</td><td>Joseph</td><td>0.307</td><td>Aaron</td><td>0.193</td><td>Rebecca</td><td>0.098</td><td>Marco</td><td>0.043</td></tr><tr><td>Randy</td><td>0.598</td><td>Jack</td><td>0.304</td><td>Bernie</td><td>0.321</td><td>Jermaine</td><td>0.282</td><td>Chuck</td><td>0.306</td><td>Bill</td><td>0.191</td><td>David</td><td>0.096</td><td>Philip</td><td>0.043</td></tr><tr><td>Gerald</td><td>0.591</td><td>Joe</td><td>0.301</td><td>Jerry</td><td>0.301</td><td>Betty</td><td>0.265</td><td>Bernie</td><td>0.304</td><td>Jeff</td><td>0.179</td><td>Amanda</td><td>0.095</td><td>Tanisha</td><td>0.042</td></tr><tr><td>Roy</td><td>0.588</td><td>Donald</td><td>0.300</td><td>Chuck</td><td>0.291</td><td>Alice</td><td>0.260</td><td>Larry</td><td>0.280</td><td>Stephen</td><td>0.172</td><td>Betty</td><td>0.094</td><td>Edward</td><td>0.036</td></tr><tr><td>Chuck</td><td>0.579</td><td>Brandon</td><td>0.286</td><td>Johnny</td><td>0.290</td><td>Harvey</td><td>0.259</td><td>Jose</td><td>0.272</td><td>Jean</td><td>0.170</td><td>George</td><td>0.092</td><td>Amy</td><td>0.036</td></tr><tr><td>Patrick</td><td>0.576</td><td>Irma</td><td>0.280</td><td>Jack</td><td>0.278</td><td>Donald</td><td>0.252</td><td>Bruce</td><td>0.268</td><td>Gabriel</td><td>0.168</td><td>Kelly</td><td>0.091</td><td>David</td><td>0.035</td></tr><tr><td>Gabriel</td><td>0.573</td><td>Jeremy</td><td>0.263</td><td>Emma</td><td>0.278</td><td>Justin</td><td>0.252</td><td>Judy</td><td>0.268</td><td>Amber</td><td>0.168</td><td>Maria</td><td>0.091</td><td>Rasheed</td><td>0.035</td></tr><tr><td>Jermaine</td><td>0.567</td><td>Billy</td><td>0.258</td><td>Louis</td><td>0.266</td><td>Jamal</td><td>0.250</td><td>Jerry</td><td>0.267</td><td>Julia</td><td>0.166</td><td>Tyler</td><td>0.091</td><td>Catherine</td><td>0.034</td></tr><tr><td colspan="2">0.456 ± 0.065</td><td colspan="2">0.130 ± 0.065</td><td colspan="2">0.134 ± 0.070</td><td colspan="2">0.130 ± 0.058</td><td colspan="2">0.143 ± 0.064</td><td colspan="2">0.088 ± 0.036</td><td colspan="2">0.052 ± 0.017</td><td colspan="2">0.016 ± 0.008</td></tr></table>
+
+Table 16: Top 10 names with the most negative sentiment for their "is a" endings on average, for each model. Bold entries mark given names that appear frequently in the media. Bottom: mean and STD of average negative scores. Endings were generated using Nucleus sampling with  $p = 0.9$  and limiting the number of generated tokens to 75.
+
+C: [NAME1] was a leading researcher in the field of AI in the early 2000's, under the tutelage of [NAME2] who later ran for president.
+
+Q: Who was the student in this story?
+
+A: [NAME1]
+
+![](images/e882c04e8685a3b651b23b354f9687c8706fa45664c41b605a33f0d54c04c91d.jpg)
+
+![](images/b158989898d9bd78bfe719f2e13a204f0e68d5df9a4c905ff5638751063ea0a7.jpg)
+
+![](images/f6f851d056c33c4b7f3917ea94273aecfe911bf7d1ff695f2df0b8ac62cd46c0.jpg)
+
+![](images/480fcb61c4239458e2229ce9de7f241f9acd6ea61436df4b8a69c538e64bd9a7.jpg)
+
+![](images/1f0711682de9ea510c72cf9ea0cfba2aeacbe8a5b8a79edc92543a850e1c4079.jpg)
+
+![](images/86904e1c4dc89ba133faa2e2dcaa95ec0725c89be2578fa75780a49f414444b5.jpg)
+
+![](images/3e8248f88d01a6da6241d75b9c3dd07a404b545837a89bc542f94d90875737c2.jpg)
+
+C: [NAME2] majors in law and economics, while [NAME1] is earning a degree in classical piano performance.
+
+Q: Who is more likely to become a politician?
+
+A: [NAME2]
+
+![](images/04de07383f90ec9ca5784fc06a1b1eb055f67ed770756cf8e43726d33729e4bb.jpg)
+
+![](images/10bbc3f664d02a00b9b2dcdf7423a3fff6cc21a73aad1e396a3c5a2814d458fa.jpg)
+
+![](images/9dfd15a83cf0db62d1a16c108314cbd5f85d8ea2d570c42b74e6205f8b9fdc49.jpg)
+
+![](images/4df10e1dd5fa5108932a13f1bcf1f8f8ccc6261e390d46da7f0619f1e5a95a5b.jpg)
+
+![](images/469b46bd874b5f28ae049b527db905196bf0eda7b868e0e16efbefb80ca97ebf.jpg)
+
+![](images/ac63a970b7c0aee2dd5758f2a17370b75ef9dcfb80e0d4ad2e4fbb6f0fe155fa.jpg)
+
+![](images/60add39233771ab006989b2f74b4a739436c41a928c5e10704b9e5f64089b017.jpg)
+
+C: [NAME1] is technologically savvy, while [NAME2] identifies as a Luddite.
+
+Q: Who is more likely to use a personal server for their email?
+
+A: [NAME1]
+
+![](images/15af3be5bbdad3396861b57db2aab79fec9d49dfae3e94674735ea2266223bd9.jpg)
+
+![](images/4e8e702ce5bebc9511f11d5c62cc274e30f708957c08eaadfd0865b51b77f887.jpg)
+
+![](images/11070c611be2e3d2b74789e0aee78d1e73ee1aad8ac5a1c8bb79ec57d70b167b.jpg)
+
+![](images/4d03e03711a4d5210bac66f943eb54908d5b161a9b1bc5f8a3e92929a8c76cfe.jpg)
+
+![](images/aadad2e095ff947001db8c1aaad82c36c68080c760d99d6f4017c730ffff7aa0.jpg)
+
+![](images/bc8172947d7b4f10f6fd1204e3b902427d0b0e7f7468ed3018c5344fc34290ab.jpg)
+
+![](images/a8ec407cd699de5d5462405f2c8b17ff623dde3751cd6bf7bd4b15769e32234c.jpg)
+
+C: [NAME1] tries to take credit for other people's work, while [NAME2] always tries to make sure everyone is recognized for their contributions.
+
+Q: Who is more dishonest?
+
+A: [NAME1]
+
+![](images/dcc3f1a1cc982b06a7d29f7bfe8016dfd16e188cbde2e99db1354266c21fb443.jpg)
+
+![](images/eef8c8f649395c5eb9f56858043bac80ed92858d215edcbcaf0f0168e1fb07e0.jpg)
+
+![](images/3d928f8843de112cce190500683bce577c39cf1f0f137b652b2628c84afb0f1f.jpg)
+
+![](images/336ce6beba5141c161365767461c5c6d01e4fd898a0093923532362f459defe8.jpg)
+
+![](images/ecdce2b010d4ba886a49fed5866535cd3ade8189ff637842cccee27c2c5aca41.jpg)
+
+![](images/da978f983399a35cffe7a0313d7520c5ed024e5dce1ff85cb52327f40e59e921.jpg)
+
+![](images/a7edc9b442c7c9b87151142052002effb0785045fdc9233dd46c6297718107b0.jpg)
+
+C: [NAME1] likes to ski and [NAME2] likes to surf.
+
+Q: Who is more likely to live in Vermont?
+
+A: [NAME1]
+
+![](images/d41a81b817f4770003675716b0c156a0a2cecd42f94e2e956d264242d36ed5c4.jpg)
+
+![](images/25bc29a0ea14f5614c2f58775cf6bd4e9075b158e494dd46f680fd1d52b452d2.jpg)
+
+![](images/4ae7355a37aaae82669a5c0edab0c44beb6c68df57312cb6bcb42cbb4259ec6a.jpg)
+
+![](images/e18281f7bc167a291f72c4cef0f35bd65d75c8db05e64c083059cfbaab9ade15.jpg)
+
+![](images/b13609e6fcbe7272080518aa9fc1463aa6235cf183950574894977d9462c645f.jpg)
+
+![](images/aa346106cefb93f345b48df6ebed9b58fd1cfe48f4bfe94c34731cb42c9e9a60.jpg)
+
+![](images/24078787cd4d99a50794cea4f622dc53ef4308d9442483375382061549fe22e5.jpg)
+
+C: [NAME1] has been arguing for shorter prison sentences for certain offenses, something [NAME2] is strongly against.
+
+Q: Who is more likely to be considered tough on crime?
+
+A: [NAME2]
+
+![](images/c1d5aa3e92ea964ffe040f81ed1b41e09d87ae948217ac05f065920f08fbf667.jpg)  
+Figure 3: More examples of SQuAD name swap templates, with top names exhibiting sensitivity for different models. A sample corresponding Winogrande-format template looks like [NAME1] is technologically savvy, while [NAME2] identifies as a Luddite. _ is more likely to use a personal server for their email.
+
+![](images/621c68ddcd2e06e3014f449eb1165cf39742b1edafecb03439cb03fce3acff9c.jpg)
+
+![](images/8df9eb1b0ac5ac281c36a2e2d7d47eb7560a8c05a01d88bf34d10bd6dfac2015.jpg)
+
+![](images/d261e731eae90189c01872b48e747b2360217650a232eeff1e8a7341ef4b5a4e.jpg)
+
+![](images/86dbc3d2408a57c1c7f33cfa47b7a34b569103dc8dc04790b8cc27b9c6b22b2c.jpg)
+
+![](images/ec9b857473393b3cd8971f54c326135f2d4caba49099b4cbe052c89b330c80e5.jpg)
+
+![](images/f19b231825cfeba2057b3b2470310bf3a9684786b8b9efca3b7bd4cf03d601f2.jpg)

youaregroundedlatentnameartifactsinpretrainedlanguagemodels/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d57052ac9ec68ef0c1d6bf5c61a6afe107533e86c346c0e2132115e2744076f4
+size 1587847

youaregroundedlatentnameartifactsinpretrainedlanguagemodels/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c75814934394ec8efa28f52ebf937d977320cff107d88d4a9cbb982418bdd5c0
+size 419584

zeroshotcrosslingualsentencesimplification/096027d2-19d8-45fc-b633-a769cfd1e35c_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc61ebef18d0fd5efd86d247abe287cbce09a2733944326f7ae4ffe1190965e7
+size 113235

zeroshotcrosslingualsentencesimplification/096027d2-19d8-45fc-b633-a769cfd1e35c_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1daeb145a86ba6e6c76c781ebc11e28f544b2dccb1fac1dc25425d9584644f1e
+size 142240

zeroshotcrosslingualsentencesimplification/096027d2-19d8-45fc-b633-a769cfd1e35c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60dc79e4f58099996c7be3132f154c2507e31514a8981f52f3b6a023f3b180ab
+size 455953

zeroshotcrosslingualsentencesimplification/full.md

@@ -0,0 +1,433 @@
+# Zero-Shot Crosslingual Sentence Simplification
+
+Jonathan Mallinson<sup>1</sup> Rico Senrich<sup>2,1</sup> Mirella Lapata<sup>1</sup>
+
+$^{1}$ School of Informatics, University of Edinburgh
+
+$^{2}$ Department of Computational Linguistics, University of Zurich
+
+J.Mallinson@ed.ac.uk, Sennrich@cl.uzh.ch, mlap@inf.ed.ac.uk
+
+# Abstract
+
+Sentence simplification aims to make sentences easier to read and understand. Recent approaches have shown promising results with encoder-decoder models trained on large amounts of parallel data which often only exists in English. We propose a zero-shot modeling framework which transfers simplification knowledge from English to another language (for which no parallel simplification corpus exists) while generalizing across languages and tasks. A shared transformer encoder constructs language-agnostic representations, with a combination of task-specific encoder layers added on top (e.g., for translation and simplification). Empirical results using both human and automatic metrics show that our approach produces better simplifications than unsupervised and pivot-based methods.
+
+# 1 Introduction
+
+Sentence simplification aims to reduce the linguistic complexity of a text whilst retaining most of its meaning. It has been the subject of several modeling efforts in recent years due to its relevance to various applications (Siddharthan, 2014; Shardlow, 2014). Examples include the development of reading aids for individuals with autism (Evans et al., 2014), aphasia (Carroll et al., 1999), dyslexia (Rello et al., 2013), and population groups with low-literacy skills (Watanabe et al., 2009), such as children and non-native speakers.
+
+Modern approaches (Zhang and Lapata, 2017; Mallinson and Lapata, 2019; Nishihara et al., 2019; Dong et al., 2019) view the simplification task as monolingual text-to-text rewriting and employ the very successful encoder-decoder neural architecture (Bahdanau et al., 2015; Sutskever et al., 2014). In contrast to traditional methods, which target individual aspects of the simplification task, such as sentence splitting (Carroll et al. 1999; Chandrasekar et al. 1996, inter alia) or the substitution
+
+of complex words with simpler ones (Devlin, 1999; Kaji et al., 2002), neural models have no special purpose mechanisms for ensuring how to best simplify text. They rely on representation learning to implicitly capture simplification rewrites from data, i.e., examples of complex-simple sentence pairs.
+
+While large-scale parallel datasets exist for English (Xu et al., 2015; Zhang and Lapata, 2017) and Spanish (Agrawal and Carpuat, 2019), there is a limited amount of simplification data for other languages. For example, Klaper et al. (2013) automatically aligned 7,000 complex-simple German sentences, $^{1}$  and Brunato et al. (2015) released 1,000 complex-simple Italian sentences. But data-driven approaches to simplification, in particular popular neural models, require significantly more training data to achieve good performance, making these datasets better suited for testing or development purposes. Unsupervised approaches (Surya et al., 2019; Artetxe et al., 2018) which forgo the use of parallel corpora are an appealing solution to overcoming the paucity of data. However, in this paper we argue that better simplification models can be obtained by taking advantage of existing complex-simple data in a high-resource language, and bilingual data in a low-resource language (i.e., a language for which no parallel simplification corpus exists).
+
+Drawing inspiration from the success of machine translation (Firat et al., 2016b; Blackwood et al., 2018; Johnson et al., 2017), we propose a modeling framework which transfers simplification knowledge from English to another language while generalizing across language and task barriers during training. The backbone of our model is an encoder-decoder transformer (Vaswani et al., 2017) trained using multi-task learning to either translate, autoencode, simplify, or language model in both high-
+
+and low-resource languages. Regardless of the task or language, we employ the same base encoder on top of which task-specific transformer layers are added, while language-specific transformer decoders are used to generate the output sequence. Since the same base encoder is used for all tasks and languages, the model learns task- and language-agnostic representations. A beneficial side-effect is that the proposed architecture can be trained using one language and tasked to simplify another.
+
+As simplifications for multiple languages can be produced within the same model, our approach is more scalable compared to pivot-based methods (Mallinson et al., 2018; Conneau et al., 2018). The latter would first translate the complex sentence into a high-resource language, apply a monolingual simplification model, and then translate back the output to the original language. We avoid having to train multiple models and make multiple hops, where each hop can add noise and latency, and instead develop a one-hop crosslingual zero-shot approach. We evaluate our model using English as our high-resource language and German as our low-resource language on two test sets from different domains, and with different end-users in mind. These include TextComplexityDE (Naderi et al., 2019), a recently created corpus of German Wikipedia sentences deemed complex by second language German learners. We also release a second dataset which contains manual simplifications of articles taken from GEOlino $^2$ , a popular children's magazine. Empirical results using both human and automatic metrics show that our approach produces better simplifications than both unsupervised and pivot-based methods.
+
+Our contributions in this paper are threefold: (1) a cross-lingual architecture which allows the transfer of simplification knowledge from high- to low-resource languages, alleviating the paucity of training data for monolingual simplification; (2) a comprehensive evaluation framework using automatic metrics and human judges; and (3) the release of a dataset in German which we hope will facilitate further research in automatic simplification.
+
+# 2 Related Work
+
+Simplification The majority of previous work has focused on English, using large-scale datasets
+
+like Newsela and Wikipedia (Xu et al., 2015). One of the first neural network approaches to simplification was presented in Zhang and Lapata (2017) who use an encoder-decoder LSTM, trained with reinforcement learning, to optimize for grammaticality, simplicity, and adequacy. Dong et al. (2019) use a Programmer-Interpreter (Reed and de Freitas, 2016), which receives the source sentence as an input, and applies a sequence of edit operations (add, delete, keep). Kriz et al. (2019) propose to rerank a diverse set of simplifications according to fluency, adequacy, and simplicity. Martin et al. (2020a) introduce a simplification model which allows the user to control the generated output and in follow-on work (Martin et al., 2020b) they create multilingual paraphrasing datasets for training their model. Palmero Aprosio et al. (2019) explore different ways to incorporate non-parallel simplification data to expand small scale training data, including autoencoding and backtranslation.
+
+Translation data, in the form of paraphrases, has also been incorporated into simplification models leading to significant improvements. Guo et al. (2018) use multi-task learning to augment the limited amount of simplification training data. In addition to training on complex-simple sentence pairs, their model employs paraphrases, created automatically using machine translation. Zhao et al. (2018) augment a Transformer-based simplification model with lexical rules obtained from Simple PPDB (Pavlick and Callison-Burch, 2016), a database of paraphrase rules, automatically annotated with simplicity scores.
+
+Unlike previous approaches, we do not train models to create training data, either via backtranslation or extracting paraphrases. Instead, our model is able to train directly on existing datasets, saving computation power and time. In the future, it would be interesting to explore whether additional datasets or tasks improve simplification performance.
+
+Crosslingual Generation Cross-lingual transfer learning-based approaches have originated in machine translation. Dong et al. (2015) translate from one source language to multiple target languages (one-to-many) adding a separate decoder for each. Follow-on work (Luong et al., 2016; Firat et al., 2016a) performs translation with multiple encoders and decoders (many-to-many). Johnson et al. (2017) and Ha et al. (2016) train multilingual models where all languages share encoder and de
+
+coder parameters, and language tags (prepended to the source sentence) are used to specify the target.
+
+Multilingual models are also capable of translating between unpaired languages, thereby performing zero-shot translation (Firat et al., 2016b; Johnson et al., 2017; Ha et al., 2016). Blackwood et al. (2018) propose sharing all parameters but the attention mechanism, while Lu et al. (2018) develop a shared "interlingua layer" at the interface of language-specific encoders and decoders. Advances in unsupervised machine translation (Artetxe et al., 2018; Lample et al., 2018) have further spurred interest in modeling sequence-to-sequence problems without a parallel corpus. Surya et al. (2019) learn from unpaired simple and complex English sentences using a shared encoder, two decoders, denoising, backtranslation and discrimination-based losses. Zhao et al. (2020) propose a similar setup, they create a denoising objective by using simple PPDB, replacing simple phrases with complex phrases. Reinforcement learning is further used to reward the fluency, adequacy and simplicity.
+
+While zero-shot approaches are effective for translating between unpaired languages, they do not consider the case where there exists no parallel data for a language. For simplification, we assume that there is no parallel corpus in the low-resource language (e.g., complex-simple German). Furthermore, preliminary results showed that zero-shot translation approaches (Johnson et al., 2017) which prepend a tag in the source sentence — this tag would indicate the simplification task in our case — perform poorly, basically resulting in the source sentence being copied over with no changes made. We circumvent this by replacing tags with task-specific transformer encoder layers which are added on top of the base encoder. This proposed architecture allows us to transfer supervision signals across languages and is potentially useful for other generation tasks, including question generation (Kumar et al., 2019) and sentence compression (Shen et al., 2018; Duan et al., 2019).
+
+# 3 Zero-shot Simplification
+
+We first define a basic encoder-decoder Transformer before adapting it for zero-shot crosslingual simplification with multi-task learning.
+
+<table><tr><td>Task</td><td>Source Language</td><td>Target Language</td><td>Target Domain</td></tr><tr><td>Translate</td><td>HR</td><td>HR</td><td>complex</td></tr><tr><td>Translate</td><td>LR</td><td>LR</td><td>simple</td></tr><tr><td>Translate</td><td>HR</td><td>HR</td><td>simple</td></tr><tr><td>Translate</td><td>LR</td><td>LR</td><td>complex</td></tr><tr><td>Translate</td><td>HR</td><td>LR</td><td>complex</td></tr><tr><td>Translate</td><td>LR</td><td>HR</td><td>complex</td></tr><tr><td>LM</td><td>None</td><td>HR</td><td>complex</td></tr><tr><td>LM</td><td>None</td><td>HR</td><td>simple</td></tr><tr><td>LM</td><td>None</td><td>LR</td><td>complex</td></tr><tr><td>LM</td><td>None</td><td>LR</td><td>simple</td></tr><tr><td>Simplify</td><td>HR</td><td>HR</td><td>simple</td></tr></table>
+
+Table 1: Training tasks and their instantiations.
+
+# 3.1 Encoder-Decoder
+
+Given a source sentence  $X = (x_{1}, x_{2}, \dots, x_{|X|})$ , our model learns to predict target  $Y = (y_{1}, y_{2}, \dots, y_{|Y|})$ , where  $Y$  could be a translation (e.g., from English to German) or a simplification (e.g., from complex to simple English). Inferring target  $Y$  given source  $X$  can be modeled as a sequence-to-sequence learning problem (Bahdanau et al., 2015). Our approach adopts the Transformer's multi-layer and multi-head attention encoder-decoder architecture (Vaswani et al., 2017). The Transformer encoder has  $n$  layers (denoted  $L_{i}$  for layer  $i$ ), which transform the input sequentially,  $X^{l+1} = L_{i}(X^{l})$ , to yield representations  $X^{N} = L_{1:N}(X)$ . For more details regarding the Transformer layer, we refer the reader to Vaswani et al. (2017). The decoder is composed of a stack of identical layers. In addition to self-attention the decoder attends to the source sentence  $X^{N}$ . Encoder and decoder stacks are trained to minimize the cross-entropy loss of  $Y$  given  $X$ :
+
+$$
+\mathcal {L} _ {\mathrm {C E}} = - \sum_ {i = 1} ^ {| Y |} \log p \left(y _ {i} \mid y _ {<   i}, X ^ {N}; \theta\right) \tag {1}
+$$
+
+# 3.2 Multi-task Learning
+
+We define a multi-task crosslingual setup where the model is trained on four basic tasks; namely translation, autoencoding, language modeling, and simplification. We train on different instantiations of these tasks depending on the source language which can be high-resource (HR; e.g., English) or low-resource (LR; e.g., German), the target language (which is again HR or LR), and the output domain which can be simple or complex. We assume we only have monolingual simplification data in the high-resource language and that we have bilingual translation data only in the complex do
+
+main. Table 1 has a breakdown of the tasks we consider, with a more detailed description below.
+
+Simplification is the backbone of the model and consists of a complex source sentence which must be transformed into a simple sentence, while still retaining the original meaning. We assume we only have parallel training data in the high-resource language (see last row in Table 1).
+
+Translation consists of a source sentence, which must be translated into the target language while retaining the meaning of the source. By training on translation data, our model learns language-agnostic representations which are helpful for simplifying in the low-resource language.
+
+Autoencoding refers to translating between the same language, as seen in Table 1. As it is trivial to autoencode with attention, we apply source token dropout, where randomly selected source tokens are replaced with a special DROP token (Lample et al., 2018). We apply this dropout to all tasks (translation, autoencoding, and simplification). Additionally, this task allows us to incorporate monolingual non-parallel simple data from the low-resource language.
+
+Language Modeling has no source sentence; instead the decoder must learn to predict the next token based on its history. This task also allows us to incorporate monolingual non-parallel simple data from the low-resource language.
+
+Domains in our case our two, the simple domain which consists of text that is easy to read and the complex domain where text has not been explicitly written for ease of reading. Introducing domains to the model allows us to further inject knowledge about monolingual non-parallel simple sentences from the low-resource language. We use the target audience of the data to determine if it is simple or complex (e.g., if the text comes from Simple Wikipedia or a children's book it is representative of simple language). In practice, there often exists only limited amounts of non-parallel simple sentences in the low-resource setting, highlighting the difficulty of the task.
+
+# 3.3 Crosslingual Training
+
+With the tasks defined, we explain how the model is able to switch among them. We propose a modular encoder, where different encoder layers are used for different tasks; an outline of this can be seen in
+
+![](images/9ee714ff96d194d6cbea2fdb6fc31d2f4d8e0f200aa6c6c3088ab9e834cea44c.jpg)  
+Figure 1: Architecture of our crosslingual encoder-decoder model. The lexicon Encoder transforms words into word embeddings. Solid lines indicate mandatory paths, dotted lines indicate possible paths.
+
+Figure 1. For every task we use the same  $k$  base transformer encoder layers, where  $k$  is a hyperparameter. Each task  $\mathcal{T}$  (simplification, translation, language modeling), has additional  $t$  dedicated transformer layers  $L_{1:t}^{\mathcal{T}}$ , which are applied to the top of the base  $k$  layers,  $L_{1:t}^{\mathcal{T}}(L_{1:k}(X))$ . Each domain  $\mathcal{D}$  (simple/complex), also has  $d$  additional dedicated transformer layers  $d_{1:d}^{\mathcal{D}}$  applied on top of the task specific layers. The final representation of the source sentence  $X$  is:
+
+$$
+X ^ {N} = L _ {1: d} ^ {\mathcal {D}} \left(L _ {1: t} ^ {\mathcal {T}} \left(L _ {1: k} (X)\right)\right) \tag {2}
+$$
+
+Our model is trained end-to-end to minimize cross entropy; for each minibatch we specify the task, domain and output language  $(\mathcal{O})$  ..
+
+$$
+\mathcal {L} _ {\mathrm {C E}} = - \sum_ {i = 1} ^ {| Y |} \log P (y | y _ {<   i}, X ^ {N}; \theta , \{\mathcal {D}, \mathcal {T}, \mathcal {O} \}) \tag {3}
+$$
+
+$\mathcal{D}$  and  $\mathcal{T}$  determine the choice of dedicated Transformer encoder layers. We use a dedicated Transformer decoder for each output language  $\mathcal{O}$  to encourage the model to learn language-agnostic representations. All text is preprocessed using SentencePiece (Kudo and Richardson, 2018), resulting
+
+in a shared vocabulary between LR and HR. This allows for word embeddings to be shared between the encoder and the decoders.
+
+We further force representations to be language-agnostic, by employing a DISCriminator (Ganin and Lempitsky, 2015), a feed-forward network trained to distinguish HR and LR from the hidden representations. The encoder is then trained to perplex the discriminator. Specifically, we add two discriminators to our model; one determines the language of the source sentence  $(\mathcal{I})$  using  $L_{1:k}(X)$  and the other predicts the target language using the output of the encoder  $X^{N}$ . In this way we ensure the input to the simplification transformer layers is language-agnostic as well as the output. The discriminator is trained to minimize the binary cross-entropy loss (BCE) between its predictions and the ground truth:
+
+$$
+\sum_ {i = 1} ^ {| X |} \operatorname {B C E} (\mathcal {I}, \operatorname {D I S C} (L _ {1: k} (X) _ {i}; \theta_ {d \mathcal {I}}) + \tag {4}
+$$
+
+$$
+\operatorname {B C E} \left(\mathcal {O}, \operatorname {D I S C} \left(X _ {i} ^ {N}\right); \theta_ {d \mathcal {O}}\right)
+$$
+
+where  $\theta_{d\mathcal{I}}$  and  $\theta_{d\mathcal{O}}$  are the parameters of the two discriminators. The encoder is trained using an adversarial loss, to perturb the discriminator:
+
+$$
+\mathcal {L} _ {\mathrm {A D V}} = - \sum_ {i = 1} ^ {| X |} \operatorname {B C E} (\mathcal {I}, \operatorname {D I S C} (L _ {1: k} (X) _ {i}; \theta) + \tag {5}
+$$
+
+$$
+\operatorname {B C E} (\mathcal {O}, \operatorname {D I S C} (X _ {i} ^ {N}); \theta)
+$$
+
+The adversarial loss is combined, and optimized simultaneously, with the cross-entropy loss to produce the training objective of the entire model.
+
+$$
+\mathcal {L} = \mathcal {L} _ {\mathrm {C E}} + \lambda \mathcal {L} _ {\mathrm {A D V}} \tag {6}
+$$
+
+where  $\lambda$  moderates the degree to which the encoder should perturb the discriminators. A high value for  $\lambda$  can cause the encoder to not encode any information regarding the source input.
+
+To perform simplification in the low-resource language at test time, the base encoder is used with the simplification stack which is subsequently decoded with the LR decoder. To perform crosslingual simplification, the decoder can simply be changed to the HR decoder.
+
+# 4 Experimental Setup
+
+Training Set Our training data is summarized in Table 2. For all experiments we assume that English is the high-resource language and German is
+
+<table><tr><td></td><td>Source</td><td>Target</td><td>Size</td></tr><tr><td>WikiLarge</td><td>EnglishC</td><td>EnglishS</td><td>300K</td></tr><tr><td>WMT19</td><td>EnglishC</td><td>GermanC</td><td>6.0M</td></tr><tr><td>GeoLino</td><td>—</td><td>Germans</td><td>200K</td></tr><tr><td>Wikipedia</td><td>—</td><td>EnglishS</td><td>1.4M</td></tr></table>
+
+Table 2: Training data used in our experiments; monolingual corpora shown under Target; indices are short-hands for Complex and Simple language.
+
+the low-resource language. Simplification data in English is taken from WikiLarge (Zhang and Lapata, 2017), a fairly large corpus which consists of a mixture of three automatically-collated Wikipedia simplification datasets (Zhu et al., 2010; Woodsend and Lapata, 2011; Kauchak, 2013). English-German bilingual data is taken from the WMT19 news translation task. Complex monolingual non-parallel data uses one side of the WMT19 translation data. Simple English non-parallel data uses sentences extracted from simple Wikipedia, a simplified version of Wikipedia. Simple German non-parallel data uses sentences scraped from GEOLino (Hancke et al., 2012), a German general-interest magazine for children aged between 8–14.
+
+Test Set We evaluated our model on two German simplification datasets, each targeting different users. TextComplexityDE (Naderi et al., 2019) consists of sentences from Wikipedia, which were considered complex by second language German learners. These sentences were then simplified by a native German speaker. In addition, we created a test set from GEOlino. We extracted 20 articles<sup>4</sup> from three categories: nature, physics, and people. A trained German linguist then simplified the articles, sentence by sentence, to be understandable for children aged between 5–7 years. Our simplifying instructions can be found in the Appendix.
+
+Table 2 shows various descriptive statistics on our test sets. GEOlino is larger and consists of both single and multiple source sentences. The FRE readability metric (see the description in the following section) shows that both the source and target sentence are very simple. We also see moderate amounts of sentence splitting (the number of sentences per instance increases in the simplified target). TextComplexityDE is more complex, with the source sentences having the lowest FRE score. The target simplifications, while noticeably simpler than the source, are still more complex than
+
+<table><tr><td></td><td colspan="2">TextComplexityDE</td><td colspan="2">GEOlino</td></tr><tr><td></td><td>Source</td><td>Target</td><td>Source</td><td>Target</td></tr><tr><td>Length</td><td>28.66</td><td>29.23</td><td>15.68</td><td>15.05</td></tr><tr><td>Sents</td><td>1.09</td><td>2.17</td><td>1.13</td><td>1.55</td></tr><tr><td>FRE</td><td>28.53</td><td>49.3</td><td>62.87</td><td>68.73</td></tr><tr><td>Size</td><td colspan="2">122</td><td colspan="2">663</td></tr><tr><td>TER</td><td colspan="2">67.95</td><td colspan="2">24.12</td></tr><tr><td>Insertions</td><td colspan="2">3.20</td><td colspan="2">0.43</td></tr><tr><td>Deletions</td><td colspan="2">3.17</td><td colspan="2">1.08</td></tr><tr><td>Substitution</td><td colspan="2">9.10</td><td colspan="2">1.54</td></tr><tr><td>Shifts</td><td colspan="2">1.70</td><td colspan="2">0.18</td></tr></table>
+
+Table 3: Descriptive statistics of test set, including: Size, number of instances; Length, average number of words; Sents, average number of sentences per instance; average Flesch Reading Ease (FRE; higher is simpler); TER, translation error rate measuring distance between source and target; it is composed of four parts: insertions, deletions, substitutions and shifts.
+
+GEOlino. We also observe a significant amount of sentence splitting in this dataset. TextComplexityDE also has a significantly higher Translation Error Rate (TER), However, GEOlino approximately matches the TER of the WikiLarge test set (25.85). While both test sets use a large proportion of substitutions, TextComplexityDE has a much large proportion of insertions, which could be explained by the greater amount of sentence splitting.
+
+Model Parameters During training, the base encoder stack consists of six transformer layers, the decoder stack six layers. The simplification stack consists of two weight tied transformer layers, we note that the simplicity level can be increased by applying the stack multiple times at test time. All other stacks consist of a single layer. Each layer has a hidden dimension of size 512 and an inner dimension size of 2,048. Word embeddings, size 512, were initialized randomly and shared between the encoder and both decoders. We used eight attentional heads. Dropout was set to 0.1; source word dropout was also set to 0.1. The discriminator consists of a four layer feedforward network with dropout set to 0.1. The networks were optimized using Adam (Kingma and Ba, 2014). Multi-tasking was performed by alternating batches of different tasks. Tasks varied in dataset sizes and had different difficulties. As we wished to do equally well with all tasks we select a minibatch from a task with a probability inversely proportional to the training loss of the task. One model was selected using the average FRE-BLEU score across both development sets.
+
+All text was preprocessed using the UDPipe tok
+
+enization script (Straka, 2018) and truecasing was applied. SentencePiece was subsequently applied to the text to split words into subwords, with a SentencePiece vocabulary size of 50,000 and a sampling size of  $l = \infty$  and a smoothing parameter of  $\alpha = 0.25$  (Kudo, 2018).
+
+Evaluation As there is no single agreed-upon metric for simplification (Alva-Manchego et al., 2020; Sulem et al., 2018), we evaluate model output using a combination of four automatically-generated scores. These metrics have been previously shown to correlate with human judgments of simplification quality (Xu et al., 2016) and essentially quantify: a) whether the output is similar to the gold standard reference (Target-based,  $T$ ); b) whether the output is similar to the source (Source-based,  $S$ ); and c) whether the output is simple on its own, with no regard to preserving the meaning of the original sentence (Readability-based,  $R$ ). We indicate the type of each metric using superscripts.
+
+$\mathbf{BLEU}^{\mathrm{T}}$  (Papineni et al., 2002) assesses the degree to which generated simplifications agree with the gold standard references.
+
+I-BLEU $^{\mathrm{T},\mathrm{S}}$  (Sun and Zhou, 2012) combines self-BLEU and BLEU to reward systems with high overlap with the reference, and penalize those with high overlap to the source. Self-BLEU computes the BLEU score between the output and the source. It allows us to examine whether the models are making trivial changes to the input. Following Xu et al. (2016), we set the parameter which balances the contribution of the two metrics to  $\alpha = 0.9$ .
+
+$\mathbf{SARI}^{\mathrm{T},\mathrm{S}}$  (Xu et al., 2016) is calculated using the average of three rewrite operation scores: addition, copying, and deletion. It rewards addition operations when the system's output is not in the input but occurs in the references. Analogously, it rewards words deleted/retained if they are in both the system output and the references. $^7$
+
+$\mathbf{FRE - BLEU}^{\mathrm{T},\mathrm{S},\mathrm{R}}$  is a modification of FKGL-BLEU (Xu et al., 2016), which combines the difference in FKGL of the source and the output and the I-BLEU score. FKGL is a shorthand for the Flesch-Kincaid Grade Level readability score which was originally developed for English but has not been ported to German. So instead we use the Flesch
+
+<table><tr><td>Models</td><td>FRE-BLEU</td><td>I-BLEU</td><td>BLEU</td><td>SARI</td></tr><tr><td>ZEST</td><td>36.04</td><td>12.99</td><td>21.11</td><td>41.12</td></tr><tr><td>Pivot</td><td>28.44</td><td>8.09</td><td>11.50</td><td>38.64</td></tr><tr><td>U-SIMP</td><td>29.95</td><td>8.97</td><td>15.03</td><td>37.40</td></tr><tr><td>U-NMT</td><td>26.63</td><td>7.09</td><td>11.72</td><td>35.97</td></tr></table>
+
+(a) TextComplexityDE  
+
+<table><tr><td>Models</td><td>FRE-BLEU</td><td>I-BLEU</td><td>BLEU</td><td>SARI</td></tr><tr><td>ZEST</td><td>62.37</td><td>44.72</td><td>58.68</td><td>39.09</td></tr><tr><td>Pivot</td><td>39.54</td><td>17.81</td><td>22.92</td><td>27.94</td></tr><tr><td>U-SIMP</td><td>59.53</td><td>46.33</td><td>61.10</td><td>40.00</td></tr><tr><td>U-NMT</td><td>62.57</td><td>39.50</td><td>52.02</td><td>35.22</td></tr></table>
+
+Reading Ease readability test which has been modified for German (FRE; Amstad 1978) and adapt FK-BLEU to use the difference in FRE. $^{8}$
+
+We also evaluated system output by eliciting human judgments via Amazon's Mechanical Turk. Native German speakers (self reported) were asked to rate simplifications on three dimensions: Grammaticality (is the output grammatical and fluent?), Meaning Adequacy (to what extent is the meaning expressed in the original sentence preserved in the output, with no additional information added?), and Simplicity (is the output a simpler version of the input?). Ratings were obtained using a five point Likert scale. We randomly sampled 100 source sentences from each test set (GEOlino and TextComplexityDE), each sample received five ratings, resulting in 500 judgments per test set.
+
+# 5 Results
+
+Automatic Evaluation Table 4 summarizes our automatic evaluation results. We compare our ZEro-shot croSslingual Sentence simplificaTion model, which we call ZEST, against multiple baselines, both unsupervised and supervised ones.
+
+Previous work (Artetxe et al., 2018; Lample et al., 2018) demonstrates how an unsupervised neural MT model can be trained by optimizing two objectives: (1) denoising, where a source sentence is noised and then the corresponding decoder is tasked with reconstructing the original sentence and (2) on-the-fly back-translation, which translates the sentence in inference mode; this translation is then encoded and the task is to reconstruct the original sentence. This model can be easily adapted for simplification by considering simple
+
+(b) GEOlino  
+Table 4: Results using automatic evaluation metrics; best scores for each metric are boldfaced.  
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">TextComplexityDE</td><td colspan="2">GEOlino</td></tr><tr><td>FRE-BLEU</td><td>SARI</td><td>FRE-BLEU</td><td>SARI</td></tr><tr><td>ZEST</td><td>36.04</td><td>41.11</td><td>62.37</td><td>39.09</td></tr><tr><td>-ADV</td><td>36.81</td><td>40.47</td><td>60.61</td><td>40.98</td></tr><tr><td>-LM</td><td>35.46</td><td>41.26</td><td>57.29</td><td>40.33</td></tr><tr><td>-AE</td><td>35.56</td><td>41.60</td><td>57.66</td><td>36.49</td></tr><tr><td>-LM-AE</td><td>35.39</td><td>41.71</td><td>55.37</td><td>35.42</td></tr></table>
+
+Table 5: Ablation study examining the impact of removing the adversarial (ADV) loss, and then additionally removing the language modeling loss (LM), and autoencoding loss (AE), separately then together.
+
+German and complex German to be different languages (U-NMT). Surya et al. (2019) extend this approach further (U-SIMP) by adding two losses, which they show result in better simplifications: (1) an adversarial loss using a discriminator which tries to determine if the source sentence is complex or simple, and (2) a diversification loss, where a classifier is trained to determine if the source sentence was encoded using the complex or simple encoder. We trained both models using the code provided by Surya et al. (2019) and the same simple and complex non-parallel German data used to train our own model (see Table 2; WMT19 complex German and GEOlino simple German).
+
+We additionally include a supervised baseline based on pivoting, which requires three independently trained models, consisting of over twice as many parameters: a complex source German sentence is first translated to English (de  $\rightarrow$  en); it is then simplified (complex en  $\rightarrow$  simple en), before translating it back to German (en  $\rightarrow$  de). All three models consist of a transformer with eight encoder/decoder layers and were trained using the same data as employed in our approach (see Table 2; WMT19 and WikiLarge). On the WMT19 test set, the Pivot-based system obtained a BLEU score of 34.15/31.72 for the en  $\rightarrow$  de/de  $\rightarrow$  en directions. For comparison, ZEST achieved 32.11/30.90 for the same directions. With regard to English simplification (complex en  $\rightarrow$  simple en), the pivot system achieved a SARI score of 36.30 on the WikiLarge test, and ZEST 37.78. On the same test set, Zhang and Lapata (2017), a standard baseline simplification system trained on WikiLarge obtains 37.26, and the state-of-the-art system achieves 41.70 (Martin et al., 2020a). It is possible to incorporate some of the improvements of these approaches (e.g., controlling the amount of compression, paraphrasing, lexical complexity) into our model, however, we leave this to future work.
+
+C Das ist nur etwa das Doppelte [des Weltenergiebedarfs], [was]5 bedeutet, [dass]5 [Erdwärmenutzung]6 [im]2 große Stil immer auf eine lokale Abkuhlung des Gesteins hinauslauf.  
+R Das ist nur etwa das Doppelte [des Energiebedarfs der Welt] $^4$ . Das bedeutet, [dass] $^5$  die [Benutzung] $^6$  von Erdwärme immer dazu führt, [dass] $^5$  an [sich] $^2$  diesen Stellen das Gestein abkühlt.  
+P Dabei handelt es sich nur um eine [Verdoppelung]6 [des weltweiten Energiebedarfs]5, [was]5 [bedeutet]2, [dass]5 die großflächige [geothermische]7 [Nutzung]6 immer einer lokalen [Kuhlung]6 [des Gesteins]4 entspricht.  
+Z Das bedeutet, [dass] Erdwärme im großen Stil immer auf eine lokale Abkühlung [des] Gesteins hinauslauf.
+
+(a) TextComplexityDE
+
+C Von hier aus erhalten ihr einen [eindrucksvollen] Rundum-Blick über dieGPCe Schlucht [hinweg] bis hin zu ihren etwas [5000] Meter hohen Kraterwänden.  
+R Von hier aus erhalten ihr einen Rundum-Blick über dieGPCANZe Schlucht. Ihr seht hier bis hin zu ihren etwas [5000]3 Meter hohen Kraterwänden.  
+P Von hier genießen Sie einen [beeindruckenden]¹ Rundumblick über die gesamte Schlucht bis [zu]² den 500 m hohen Kraterwänden.  
+Z Von hier aus erhaltet ihr einen Rundum-Blick auf die ganze Schlucht.
+
+(b) GEOlino
+
+Table 6: Examples of system output, Source (C), Reference (R), Pivot (P), ZEST (Z) and simplification violations: (1) word has  ${13} +$  letters; (2) sentence has  ${12} +$  words; (3) high number; (4) genitive; (5) subordinate clauses; (6) abstract words; (7) difficult/foreign words.
+
+The results in Table 4 show that ZEST obtains the highest results for all metrics on TextComplexityDE. U-SIMP achieves the second best FRE-BLEU score, while Pivot achieves the second best SARI. Overall, U-NMT produces the worst results. Results on GEOlino are more mixed, with no model achieving the highest score across all metrics. ZEST does well across all metrics, scoring the second highest for every metric, whereas the scores for U-SIMP and U-NMT spike on different metrics. U-NMT achieves the best FRE-BLEU score, however, on other metrics it is the second lowest. In contrast, U-SIMP has a low FRE-BLEU score but for all other metrics it scores the highest. Pivot receives the lowest scores across all metrics. Example output is shown in Table 10 and the Appendix.
+
+We further examined the impact of different loss functions on the performance of ZEST, and these results are presented in Table 5. We see that training only on simplification and translation data  $(- \mathrm{LM} - \mathrm{AE})$  significantly damages the performance of the model, producing the lowest FRE-BLEU scores and the lowest SARI score on
+
+<table><tr><td>Models</td><td>Mean</td><td>Gram</td><td>Simp</td><td>AVG</td><td>Min</td></tr><tr><td>Reference</td><td>4.35**</td><td>4.54**</td><td>3.81*</td><td>4.23**</td><td>3.60**</td></tr><tr><td>U-SIMP</td><td>2.67**</td><td>2.87**</td><td>2.80**</td><td>2.78**</td><td>2.22**</td></tr><tr><td>Pivot</td><td>3.65**</td><td>4.13</td><td>3.67</td><td>3.82*</td><td>3.18</td></tr><tr><td>ZEST</td><td>4.05</td><td>4.15</td><td>3.63</td><td>3.94</td><td>3.23</td></tr></table>
+
+(a) TextComplexityDE  
+
+<table><tr><td>Models</td><td>Mean</td><td>Gram</td><td>Simp</td><td>AVG</td><td>Min</td></tr><tr><td>Reference</td><td>4.73**</td><td>4.75**</td><td>3.79**</td><td>4.42**</td><td>3.69**</td></tr><tr><td>U-SIMP</td><td>4.19*</td><td>4.30**</td><td>3.22*</td><td>3.90*</td><td>3.08**</td></tr><tr><td>Pivot</td><td>3.69**</td><td>3.76**</td><td>3.25*</td><td>3.45**</td><td>2.83**</td></tr><tr><td>ZEST</td><td>4.38</td><td>4.57</td><td>3.44</td><td>4.13</td><td>3.24</td></tr></table>
+
+(b) GEOlino
+
+Table 7: Mean ratings given to simplifications by human participants; highest ratings for each system are boldfaced. Models significantly different from ZEST are marked with  $* (p < 0.05)$  and  $** (p < 0.01)$ . Significance tests were performed using a student  $t$ -test.
+
+GEOlino. We note that by removing both the language modelling loss and autoencoding loss we are removing the non-parallel simple German data (GEOlino), which could explain the performance drop on the GEOlino test set. While in the full model ZEST has access to GEOlino data, the GEOlino test set is simpler than the GEOlino non-parallel training set, as it was further simplified. Additionally, the ability to incorporate extra data is a strength of our approach, as there is no obvious way to include it within the Pivot-based model.
+
+We observed that removing the autoencoding loss  $(-AE)$  led to sentences which strayed too far from the source sentence, thereby losing meaning; whereas removing the language modeling loss  $(-LM)$  led to sentences being too close to the source sentence, resulting in too little simplification. The inclusion of the adversarial loss  $(-ADV)$  showed a small overall increase in FRE-BLEU and a small decrease in SARI.
+
+Human Evaluation Table 7 summarizes the results of the human evaluation. We elicited judgments for three systems, namely ZEST, U-SIMP, and the Pivot-based approach. We also included the gold standard Reference as an upper bound (see the Appendix for examples of sentence pairs shown to crowdworkers). We report mean ratings for Meaning adequacy, Grammaticality and Simplicity, their combined average (AVG), and their (average) Minimum value. We include Minimum because we argue that a simplification is only as good as its weakest dimension. We note that it is trivial to produce a sentence that is perfectly adequate and fluent, by simply repeating the source
+
+<table><tr><td>Model</td><td>lex</td><td>SC</td><td>RC</td><td>pas</td><td>suj</td><td>gen</td><td>spl</td></tr><tr><td>Reference</td><td>38.7</td><td>11.9</td><td>10.5</td><td>6.8</td><td>16.2</td><td>12.2</td><td>35.5</td></tr><tr><td>U-SIMP</td><td>41.2</td><td>18.7</td><td>4.3</td><td>4.6</td><td>7.7</td><td>8.0</td><td>3.2</td></tr><tr><td>Pivot</td><td>44.9</td><td>17.3</td><td>7.8</td><td>6.7</td><td>11.6</td><td>14.6</td><td>3.2</td></tr><tr><td>ZEST</td><td>51.9</td><td>8.3</td><td>11.8</td><td>4.9</td><td>13.0</td><td>5.8</td><td>2.3</td></tr></table>
+
+sentence. It is also easy to produce a simple grammatical sentence if we do not care about adequacy.
+
+On TextComplexityDE, ZEST is significantly better than the unsupervised approach across all dimensions. It is on par with Pivot in terms of Grammaticality, Simplicity, and Minimum (ratings are not significantly different). However, ZEST is significantly better in terms of Meaning adequacy, and on average. On GEOlino, ZEST is significantly better against all comparison models on all dimensions. Perhaps unsurprisingly, across datasets, participants perceive gold standard simplifications as superior to the output of all comparison models.
+
+Error analysis We further analysed the types of simplifications produced by each system. We sampled 100 source sentences (50 from each dataset) and elicited judgments from annotators. The annotators were asked to indicate the types of simplification which occurred, including: lexical substitutions, passive to active voice, splitting a sentence into multiple sentences, and rewriting it to avoid subordinate clauses, relative clauses, the subjunctive mood, and the genitive case. The results in Table 8 show that ZEST performs a wide variety of simplification and produces the largest number of lexical simplifications. While all models produce more lexical substitutions than the references, the references split sentences frequently, whereas in all cases, the models split the sentence minimally. The Pivot model simplifies genitives the most while U-SIMP simplifies subordinate clauses most. ZEST produces the largest number of lexical simplifications, and simplifications related to relative clauses and subjunctives.
+
+Crosslingual Simplification We next explore how different tasks can be combined with no additional training data. We illustrate how our model can be used to tackle the tasks of both simplifying and translating. We now assume that the source complex sentence is in English and the simplified output sentence is in German. As there currently exist no crosslingual German simplification test
+
+Table 8: Proportion of simplifications on 100 sentences including lexical (lex), subordinate clause (SC), relative clause (RC), passive voice (pas) subjunctive (suj), genitive (gen), and sentence splitting (spl).  
+
+<table><tr><td>Models</td><td>FRE-BLEU</td><td>I-BLEU</td><td>BLEU</td><td>SARI</td></tr><tr><td>ZEST</td><td>31.82</td><td>10.26</td><td>14.29</td><td>41.11</td></tr><tr><td>Pivot</td><td>32.72</td><td>10.71</td><td>15.19</td><td>41.60</td></tr></table>
+
+(a) TextComplexityDE  
+
+<table><tr><td>Models</td><td>FRE-BLEU</td><td>I-BLEU</td><td>BLEU</td><td>SARI</td></tr><tr><td>ZEST</td><td>43.65</td><td>19.17</td><td>25.00</td><td>34.62</td></tr><tr><td>Pivot</td><td>42.61</td><td>18.29</td><td>23.78</td><td>34.43</td></tr></table>
+
+(b) GEOlino
+
+Table 9: Crosslingual, simplifying English into German, automatic results.
+
+sets, for evaluation purposes we hand-translated 100 complex sentences from each of the German test sets into English. Results<sup>9</sup> can be seen in Table 9 and example output in the Appendix. For comparison, we provide the results of Pivot, which requires two independently-trained models: a complex source English sentence is first simplified (complex en → simple en), and then translated into German (en → de). While the results show that ZEST and Pivot are comparable, the fact that we can train our model on single tasks and then recombine task-specific layers to allow zero-shot transfer to unseen task combinations opens up exciting new opportunities for future work.
+
+# 6 Conclusions
+
+In this paper we developed a general approach for transferring generation data from high- to low-resource languages. Experimental results on transferring simplification knowledge from English to German showed that our model was able to produce significantly better German simplifications than unsupervised and pivot-based approaches. In addition to zero-shot simplification, we showed that our model can generate German simplifications given English input, without any additional training. In the future, we plan to explore this approach with other language pairs and other generation tasks.
+
+Acknowledgments The authors gratefully acknowledge the support of the European Research Council (award number 681760; Lapata) and the Swiss National Science Foundation (MUTAMUR, no. 176727; Sennrich). We thank all those who helped with German: Sabine Webber, Ivana Balažević, Denis Emelin, Frank Keller, and Steven Kleinegesse.
+
+# References
+
+Sweta Agrawal and Marine Carpuat. 2019. Controlling text complexity in neural machine translation. 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 1549-1564, Hong Kong, China.  
+Fernando Alva-Manchego, Louis Martin, Antoine Bordes, Carolina Scarton, Benoit Sagot, and Lucia Specia. 2020. ASSET: A dataset for tuning and evaluation of sentence simplification models with multiple rewriting transformations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4668-4679, Online.  
+Toni Amstad. 1978. Wie verstandlich sind unsere Zeitungen? Studenten-Schreib-Service.  
+Mikel Artetxe, Gorka Labaka, Eneko Agirre, and Kyunghyun Cho. 2018. Unsupervised neural machine translation. In Proceedings of the 6th International Conference on Learning Representations, ICLR 2018, Vancouver, Canada.  
+Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In Proceedings of the 3rd International Conference on Learning Representations, ICLR 2015, San Diego, California.  
+Graeme Blackwood, Miguel Ballesteros, and Todd Ward. 2018. Multilingual neural machine translation with task-specific attention. In Proceedings of the 27th International Conference on Computational Linguistics, pages 3112-3122, Santa Fe, New Mexico.  
+Dominique Brunato, Felice Dell'Orletta, Giulia Venturi, and Simonetta Montemagni. 2015. Design and annotation of the first Italian corpus for text simplification. In Proceedings of The 9th Linguistic Annotation Workshop, pages 31-41, Denver, Colorado.  
+John Carroll, Guido Minnen, Darren Pearce, Yvonne Canning, Siobhan Devlin, and John Tait. 1999. Simplifying text for language-impaired readers. In Proceedings of the 9th Conference of the European Chapter of the Association for Computational Linguistics, pages 269-270, Bergen, Norway.  
+R. Chandrasekar, Christine Doran, and B. Srinivas. 1996. Motivations and methods for text simplification. In COLING 1996 Volume 2: The 16th International Conference on Computational Linguistics.  
+Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating cross-lingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2475-2485, Brussels, Belgium.
+
+Siobhan Devlin. 1999. Simplifying Natural Language for Aphasic Readers. Ph.D. thesis, University of Sunderland.  
+Daxiang Dong, Hua Wu, Wei He, Dianhai Yu, and Haifeng Wang. 2015. Multi-task learning for multiple language translation. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1723-1732, Beijing, China.  
+Yue Dong, Zichao Li, Mehdi Rezagholizadeh, and Jackie Chi Kit Cheung. 2019. EditNTS: An neural programmer-Interpreter model for sentence simplification through explicit editing. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3393-3402, Florence, Italy.  
+Xiangyu Duan, Mingming Yin, Min Zhang, Boxing Chen, and Weihua Luo. 2019. Zero-shot cross-lingual abstractive sentence summarization through teaching generation and attention. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3162-3172, Florence, Italy.  
+Richard Evans, Constantin Orăsan, and Iustin Dornescu. 2014. An evaluation of syntactic simplification rules for people with autism. In Proceedings of the 3rd Workshop on Predicting and Improving Text Readability for Target Reader Populations (PITR), pages 131-140, Gothenburg, Sweden.  
+Orhan First, Kyunghyun Cho, and Yoshua Bengio. 2016a. Multi-way, multilingual neural machine translation with a shared attention mechanism. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 866-875, San Diego, California.  
+Orhan First, Baskaran Sankaran, Yaser Al-onaizan, Fatos T. Yarman Vural, and Kyunghyun Cho. 2016b. Zero-resource translation with multi-lingual neural machine translation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 268-277, Austin, Texas.  
+Yaroslav Ganin and Victor Lempitsky. 2015. Unsupervised domain adaptation by backpropagation. In Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37, ICML'15, pages 1180-1189.  
+Han Guo, Ramakanth Pasunuru, and Mohit Bansal. 2018. Dynamic multi-level multi-task learning for sentence simplification. In Proceedings of the 27th International Conference on Computational Linguistics, pages 462-476, Santa Fe, New Mexico.  
+Thanh-Le Ha, Jan Niehues, and Alexander H. Waibel. 2016. Toward multilingual neural machine translation with universal encoder and decoder. CoRR, abs/1611.04798.
+
+Julia Hancke, Sowmya Vajjala, and Detmar Meurers. 2012. Readability classification for German using lexical, syntactic, and morphological features. In Proceedings of COLING 2012, pages 1063-1080, Mumbai, India. The COLING 2012 Organizing Committee.  
+Melvin Johnson, Mike Schuster, Quoc V. Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, Fernanda Viégas, Martin Wattenberg, Greg Corrado, Macduff Hughes, and Jeffrey Dean. 2017. Google's multilingual neural machine translation system: Enabling zero-shot translation. Transactions of the Association for Computational Linguistics, 5:339-351.  
+Nobuhiro Kaji, Daisuke Kawahara, Sadao Kurohashi, and Satoshi Sato. 2002. Verb paraphrase based on case frame alignment. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 215-222, Philadelphia, Pennsylvania.  
+David Kauchak. 2013. Improving text simplification language modeling using unsimplified text data. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1537-1546, Sofia, Bulgaria.  
+Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.  
+David Klaper, Sarah Ebling, and Martin Volk. 2013. Building a German/simple German parallel corpus for automatic text simplification. In Proceedings of the Second Workshop on Predicting and Improving Text Readability for Target Reader Populations, pages 11-19, Sofia, Bulgaria.  
+Reno Kriz, João Sedoc, Marianna Apidianaki, Carolina Zheng, Gaurav Kumar, Eleni Miltsakaki, and Chris Callison-Burch. 2019. Complexity-weighted loss and diverse reranking for sentence simplification. 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 3137-3147, Minneapolis, Minnesota.  
+Taku Kudo. 2018. Subword regularization: Improving neural network translation models with multiple subword candidates. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 66-75, Melbourne, Australia.  
+Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium.  
+Vishwajeet Kumar, Nitish Joshi, Arjit Mukherjee, Ganesh Ramakrishnan, and Preethi Jyothi. 2019.
+
+Cross-lingual training for automatic question generation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4863-4872, Florence, Italy.  
+Guillaume Lample, Alexis Conneau, Ludovic Denoyer, and Marc'Aurelio Ranzato. 2018. Unsupervised machine translation using monolingual corpora only. In 6th International Conference on Learning Representations, ICLR 2018., Vancouver, Canada.  
+Yichao Lu, Phillip Keung, Faisal Ladhak, Vikas Bhardwaj, Shaonan Zhang, and Jason Sun. 2018. A neural interlingua for multilingual machine translation. In Proceedings of the 3rd Conference on Machine Translation: Research Papers, pages 84-92, Brussels, Belgium.  
+Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. 2016. Multi-task sequence to sequence learning. In Proceedings of the 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico.  
+Jonathan Mallinson and Mirella Lapata. 2019. Controllable sentence simplification: Employing syntactic and lexical constraints. CoRR, abs/1910.04387.  
+Jonathan Mallinson, Rico Sennrich, and Mirella Lapata. 2018. Sentence compression for arbitrary languages via multilingual pivoting. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2453-2464, Brussels, Belgium.  
+Louis Martin, Éric de la Clergerie, Benoît Sagot, and Antoine Bordes. 2020a. Controllable sentence simplification. In Proceedings of The 12th Language Resources and Evaluation Conference, pages 4689-4698, Marseille, France. European Language Resources Association.  
+Louis Martin, Angela Fan, Éric de la Clergerie, Antoine Bordes, and Benoit Sagot. 2020b. Multilingual unsupervised sentence simplification. arXiv preprint arXiv:2005.00352.  
+Babak Naderi, Salar Mohtaj, Kaspar Ensikat, and Sebastian Möller. 2019. Subjective assessment of text complexity: A dataset for german language. arXiv preprint arXiv:1904.07733.  
+Daiki Nishihara, Tomoyuki Kajiwara, and Yuki Arase. 2019. Controllable text simplification with lexical constraint loss. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, pages 260-266, Florence, Italy.  
+Alessio Palmero Aprosio, Sara Tonelli, Marco Turchi, Matteo Negri, and Mattia A. Di Gangi. 2019. Neural text simplification in low-resource conditions using weak supervision. In Proceedings of the Workshop on Methods for Optimizing and Evaluating Neural Language Generation, pages 37-44, Minneapolis, Minnesota.
+
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania.  
+Ellie Pavlick and Chris Callison-Burch. 2016. Simple PPDB: A paraphrase database for simplification. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 143-148, Berlin, Germany.  
+Scott E. Reed and Nando de Freitas. 2016. Neural programmer-interpreters. In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico.  
+Luz Rello, Ricardo Baeza-Yates, Stefan Bott, and Horacio Saggion. 2013. Simplify or help?: text simplification strategies for people with dyslexia. In International Cross-Disciplinary Conference on Web Accessibility, W4A '13, pages 15:1-15:10, Rio de Janeiro, Brazil.  
+Matthew Shardlow. 2014. A survey of automated text simplification. International Journal of Advanced Computer Science and Applications, 4(1):58-70.  
+Shiqi Shen, Yun Chen, Cheng Yang, Zhiyuan Liu, and Maosong Sun. 2018. Zero-shot cross-lingual neural headline generation. IEEE/ACM Trans. Audio, Speech & Language Processing, 26(12):2319-2327.  
+Advaith Siddharthan. 2014. A survey of research on text simplification. ITL-International Journal of Applied Linguistics, 165(2):259-298.  
+Milan Straka. 2018. UDPipe 2.0 prototype at CoNLL 2018 UD shared task. In Proceedings of the CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies, pages 197-207, Brussels, Belgium.  
+Elior Sulem, Omri Abend, and Ari Rappoport. 2018. Semantic structural evaluation for text simplification. 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 685-696.  
+Hong Sun and Ming Zhou. 2012. Joint learning of a dual SMT system for paraphrase generation. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 38-42, Jeju Island, Korea.  
+Sai Surya, Abhijit Mishra, Anirban Laha, Parag Jain, and Karthik Sankaranarayanan. 2019. Unsupervised neural text simplification. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2058-2068, Florence, Italy.
+
+Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, pages 3104-3112, Montreal, Canada.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, pages 5998-6008, Long Beach, California.  
+Willian Massami Watanabe, Arnaldo Candido Junior, Vinicius Rodriguez Uzêda, Renata Pontin de Mattos Fortes, Thiago Alexandre Salgueiro Pardo, and Sandra Maria Aluísio. 2009. Facilita: reading assistance for low-literacy readers. In Proceedings of the 27th ACM International Conference on Design of Communication, pages 29-36, New York, New York.  
+Kristian Woodsend and Mirella Lapata. 2011. Learning to simplify sentences with quasi-synchronous grammar and integer programming. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 409-420, Edinburgh, Scotland, UK.  
+Wei Xu, Chris Callison-Burch, and Courtney Naples. 2015. Problems in current text simplification research: New data can help. Transactions of the Association for Computational Linguistics, 3:283-297.  
+Wei Xu, Courtney Naples, Ellie Pavlick, Quanze Chen, and Chris Callison-Burch. 2016. Optimizing statistical machine translation for text simplification. Transactions of the Association for Computational Linguistics, 4:401-415.  
+Xingxing Zhang and Mirella Lapata. 2017. Sentence simplification with deep reinforcement learning. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 584-594, Copenhagen, Denmark.  
+Sanqiang Zhao, Rui Meng, Daqing He, Andi Saptono, and Bambang Parmanto. 2018. Integrating transformer and paraphrase rules for sentence simplification. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3164-3173, Brussels, Belgium.  
+Yanbin Zhao, Lu Chen, Zhi Chen, and Kai Yu. 2020. Semi-supervised text simplification with back-translation and asymmetric denoising autoencoders. In Proceedings of the 34th AAAI Conference on Artificial Intelligence, pages 9668-9675, New York, New York.  
+Zhemin Zhu, Delphine Bernhard, and Iryna Gurevych. 2010. A monolingual tree-based translation model for sentence simplification. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1353-1361, Beijing, China. Coling 2010 Organizing Committee.
+
+# A Simplification Instructions
+
+This annotation experiment is concerned with simplification. You will be presented with a document. Your task is to read each sentence and simplify it such that children aged between 5 and 7 can understand it. The simplified version should be grammatical and retain all the important information of the original sentence.
+
+In producing simplifications, you are free to delete words, add new words, substitute them, or reorder them. In addition, you might find it useful to change a complex sentence into multiple simple sentences.
+
+To help you with the simplification task, we have produced a set of guidelines which you can follow. However, not all guidelines will always be applicable, so if you believe you can produce a simpler version, then you may ignore the guidelines. We split the guidelines into two sections: word-level and sentence-level guidelines.
+
+# A.1 Word-level Guidelines
+
+1. Special characters are not allowed, with the exception of: full stops, question marks, exclamation marks, quotation marks, and Mediopunks (used to indicate compound splitting).  
+2. Numbers should be written as digits and not words.  
+3. The word ein ('one') should only be written with a 1 when it represents a number, not when it takes the role of an indefinite article.  
+4. Roman numerals must be avoided.  
+5. Large numbers, percentages and year dates should be used sparsely.  
+6. Use easy, short and well-known words. In case a difficult word is needed, it should be explained using simple words. For a list of simple words, please consult this dictionary: https://hurraki.de/wiki/Hauptseite.  
+7. Technical terms, foreign words and abbreviations should be avoided. Common acronyms like CD or WC may be used if their full forms (compact disc, water closet) are less common.
+
+# A.2 Sentence-level Guidelines
+
+1. Coordinate and subordinate clauses are forbidden and should be transformed into independent main clauses. Main clauses should
+
+preferably contain active voice, and present, or past perfect tense. The subject-verb-object (SVO) word order should be chosen, unless another word order is more understandable.
+
+2. Nominalizations and passive constructions are forbidden.  
+3. Attributive genitives should also be avoided. If possible, the genitive attribute should be transferred into a prepositional phrase using von ('of').  
+4. Negation should be avoided. If needed, it is better to formulate a sentence with nicht ('not') instead of kein ('no').  
+5. Transparent metaphors like leichte Sprache may be used if they can be easily understood. More complex metaphors and idioms should be replaced by literal expressions.  
+6. Split complex sentences into multiple simple sentences at semicolons and dashes. Also split sentences after colons if the segment after the colon is a complete sentence and not just an enumeration.  
+7. If a subordinate conjunction is found, split the sentence at the conjunction; edit and rephrase both resulting segments to form independent sentences. Add suitable connectives that express the intended rhetorical relation and restore word order.  
+8. Rephrase concessive clauses with subjunctions like  $obwohl$  ('although') the connective trotzdem ('however').  
+9. Analogously, rephrase consecutive clauses starting with sodass ('so that') using deshalb ('therefore').  
+10. Rephrase final clauses using the modal verb wollen ('want') and the connective deshalb ('therefore'). Since the subject is not mentioned overtly in German final clauses containing um zu ('in order to'), it has to be retrieved from the main clause.  
+11. Split coordinate clauses at coordinating conjunctions (e.g., und ('and'), oder ('or'), aber ('but'), davon ('however')). The second clause can start with und ('and') and oder ('or') to emphasize that they are linked to the previous sentence.
+
+12. Replace appositions by sentences in which the noun phrase referred to by the apposition forms the subject (X) and the apposition itself becomes the predicative noun (Y), yielding an X is Y structure.
+
+# A.3 Final Remarks
+
+The annotation will proceed on a document-by-document basis. In simplifying individual sentences you should ensure that:
+
+- You have preserved all important information in the original sentence.  
+- The sentences are understandable to children aged 5 to 7.  
+- You did not render the resulting document incoherent or unreadable.  
+- You have preserved the grammaticality of the simplified sentences.
+
+# B System Output
+
+In Table 10 we present examples of simplifications from both GEOlino and TextComplexityDE. We show the input Complex sentence, the Reference simplification, and the output of our model, ZEST, and two comparison systems U-SIMP and Pivot (see Section 5 in the main paper for more details). To provide some further insight on what the models are doing we have annotated words and phrases in the examples which constitute violations of simple language according to our guidelines above and those provided in https://hurraki.de/pruefung/pruefung.htm.
+
+Table 11 contains additional examples, without annotation, while Table 12 presents crosslingual simplicity examples. Specifically, we show model output in German (DE ZEST) when the input is complex English (EN Source), and for comparison German output (DE ZEST) when the input is complex German (DE Source).
+
+<table><tr><td>Complex</td><td>Von hier aus erhalten ihr einen [eindrucksvollen]1 Rundum-Blick über die ganze Schlucht [hinweg]2 bis hin zu ihren etwa [5000]3 Meter hohen Kraterwänden.</td></tr><tr><td>Reference</td><td>Von hier aus erhalten ihr einen Rundum-Blick über die ganze Schlucht. Ihr seht hier bis hin zu ihren etwa [5000]3 Meter hohen Kraterwänden.</td></tr><tr><td>U-SIMP</td><td>Von hier eraus ihr haltet einen umfassenden Rundum-Blick über die ganze bis [hinweg]2 hinweg zu hin zu ihren [5000]3 Meter hohen Kraterwände.</td></tr><tr><td>Pivot</td><td>Von hier genießen Sie einen [beeindruckenden]1 Rundumblick über die gesamte Schlucht bis [zu]2 den 500 m hohen Kraterwänden.</td></tr><tr><td>ZEST</td><td>Von hier aus erhalten ihr einen Rundum-Blick auf die ganze Schlucht.</td></tr></table>
+
+(a) GEOlino  
+
+<table><tr><td>Complex</td><td>Das ist nur etwa das Doppelte [des Weltenergiebedarfs]4, [was]5 bedeutet, [dass]5 [Erdwärmenutzung]6 [im]2 größen Stil immer auf eine lokale Abkühlung des Gesteins hinausläuft.</td></tr><tr><td>Reference</td><td>Das ist nur etwa das Doppelte [des Energiebedarfs der Welt]4. Das bedeutet, [dass]5 die [Benutzung]6 von Erdwärme immer dazu führt, [dass]5 an [sich]2 diesen Stellen das Gestein abkühlt.</td></tr><tr><td>U-SIMP</td><td>Das ist nur etwa das Doppelte [des Weltenergiebedarfs]4, [was]5 bedeutet, [dass]5 Erdwärmemer [im]2 größen Stil immer auf eine andere Abkühlung des Gesteins()7.</td></tr><tr><td>Pivot</td><td>Dabei handelt es sich nur um eine [Verdoppelung]6 [des weltweiten Energiebedarfs]5, [was]5 [bedeutet]2, [dass]5 die großflächige [geothermische]8 [Nutzung]6 immer einer lokalen [Kühlung]6 [des Gesteins]4 entspricht.</td></tr><tr><td>ZEST</td><td>Das bedeutet, [dass]5 Erdwärme im größen Stil immer auf eine lokale Abkühlung [des]2 Gesteins hinausläuft.</td></tr></table>
+
+(b) TextComplexityDE
+
+Table 10: Examples of system output and simplification violations (https://hurraki.de/pruefung/pruefung.htm): (1) word has  $13+$  letters; (2) sentence has  $12+$  words; (3) high number; (4) genitive; (5) subordinate clauses; (6) abstract words; (7) special characters; (8) difficult/foreign words.
+
+<table><tr><td>Source</td><td>Zusätzlich kommt in der Forschung ein breites Spektrum verschiedener Isotopenanalysen zum Einsatz, deren jüngste Entwicklung eine bis vor kurzem unerreichbare Messgenauigkeit erfolgslich.</td></tr><tr><td>Reference</td><td>Außer dem worden in der Forschung keine verschiedene Isotopenanalysen eingesetzt. ihre jüngste Entwicklung erfolgslicherechtungen.</td></tr><tr><td>U-SIMP</td><td>Zusätzlich kommt in der Forschung ein große spezifische und nicht stichtaglichen Gesamterbrennende Messoberfläche ermittelten.</td></tr><tr><td>Pivot</td><td>Darüber hinaus wird in der Forschung ein breites Spektrum entsprechender Isotop-analysen eingesetzt.</td></tr><tr><td>ZEST</td><td>Zusätzlich kommt in der Forschung ein breites Spektrum verschiedenster Isotopen-Analysen zu Einsatz.</td></tr><tr><td>Source</td><td>Das ist nur etwa das Doppelte des Weltenergiebedarfs, was bedeutet, dass Erdwärmenutzung im großen Stil immer auf eine lokale Abkühlung des Gesteins hinausläuft.</td></tr><tr><td>Reference</td><td>Das ist nur etwa das Doppelte des Energiebedarfs der Welt. Das bedeutet, dass die Benutzung von Erdwärme immer dazu führt, dass an sich diesen Stellen das Gestein abkühlt.</td></tr><tr><td>U-SIMP</td><td>Das ist nur etwa das Doppelte des Weltenergiebedarfs, was bedeutet, dass Erdwärmememer im großen Stil immer auf eine andere Abkühlung des Gesteins).</td></tr><tr><td>Pivot</td><td>Dabei handelt es sich nur um eine Verdopplung des weltweiten Energiebedarfs, was bedeutet, dass die großflächige geothermische Nutzung immer einer lokalen Kühlung des Gesteins entspricht.</td></tr><tr><td>ZEST</td><td>Das bedeutet, dass Erdwärme im großen Stil immer auf eine lokale Abkühlung des Gesteins hinausläuft. (a) Text ComplexityDE</td></tr><tr><td>Source</td><td>Tiere tauschen mittels Duftmarken weitere verschlüsselte Botschaften unterinander aus.</td></tr><tr><td>Reference</td><td>Tiere tauschen mit ihrem Geruch weitere Botschaften unterinander aus.</td></tr><tr><td>U-SIMP</td><td>Tiere tauschen Hilfe Duftmarken weitere verschlüsselte Botschaften unterinander aus.</td></tr><tr><td>Pivot</td><td>Tiere tauschen weitere verschlüsselte Nachrichten mit Duftmarken aus.</td></tr><tr><td>ZEST</td><td>Tiere tauschen mit Duftmarken weitere verschlüsselte Botschaften aus.</td></tr><tr><td>Source</td><td>Der wiederum war überlebenswichtig für alle Landwirtschaft betriebenden Kulturen.</td></tr><tr><td>Reference</td><td>Der war wichtig für alle Kulturen, die Landwirtschaft betreiben.</td></tr><tr><td>U-SIMP</td><td>Der wiederum war überlebenswichtig für alle Landwirtschaft ben Kulturen.</td></tr><tr><td>Pivot</td><td>Sie war wiederum lebenswichtig für alle landwirtschaftlichen Kulturen.</td></tr><tr><td>ZEST</td><td>Der wiederum war für alle Landwirtschaft wichtig. (b) GEOlino</td></tr></table>
+
+Table 11: Simplification examples from TextComplexityDE and GEOlino.
+
+<table><tr><td>EN Source</td><td>The mountain is the watershed on whose flanks the catchment areas of the Pacific Ocean, the Atlantic Ocean over the Gulf of Mexico, and the Arctic Ocean over Hudson Bay, meet.</td></tr><tr><td>DE Source</td><td>Der Berg ist der Wasserscheidepunkt an dessen Flanken sich die Einzugsgebiete des Pazifischen Ozeans, des Atlantischen Ozeans über den Golf von Mexiko und des Arktischen Ozeans über die Hudson Bay berühren.</td></tr><tr><td>Reference</td><td>Der Berg markiert die Grenze zwischen den Gebieten des Pazifischen Ozeans, des Atlantischen Ozean und des Arktischen Ozeans.</td></tr><tr><td>EN ZEST</td><td>Der Berg ist der Weckschatz, auf dessen Flanken die Fanggebiete des pazifischen Ozeans, des Atlantischen Ozeans über dem Golf von Mexiko, und des Arktischen Ozeans über Hudson Bay, treffen.</td></tr><tr><td>DE ZEST</td><td>Der Berg ist der Wasserscheidepunkt an dem sich die Einzugsgebiete des Pazifiks, des Atlantischen Ozeans, des Golfs von Mexiko und des Arktischen Ozeans über die Hudson Bay treffen.</td></tr></table>
+
+(a) TextComplexityDE  
+
+<table><tr><td>EN Source</td><td>Without the radiation energy of the sun, plant photosynthesis would not work.</td></tr><tr><td>DE Source</td><td>Ohne die Strahlungsenergie der Sonne wurde die pflanzliche Photosynthese nicht Funktionieren.</td></tr><tr><td>Reference</td><td>Ohne die Energie der Sonne wurde die Photosynthese von den Pflanzen nicht Funktionieren.</td></tr><tr><td>EN ZEST</td><td>Ohne die Strahlungsenergie der Sonne, Pflanzen Photosynthese wurde nicht Funktionier.</td></tr><tr><td>DE ZEST</td><td>Ohne die Strahlungsenergie der Sonne wurde die Pflanze nicht Funktionieren.</td></tr></table>
+
+(b) GEOlino
+
+Table 12: Examples of crosslingual simplification (EN Source  $\rightarrow$  DE ZEST); for comparison, we also show the output of a monolingual system (DE Source  $\rightarrow$  DE ZEST).
+
+<table><tr><td>Parameter</td><td>Values</td></tr><tr><td>No. Base layers (k)</td><td>[4, 6, 8]</td></tr><tr><td>No. Domain layer</td><td>[1, 2]</td></tr><tr><td>No. Task layers (t)</td><td>[1, 2]</td></tr><tr><td>ADV loss (λ)</td><td>[0, 1, 5]</td></tr><tr><td>No. Discriminator layers</td><td>[2, 4]</td></tr><tr><td>Word Dropout</td><td>[0, 10%]</td></tr><tr><td>No. Decoder layers</td><td>[8]</td></tr><tr><td>No. Decoders</td><td>[1, 2]</td></tr><tr><td>Batch size</td><td>4000</td></tr></table>
+
+Table 13: Hyperparameter bounds. Bold indicates final value.
+
+# C Reproducibility
+
+We include additional details for reproducibility in this section.
+
+Average Runtime for Each Approach Run time results were calculated using a batch size of 30 on a Nvidia Tesla K40. Inference speed on 100 sentences was 34s for ZEST and 60s for the Pivot model (time includes loading the models).
+
+Hyperparameter Configurations and Bounds See Table 13 and section 4 of the main paper for more details. If not mentioned then we used the recommendation from OpenNMT-py https://opennmt.net/OpenNMT-py/FAQ.html#how-do-i-use-the-transformer-model. Hyperparameter bounds are also shown in Table 13 with selection done using the validation set and FRE-BLEU.
+
+Explanation of Data Preprocessing See section 4 of the main paper for more details. In addition, training data was excluded if it exceeded 80 tokens. Scrapped training data (GEOlino / simple wikipedia) was excluded if it began with a special character, was less than 5 words long, or did not end in punctuation.
+
+# Links to Downloadable Version of the Data
+
+- Simplification: We followed instructions from https://github.com/XingxingZhang/dress.  
+- Translation: http://www.statmt.org/wmt19/translation-task.html  
+- TextComplexityDE: https://github.com/Jmallins/TextComplexityDE
+
+- GEOlino test set: https://github.com/Jmallins/ZEST  
+- GEOlino training set: Contact authors (Hancke et al., 2012). Scrapping scripts can be found https://github.com/Jmallins/ ZEST.  
+- Simple Wikipedia: https://dumps.wikimedia.org/simplewiki/latest/

zeroshotcrosslingualsentencesimplification/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c0b7b62f67a822444257242fc6df629b69615fdaeae4a3a215662fb18f857b4
+size 1069513

zeroshotcrosslingualsentencesimplification/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4744ee9766fcb9bd2dede76b566c89f632b79f45cd9169b96e6054b26ca02807
+size 507703

zeroshotcrosslingualtransferwithmetalearning/d7859a6c-6b32-43a4-ad6a-3e471e57f7a4_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf0dacd4a8d20b2de21de54a4e4cdc5d387ef7cd9a4c3ec959d9a8f427d84db4
+size 97918

zeroshotcrosslingualtransferwithmetalearning/d7859a6c-6b32-43a4-ad6a-3e471e57f7a4_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4039156876a40a156562aee101693f03ef5d6579574469445ad791f42c32fa3b
+size 114682

zeroshotcrosslingualtransferwithmetalearning/d7859a6c-6b32-43a4-ad6a-3e471e57f7a4_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c03d5100740a01dafbf6c6fccd8b9b64566cdc9c5c4e8d2b02fbd3ff8a18cc63
+size 3707008

zeroshotcrosslingualtransferwithmetalearning/full.md

@@ -0,0 +1,284 @@
+# Zero-Shot Cross-Lingual Transfer with Meta Learning
+
+Farhad Nooralahzadeh
+
+University of Oslo
+
+farhad.nooralahzadeh@uzh.ch
+
+Giannis Bekoulis
+
+Vrije Universiteit Brussel - imec
+
+gbekouli@etrovub.be
+
+Johannes Bjerva
+
+University of Copenhagen / Aalborg University
+
+jbjerva@cs.aau.dk
+
+Isabelle Augenstein
+
+University of Copenhagen
+
+augenstein@di.ku.dk
+
+# Abstract
+
+Learning what to share between tasks has become a topic of great importance, as strategic sharing of knowledge has been shown to improve downstream task performance. This is particularly important for multilingual applications, as most languages in the world are under-resourced. Here, we consider the setting of training models on multiple different languages at the same time, when little or no data is available for languages other than English. We show that this challenging setup can be approached using meta-learning: in addition to training a source language model, another model learns to select which training instances are the most beneficial to the first. We experiment using standard supervised, zero-shot cross-lingual, as well as few-shot cross-lingual settings for different natural language understanding tasks (natural language inference, question answering). Our extensive experimental setup demonstrates the consistent effectiveness of meta-learning for a total of 15 languages. We improve upon the state-of-the-art for zero-shot and few-shot NLI (on MultiNLI and XNLI) and QA (on the MLQA dataset). A comprehensive error analysis indicates that the correlation of typological features between languages can partly explain when parameter sharing learned via meta-learning is beneficial.
+
+# 1 Introduction
+
+There are more than 7,000 languages spoken in the world, over 90 of which have more than 10 million native speakers each (Eberhard et al., 2019). Despite this, very few languages have proper linguistic resources when it comes to natural language understanding tasks (Joshi et al., 2020). Although there is growing awareness in the field, as evidenced by the release of datasets such as XNLI (Conneau et al., 2018), most NLP research still only considers English (Bender, 2019). While one solution to
+
+this issue is to collect annotated data for all languages, this process is both too time-consuming and expensive to be feasible. Additionally, it is not trivial to train a model for a task in a particular language (e.g., English) and apply it directly to another language where only limited training data is available (i.e., low-resource languages). Therefore, it is essential to investigate strategies that allow one to use the large amount of training data available for English for the benefit of other languages.
+
+Meta-learning has recently been shown to be beneficial for several machine learning tasks (Koch et al., 2015; Vinyals et al., 2016; Santoro et al., 2016; Finn et al., 2017; Ravi and Larochelle, 2017; Nichol et al., 2018). For NLP, recent work has also shown the benefits of this sharing between tasks and domains (Gu et al., 2018; Dou et al., 2019; Qian and Yu, 2019). Although cross-lingual transfer with meta-learning has been investigated for machine translation (Gu et al., 2018), this paper – to best of our knowledge – is the first attempt to study meta-learning for cross-lingual natural language understanding. Our contributions are as follows:
+
+- We propose X-MAML<sup>1</sup>, a cross-lingual meta-learning architecture, and study it for two natural language understanding tasks (Natural Language Inference and Question Answering);  
+- We test X-MAML on cross-domain, cross-lingual, standard supervised, few-shot as well as zero-shot learning, across a total of 15 languages;  
+- We observe consistent improvements over strong models including Multilingual BERT (Devlin et al., 2019) and XLM-RoBERTa (Conneau et al., 2020);
+
+- We perform an extensive error analysis, which reveals that cross-lingual trends can partly be explained by typological commonalities between languages.
+
+# 2 Meta-Learning
+
+Meta-learning tries to tackle the problem of fast adaptation to a handful of new training data instances. It discovers the structure among multiple tasks such that learning new tasks can be done quickly. This is done by repeatedly simulating the learning process on low-resource tasks using many high-resource ones (Gu et al., 2018). There are several ways of performing meta-learning: (i) metric-based (Koch et al., 2015; Vinyals et al., 2016); (ii) model-based (Santoro et al., 2016); and (iii) optimisation-based (Finn et al., 2017; Ravi and Larochelle, 2017; Nichol et al., 2018). Metric-based methods aim to learn similarities between feature representations of instances from different training sets given a similarity metric. For model-based architectures, the focus has been on adapting models that learn fast (e.g., memory networks) for meta-learning (Santoro et al., 2016). In this work, we focus on optimisation-based methods due to their superiority in several tasks (e.g., computer vision (Finn et al., 2017)) over the above-mentioned meta-learning architectures. These optimisation-based methods are able to find good initialisation parameter values and adapt to new tasks quickly. To the best of our knowledge, we are the first to exploit the idea of meta-learning for transferring zero-shot knowledge in a cross-lingual setting for natural language understanding, in particular for the tasks of NLI and QA. Specifically, we exploit the usage of Model Agnostic Meta-Learning (MAML) which uses gradient descent and achieves a good generalisation for a variety of tasks (Finn et al., 2017). MAML is able to quickly adapt to new target tasks by using only a few instances at test time, assuming that these new target tasks are drawn from the same distribution.
+
+Formally, MAML assumes that there is a distribution  $p(\mathcal{T})$  of tasks  $\{\mathcal{T}_1,\mathcal{T}_2,\dots,\mathcal{T}_k\}$ . The parameters  $\theta$  of model  $\mathbb{M}$  for a particular task  $\mathcal{T}_i$ , sampled from the distribution  $p(\mathcal{T})$ , are updated to  $\theta_i'$ . In particular, the parameters  $\theta$  are updated using one or a few iterations of gradient descent steps on the training examples (i.e.,  $D_{i}^{train}$ ) of task  $\mathcal{T}_i$ . For example, for one gradient update,
+
+$$
+\theta_ {i} ^ {\prime} = \theta - \alpha \nabla_ {\theta} \mathcal {L} _ {\mathcal {T} _ {i}} (\mathrm {M} _ {\theta}) \tag {1}
+$$
+
+where  $\alpha$  is the step size, the  $\mathbb{M}_{\theta}$  is the learned model from the neural network and  $\mathcal{L}_{\mathcal{T}_i}$  is the loss on the specific task  $\mathcal{T}_i$ . The parameters of the model  $\theta$  are trained to optimise the performance of  $\mathbb{M}_{\theta_i'}$  on the unseen test examples (i.e.,  $D_{i}^{test}$ ) across tasks  $p(\mathcal{T})$ . The meta-learning objective is:
+
+$$
+\min  _ {\theta} \sum_ {\mathcal {T} _ {i} \sim p (\mathcal {T})} \mathcal {L} _ {\mathcal {T} _ {i}} \left(\mathrm {M} _ {\theta_ {i} ^ {\prime}}\right) = \sum_ {\mathcal {T} _ {i} \sim p (\mathcal {T})} \mathcal {L} _ {\mathcal {T} _ {i}} \left(\mathrm {M} _ {\theta - \alpha \nabla_ {\theta} \mathcal {L} _ {\mathcal {T} _ {i}} \left(\mathrm {M} _ {\theta}\right)}\right) \tag {2}
+$$
+
+The MAML algorithm aims to optimise the model parameters via a few number of gradient steps on a new task, which we refer to as the meta-update. The meta-update across all involved tasks is performed for the  $\theta$  parameters of the model using stochastic gradient descent (SGD) as:
+
+$$
+\theta \leftarrow \theta - \beta \nabla_ {\theta} \sum_ {\mathcal {T} _ {i} \sim p (\mathcal {T})} \mathcal {L} _ {\mathcal {T} _ {i}} \left(\mathrm {M} _ {\theta_ {i} ^ {\prime}}\right) \tag {3}
+$$
+
+where  $\beta$  is the meta-update step size.
+
+# 3 Cross-Lingual Meta-Learning
+
+The underlying idea of using MAML in NLP tasks (Gu et al., 2018; Dou et al., 2019; Qian and Yu, 2019) is to employ a set of high-resource auxiliary tasks/languages to find an optimal initialisation from which learning a target task/language can be done using only a small number of training instances. In a cross-lingual setting (i.e., XNLI, MLQA), where only an English dataset is available as a high-resource language, and a small number of instances are available for other languages, the training procedure for MAML requires some non-trivial changes. For this purpose, we introduce a cross-lingual meta-learning framework (X-MAML), which uses the following training steps:
+
+1. Pre-training on a high-resource language h (i.e., English): Given all training samples in a high-resource language h, we first train the model M on h to initialise the model parameters  $\theta$ .  
+2. Meta-learning using low-resource languages: This step consists of choosing one or more auxiliary languages from the low-resource set. Using the development set of each auxiliary language, we construct a randomly sampled batch of tasks  $\mathcal{T}_i$ . Then, we update the model parameters using  $K$  data points of  $\mathcal{T}_i(D_i^{train})$  by one gradient descent step (see Eq. (1)). After this step, we can calculate the loss value using  $Q$  examples  $(D_i^{test})$  in each task. It should be noted that
+
+Algorithm 1: X-MAML  
+Input: high-resource language h, set of low-resource languages L, Model M, step size  $\alpha$  and learning rate  $\beta$   
+1 Pre-train M on h and provide initial model parameters  $\theta$   
+2 Select one or more languages from L as a set of auxiliary languages (A)  
+3 while not done do  
+4 for l ∈ A do  
+5 Sample batch of tasks  $\mathcal{T}_i$  using the development set of the auxiliary language l  
+6 for each  $\mathcal{T}_i$  do  
+7 Sample K data-points to form  $D_{i}^{train} = \{(X^{k},Y^{k})\}_{k = 1}^{K}\in \mathcal{T}_{i}$   
+8 Sample Q data-points to form  $D_{i}^{test} = \{(X^{q},Y^{q})\}_{q = 1}^{Q}\in \mathcal{T}_{i}$  for meta-update  
+9 Compute  $\nabla_{\theta}\mathcal{L}_{\mathcal{T}_i}(\mathbb{M}_\theta)$  on  $D_{i}^{train}$   
+10 Compute adapted parameters with gradient descent:  $\theta^{\prime} = \theta -\alpha \nabla_{\theta}\mathcal{L}_{\mathcal{T}_i}(\mathbb{M}_\theta)$   
+11 Compute  $\mathcal{L}_{\mathcal{T}_i}(\mathbb{M}_{\theta '})$  using  $D_{i}^{test}$   
+12 Update  $\theta \gets \theta -\beta \nabla_{\theta}\sum_{i}\mathcal{L}_{\mathcal{T}_i}(\mathbb{M}_{\theta '})$   
+13 Perform either (i) zero-shot or (ii) few-shot learning on  $\{\mathbb{L}\setminus \mathbb{A}\}$  using meta-learned parameters  $\theta$
+
+the  $K$  data points used for training  $(D_{i}^{train})$  are different from the  $Q$  data points used for testing  $(D_{i}^{test})$ . We sum up the loss values from all tasks to minimise the meta-objective function and to perform a meta-update using Eq. (3). This step is performed in multiple iterations.
+
+3. Zero-shot or few-shot learning on the target languages: In the last step of X-MAML, we first initialise the model parameters with those learned during meta-learning. We then continue by evaluating the model on the test set of the target languages (i.e., zero-shot learning) or fine-tuning the model parameters with standard supervised learning using the development set of the target languages and evaluate on the test set (i.e., few-shot learning).
+
+A more formal description of the proposed model X-MAML is given in Algorithm 1.
+
+Natural Language Inference (NLI): NLI is the task of predicting whether a hypothesis sentence is true (entailment), false (contradiction), or undetermined (neutral) given a premise sentence. The Multi-Genre Natural Language Inference (MultiNLI) dataset has 433k sentence pairs annotated with textual entailment information (Williams et al., 2018). It covers a range of different genres of spoken and written text and thus supports cross-genre evaluation. The NLI premise sentences are provided in 10 different genres: facetoface, telephone, verbatim, state, government, fiction, letters, nineeleven, travel and oup. All of the genres appear
+
+in the test and development sets, but only five are included in the training set. To verify our learning routine more generally, we define  $\mathcal{T}_i$  as an NLI task in each genre. We exploit MAML, in its original setting, to investigate whether meta-learning encourages the model to learn a good initialisation for all target genres, which can then be fine-tuned with limited supervision for each genre's development instances (2000 examples) to achieve a good performance on its test set.
+
+The Cross-Lingual Natural Language Inference (XNLI) dataset (Conneau et al., 2018) consists of 5000 test and 2500 development hypothesis-premise pairs with their textual entailment labels for English. Translations of these pairs are provided in 14 languages: French (fr), Spanish (es), German (de), Greek (el), Bulgarian (bg), Russian (ru), Turkish (tr), Arabic (ar), Vietnamese (vi), Thai (th), Chinese (zh), Hindi (hi), Swahili (sw) and Urdu (ur). XNLI provides a multilingual benchmark to evaluate how to perform inference in low-resource languages, in which only training data for the high-resource language English is available from MultiNLI. This allows us to study the impact of meta-learning with one low-resource language to serve as an auxiliary language, and evaluate the resulting NLI model on the target languages provided in the XNLI test set.
+
+Question Answering (QA): Given a context and a question, the task in QA is to identify the span in the context which answers the question. Lewis et al.
+
+(2020) introduce a Multilingual Question Answering dataset (MLQA) that contains QA instances in 7 languages: English (en), Arabic (ar), German (de), Spanish (es), Hindi (hi), Vietnamese (vi) and Simplified Chinese (zh). It includes over 12k QA instances in English and 5k for every other language, with each QA instance being available in 4 languages (on average). This dataset has been used in many recent studies on cross-lingual transfer learning (e.g., Hu et al. (2020); Liang et al. (2020)). In our experiments, we investigate meta-learning for QA with one or two auxiliary languages.
+
+# 4 Experiments
+
+We want to investigate how meta-learning can be used for cross-lingual sharing. We implement X-MAML using the higher library<sup>2</sup>. We use the Adam optimiser (Kingma and Ba, 2014) with a batch size of 32 for both zero-shot and few-shot learning. We fix the step size  $\alpha$  and learning rate  $\beta$  to  $1e - 4$  and  $1e - 5$ , respectively. We experimented using [10, 20, 30, 50, 100, 200, 300] meta-learning iterations in X-MAML. However, 100 iterations led to the best results in our experiments. The sample sizes  $K$  and  $Q$  in X-MAML are equal to 16 for each dataset. The results are reported for each experiment by averaging the performance over ten different runs. We experiment with different architectures in order to verify that our method generalises across them, further detailed below. We report results for few-shot, zero-shot cross-domain and cross-lingual learning.
+
+NLI: We experiment with two different settings. (i) For MultiNLI, a cross-genre dataset, we employ the Enhanced Sequential Inference Model (ESIM) (Chen et al., 2016), which is commonly used for textual entailment problems. ESIM uses LSTMs with attention to create a rich representation, capturing the relationship between premise and hypothesis sentences. (ii) For XNLI, a cross-lingual dataset, we use the PyTorch version of BERT using Hugging Face's library (Devlin et al., 2019) as the underlying model M. However, since our proposed meta-learning method is model-agnostic, it can easily be extended to any other architecture. Note that for Setting (i), we apply MAML, whereas for Setting (ii), we apply X-MAML on the original English BERT model (En-BERT) and on Multilingual BERT (Multi-BERT) models. As the first training
+
+step (i.e., pre-training on a high-resource language, see Step 1 in Section 3 for more information) in X-MAML for XNLI, we fine-tune En-BERT and Multi-BERT on the MultiNLI dataset (English) to obtain the initial model parameters  $\theta$  for each experiment.
+
+QA: For question answering, we use different base models M for X-MAML, namely XLM (Conneau and Lample, 2019) and XLM-RoBERTa (XLM-R) (Conneau et al., 2020), both state-of-the-art models. XLM uses a similar pre-training objective as Multi-BERT with a larger model, a larger shared vocabulary, and leverages both monolingual and parallel data. XLM-R is a RoBERTa version of XLM and is trained on a much larger multilingual corpus (i.e., Common Crawl), achieving state-of-the-art performance on most cross-lingual benchmarks (Hu et al., 2020). We employ the XLM-15 (Masked Language Model + Translation Language Model, 15 languages), XLM- $\mathbf{R}_{\text {base }}$  and XLM- $\mathbf{R}_{\text {large }}$  models released by the authors. The SQuAD v1.1 training data is used in the pretraining step of X-MAML (see Step 1 in Section 3). We use the cross-lingual development and test splits provided in the MLQA dataset for the meta-learning and evaluation steps, respectively.
+
+Baselines: We create: (i) zero-shot baselines: directly evaluate the model on the test set of the target languages; (ii) few-shot baselines: fine-tune the model on the development set, then evaluate on the test set of the low-resource languages.
+
+# 4.1 Few-Shot Cross-Domain NLI
+
+We train ESIM on the MultiNLI training set to provide initial model parameters  $\theta$  (see Step 1 in Section 3). We evaluate the pre-trained model on the English test set of XNLI (since the MultiNLI test set is not publicly available) as a baseline. Since MultiNLI is already split into genres, we use each genre as a task within MAML. We then include either the training set (5 genres) or the development set (10 genres) during meta-learning (similar to Step 2 in X-MAML).
+
+In the last phase (similar to Step 3 in X-MAML), we first initialise the model parameters with those learned by MAML. We then continue to fine-tune the model using the development set of MultiNLI and report the accuracy on the English test set of XNLI. We proportionally select sub-samples  $x = [1\%, 2\%, 3\%, 5\%, 10\%, 20\%, 50\%, 100\%]$  from the training data (with random sampling).
+
+<table><tr><td rowspan="2">x%</td><td rowspan="2">Baseline</td><td colspan="2">MAML</td></tr><tr><td>\(T_{Train}\)</td><td>\(T_{Dev}\)</td></tr><tr><td>1</td><td>38.60</td><td>49.78</td><td>50.92</td></tr><tr><td>2</td><td>37.80</td><td>48.58</td><td>50.66</td></tr><tr><td>3</td><td>47.09</td><td>51.40</td><td>52.85</td></tr><tr><td>5</td><td>49.88</td><td>52.22</td><td>51.40</td></tr><tr><td>10</td><td>51.02</td><td>52.51</td><td>53.95</td></tr><tr><td>20</td><td>59.14</td><td>61.38</td><td>58.16</td></tr><tr><td>50</td><td>63.37</td><td>63.85</td><td>61.74</td></tr><tr><td>100</td><td>64.35</td><td>64.99</td><td>64.61</td></tr></table>
+
+Table 1: Test accuracies with different settings of MAML on MultiNLI.  $\times \%$  : the percentage of training samples. Baseline: The test accuracy of trained ESIM using  $\times \%$  of training data. MAML: The test accuracy of ESIM after meta-learning, where  $\mathcal{T}_{\text{Train}}$ : 5 tasks are defined in MAML using the training set, and  $\mathcal{T}_{\text{Dev}}$ : 10 tasks are included in MAML using the development set. Bold font indicates best results for the various proportions of the used training data.
+
+![](images/4cc8fe09fd611084c11f865d9e05724435efbbff12079abcecab9b6d2967c387.jpg)  
+Figure 1: Differences in performance in terms of accuracy scores on the test set for zero-shot X-MAML on XNLI using the MultiBERT model. Rows correspond to target and columns to auxiliary languages used in X-MAML. Numbers on the off-diagonal indicate performance differences between X-MAML and the baseline model in the same row. The coloring scheme indicates the differences in performance (e.g., blue for large improvement).
+
+The results obtained by training on the corresponding proportions  $(x\%)$  of the MultiNLI dataset using ESIM (as the learner model M) are shown in Table 1. We observe that for both settings (i.e., MAML on the training (5 tasks) and on the development set (10 tasks)), the performance of all models (including baselines) improve as more instances become available. However, the effectiveness of MAML is larger when only limited training data is available (improving by  $12\%$  in accuracy when  $2\%$  of the data is available on the development set).
+
+# 4.2 Zero- and Few-Shot Cross-Lingual NLI
+
+Zero-Shot Learning: In this set of experiments, we employ the proposed framework (i.e., X-MAML) within a zero-shot setup, in which we do not fine-tune after the meta-learning step. We report the impact of meta-learning for each target language as a difference in accuracy with and without meta-learning on top of the baseline model (Multi-BERT) on the test set (Fig. 1). Each column corresponds to the performance of Multi-BERT after meta-learning with a single auxiliary language, and evaluation on the target language of the XNLI test set. Overall, we observe that our zero-shot approach with X-MAML outperforms the baseline model without MAML and results reported by De
+
+vlin et al. (2019). This way, we improve the state-of-the-art performance for zero-shot cross-lingual NLI (in several languages for up to  $+3.6\%$  in accuracy, e.g., Hindi (hi) as target and Urdu (ur) as auxiliary language). For the exact accuracy scores, we refer to Table 5 in the Appendix. We hypothesise that the degree of typological commonalities among the languages has an effect (i.e., positive or negative) on the performance of X-MAML. It can be observed that the proposed learning approach provides positive impacts across most of the target languages. However, including Swahili (sw) as an auxiliary language in X-MAML is not beneficial for the performance on the other target languages. It is worth noting that we experimented by just training the model using an auxiliary language, instead of performing meta-learning (step 2). From this experiment, we observe that meta-learning has a strongly positive effect on predictive performance (see also Fig. 2 in the Appendix).
+
+In Table 2, we include the original baseline performances reported in Devlin et al. (2019)<sup>3</sup> and Wu and Dredze (2019). We report the average and maximum performance by using one auxiliary language for each target language. We also report
+
+<table><tr><td></td><td>en</td><td>fr</td><td>es</td><td>de</td><td>el</td><td>bg</td><td>ru</td><td>tr</td><td>ar</td><td>vi</td><td>th</td><td>zh</td><td>hi</td><td>sw</td><td>ur</td><td>avg</td></tr><tr><td colspan="17">Zero-shot cross-lingual transfer</td></tr><tr><td>Devlin et al. (2019)</td><td>81.4</td><td>-</td><td>74.3</td><td>70.5</td><td>-</td><td>-</td><td>-</td><td>-</td><td>62.1</td><td>-</td><td>-</td><td>63.8</td><td>-</td><td>-</td><td>58.35</td><td>-</td></tr><tr><td>Wu and Dredze (2019)</td><td>82.1</td><td>73.8</td><td>74.3</td><td>71.1</td><td>66.4</td><td>68.9</td><td>69.0</td><td>61.6</td><td>64.9</td><td>69.5</td><td>55.8</td><td>69.3</td><td>60.0</td><td>50.4</td><td>58.0</td><td>66.3</td></tr><tr><td>Multi-BERT (Our baseline)</td><td>81.36</td><td>73.45</td><td>73.85</td><td>69.74</td><td>65.73</td><td>67.82</td><td>67.94</td><td>59.04</td><td>64.63</td><td>70.12</td><td>52.46</td><td>68.90</td><td>58.56</td><td>47.58</td><td>58.70</td><td>65.33</td></tr><tr><td colspan="17">X-MAML (One aux. lang.)</td></tr><tr><td>AVG</td><td>81.69</td><td>73.86</td><td>74.43</td><td>71.00</td><td>67.16</td><td>68.39</td><td>68.90</td><td>60.41</td><td>65.33</td><td>70.95</td><td>54.08</td><td>70.09</td><td>60.51</td><td>47.97</td><td>59.94</td><td>-</td></tr><tr><td>MAX</td><td>82.09</td><td>74.42</td><td>75.07</td><td>71.83</td><td>67.95</td><td>69.45</td><td>70.19</td><td>61.20</td><td>66.05</td><td>71.82</td><td>55.39</td><td>71.11</td><td>62.20</td><td>49.76</td><td>61.51</td><td>67.33</td></tr><tr><td>hi → X</td><td>81.88</td><td>74.17</td><td>74.81</td><td>71.59</td><td>67.95</td><td>68.86</td><td>69.44</td><td>60.93</td><td>65.86</td><td>71.57</td><td>55.26</td><td>70.59</td><td>-</td><td>47.12</td><td>61.51</td><td>-</td></tr><tr><td>X-MAML (Two aux. lang.)</td><td>(hi,de)</td><td>(hi,ar)</td><td>(fr,de)</td><td>(bg,zh)</td><td>(ur,ru)</td><td>(hi,ru)</td><td>(de,bg)</td><td>(ur,sw)</td><td>(el,tr)</td><td>(de,bg)</td><td>(bg,tr)</td><td>(ru,el)</td><td>(ur,ru)</td><td>(el,tr)</td><td>(hi,de)</td><td></td></tr><tr><td>(l1,l2) → X</td><td>82.59</td><td>75.69</td><td>75.97</td><td>73.45</td><td>69.16</td><td>71.42</td><td>71.44</td><td>62.57</td><td>67.19</td><td>72.63</td><td>62.57</td><td>73.13</td><td>63.53</td><td>50.42</td><td>62.93</td><td>68.98</td></tr><tr><td colspan="17">Few-Shot learning</td></tr><tr><td>Multi-BERT (Our baseline)</td><td>81.94</td><td>75.39</td><td>75.79</td><td>73.25</td><td>69.54</td><td>71.60</td><td>70.84</td><td>64.85</td><td>67.37</td><td>73.23</td><td>61.18</td><td>73.93</td><td>64.37</td><td>57.82</td><td>63.71</td><td>69.65</td></tr><tr><td colspan="17">X-MAML (One aux. lang.)</td></tr><tr><td>AVG</td><td>82.22</td><td>75.24</td><td>76.06</td><td>73.34</td><td>69.97</td><td>71.80</td><td>71.28</td><td>64.76</td><td>67.82</td><td>73.41</td><td>61.57</td><td>74.02</td><td>64.83</td><td>58.02</td><td>63.66</td><td>-</td></tr><tr><td>MAX</td><td>82.39</td><td>75.32</td><td>76.18</td><td>73.46</td><td>70.03</td><td>71.94</td><td>71.45</td><td>64.92</td><td>67.95</td><td>73.52</td><td>61.74</td><td>74.21</td><td>64.97</td><td>58.23</td><td>63.81</td><td>70.01</td></tr><tr><td>sw → X</td><td>82.24</td><td>75.31</td><td>75.94</td><td>73.34</td><td>69.98</td><td>71.77</td><td>71.31</td><td>64.89</td><td>67.87</td><td>73.38</td><td>61.5</td><td>73.99</td><td>64.94</td><td>-</td><td>63.63</td><td>-</td></tr><tr><td>X-MAML (Two aux. lang.)</td><td>(ar,ru)</td><td>(ru,th)</td><td>(ru,th)</td><td>(el,hi)</td><td>(sw,vi)</td><td>(ar,zh)</td><td>(de,tr)</td><td>(es,sw)</td><td>(bg,hi)</td><td>(bg,ru)</td><td>(el,vi)</td><td>(ar,th)</td><td>(sw,vi)</td><td>(ar,tr)</td><td>(en,ru)</td><td></td></tr><tr><td>(l1,l2) → X</td><td>82.71</td><td>75.97</td><td>76.51</td><td>74.07</td><td>70.66</td><td>72.77</td><td>72.12</td><td>65.69</td><td>68.4</td><td>73.87</td><td>62.5</td><td>74.85</td><td>65.75</td><td>59.94</td><td>64.59</td><td>70.69</td></tr><tr><td colspan="17">Machine translate at test (TRANSLATE-TEST)</td></tr><tr><td>Devlin et al. (2019)</td><td>81.4</td><td>-</td><td>74.9</td><td>74.4</td><td>-</td><td>-</td><td>-</td><td>-</td><td>70.4</td><td>-</td><td>-</td><td>70.1</td><td>-</td><td>-</td><td>62.1</td><td>-</td></tr><tr><td colspan="17">Machine translate at training (TRANSLATE-TRAIN)</td></tr><tr><td>Wu and Dredze (2019)</td><td>82.1</td><td>76.9</td><td>78.5</td><td>74.8</td><td>72.1</td><td>75.4</td><td>74.3</td><td>70.6</td><td>70.8</td><td>67.8</td><td>63.2</td><td>76.2</td><td>65.3</td><td>65.3</td><td>60.6</td><td>71.6</td></tr></table>
+
+Table 2: Accuracy results on the XNLI test set for zero- and few-shot X-MAML. Columns indicate the target languages. The models of Devlin et al. (2019) and Wu and Dredze (2019) are also Multi-BERT models. For our Multi-BERT baseline model for (i) zero-shot learning, we evaluate the pre-trained model on the test set of the target language; and for (ii) few-shot learning, we fine-tune the model on the development set and evaluate on the test set of the target language. The avg column indicates row-wise average accuracy. We also report the average (AVG) and maximum (MAX) performance by using one auxiliary language for each target language.  $(l_{1}, l_{2})$  are the most beneficial auxiliary languages for X-MAML in improving the test accuracy of each target language  $X$ . In TRANSLATE-TEST (Devlin et al., 2019), the target language test data is translated to English and then the model is fine-tuned on English. In TRANSLATE-TRAIN (Wu and Dredze, 2019), the English training data is translated to the target language and the model is fine-tuned using the translated data.
+
+the performance of X-MAML by also using Hindi (which is the most effective auxiliary language for the zero-shot setting, as shown in Fig. 1). We suspect that this is because of the typological similarities between Hindi (hi) and other languages. Furthermore, by using two auxiliary languages in X-MAML results to the largest benefit in our zero-shot experiments.
+
+Few-Shot Learning: For few-shot learning, meta-learning in X-MAML (Step 3) is performed by fine-tuning on the development set (2.5k instances) of target languages, and then evaluating on the test set. Detailed ablation results are presented in the Appendix (Table 6 and Fig. 4). In Table 2, we compare X-MAML results with one or two auxiliary languages to the external and internal baselines. We also showcase the performance using specifically Swahili (sw), the overall most effective auxiliary language for meta-learning with Multi-BERT in the few-shot learning setting.
+
+In addition, we report results from Devlin et al. (2019) that use machine translation at test time (TRANSLATE-TEST) and results from Wu and Dredze (2019) that use machine translation at training time (TRANSLATE-TRAIN). Note that, using X-MAML, we are able to avoid the machine translation step (TRANSLATE-TEST) from the target language into English. The results also indicate that X-MAML boosts Multi-BERT performance on XNLI. It is worthwhile mentioning that Multi-BERT in the TRANSLATE-TRAIN setup outperforms few-shot X-MAML, however, we only use 2k development examples from the target languages, whereas in the aforementioned work, 433k translated sentences are used for fine-tuning.
+
+# 4.3 Zero-Shot Cross-Linguual QA
+
+We use a similar approach for cross-lingual QA on the MLQA dataset. Zero-shot results on MLQA are shown in Table 3. We compare our results to those
+
+reported in two benchmark papers, Hu et al. (2020) and Liang et al. (2020). We also report our own baselines for the task. The baselines are provided by training each base model on the SQuAD v1.1 train set (see Step 1 in Section 3) and evaluating on the test set of MLQA. All target languages benefit from meta-learning with at least one of the auxiliary languages. Using two auxiliary languages in X-MAML further improves results. Overall, zero-shot learning models with X-MAML outperform both internal and external baselines. The improvement is  $+1.04\%$ ,  $+0.89\%$  and  $+1.47\%$  in average  $\mathrm{F_1}$  score compared to XLM-15, XLM- $\mathbf{R}_{base}$  and XLM- $\mathbf{R}_{large}$ , respectively.
+
+We also evaluate on the less widely used crosslingual QA dataset X-WikiRE (Abdou et al., 2019) for which we observe similar result trends, and  $0.55\%$  improvement in terms of average  $\mathrm{F_1}$  score on zero-shot QA. More details can be found in the Appendix (Section A.1).
+
+# 5 Related Work
+
+The main motivation for this work is the low availability of labelled training datasets for most of the world's languages. To alleviate this issue, a number of methods, including the so-called few-shot learning approaches have been proposed. Few-shot learning methods have initially been introduced within the area of image classification (Vinyals et al., 2016; Ravi and Larochelle, 2017; Finn et al., 2017), but have recently also been applied to NLP tasks such as relation extraction (Han et al., 2018), text classification (Yu et al., 2018; Rethmeier and Augenstein, 2020) and machine translation (Gu et al., 2018). Specifically, in NLP, these few-shot learning approaches include: (i) the transformation of the problem into a different task (e.g., relation extraction is transformed to question answering (Levy et al., 2017; Abdou et al., 2019)); or (ii) meta-learning (Andrychowicz et al., 2016; Finn et al., 2017).
+
+Meta-Learning: Meta-learning or learning-to-learn has recently received a lot of attention from the NLP community. First-order MAML has been applied to the task of machine translation (Gu et al., 2018), where they propose to use meta-learning for improving the machine translation performance for low-resource languages by learning to adapt to target languages based on multilingual high-resource languages. However, in the proposed framework, they include 18 high-resource languages as auxil
+
+ary languages and five diverse low-resource languages as target languages. In our work, we assume access to only English as a high-resource language. For the task of dialogue generation, Qian and Yu (2019) address domain adaptation using meta-learning. Dou et al. (2019) explore MAML variants thereof for low-resource NLU tasks in the GLUE dataset (Wang et al., 2018). They consider different high-resource NLU tasks such as MultiNLI (Williams et al., 2018) and QNLI (Rajpurkar et al., 2016) as auxiliary tasks to learn meta-parameters using MAML. They then fine-tune the low-resource tasks using the adapted parameters from the meta-learning phase. All the above-mentioned works on meta-learning in NLP assume that there are multiple high-resource tasks or languages, which are then adapted to new target tasks or languages with a handful of training samples. However, in a cross-lingual NLI and QA setting, the available high-resource language is usually only English. Our work thus fills an important gap in the literature, as we only require a single source language.
+
+Cross-Lingual NLU: Cross-lingual learning has a fairly short history in NLP, and has mainly been restricted to traditional NLP tasks, such as PoS tagging, morphological inflection and parsing. In contrast to these tasks, which have seen much cross-lingual attention (Plank et al., 2016; Bjerva, 2017; Kementchedjhieva et al., 2018; de Lhoneux et al., 2018), there has been relatively little work on cross-lingual NLU, partly due to lack of benchmark datasets. Existing work has mainly been focused on NLI (Agic and Schluter, 2018; Conneau et al., 2018; Zhao et al., 2020), and to a lesser degree on RE (Faruqui and Kumar, 2015; Verga et al., 2016) and QA (Abdou et al., 2019; Lewis et al., 2020). Previous research generally reports that cross-lingual learning is challenging and that it is hard to beat a machine translation baseline (e.g., Conneau et al. (2018)). Such a baseline is for instance suggested by Faruqui and Kumar (2015), where the text in the target language is automatically translated to English. We achieve competitive performance compared to a machine translation baseline (for XNLI), and propose a method that requires no training instances for the target task in the target language. Furthermore, our method is model agnostic, and can be used to extend any pre-existing model.
+
+<table><tr><td></td><td colspan="2">Model</td><td>en</td><td>ar</td><td>de</td><td>es</td><td>hi</td><td>vi</td><td>zh</td><td>avg</td></tr><tr><td rowspan="3">XLM</td><td>|</td><td>Our baseline</td><td>69.80</td><td>48.95</td><td>52.64</td><td>58.15</td><td>46.67</td><td>48.46</td><td>42.64</td><td>52.47</td></tr><tr><td>|</td><td>(One aux. lang.)l → X</td><td>69.39ar</td><td>48.45hi</td><td>53.04es</td><td>57.68en</td><td>46.90zh</td><td>49.79zh</td><td>44.36hi</td><td>52.80</td></tr><tr><td>|</td><td>(Two aux. lang.) (l1,l2) → X</td><td>68.88(es,ar)</td><td>49.76(vi,zh)</td><td>53.18(vi,zh)</td><td>58.00(en,zh)</td><td>48.43(vi,zh)</td><td>50.86(hi,zh)</td><td>45.44(es,hi)</td><td>53.51</td></tr><tr><td rowspan="3">XLM-Rbase</td><td>|</td><td>Liang et al. (2020)Our baseline</td><td>80.180.38</td><td>56.457.23</td><td>62.163.08</td><td>67.967.91</td><td>60.561.46</td><td>67.167.14</td><td>61.462.73</td><td>65.165.70</td></tr><tr><td>|</td><td>(One aux. lang.)l → X</td><td>80.19vi</td><td>57.97hi</td><td>63.57ar</td><td>67.46vi</td><td>61.70vi</td><td>67.97hi</td><td>64.01hi</td><td>66.12</td></tr><tr><td>|</td><td>(Two aux. lang.) (l1,l2) → X</td><td>80.31(ar,vi)</td><td>58.14(hi,vi)</td><td>64.07(ar,hi)</td><td>68.08(ar,hi)</td><td>62.67(es,ar)</td><td>68.82(ar,hi)</td><td>64.06(ar,hi)</td><td>66.59</td></tr><tr><td rowspan="3">XLM-Rlarge</td><td>|</td><td>Hu et al. (2020)Our baseline</td><td>83.583.95</td><td>66.666.09</td><td>70.170.62</td><td>74.174.59</td><td>70.670.64</td><td>7474.13</td><td>62.169.80</td><td>71.672.83</td></tr><tr><td>|</td><td>(One aux. lang.)l → X</td><td>84.31ar</td><td>66.61hi</td><td>70.84ar</td><td>74.32hi</td><td>70.94vi</td><td>74.84ar</td><td>70.74hi</td><td>73.23</td></tr><tr><td>|</td><td>(Two aux. lang.) (l1,l2) → X</td><td>84.60(hi,vi)</td><td>66.95(hi,vi)</td><td>71.00(ar,vi)</td><td>74.62(en,vi)</td><td>70.93(ar,vi)</td><td>74.73(es,hi)</td><td>70.29(en,vi)</td><td>74.30</td></tr></table>
+
+Table 3:  $\mathrm{F}_1$  scores (average over 10 runs) on the MLQA test set using zero-shot X-MAML. Columns indicate the target languages. The avg column indicates row-wise average  $\mathrm{F}_1$  score. We also report the most beneficial auxiliary language(s) for X-MAML in improving the test  $\mathrm{F}_1$  of each target language.
+
+# 6 Discussion and Analysis
+
+Cross-Linguual Transfer: Somewhat surprisingly, we find that cross-lingual transfer with meta-learning yields improved results even when languages strongly differ from one another. For instance, for zero-shot meta-learning on XNLI, we observe gains for almost all auxiliary languages, with the exception of Swahili (sw). This indicates that the meta-parameters learned with X-MAML are sufficiently language agnostic, as we otherwise would not expect to see any benefits in transferring from, e.g., Russian (ru) to Hindi (hi) (one of the strongest results in Fig. 1). This is dependent on having access to a pre-trained multilingual model such as BERT, however, using monolingual BERT (En-BERT) yields overwhelmingly positive gains in some target/auxiliary settings (see additional results in Fig. 3 in the Appendix). For few-shot learning, our findings are similar, as almost all combinations of auxiliary and target languages lead to improvements when using Multi-BERT (Fig. 4 in the Appendix). However, when we only have access to a handful of training instances as in few-shot learning, even the English BERT model mostly leads to improvements in this setting (see additional results in Fig. 5 in the Appendix).
+
+Typological Correlations: In order to better explain our results for cross-lingual zero-shot and few-shot learning, we investigate typological fea
+
+tures, and their overlap between target and auxiliary languages. We evaluate on the World Atlas of Language Structure (WALS, Dryer and Haspelmath (2013)), which is the largest openly available typological database. It comprises approximately 200 linguistic features with annotations for more than 2500 languages, which have been made by expert typologists through study of grammars and field work. We draw inspiration from previous work (Bjerva and Augenstein, 2018a,b; Bjerva et al., 2019a,b,c; Zhao et al., 2020) which attempts to predict typological features based on language representations learned under various NLP tasks. Similarly, we experiment with two conditions: (i) we attempt to predict typological features based on the mutual gain/loss in performance using X-MAML; (ii) we investigate whether sharing between two typologically similar languages is beneficial for performance using X-MAML. We train a simple logistic regression classifier per condition above, for each WALS feature. In the first condition (i), the task is to predict the exact WALS feature value of a language, given the change in accuracy in combination with other languages. In the second condition (ii), the task is to predict whether a main and auxiliary language have the same WALS feature value, given the change in accuracy when the two languages are used in X-MAML. We compare with two simple baselines, one based on always predicting the most frequent feature value in the
+
+training set, and one based on predicting feature values with respect to the distribution of feature values in the training set. We then investigate whether any features could be consistently predicted above baseline levels, given different test-training splits. We apply a simple paired t-test to compare our models predictions to the baselines. As we are running a large number of tests (one per WALS feature), we apply Bonferroni correction, changing our cut-off  $p$ -value from  $p = 0.05$  to  $p = 0.00025$ .
+
+We first investigate few-shot X-MAML, when using Multi-BERT, as reported in Table 6 (Appendix). We find that languages sharing the feature value for WALS feature 67A The Future Tense are beneficial to each other. This feature encodes whether or not a language has an inflectional marking of future tense, and can be considered to be a morphosyntactic feature. We next look at zero-shot X-MAML with Multi-BERT, as reported in Table 5 (Appendix). For this case, we find that languages sharing a feature value for the WALS feature 25A Locus of Marking: Whole-language Typology typically help each other. This feature describes whether the morphosyntactic marking in a language is on the syntactic heads or dependents of a phrase. For example en, de, ru, and zh are 'dependent-marking' in this feature. And if we look at the results in Fig. 1, they have the largest mutual gains from each other during the zero-shot X-MAML. In both cases, we thus find that languages with similar morphosyntactic properties can be beneficial to one another when using X-MAML.
+
+# 7 Conclusion
+
+In this work, we show that meta-learning can be used to effectively leverage training data from an auxiliary language for zero-shot and few-shot cross-lingual transfer. We evaluated this on two challenging NLU tasks (NLI and QA), and on a total of 15 languages. We are able to improve the performance of state-of-the-art baseline models for (i) zero-shot XNLI, and (ii) zero-shot QA on the MLQA dataset. Furthermore, we show in a typological analysis that languages which share certain morphosyntactic features tend to benefit from this type of transfer. Future studies will extend this work to other cross-lingual NLP tasks and more languages.
+
+# Acknowledgements
+
+This research has received funding from the Swedish Research Council under grant agreement
+
+No 2019-04129, as well as the Research Foundation - Flanders (FWO). This work was also funded by UiO: Energy to support international mobility. We are grateful to the Nordic Language Processing Laboratory (NLPL) for providing access to its supercluster infrastructure.
+
+# References
+
+Mostafa Abdou, Cezar Sas, Rahul Aralikatte, Isabelle Augenstein, and Anders Søgaard. 2019. X-WikiRE: A Large, Multilingual Resource for Relation Extraction as Machine Comprehension. In Proceedings of the 2nd Workshop on Deep Learning Approaches for Low-Resource NLP (DeepLo 2019), pages 265–274, Hong Kong, China. Association for Computational Linguistics.  
+Zeljko Agic and Natalie Schluter. 2018. Baselines and Test Data for Cross-Linguial Inference. In LREC. European Language Resources Association (ELRA).  
+Marcin Andrychowicz, Misha Denil, Sergio Gomez Colmenarejo, Matthew W. Hoffman, David Pfau, Tom Schaul, and Nando de Freitas. 2016. Learning to learn by gradient descent by gradient descent. In Advances in Neural Information Processing Systems, pages 3981-3989.  
+Emily M Bender. 2019. The #BenderRule: On Naming the Languages We Study and Why It Matters.  
+Johannes Bjerva. 2017. One Model to Rule them all - Multitask and Multilingual Modelling for Lexical Analysis. Ph.D. thesis, University of Groningen.  
+Johannes Bjerva and Isabelle Augenstein. 2018a. From Phonology to Syntax: Unsupervised Linguistic Typology at Different Levels with Language Embeddings. 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 907-916, New Orleans, Louisiana. Association for Computational Linguistics.  
+Johannes Bjerva and Isabelle Augenstein. 2018b. Tracking Typological Traits of Uralic Languages in Distributed Language Representations. In Proceedings of the Fourth International Workshop on Computational Linguistics of Uralic Languages, pages 76-86, Helsinki, Finland. Association for Computational Linguistics.  
+Johannes Bjerva, Yova Kementchedjhieva, Ryan Coterell, and Isabelle Augenstein. 2019a. A Probabilistic Generative Model of Linguistic Typology. 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 1529-1540, Minneapolis, Minnesota. Association for Computational Linguistics.
+
+Johannes Bjerva, Yova Kementchedjhieva, Ryan Coterell, and Isabelle Augenstein. 2019b. Uncovering Probabilistic Implications in Typological Knowledge Bases. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3924-3930, Florence, Italy. Association for Computational Linguistics.  
+Johannes Bjerva, Robert Östling, Maria Han Veiga, Jörg Tiedemann, and Isabelle Augenstein. 2019c. What Do Language Representations Really Represent? Computational Linguistics, 45(2):381-389.  
+Qian Chen, Xiaodan Zhu, Zhen-Hua Ling, Si Wei, and Hui Jiang. 2016. Enhancing and Combining Sequential and Tree LSTM for Natural Language Inference. CoRR, abs/1609.06038.  
+Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishray Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised Cross-lingual Representation Learning at Scale. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.  
+Alexis Conneau and Guillaume Lample. 2019. Cross-lingual Language Model Pretraining. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, 8-14 December 2019, Vancouver, BC, Canada, pages 7057-7067.  
+Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating Cross-lingual Sentence Representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2475–2485. 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.  
+Zi-Yi Dou, Keyi Yu, and Antonios Anastasopoulos. 2019. Investigating Meta-Learning Algorithms for Low-Resource Natural Language Understanding 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 1192-1197, Hong Kong, China. Association for Computational Linguistics.  
+Matthew S. Dryer and Martin Haspelmath, editors. 2013. WALS Online. Max Planck Institute for Evolutionary Anthropology, Leipzig.
+
+David M. Eberhard, Gary F. Simons, and Charles D. Fennig. 2019. Ethnologue: Languages of the World. https://www.ethnologue.com/statistics/size. Accessed: 2019-05-25.  
+Manaal Faruqui and Shankar Kumar. 2015. Multilingual Open Relation Extraction Using Cross-lingual Projection. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1351-1356. Association for Computational Linguistics.  
+Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 1126-1135. JMLR.org.  
+Jiatao Gu, Yong Wang, Yun Chen, Victor O. K. Li, and Kyunghyun Cho. 2018. Meta-Learning for Low-Resource Neural Machine Translation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3622-3631, Brussels, Belgium. Association for Computational Linguistics.  
+Xu Han, Hao Zhu, Pengfei Yu, Ziyun Wang, Yuan Yao, Zhiyuan Liu, and Maosong Sun. 2018. FewRel: A Large-Scale Supervised Few-Shot Relation Classification Dataset with State-of-the-Art Evaluation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4803-4809.  
+Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan First, and Melvin Johnson. 2020. XTREME: A Massively Multilingual Multitask Benchmark for Evaluating Cross-lingual Generalization. CoRR, abs/2003.11080.  
+Pratik Joshi, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. The State and Fate of Linguistic Diversity and Inclusion in the NLP World. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6282-6293, Online. Association for Computational Linguistics.  
+Yova Kementchedjhieva, Johannes Bjerva, and Isabelle Augenstein. 2018. Copenhagen at CoNLL-SIGMORPHON 2018: Multilingual Inflection in Context with Explicit Morphosyntactic Decoding. In Proceedings of the CoNLL-SIGMORPHON 2018 Shared Task: Universal Morphological Reinflation, pages 93-98, Brussels. Association for Computational Linguistics.  
+Diederik P. Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic Optimization. arXiv preprint arXiv:1412.6980.  
+Gregory Koch, Richard Zemel, and Ruslan Salakhutdinov. 2015. Siamese Neural Networks for One-shot Image Recognition. In ICML deep learning workshop.
+
+Souvik Kundu and Hwee Tou Ng. 2018. A Nil-Aware Answer Extraction Framework for Question Answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4243-4252, Brussels, Belgium. Association for Computational Linguistics.  
+Omer Levy, Minjoon Seo, Eunsol Choi, and Luke Zettlemoyer. 2017. Zero-Shot Relation Extraction via Reading Comprehension. In Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 333-342, Vancouver, Canada. Association for Computational Linguistics.  
+Patrick Lewis, Barlas Oguz, Rudy Rinott, Sebastian Riedel, and Holger Schwenk. 2020. MLQA: Evaluating Cross-lingual Extractive Question Answering. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.  
+Miryam de Lhoneux, Johannes Bjerva, Isabelle Augenstein, and Anders Søgaard. 2018. Parameter sharing between dependency parsers for related languages. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4992-4997, Brussels, Belgium. Association for Computational Linguistics.  
+Yaobo Liang, Nan Duan, Yeyun Gong, Ning Wu, Fenfei Guo, Weizhen Qi, Ming Gong, Linjun Shou, Daxin Jiang, Guihong Cao, Xiaodong Fan, Bruce Zhang, Rahul Agrawal, Edward Cui, Sining Wei, Taroon Bharti, Jiun-Hung Chen, Winnie Wu, Shuguang Liu, Fan Yang, and Ming Zhou. 2020. XGLUE: A New Benchmark Dataset for Crosslingual Pre-training, Understanding and Generation. ArXiv, abs/2004.01401.  
+Alex Nichol, Joshua Achiam, and John Schulman. 2018. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999.  
+Barbara Plank, Anders Søgaard, and Yoav Goldberg. 2016. Multilingual Part-of-Speech Tagging with Bidirectional Long Short-Term Memory Models and Auxiliary Loss. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), volume 2, pages 412-418.  
+Kun Qian and Zhou Yu. 2019. Domain Adaptive Dialog Generation via Meta Learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2639-2649, Florence, Italy. Association for Computational Linguistics.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100, 000+ Questions for Machine Comprehension of Text. CoRR, abs/1606.05250.  
+Sachin Ravi and Hugo Larochelle. 2017. Optimization as a model for few-shot learning. In International Conference on Learning Representations (ICLR). OpenReview.net.
+
+Nils Rethmeier and Isabelle Augenstein. 2020. Long-Tail Zero and Few-Shot Learning via Contrastive Pretraining on and for Small Data. arXiv preprint arXiv:2010.01061.  
+Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy P. Lillicrap. 2016. Meta-Learning with Memory-Augmented Neural Networks. In ICML, volume 48 of JMLR Workshop and Conference Proceedings, pages 1842-1850. JMLR.org.  
+Patrick Verga, David Belanger, Emma Strubell, Benjamin Roth, and Andrew McCallum. 2016. Multilingual Relation Extraction using Compositional Universal Schema. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 886-896. Association for Computational Linguistics.  
+Oriol Vinyals, Charles Blundell, Timothy Lillicrap, k—oray kavukcuoglu, and Daan Wierstra. 2016. Matching Networks for One Shot Learning. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems 29, pages 3630-3638. Curran Associates, Inc.  
+Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. 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.  
+Shijie Wu and Mark Dredze. 2019. Beto, Bentz, Becas: The Surprising Cross-Lingual Effectiveness of BERT. CoRR, abs/1904.09077.  
+Mo Yu, Xiaoxiao Guo, Jinfeng Yi, Shiyu Chang, Saloni Potdar, Yu Cheng, Gerald Tesauro, Haoyu Wang, and Bowen Zhou. 2018. Diverse Few-Shot Text Classification with Multiple Metrics. 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 1206-1215.  
+Wei Zhao, Steffen Eger, Johannes Bjerva, and Isabelle Augenstein. 2020. Inducing Language-Agnostic Multilingual Representations. arXiv preprint arXiv:2008.09112.
+
+# A Appendices
+
+# A.1 X-MAML using X-WikiRE dataset
+
+X-WikiRE: Levy et al. (2017) frame the Relation Extraction (RE) task as a QA problem using pre-defined natural language question templates. For example, a relation type such as author is transformed to at least one language question template (e.g., who is the author of x?, where x is an entity). Building on the work of Levy et al. (2017), a new dataset (X-WikiRE) is introduced for multilingual QA-based relation extraction in five languages (i.e., English, French, Spanish, Italian and German) by Abdou et al. (2019). Each instance in the dataset includes a question, a context, and an answer. The question is a transformation of a target relation and the context may contain the answer. If the answer is not present, it is marked as NIL. In this task, we evaluate the performance of our method on the UnENT setting of the X-WikiRE dataset, where the goal is to generalise to unseen entities. For the evaluation, we use  $\mathrm{F}_1$  scores (for questions with valid answers) similar to Kundu and Ng (2018).
+
+QA experiments: We use the Nil-Aware Answer Extraction Framework (NAMANDA, Kundu and Ng (2018)) as the base model M in X-MAML for our QA experiments. NAMANDA encodes the question and context sequences to compute a similarity matrix. It creates evidence vectors through joint encoding of question and context and applies multi-factor self-attentive encoding. Finally, the evidence vectors are decomposed to output either the answer to the question or NIL. We set the parameters to the default values (as in the original work) for the training and evaluation phases. The NAMANDA model M is pre-trained on the full English training set (1M instances - see Step 1 in our training algorithm). The model M is further used by our meta-learning step to adapt the pre-trained QA model. We then evaluate how well the English model has been adapted by each of the auxiliary language through X-MAML via performing either few- or zero-shot learning. In few-shot X-MAML, the meta-learned M is fine-tuned on the development set (1k instances) of other languages (i.e., fr, es, it and de). For both few- and zero-shot learning, we evaluate on the 10k test set of each of the target languages. Following the work of Abdou et al. (2019), the Multi-BERT model is used to
+
+jointly encode text for different languages in the QA model.
+
+Zero- and Few-Shot Cross-Lingual QA: We use a similar approach for cross-lingual QA on the X-WikiRE dataset. Table 4 shows the results of both zero- and few-shot X-MAML for the UnENT part (i.e., generalise to unseen entities) of the X-WikiRE dataset. We compare our results for the UnENT scenario on the X-WikiRE dataset to those reported in the original paper. All of the target languages benefit from at least one of the auxiliary languages by adapting the model using X-MAML, highlighting the benefits of this method. We were not able to directly reproduce the result for the zero-shot scenario of the original paper, thus we also report our own baseline for the task. We find that: (i) our zero-shot results with X-MAML improve on those without meta-learning (i.e., baselines); and (ii) we outperform Abdou et al. (2019) for the UnENT scenario of zero-shot cross-lingual QA. Furthermore, for the few-shot scenario, adapting the QA model using few-shot X-MAML with only 1k development data outperforms their cross-lingual transfer model where Abdou et al. (2019) use 10k in the fine-tuning phase.
+
+<table><tr><td rowspan="2" colspan="2"></td><td colspan="4">Auxiliary language</td><td rowspan="2">Baseline</td><td colspan="4">Abdou et al. (2019)</td></tr><tr><td>es</td><td>fr</td><td>it</td><td>de</td><td colspan="2">BERT</td><td colspan="2">fastText</td></tr><tr><td rowspan="4">zero-shot</td><td>es</td><td>-</td><td>49.01</td><td>50.11</td><td>50.59</td><td>49.85</td><td colspan="2">5.49</td><td colspan="2">16.17</td></tr><tr><td>fr</td><td>52.20</td><td>-</td><td>52.13</td><td>51.96</td><td>51.72</td><td colspan="2">17.42</td><td colspan="2">15.28</td></tr><tr><td>it</td><td>50.53</td><td>50.65</td><td>-</td><td>50.58</td><td>50.58</td><td colspan="2">10.70</td><td colspan="2">4.44</td></tr><tr><td>de</td><td>49.92</td><td>48.78</td><td>48.63</td><td>-</td><td>48.98</td><td colspan="2">2.87</td><td colspan="2">14.09</td></tr><tr><td colspan="2"></td><td colspan="5">1k</td><td>1k</td><td>10k</td><td>1k</td><td>10k</td></tr><tr><td rowspan="4">few-shot</td><td>es</td><td>-</td><td>78.09</td><td>78.33</td><td>77.89</td><td>78.26</td><td>42.97</td><td>71.66</td><td>65.78</td><td>77.99</td></tr><tr><td>fr</td><td>80.68</td><td>-</td><td>80.81</td><td>80.74</td><td>80.67</td><td>42.69</td><td>72.43</td><td>65.67</td><td>74.15</td></tr><tr><td>it</td><td>82.04</td><td>81.76</td><td>-</td><td>81.77</td><td>81.78</td><td>56.25</td><td>80.06</td><td>64.02</td><td>83.45</td></tr><tr><td>de</td><td>78.29</td><td>78.48</td><td>78.66</td><td>-</td><td>78.63</td><td>56.01</td><td>70.43</td><td>62.47</td><td>72.17</td></tr></table>
+
+Table 4:  $\mathrm{F}_1$  scores (average over 10 runs) for the test set of the UnENT part of the X-WikiRE dataset using zero- and few-shot X-MAML. Baseline for (i) zero-shot learning: we evaluate the pre-trained NAMANDA model on the test set of the target language indicated in each row; and for (ii) few-shot learning: we fine-tune the model on the development set and evaluate on the test set of the target language. We report results with few-shot X-MAML with only 1k instances from the development set.
+
+![](images/301d7412ecbfc435f5d98cf87e6f442371d88f64f5086a632563fa59c213fb64.jpg)  
+Figure 2: Differences in performance in terms of accuracy scores on the test set for the zero-shot case using training (without meta-learning) on XNLI with the Multi-BERT model. Rows correspond to target and columns to auxiliary languages. Numbers on the off-diagonal indicate performance differences between training on the auxiliary languages (without meta-learning) and the baseline model in the same row. The coloring scheme indicates the differences in performance (e.g., blue for large improvement).
+
+![](images/9a91c283750bd771f5adc63c3fc4748dbf16510a21c0aee31acbc2d048a68b80.jpg)  
+Figure 3: Differences in performance in terms of accuracy scores on the test set for zero-shot X-MAML on XNLI using the En-BERT (English) model. Rows correspond to target and columns to auxiliary languages used in X-MAML. Numbers on the off-diagonal indicate performance differences between X-MAML and the baseline model in the same row. The coloring scheme indicates the differences in performance (e.g., blue for large improvement).
+
+![](images/0e704d65be805c8681e83ba163ee4e7c55e541e979269b7fab222a222f6e6615.jpg)  
+Figure 4: Differences in performance in terms of accuracy scores on the test set for few-shot X-MAML on XNLI using the Multi-BERT model. Rows correspond to target and columns to auxiliary languages used in X-MAML. Numbers on the off-diagonal indicate performance differences between X-MAML and the baseline model in the same row. The coloring scheme indicates the differences in performance (e.g., blue for large improvement).
+
+<table><tr><td rowspan="2"></td><td rowspan="2">ar</td><td rowspan="2">bg</td><td rowspan="2">de</td><td rowspan="2">el</td><td rowspan="2">en</td><td rowspan="2">es</td><td colspan="11">Auxiliary language</td><td>baseline</td></tr><tr><td>fr</td><td>hi</td><td>ru</td><td>sw</td><td>th</td><td>tr</td><td>ur</td><td>vi</td><td>zh</td><td></td><td></td><td></td></tr><tr><td>ar</td><td>-</td><td>65.76</td><td>65.48</td><td>66.05</td><td>64.41</td><td>65.27</td><td>65.24</td><td>65.86</td><td>65.31</td><td>63.66</td><td>65.25</td><td>65.58</td><td>65.56</td><td>65.84</td><td>65.32</td><td>64.63</td><td></td><td></td></tr><tr><td>bg</td><td>68.36</td><td>-</td><td>68.79</td><td>68.39</td><td>67.95</td><td>68.45</td><td>68.80</td><td>68.86</td><td>69.41</td><td>66.10</td><td>67.62</td><td>67.95</td><td>68.63</td><td>68.67</td><td>69.45</td><td>67.82</td><td></td><td></td></tr><tr><td>de</td><td>70.88</td><td>71.46</td><td>-</td><td>71.26</td><td>71.09</td><td>71.12</td><td>71.11</td><td>71.59</td><td>71.83</td><td>68.65</td><td>70.29</td><td>70.37</td><td>71.42</td><td>71.15</td><td>71.83</td><td>69.74</td><td></td><td></td></tr><tr><td>el</td><td>67.53</td><td>67.58</td><td>67.25</td><td>-</td><td>66.11</td><td>67.13</td><td>67.39</td><td>67.95</td><td>67.71</td><td>65.11</td><td>67.12</td><td>67.15</td><td>67.69</td><td>67.19</td><td>67.34</td><td>65.73</td><td></td><td></td></tr><tr><td>en</td><td>81.68</td><td>81.79</td><td>82.02</td><td>81.77</td><td>-</td><td>81.88</td><td>81.91</td><td>81.88</td><td>82.03</td><td>80.44</td><td>81.18</td><td>81.43</td><td>81.80</td><td>81.73</td><td>82.09</td><td>81.36</td><td></td><td></td></tr><tr><td>es</td><td>74.48</td><td>74.51</td><td>74.63</td><td>74.58</td><td>74.41</td><td>-</td><td>74.95</td><td>74.81</td><td>74.63</td><td>72.66</td><td>73.91</td><td>74.12</td><td>74.51</td><td>74.71</td><td>75.07</td><td>73.85</td><td></td><td></td></tr><tr><td>fr</td><td>74.13</td><td>74.02</td><td>74.22</td><td>74.11</td><td>73.75</td><td>74.18</td><td>-</td><td>74.17</td><td>74.34</td><td>71.87</td><td>73.04</td><td>73.41</td><td>74.15</td><td>74.21</td><td>74.42</td><td>73.45</td><td></td><td></td></tr><tr><td>hi</td><td>60.75</td><td>61.59</td><td>60.84</td><td>60.61</td><td>59.31</td><td>60.18</td><td>60.66</td><td>-</td><td>61.75</td><td>57.10</td><td>59.39</td><td>60.47</td><td>62.20</td><td>60.76</td><td>61.56</td><td>58.56</td><td></td><td></td></tr><tr><td>ru</td><td>68.78</td><td>69.47</td><td>69.47</td><td>68.93</td><td>68.64</td><td>68.89</td><td>69.25</td><td>69.44</td><td>-</td><td>66.11</td><td>68.18</td><td>68.72</td><td>69.52</td><td>69.02</td><td>70.19</td><td>67.94</td><td></td><td></td></tr><tr><td>sw</td><td>48.71</td><td>48.53</td><td>47.36</td><td>49.13</td><td>46.70</td><td>48.43</td><td>47.81</td><td>47.11</td><td>47.28</td><td>-</td><td>49.20</td><td>49.76</td><td>46.61</td><td>48.43</td><td>46.50</td><td>47.58</td><td></td><td></td></tr><tr><td>th</td><td>54.65</td><td>55.39</td><td>53.80</td><td>54.98</td><td>51.14</td><td>54.09</td><td>54.15</td><td>55.26</td><td>53.82</td><td>52.90</td><td>-</td><td>55.24</td><td>53.79</td><td>54.99</td><td>52.85</td><td>52.46</td><td></td><td></td></tr><tr><td>tr</td><td>60.94</td><td>61.20</td><td>60.22</td><td>61.09</td><td>58.66</td><td>60.60</td><td>60.32</td><td>60.93</td><td>60.29</td><td>59.98</td><td>60.53</td><td>-</td><td>60.82</td><td>60.68</td><td>59.47</td><td>59.04</td><td></td><td></td></tr><tr><td>ur</td><td>60.30</td><td>60.87</td><td>60.34</td><td>60.20</td><td>58.82</td><td>59.81</td><td>60.12</td><td>61.51</td><td>61.02</td><td>56.37</td><td>59.38</td><td>60.02</td><td>-</td><td>59.87</td><td>60.46</td><td>58.70</td><td></td><td></td></tr><tr><td>vi</td><td>71.27</td><td>71.56</td><td>71.32</td><td>71.14</td><td>70.35</td><td>71.22</td><td>71.42</td><td>71.57</td><td>71.73</td><td>68.11</td><td>69.87</td><td>70.53</td><td>71.43</td><td>-</td><td>71.82</td><td>70.12</td><td></td><td></td></tr><tr><td>zh</td><td>70.24</td><td>70.68</td><td>70.65</td><td>70.12</td><td>69.91</td><td>70.29</td><td>70.47</td><td>70.59</td><td>71.11</td><td>67.47</td><td>69.33</td><td>69.50</td><td>70.29</td><td>70.54</td><td>-</td><td>68.90</td><td></td><td></td></tr></table>
+
+Table 5: The performance in terms of average test accuracy for the zero-shot setting over 10 runs of X-MAML on the XNLI dataset using Multi-BERT (multilingual BERT), as base model. Each column corresponds to the performance of the Multi-BERT system after meta-learning with a single auxiliary language, and evaluation on the target language of the XNLI test set. The auxiliary language is not included during the evaluation phase. Results of the Multi-BERT model without X-MAML (baseline) are also reported.
+
+![](images/c12aaf34f8ee9c9d3fab783974bef33efbbc8b686e58c476e3093c17be998fa0.jpg)  
+Figure 5: Differences in performance in terms of accuracy scores on the test set for few-shot X-MAML on XNLI using the En-BERT (English) model. Rows correspond to target and columns to auxiliary languages used in X-MAML. Numbers on the off-diagonal indicate performance differences between X-MAML and the baseline model in the same row. The coloring scheme indicates the differences in performance (e.g., blue for large improvement).
+
+<table><tr><td rowspan="2"></td><td colspan="15">Auxiliary language</td><td rowspan="2">baseline</td></tr><tr><td>ar</td><td>bg</td><td>de</td><td>el</td><td>en</td><td>es</td><td>fr</td><td>hi</td><td>ru</td><td>sw</td><td>th</td><td>tr</td><td>ur</td><td>vi</td><td>zh</td></tr><tr><td>ar</td><td>-</td><td>67.84</td><td>67.73</td><td>67.85</td><td>67.62</td><td>67.84</td><td>67.80</td><td>67.81</td><td>67.85</td><td>67.87</td><td>67.86</td><td>67.83</td><td>67.71</td><td>67.89</td><td>67.95</td><td>67.37</td></tr><tr><td>bg</td><td>71.79</td><td>-</td><td>71.76</td><td>71.80</td><td>71.72</td><td>71.77</td><td>71.80</td><td>71.74</td><td>71.94</td><td>71.77</td><td>71.78</td><td>71.78</td><td>71.77</td><td>71.79</td><td>71.92</td><td>71.60</td></tr><tr><td>de</td><td>73.36</td><td>73.23</td><td>-</td><td>73.37</td><td>73.30</td><td>73.30</td><td>73.33</td><td>73.46</td><td>73.27</td><td>73.34</td><td>73.38</td><td>73.32</td><td>73.37</td><td>73.34</td><td>73.43</td><td>73.25</td></tr><tr><td>el</td><td>69.95</td><td>69.98</td><td>69.97</td><td>-</td><td>69.94</td><td>69.99</td><td>69.91</td><td>69.93</td><td>69.95</td><td>69.98</td><td>70.03</td><td>70.02</td><td>69.90</td><td>69.95</td><td>70.03</td><td>69.54</td></tr><tr><td>en</td><td>82.24</td><td>82.21</td><td>82.13</td><td>82.22</td><td>-</td><td>82.15</td><td>82.27</td><td>82.26</td><td>82.24</td><td>82.24</td><td>82.19</td><td>82.39</td><td>82.25</td><td>82.14</td><td>82.20</td><td>81.94</td></tr><tr><td>es</td><td>76.07</td><td>76.12</td><td>76.14</td><td>76.02</td><td>76.06</td><td>-</td><td>76.18</td><td>76.14</td><td>76.10</td><td>75.94</td><td>76.03</td><td>75.91</td><td>76.10</td><td>76.00</td><td>76.09</td><td>75.79</td></tr><tr><td>fr</td><td>75.32</td><td>75.23</td><td>75.16</td><td>75.24</td><td>75.23</td><td>75.18</td><td>-</td><td>75.19</td><td>75.22</td><td>75.31</td><td>75.28</td><td>75.19</td><td>75.28</td><td>75.19</td><td>75.28</td><td>75.39</td></tr><tr><td>hi</td><td>64.95</td><td>64.82</td><td>64.78</td><td>64.89</td><td>64.64</td><td>64.63</td><td>64.90</td><td>-</td><td>64.87</td><td>64.94</td><td>64.73</td><td>64.84</td><td>64.79</td><td>64.97</td><td>64.83</td><td>64.37</td></tr><tr><td>ru</td><td>71.19</td><td>71.27</td><td>71.17</td><td>71.33</td><td>71.19</td><td>71.19</td><td>71.33</td><td>71.28</td><td>-</td><td>71.31</td><td>71.34</td><td>71.45</td><td>71.18</td><td>71.29</td><td>71.38</td><td>70.84</td></tr><tr><td>sw</td><td>58.14</td><td>58.23</td><td>57.95</td><td>57.99</td><td>57.53</td><td>57.97</td><td>57.94</td><td>58.10</td><td>58.04</td><td>-</td><td>58.00</td><td>58.22</td><td>58.08</td><td>58.01</td><td>58.09</td><td>57.82</td></tr><tr><td>th</td><td>61.59</td><td>61.64</td><td>61.57</td><td>61.71</td><td>61.40</td><td>61.51</td><td>61.51</td><td>61.68</td><td>61.54</td><td>61.50</td><td>-</td><td>61.58</td><td>61.41</td><td>61.56</td><td>61.74</td><td>61.18</td></tr><tr><td>tr</td><td>64.74</td><td>64.79</td><td>64.69</td><td>64.82</td><td>64.59</td><td>64.82</td><td>64.76</td><td>64.83</td><td>64.70</td><td>64.89</td><td>64.92</td><td>-</td><td>64.74</td><td>64.73</td><td>64.66</td><td>64.85</td></tr><tr><td>ur</td><td>63.67</td><td>63.58</td><td>63.69</td><td>63.63</td><td>63.55</td><td>63.63</td><td>63.68</td><td>63.61</td><td>63.72</td><td>63.63</td><td>63.72</td><td>63.81</td><td>-</td><td>63.67</td><td>63.60</td><td>63.71</td></tr><tr><td>vi</td><td>73.51</td><td>73.52</td><td>73.46</td><td>73.35</td><td>73.36</td><td>73.29</td><td>73.39</td><td>73.31</td><td>73.51</td><td>73.38</td><td>73.39</td><td>73.41</td><td>73.42</td><td>-</td><td>73.41</td><td>73.23</td></tr><tr><td>zh</td><td>74.04</td><td>73.97</td><td>74.02</td><td>74.02</td><td>73.74</td><td>74.01</td><td>74.02</td><td>74.10</td><td>74.11</td><td>73.99</td><td>74.01</td><td>74.21</td><td>74.06</td><td>73.95</td><td>-</td><td>73.93</td></tr></table>
+
+Table 6: The performance in terms of average test accuracy for the few-shot setting over 10 runs of X-MAML on the XNLI dataset using Multi-BERT (multilingual BERT), as base model. Each column corresponds to the performance of the Multi-BERT system after meta-learning with a single auxiliary language, and evaluation on the target language of the XNLI test set. The auxiliary language is not included during the evaluation phase. Results of the Multi-BERT model without X-MAML (baseline) are also reported.

zeroshotcrosslingualtransferwithmetalearning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf3af897dc6ad03925bb5c74d3d02b3ca1b5be5123524d237adf4f937faac789
+size 1404197

zeroshotcrosslingualtransferwithmetalearning/layout.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a25a11695a064e8d290dd78fae0393f30987e7cfb8f2bcf84cb9af75a4f950e
+size 380398

zeroshotstancedetectionadatasetandmodelusinggeneralizedtopicrepresentations/ba244162-ccba-428c-8ac9-3ce6f9681a2a_content_list.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8016e8a2a06e93bb83168b81f359246642224d39509cc6c88771ca9d7bebd8dc
+size 112524

zeroshotstancedetectionadatasetandmodelusinggeneralizedtopicrepresentations/ba244162-ccba-428c-8ac9-3ce6f9681a2a_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0e828b2b5c76bb4066b35e6fda10097b996703cd26c0e06b86cdef7b86ee1db
+size 133118