Computational Historical Linguistics and Language Diversity in South Asia

Arora, Aryaman; Farris, Adam; Basu, Samopriya; Kolichala, Suresh

doi:10.18653/v1/2022.acl-long.99

Cited by 4 publications

(6 citation statements)

References 26 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…While the proportion of missing values was not manipulated in the Experiments, exploratory analysis found that larger non-missing proportions were more favorable: increasing the users assigned per utterance was beneficial for fixed total user pool; but for fixed number of users per utterance, increasing the user pool was detrimental. While having more data is statistically expedient, it also imposes a greater burden on the user (Shuster et al, 2022). In these Experiments, having 10% of the inter-rater matrix filled meant that each user labeled on average 20 utterances, which is 5x what the CV-based approach (Ju et al, 2022) would have required.…”

Section: Discussionmentioning

confidence: 99%

“…For chatbots, it is not enough to learn to generate coherent utterances-many coherent utterances are undesirable, such as offensive remarks. Postdeployment, to have the chatbot continuously improve its judgment, one way is to source training examples from feedback by live organic users (Shuster et al, 2022;. Though more realistic than crowdworkers, among organic users are trolls, whose erroneous feedback fosters bad behavior (Wolf et al, 2017).…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Learning to love diligent trolls: Accounting for rater effects in the dialogue safety task

Ilagan

2023

Findings of the Association for Computational Linguistics: EMNLP 2023

View full text Add to dashboard Cite

Chatbots have the risk of generating offensive utterances, which must be avoided. Postdeployment, one way for a chatbot to continuously improve is to source utterance/label pairs from feedback by live users. However, among users are trolls, who provide training examples with incorrect labels. To de-troll training data, previous work removed training examples that have high user-aggregated cross-validation (CV) error. However, CV is expensive; and in a coordinated attack, CV may be overwhelmed by trolls in number and in consistency among themselves. In the present work, I address both limitations by proposing a solution inspired by methodology in automated essay scoring (AES): have multiple users rate each utterance, then perform latent class analysis (LCA) to infer correct labels. As it does not require GPU computations, LCA is inexpensive. In experiments, I found that the AES-like solution can infer training labels with high accuracy when trolls are consistent, even when trolls are the majority.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Learning to love diligent trolls: Accounting for rater effects in the dialogue safety task

Ilagan

2023

Findings of the Association for Computational Linguistics: EMNLP 2023

View full text Add to dashboard Cite

show abstract

“…Our work specifically focuses on improving TOD performance, and hence we caution users against any potential unsafe/toxic/offensive responses generated from the models. Without safety guardrails such as Arora et al (2022); Lu et al (2022), we do not advocate using any of our trained models in production settings.…”

Section: Ethical Considerationsmentioning

confidence: 99%

“…The reward function R would represent the "utility" of each action contributing towards the overall performance, such as task success in TOD. Typically, this is modeled by using R(s, a) = 0 for non-terminal states, and R(s T , a) for terminal states which can be computed by combining scores such as task success and BLEU (Ramamurthy et al, 2022;Arora et al, 2022). The transition function P : S × A → S would deterministically append the action a t to the current state s t so that s t+1 = (c 0 , .…”

Section: Introductionmentioning

confidence: 99%

KRLS: Improving End-to-End Response Generation in Task Oriented Dialog with Reinforced Keywords Learning

Yu,

Wu,

Qian

et al. 2023

Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

View full text Add to dashboard Cite

In task-oriented dialogs (TOD), reinforcement learning (RL) algorithms train a model to directly optimize response for task-related metrics. However, RL needs to perform exploration, which can be time-consuming due to the slow auto-regressive sequence generation process. We investigate an approach to create a more efficient RL-based algorithm to improve TOD performance in an offline setting. First, we use a faster generation procedure that samples from independent next-word distributions after training the language model (LM) with supervised learning. We then introduce a finegrained reward function to help the model focus on learning key information in a dialog, by measuring the importance and semantic closeness of each generated token. Experiments on the MultiWoZ dataset show our new training algorithm, Keywords Reinforcement Learning with Next-word Sampling (KRLS), achieves state-of-the-art performance on the end-to-end response generation task, with a 15% training time reduction compared to a standard RL algorithm using auto-regressive generation 1 .

show abstract

“…Chinese [7-9, 16, 42, 50] and Japanese [37]), but it consumes considerable computing resources for each language, which is not eco-friendly. In addition, it is unlikely to collect massive dialogue sessions of some languages due to the data scarcity and/or scatteredness problem [5]. Another line of investigation has focused on adaptation from pre-trained English models [1,2].…”

mentioning

confidence: 99%