Chunk-level Retrieval Eval
Collection
Contextualized chunk-level (query2chunk) retrieval eval for the Citras benchmark. • 19 items • Updated
chunk_id stringlengths 3 7 | chunk stringlengths 1 823 | source_url stringclasses 416
values | title stringclasses 416
values | chunk_idx int64 0 294 | chunk_start_char int64 0 139k | chunk_end_char int64 303 139k |
|---|---|---|---|---|---|---|
0_0 | End-to-End Trainable Non-Collaborative Dialog System
End-to-end task-oriented dialog models have achieved promising performance on collaborative tasks where users willingly coordinate with the system to complete a given task. While in non-collaborative settings, for example, negotiation and persuasion, users and syste... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 0 | 0 | 666 |
0_1 | Building upon TransferTransfo (Wolf et al. 2019), we propose an end-to-end neural network model to generate diverse coherent responses. Our model utilizes intent and semantic slots as the intermediate sentence representation to guide the generation process. In addition, we design a filter to select appropriate respons... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 1 | 666 | 1,305 |
0_2 | Both automatic and human evaluations suggest that our model outperforms multiple baselines in these two non-collaborative tasks.
Introduction
Considerable progress has been made building end-to-end dialog systems for collaborative tasks in which users cooperate with the system to achieve a common goal. Examples of c... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 2 | 1,305 | 2,001 |
0_3 | Examples of such tasks include deceiving attackers, persuading users to donate to a cause BIBREF1, and negotiating a product price BIBREF2, BIBREF3. In these tasks, users often perform complex actions that are beyond a simple set of pre-defined intents. In order to reach a common state, the user and the system need to... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 3 | 2,001 | 2,587 |
0_4 | An example of an on-task system response that the system could have made is “Do you want to make a donation?", which sticks to the task but neglects users' question. However, a better response to such an off-task question is “War is destructive and pitiless, but you can donate to help child victims of war." This respo... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 4 | 2,587 | 3,195 |
0_5 |
To tackle the issue of incoherent system responses to off-task content, previous studies have built hybrid systems to interleave off-task and on-task content. BIBREF4 used a rule-based dialog manager for on-task content and a neural model for off-task content, and trained a reinforcement learning model to select betw... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 5 | 3,195 | 3,853 |
0_6 | For on-task information, we directly use task-related intents for representation. Off-task information, on the other hand, is too general to categorize into specific intents, so we choose dialog acts that convey syntax information. These acts, such as “open question" are general to all tasks.
Previous studies use tem... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 6 | 3,853 | 4,420 |
0_7 | We propose Multiple Intents and Semantic Slots Annotation Neural Network (MISSA) to combine the advantages of both template and generation models and takes advantage from the hierarchical annotation at the same time. MISSA follows the TransferTransfo framework BIBREF0 with three modifications: (i) We first concurrentl... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 7 | 4,420 | 4,881 |
0_8 | Specifically, we generate responses conditioned on the above intermediate representation (intents and slots); (iii) Finally, we generate multiple responses with the nucleus sampling strategy BIBREF5 and then apply a response filter, which contains a set of pre-defined constraints to select coherent responses. The cons... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 8 | 4,881 | 5,496 |
0_9 | As non-collaborative tasks are still relatively new to the study of dialog systems, there are insufficiently many meaningful datasets for evaluation and we hope this provides a valuable example. We evaluate MISSA on the newly collected AntiScam dataset and an existing PersuasionForGood dataset. Both automatic and huma... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 9 | 5,496 | 5,892 |
0_10 |
In summary, our contributions include: (i) We design a hierarchical intent annotation scheme and a semantic slot annotation scheme to annotate the non-collaborative dialog dataset, we also propose a carefully-designed AntiScam dataset to facilitate the research of non-collaborative dialog systems. (ii) We propose a m... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 10 | 5,892 | 6,469 |
0_11 | Furthermore, we also build a persuasion dialog system to persuade people to donate to charities. We release the code and data.
Related Work
The interest in non-collaborative tasks has been increasing and there have already been several related datasets. For instance, BIBREF1 wang2019persuasion collected conversation... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 11 | 6,469 | 6,976 |
0_12 | There are many other non-collaborative tasks, such as the turn-taking game BIBREF6, the multi-party game BIBREF7 and item splitting negotiation BIBREF8. Similar to the AntiScam dataset proposed in this paper, these datasets contain off-task content and can be used to train non-collaborative dialog systems. However, si... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 12 | 6,976 | 7,477 |
0_13 | Therefore, we propose the AntiScam dataset, which is designed to interleave the on-task and off-task contents in the conversation, and can serve as a benchmark dataset for similar non-collaborative tasks.
To better understand user utterances and separate on-task and off-task content within a conversation, previous wo... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 13 | 7,477 | 8,031 |
0_14 | BIBREF11 gupta2018semantic used a hierarchical annotation scheme for semantic parsing. Inspired by these studies, our idea is to annotate the intent and semantic slot separately in non-collaborative tasks. We propose a hierarchical intent annotation scheme that can be adopted by all non-collaborative tasks. With this ... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 14 | 8,031 | 8,470 |
0_15 |
Traditional task-oriented dialog systems BIBREF12 are usually composed of multiple independent modules, for example, natural language understanding, dialog state tracking BIBREF13, BIBREF14, dialog policy manager BIBREF15, and natural language generation BIBREF16. Conversational intent is adopted to capture the meani... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 15 | 8,470 | 9,085 |
0_16 | The major defect of a separately trained pipeline is the laborious dialog state design and annotation. In order to mitigate this problem, recent work has explored replacing independent modules with end-to-end neural networks BIBREF18, BIBREF19, BIBREF20. Our model also follows this end-to-end fashion.
Over the last f... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 16 | 9,085 | 9,622 |
0_17 | Recent improvements build on top of the transformer and pre-trained language models BIBREF24, BIBREF25, BIBREF26, obtained state-of-the-art results on the Persona-Chat dataset BIBREF0. Pre-trained language models are proposed to build task-oriented dialog systems to drive the progress on leveraging large amounts of av... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 17 | 9,622 | 10,179 |
0_18 |
Another line of work interleaves on-task and off-task content by building a hybrid dialog system that combines a task-oriented model and a non-task-oriented model BIBREF4, BIBREF29. In these studies, task-oriented systems and non-task-oriented systems are designed separately and both systems generate candidate respon... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 18 | 10,179 | 10,765 |
0_19 |
Non-Collaborative Task Annotation Scheme
To decouple syntactic and semantic information in utterances and provide detailed supervision, we design a hierarchical intent annotation scheme for non-collaborative tasks. We first separate on-task and off-task intents. As on-task intents are key actions that can vary among... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 19 | 10,765 | 11,302 |
0_20 | The advantage of this hierarchical annotation scheme is apparent when starting a new non-collaborative task: we only need to focus on designing the on-task categories and semantic slots which are the same as traditional task-oriented dialog systems. Consequently, we don't have to worry about the off-task annotation de... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 20 | 11,302 | 11,932 |
0_21 | Since our AntiScam focuses on understanding and reacting towards elicitations, we define elicitation, providing_information and refusal as on-task intents. In the PersuasionForGood dataset, we define nine on-task intents in Table TABREF2 based on the original PersuasionForGood dialog act annotation scheme. All these i... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 21 | 11,932 | 12,464 |
0_22 | General intents are more closely related to the syntactic meaning of the sentence (open_question, yes_no_question, positive_answer, negative_answer, responsive_statement, and nonresponsive_statement) while social intents are common social actions (greeting, closing, apology, thanking,respond_to_thank, and hold).
For ... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 22 | 12,464 | 13,063 |
0_23 | Following BIBREF1, we segment each conversation turn into single sentences and then annotate each sentence rather than turns.
Datasets
We test our approach on two non-collaborative task datasets: the AntiScam dataset and the PersuasionForGood dataset BIBREF1. Both datasets are collected from the Amazon Mechanical Tu... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 23 | 13,063 | 13,666 |
0_24 | We chose a popular Amazon customer service scam scenario to collect dialogs between users and attackers who aim to collect users information. We posted a role-playing task on the Amazon Mechanical Turk platform and collected a typing conversation dataset named AntiScam. We collected 220 human-human dialogs. The averag... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 24 | 13,666 | 14,231 |
0_25 | We recruited two expert annotators who have linguistic training to annotate 3,044 sentences in 100 dialogs, achieving a 0.874 averaged weighted kappa value.
Datasets ::: PersuasionForGood Dataset
The PersuasionForGood dataset BIBREF1 was collected from typing conversations on Amazon Mechanical Turk platform. Two wor... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 25 | 14,231 | 14,757 |
0_26 | The dataset consists of 1,017 dialogs, where 300 dialogs are annotated with dialog acts. The average conversation length is 10.43, the vocabulary size is 8,141. Since the original PersuasionForGood dataset is annotated with dialog acts, we select the on-task dialog acts as on-task intents shown in Table TABREF2, and c... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 26 | 14,757 | 15,250 |
0_27 | BIBREF0 wolf2019transfertransfo fine-tuned the generative pre-training model (GPT) BIBREF32 with the PERSONA-CHAT dataset BIBREF33 in a multi-task fashion, where the language model objective is combined with a next-utterance classification task. | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 27 | 15,250 | 15,496 |
0_28 | The language model's objective is to maximize the following likelihood for a given sequence of tokens, $X = \lbrace x_1,\dots ,x_n\rbrace $:
The authors also trained a classifier to distinguish the correct next-utterance appended to the input human utterances from a set of randomly selected utterance distractors. In ... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 28 | 15,496 | 16,149 |
0_29 | We keep all the embeddings in the framework and train the language model and next-utterance classification task in a multi-task fashion following TransferTransfo.
We make two major changes: (1) To address the problem that TransferTransfo is originally designed for an open domain without explicit intents and regulatio... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 29 | 16,149 | 16,840 |
0_30 | Generating diverse responses has proven to be an enduring challenge. To increase response diversity, we sample multiple generated responses and choose an appropriate one according to a set of pre-defined rules.
Model ::: Intent and Semantic Slot Classifiers
We train MISSA in a multi-task fashion. In addition to the ... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 30 | 16,840 | 17,341 |
0_31 | The intent classifier and semantic slot classifier for human utterances capture the semantic and syntactic meaning of human utterances, providing information to select the appropriate response among response candidates while the classifiers for the system intents and semantic slots are designed to help select an appro... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 31 | 17,341 | 17,879 |
0_32 | $h^l_{t-1}$ is the hidden states at the end of last sentence in turn $t-1$, $h^i_{t}$ is the last hidden states at the end of $i$-th sentence in turn $t$. $W_{2h}$ are weights learned during training.
MISSA is able to classify multiple intents and multiple semantic slots in a single utterance with these classifiers. ... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 32 | 17,879 | 18,257 |
0_33 | Specifically, we set a special token $<$sep$>$ at the end of each sentence in an utterance (an utterance can consist of multiple sentences). Next, we pass the token's position information to the transformer architecture and obtain the representation of the position (represented as colored position at last layer in Fig... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 33 | 18,257 | 18,859 |
0_34 | As shown in Figure FIGREF6, the loss function ${\mathcal {L}}$ for the model combines all the task losses:
where ${\mathcal {L}_{LM}}$ is the language model loss, ${\mathcal {L}_{I_h}}$, ${\mathcal {L}_{S_h}}$, ${\mathcal {L}_{I_s}}$, and ${\mathcal {L}_{S_s}}$ are losses of intent and slots classifiers, ${\mathcal {... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 34 | 18,859 | 19,228 |
0_35 | $\lambda _{LM}$, $\lambda _{I_h}$, $\lambda _{S_h}$, $\lambda _{I_s}$, $\lambda _{S_s}$, and $\lambda _{nup}$ are the hyper-parameters that control the relative importance of every loss.
Model ::: Response Generation
MISSA can generate multiple sentences in a single system turn. Therefore, we perform system generati... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 35 | 19,228 | 19,591 |
0_36 | More specifically, during the training phase, in addition to inserting a special $<$sep$>$ token at the end of each sentence, we also insert the intent of the system response as special tokens at the head of each sentence in the system response. For example, in Figure FIGREF6, we insert a $<$pos_ans$>$ token at the he... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 36 | 19,591 | 20,091 |
0_37 | During the testing phase, the model first generates a special intent token, then after being conditioned on this intent token, the model keeps generating a sentence until it generates a $<$sep$>$ token. After that, the model continues to generate another intent token and another sentence until it generates an $<$eos$>... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 37 | 20,091 | 20,712 |
0_38 | These corner cases may lead to fatal results in high-risk tasks, for example, health care and education. To improve the robustness of MISSA and improve its ability to generalize to more tasks, we add a response filtering module after the generation. With the nucleus sampling strategy BIBREF5, MISSA is able to generate... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 38 | 20,712 | 21,377 |
0_39 | The filtering module is easily adaptable to different domains or specific requirements, which makes our dialog system more controllable.
Experiments
We evaluate MISSA on two non-collaborative task datasets. AntiScam aims to build a dialog system that occupies the attacker's attention and elicits the attacker's infor... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 39 | 21,377 | 21,950 |
0_40 |
Experiments ::: Baseline Models
We compare MISSA mainly with two baseline models:
TransferTransfo The vanilla TransferTransfo framework is compared with MISSA to show the impact and necessity of adding the intent and slot classifiers. We follow the original TransferTransfo design BIBREF0 and train with undelexicali... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 40 | 21,950 | 22,517 |
0_41 | If the classifier decides that the utterance is on-task, we choose the response from MISSA; otherwise, we choose the response from vanilla TransferTransfo baseline.
In addition, we perform ablation studies on MISSA to show the effects of different components.
MISSA-sel denotes MISSA without response filtering.
MISS... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 41 | 22,517 | 23,105 |
0_42 | We choose perplexity to evaluate the model performance.
Response-Intent Prediction (RIP) $\&$ Response-Slot Prediction (RSP) Different from open-domain dialog systems, we care about the intents of the system response in non-collaborative tasks as we hope to know if the system response satisfies user intents. For exam... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 42 | 23,105 | 23,658 |
0_43 | Since our baselines are more suited for social chat as they cannot produce system intents, we use the system intent and slot classifiers trained in our model to predict their responses' intents and slots. The intent predictor achieves a $84\%$ accuracy and the semantic slot predictor achieves $77\%$ on the AntiScam da... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 43 | 23,658 | 24,154 |
0_44 |
Extended Response-Intent Prediction (ERIP) $\&$ Extended Response-Slot Prediction (ERSP) With Response-Intent Prediction, we verify the predicted intents to evaluate the coherence of the dialog. However, the real mapping between human-intent and system-intent is much more complicated as there might be multiple accept... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 44 | 24,154 | 24,635 |
0_45 | Specifically, we estimate the transition probability $p(I_i|I_j)$ by counting the frequency of all the bi-gram human-intent and system-intent pairs in the training data. During the test stage, if the predicted intent matches the ground truth, we set the score as 1, otherwise we set the score as $p(I_{predict}|I_i)$ wh... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 45 | 24,635 | 25,125 |
0_46 |
Experiments ::: Human Evaluation Metrics
Automatic metrics only validate the system’s performance on a single dimension at a time. The ultimate holistic evaluation should be conducted by having the trained system interact with human users. Therefore we also conduct human evaluations for the dialog system built on An... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 46 | 25,125 | 25,694 |
0_47 | Each time, volunteers are required to use similar sentences and strategies to interact with all five models and score each model based on the metrics listed below at the end of the current round. Each model receives a total of 45 human ratings, and the average score is reported as the final human-evaluation score. In ... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 47 | 25,694 | 26,254 |
0_48 |
Coherence Different from single sentence's fluency, coherence focuses more on the logical consistency between sentences in each turn.
Engagement In the anti-scam scenario, one of our missions is to keep engaging with the attackers to waste their time. So we directly ask volunteers (attackers) to what extend they wou... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 48 | 26,254 | 26,884 |
0_49 |
Task Success Score (TaskSuc) The other goal of the anti-scam system is to elicit attacker's personal information. We count the average type of information (name, address and phone number) that the system obtained from attackers as the task success score.
Results and Analysis
Table TABREF19 presents the main experim... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 49 | 26,884 | 27,377 |
0_50 | We observe that MISSA outperforms two baseline models (TransferTransfo and hybrid model) on almost all the metrics on both datasets. For further analysis, examples of real dialogs from the human evaluation are presented in Table TABREF21.
Compared to the first TransferTransfo baseline, MISSA outperforms the TransferT... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 50 | 27,377 | 27,937 |
0_51 | MISSA also has a higher task success score (1.294) than TransferTransfo (1.025), which indicates that it elicits information more strategically. In the top two dialogs (A and B) that are shown in Table TABREF21, both attackers were eliciting a credit card number in their first turns. TransferTransfo directly gave away... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 51 | 27,937 | 28,375 |
0_52 | Furthermore, in the next turn, TransferTransfo ignored the context and asked an irrelevant question “what is your name?” while MISSA was able to generate the response “why can't you use my address?”, which is consistent to the context. We suspect the improved performance of MISSA comes from our proposed annotation sch... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 52 | 28,375 | 28,953 |
0_53 | As shown in the bottom two dialogs in Table TABREF21, attackers in both dialogs introduced their names in their first utterances. MISSA recognized attacker's name, while the hybrid model did not. We suspect it is because the hybrid model does not have the built-in semantic slot predictor. In the second turn, both atta... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 53 | 28,953 | 29,588 |
0_54 | We suspect that the hybrid model's bad performance on the off-task content leads to its low coherence rating (2.76) and short dialog length (8.2).
To explore the influence of the intent-based conditional response generation method and the designed response filter, we perform an ablation study. The results are shown i... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 54 | 29,588 | 29,925 |
0_55 | We find that MISSA has higher fluency score and coherence score than MISSA-con (4.18 vs 3.78 for fluency, and 3.75 vs 3.68 for coherence), which suggests that conditioning on the system intent to generate responses improves the quality of the generated sentences. Compared with MISSA-sel, MISSA achieves better performa... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 55 | 29,925 | 30,351 |
0_56 | This is because the response filter removed all the incoherent responses, which makes the attacker more willing to keep chatting. The ablation study shows both the conditional language generation mechanism and the response filter are essential to MISSA's good performance.
We also apply our method to the PersuasionFor... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 56 | 30,351 | 30,943 |
0_57 | Particularly, MISSA achieves the lowest perplexity, which confirms that using conditional response generation leads to high quality responses. Compared with the result on AntiScam dataset, MISSA-con performs the best in terms of RIP and ERIP. We suspect the underlying reason is that there are more possible responses w... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 57 | 30,943 | 31,422 |
0_58 |
Conclusion and Future Work
We propose a general dialog system pipeline to build non-collaborative dialog systems, including a hierarchical annotation scheme and an end-to-end neural response generation model called MISSA. With the hierarchical annotation scheme, we can distinguish on-task and off-task intents. MISSA... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 58 | 31,422 | 32,014 |
0_59 | MISSA outperforms all baseline methods in terms of fluency, coherency, and user engagement on both the newly proposed anti-scam task and an existing persuasion task. However, MISSA still produces responses that are not consistent with their distant conversation history as GPT can only track a limited history span. In ... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 59 | 32,014 | 32,528 |
0_60 | The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes not withstanding any copyright notation therein. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or impl... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 60 | 32,528 | 33,178 |
0_61 | We give both workers specific personal data. Instructions are shown in Table TABREF24. The “attacker” additionally receives training on how to elicit information from people. Workers cannot see their partners' instructions.
There are two tasks for the users: firstly, users are required to chat with their partners and... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 61 | 33,178 | 33,725 |
0_62 | We give a bonus to users if they detect the attackers and elicit real information from the attackers, including the attacker's name, address and phone number. Since one worker can only participate once in the task, they do not know their partners are always attackers.
We provide real user information including the us... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 62 | 33,725 | 34,344 |
0_63 | We give a bonus to attackers if they elicit correct information from users, including the user's address, credit card number, CVS and expiration date. Each worker can only participate once to prevent workers from knowing their partner's information and goals in advance. We collected 220 human-human dialogs. The averag... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 63 | 34,344 | 34,909 |
0_64 |
We recruited two expert annotators who have linguistic training to annotate 3,044 sentences in 100 dialogs, achieving a 0.874 averaged weighted kappa value. Table TABREF2 shows that there is a vast amount of off-task content in the dataset, which confirms the necessity of a hierarchical on-task/off-task annotation sc... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 64 | 34,909 | 35,495 |
0_65 | Moreover, users also ask more open_questions (173 vs 54) and yes_no_questions (165 vs 117) for off-task content because they are instructed to prolong the conversation after detecting the attacker. Furthermore, attackers and users both have a massive amount of social content (292 in total and 252 in total), suggesting... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 65 | 35,495 | 35,997 |
0_66 | We use an Adam optimizer with a learning rate of 6.25e-5 and $L2$ weight decay of $0.01$, we set the coefficient of language modeling loss to be 2, the coefficient of intent and slot classifiers to be 1, and the coefficient of next-utterance classifier to be 1. We first pre-train the model on the PERSONA-CHAT dataset. | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 66 | 35,997 | 36,317 |
0_67 | When fine-tuning on the AntiScam and the PersuasionForGood datasets, we use $80\%$ data for training, $10\%$ data for validation, and $10\%$ data for testing. Since the original PersuasionForGood dataset is annotated with intents, we separate the original on-task and off-task intents, which are shown in Table TABREF2.... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 67 | 36,317 | 36,865 |
0_68 |
Appendix ::: Example Dialog
An example of human-human chat on AntiScam dataset is shown in Table TABREF25.
Table 1: Hierarchical intent annotation scheme on both ANTISCAM dataset and PERSUASIONFORGOOD dataset. The On-task intents are task-specific while the Off-task intents are general for different non-collaborati... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 68 | 36,865 | 37,421 |
0_69 | We concatenate the last hidden states at <sep> tokens with the last hidden states at the end of the last utterance to predict intents and semantic slots for corresponding sentences. We can predict multiple intents and semantic slots for each human utterance and system response. During testing, the appended response an... | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 69 | 37,421 | 38,047 |
0_70 |
Table 6: Instructions for attackers and users on Amazon Mechanical Turk.
Table 7: An example human-human dialog in ANTISCAM dataset. All the slot values have been replaced with slot tokens. | https://arxiv.org/abs/1911.10742 | End-to-End Trainable Non-Collaborative Dialog System | 70 | 38,047 | 38,239 |
1_0 | OpenTapioca: Lightweight Entity Linking for Wikidata
We propose a simple Named Entity Linking system that can be trained from Wikidata only. This demonstrates the strengths and weaknesses of this data source for this task and provides an easily reproducible baseline to compare other systems against. Our model is light... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 0 | 0 | 542 |
1_1 |
Most of the entity linking literature focuses on target knowledge bases which are derived from Wikipedia, such as DBpedia BIBREF0 or YAGO BIBREF1 . These bases are curated automatically by harvesting information from the info-boxes and categories on each Wikipedia page and are therefore not editable directly.
Wikida... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 1 | 542 | 1,033 |
1_2 | As these new approaches to entity linking also introduce novel learning methods, it is hard to tell apart the benefits that come from the new models and those which come from the choice of knowledge graph and the quality of its data.
We review the main differences between Wikidata and static knowledge bases extracted... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 2 | 1,033 | 1,638 |
1_3 | OpenTapioca can be trained easily from a Wikidata dump only, and can be efficiently kept up to date in real time as Wikidata evolves. We also propose tools to adapt existing entity linking datasets to Wikidata, and offer a new entity linking dataset, consisting of affiliation strings extracted from research articles.
... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 3 | 1,638 | 2,205 |
1_4 |
Wikidata stores information about the world in a collection of items, which are structured wiki pages. Items are identified by ther Q-id, such as Q40469, and they are made of several data fields. The label stores the preferred name for the entity. It is supported by a description, a short phrase describing the item t... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 4 | 2,205 | 2,791 |
1_5 | They can be backed by references and be made more precise with qualifiers, which all rely on a controlled vocabulary of properties (similar to RDF predicates). Finally, items can have site links, connecting them to the corresponding page for the entity in other Wikimedia projects (such as Wikipedia). Note that Wikidat... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 5 | 2,791 | 3,387 |
1_6 |
Our goal is to evaluate the usefulness of this crowdsourced structured data for entity linking. We will therefore refrain from augmenting it with any external data (such as phrases and topical information extracted from Wikipedia pages), as is generally done when working with DBpedia or YAGO. By avoiding a complex ma... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 6 | 3,387 | 3,957 |
1_7 | This means that users are able to fix or improve the knowledge graph, for instance by adding a missing alias on an item, and immediately see the benefits on their entity linking task. This constrasts with all other systems we are aware of, where the user either cannot directly intervene on the underlying data, or ther... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 7 | 3,957 | 4,617 |
1_8 | Given a text to annotate, candidate entities are generated by looking for occurrences of their surface forms in the text. Because of homonymy, many of these candidate occurrences turn out to be false matches, so a classifier is used to predict their correctness. We can group the features they tend to use in the follow... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 8 | 4,617 | 5,282 |
1_9 | It is also possible to crawl the web for Wikipedia links to improve the coverage, often at the expense of data quality BIBREF10 .
Beyond collecting a set of possible surface forms, these approaches count the number of times an entity $e$ was mentioned by a phrase $w$ . This makes it possible to use a Bayesian methodo... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 9 | 5,282 | 5,768 |
1_10 |
In Wikidata, items have labels and aliases in multiple languages. As this information is directly curated by editors, these phrases tend to be of high quality. However, they do not come with occurence counts. As items link to each other using their Wikidata identifiers only, it is not possible to compare the number o... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 10 | 5,768 | 6,337 |
1_11 | For instance Curry is the English label of both the item about the Curry programming language (Q2368856) and the item about the village in Alaska (Q5195194), and the description field is used to disambiguate them.
Manual curation of surface forms implies a fairly narrow coverage, which can be an issue for general pur... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 11 | 6,337 | 6,677 |
1_12 | For instance, people are commonly refered to with their given or family name only, and these names are not systematically added as aliases: at the time of writing, Trump is an alias for Donald Trump (Q22686), but Cameron is not an alias for David Cameron (Q192). As a Wikidata editor, the main incentive to add aliases ... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 12 | 6,677 | 7,132 |
1_13 | Aliases are not designed to offer a complete set of possible surface forms found in text: for instance, adding common mispellings of a name is discouraged.
Although Wikidata makes it impossible to count how often a particular label or alias is used to refer to an entity, these surface forms are carefully curated by t... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 13 | 7,132 | 7,578 |
1_14 | Having no access to such a probability distribution, we choose to approximate this quantity by $\frac{p(e)}{p(d[s])}$ , where $p(e)$ is the probability that $e$ is linked to, and $p(d[s])$ is the probability that $d[s]$ occurs in a text. In other words, we estimate the popularity of the entity and the commonness of th... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 14 | 7,578 | 8,070 |
1_15 | The PageRank is computed on the entire Wikidata using statement values and qualifiers as edges.
The probability $p(d[s])$ is estimated by a simple unigram language model that can be trained either on any large unannotated dataset. | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 15 | 8,070 | 8,302 |
1_16 |
The local compatibility is therefore represented by a vector of features $F(e,w)$ and the local compatibility is computed as follows, where $\lambda $ is a weights vector: $
F(e,w) &= ( -\log p(d[s]), \log p(e) , n_e, s_e, 1 ) \\
p(e|d[s]) &\propto e^{F(e,w) \cdot \lambda }
$
Topic similarity
The compatibility of ... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 16 | 8,302 | 8,785 |
1_17 | BIBREF12 or keyword extraction BIBREF13 , BIBREF14 , BIBREF9 .
Wikidata items only consist of structured data, except in their descriptions. This makes it difficult to compute topical information using the methods above. Vector-based representations of entities can be extracted from the knowledge graph alone BIBREF15... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 17 | 8,785 | 9,382 |
1_18 | This requires access to large amounts of text both to train the word vectors and to derive the entity vectors from them. These vectors have been shown to encode significant semantic information by themselves BIBREF19 , so we refrain from using them in this study.
Mapping coherence
Entities mentioned in the same cont... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 18 | 9,382 | 10,048 |
1_19 | This is often done by first defining a pairwise relatedness score between the target entities. | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 19 | 10,048 | 10,143 |
1_20 | For instance, a popular metric introduced by BIBREF20 considers the set of wiki links $|a|, |b|$ made from or to two entities $a$ , $b$ and computes their relatedness: $ \text{rel}(a,b) = 1 - \frac{\log (\max (|a|,|b|)) - \log (|a| \cap |b|)}{\log (|K|) - \log (\min (|a|,|b|))}$
where $|K|$ is the number of entities... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 20 | 10,143 | 10,486 |
1_21 |
When linking to Wikidata instead of Wikipedia, it is tempting to reuse these heuristics, replacing wikilinks by statements. However, Wikidata's linking structure is quite different from Wikipedia: statements are generally a lot sparser than links and they have a precise semantic meaning, as editors are restricted by ... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 21 | 10,486 | 11,056 |
1_22 | Most approaches build a graph of candidate entities, where edges indicate semantic relatedness: the difference between the heuristics lie in the way this graph is used for the matching decisions. BIBREF21 use an approximate algorithm to find the densest subgraph of the semantic graph. This determines choices of entiti... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 22 | 11,056 | 11,624 |
1_23 |
OpenTapioca: an entity linking model for Wikidata
We propose a model that adapts previous approaches to Wikidata. Let $d$ be a document (a piece of text). A spot $s \in d$ is a pair of start and end positions in $d$ . It defines a phrase $d[s]$ , and a set of candidate entities $E[s]$ : those are all Wikidata items ... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 23 | 11,624 | 11,981 |
1_24 | Given two spots $s, s^{\prime }$ we denote by $|s - s^{\prime }|$ the number of characters between them. We build a binary classifier which predicts for each $s \in d$ and $e \in E[s]$ if $s \in d$0 should be linked to $s \in d$1 .
Semantic similarity
The issue with the features above is that they ignore the context... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 24 | 11,981 | 12,408 |
1_25 | The general idea is to define a graph on the candidate entities, linking candidate entities which are semantically related, and then find a combination of candidate entities which have both high local compatibility and which are densely related in the graph.
For each pair of entities $e, e^{\prime }$ we define a simi... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 25 | 12,408 | 12,830 |
1_26 | Consider a one-step random walks starting on $e$ , with probability $\beta $ to stay on $e$ and probability $\frac{1-\beta }{|l(e)|}$ to reach one of the linked items. We define $s(e,e^{\prime })$ as the probability that two such one-step random walks starting from $e$ and $s(e,e^{\prime })$0 end up on the same item. | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 26 | 12,830 | 13,149 |
1_27 | This can be computed explicitly as $s(e,e^{\prime })$1
We then build a weighted graph $G_d$ whose vertices are pairs $(s \in d, e \in E[s])$ . In other words, we add a vertex for each candidate entity at a given spot. We fix a maximum distance $D$ for edges: vertices $(s,e)$ and $(s^{\prime },e^{\prime })$ can only ... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 27 | 13,149 | 13,533 |
1_28 | In this case, we define the weight of such an edge as $(\eta + s(e,e^{\prime }))\frac{D - |s - s^{\prime }|}{D}$ , where $\eta $ is a smoothing parameter. In other words, the edge weight is proportional to the smoothed similarity between the entities, discounted by the distance between the mentions.
The weighted grap... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 28 | 13,533 | 13,999 |
End of preview. Expand in Data Studio
README.md exists but content is empty.
- Downloads last month
- 23