Reinforcement Learning from Human Feedback (RLHF) leverages human preference data to train language models to align more closely with human essence. These human preference data, however, are labeled at the sequence level, creating a mismatch between sequence-level preference labels and tokens, which are autoregressively generated from the language model. Although several recent approaches have tried to provide token-level (i.e., dense) rewards for each individual token, these typically rely on predefined discrete reward values (e.g., positive: +1, negative: -1, neutral: 0), failing to account for varying degrees of preference inherent to each token. To address this limitation, we introduce TLCR (Token-Level Continuous Reward) for RLHF, which incorporates a discriminator trained to distinguish positive and negative tokens, and the confidence of the discriminator is used to assign continuous rewards to each token considering the context. Extensive experiments show that our proposed TLCR leads to consistent performance improvements over previous sequence-level or token-level discrete rewards on open-ended generation benchmarks.

The growing emphasis on fairness in speech-processing tasks requires datasets with speakers from diverse subgroups that allow training and evaluating fair speech technology systems. However, creating such datasets through manual annotation can be costly. To address this challenge, we present a semi-automated dataset creation pipeline that leverages large language models. We use this pipeline to generate a dataset of speakers identifying themself or another speaker as belonging to a particular race, ethnicity, or national origin group. We use OpenaAI’s GPT-4 to perform two complex annotation tasks- separating files relevant to our intended dataset from the irrelevant ones (filtering) and finding and extracting information on identifications within a transcript (tagging). By evaluating GPT-4’s performance using human annotations as ground truths, we show that it can reduce resources required by dataset annotation while barely losing any important information. For the filtering task, GPT-4 had a very low miss rate of 6.93%. GPT-4’s tagging performance showed a trade-off between precision and recall, where the latter got as high as 97%, but precision never exceeded 45%. Our approach reduces the time required for the filtering and tagging tasks by 95% and 80%, respectively. We also present an in-depth error analysis of GPT-4’s performance.

Unsupervised speech recognition ({pasted macro ‘ASRU’}/) is the problem of learning automatic speech recognition (ASR) systems from unpaired speech-only and text-only corpora. While various algorithms exist to solve this problem, a theoretical framework is missing to study their properties and address such issues as sensitivity to hyperparameters and training instability. In this paper, we proposed a general theoretical framework to study the properties of {pasted macro ‘ASRU’}/ systems based on random matrix theory and the theory of neural tangent kernels. Such a framework allows us to prove various learnability conditions and sample complexity bounds of {pasted macro ‘ASRU’}/. Extensive {pasted macro ‘ASRU’}/ experiments on synthetic languages with three classes of transition graphs provide strong empirical evidence for our theory (code available at https://github.com/cactuswiththoughts/UnsupASRTheory.gitcactuswiththoughts/UnsupASRTheory.git).

Existing supervised sign language recognition systems rely on an abundance of well-annotated data. Instead, an unsupervised speech-to-sign language recognition (SSR-U) system learns to translate between spoken and sign languages by observing only non-parallel speech and sign-language corpora. We propose speech2sign-U, a neural network-based approach capable of both character-level and word-level SSR-U. Our approach significantly outperforms baselines directly adapted from unsupervised speech recognition (ASR-U) models by as much as 50% recall@10 on several challenging American sign language corpora with various levels of sample sizes, vocabulary sizes, and audio and visual variability. The code is available at https://github.com/cactuswiththoughts/UnsupSpeech2Sign.gitcactuswiththoughts/UnsupSpeech2Sign.git.

Automatic Speech Recognition (ASR) systems have attained unprecedented performance with large speech models pre-trained based on self-supervised speech representation learning. However, these pre-trained speech models suffer from representational bias as they tend to better represent those prominent accents (i.e., native (L1) English accent) in the pre-training speech corpus than less represented accents, resulting in a deteriorated performance for non-native (L2) English accents. Although there have been some approaches to mitigate this issue, all of these methods require updating the pre-trained model weights. In this paper, we propose Information Theoretic Adversarial Prompt Tuning (INTapt), which introduces prompts concatenated to the original input that can re-modulate the attention of the pre-trained model such that the corresponding input resembles a native (L1) English speech without updating the backbone weights. INTapt is trained simultaneously in the following two manners: (1) adversarial training to reduce accent feature dependence between the original input and the prompt-concatenated input and (2) training to minimize CTC loss for improving ASR performance to a prompt-concatenated input. Experimental results show that INTapt improves the performance of L2 English and increases feature similarity between L2 and L1 accents.

Exemplification modeling is a task where the goal is to produce a viable example sentence that uses a target word with a target definition. The task is non-trivial for polysemous words, and previous works have only explored settings where ample labeled training data is available. In this paper, we demonstrate that exemplification modeling can be performed without a large labeled training corpus by either changing the format of the task (one-shot) or prompting large language models (few-shot), and ablate key components of our proposed one-shot and few-shot systems. We provide extensive automatic and human evaluations of model performance and find that our proposed one-shot and few-shot approaches perform similarly to a fully supervised baseline. We compare and contrast each method in terms of labeled training dataset size, performance, and model size, and find that each technique has at least one tradeoff that another approach does not.

In this paper we focus on patterns of colexification (co-expressions of form-meaning mapping in the lexicon) as an aspect of lexical-semantic organization, and use them to build large scale synset graphs across BabelNet’s typologically diverse set of 499 world languages. We introduce and compare several approaches: monolingual and cross-lingual colexification graphs, popular distributional models, and fusion approaches. The models are evaluated against human judgments on a semantic similarity task for nine languages. Our strong empirical findings also point to the importance of universality of our graph synset embedding representations with no need for any language-specific adaptation when evaluated on the lexical similarity task. The insights of our exploratory investigation of large-scale colexification graphs could inspire significant advances in NLP across languages, especially for tasks involving languages which lack dedicated lexical resources, and can benefit from language transfer from large shared cross-lingual semantic spaces.

Phonemes are defined by their relationship to words: changing a phoneme changes the word. Learning a phoneme inventory with little supervision has been a longstanding challenge with important applications to under-resourced speech technology. In this paper, we bridge the gap between the linguistic and statistical definition of phonemes and propose a novel neural discrete representation learning model for self-supervised learning of phoneme inventory with raw speech and word labels. Under mild assumptions, we prove that the phoneme inventory learned by our approach converges to the true one with an exponentially low error rate. Moreover, in experiments on TIMIT and Mboshi benchmarks, our approach consistently learns a better phoneme-level representation and achieves a lower error rate in a zero-resource phoneme recognition task than previous state-of-the-art self-supervised representation learning algorithms.

Word Sense Disambiguation (WSD) is an NLP task aimed at determining the correct sense of a word in a sentence from discrete sense choices. Although current systems have attained unprecedented performances for such tasks, the nonuniform distribution of word senses during training generally results in systems performing poorly on rare senses. To this end, we consider data augmentation to increase the frequency of these least frequent senses (LFS) to reduce the distributional bias of senses during training. We propose Sense-Maintained Sentence Mixup (SMSMix), a novel word-level mixup method that maintains the sense of a target word. SMSMix smoothly blends two sentences using mask prediction while preserving the relevant span determined by saliency scores to maintain a specific word’s sense. To the best of our knowledge, this is the first attempt to apply mixup in NLP while preserving the meaning of a specific word. With extensive experiments, we validate that our augmentation method can effectively give more information about rare senses during training with maintained target sense label.

Although Question-Answering has long been of research interest, its accessibility to users through a speech interface and its support to multiple languages have not been addressed in prior studies. Towards these ends, we present a new task and a synthetically-generated dataset to do Fact-based Visual Spoken-Question Answering (FVSQA). FVSQA is based on the FVQA dataset, which requires a system to retrieve an entity from Knowledge Graphs (KGs) to answer a question about an image. In FVSQA, the question is spoken rather than typed. Three sub-tasks are proposed: (1) speech-to-text based, (2) end-to-end, without speech-to-text as an intermediate component, and (3) cross-lingual, in which the question is spoken in a language different from that in which the KG is recorded. The end-to-end and cross-lingual tasks are the first to require world knowledge from a multi-relational KG as a differentiable layer in an end-to-end spoken language understanding task, hence the proposed reference implementation is called Worldly-Wise (WoW).WoW is shown to perform end-to-end cross-lingual FVSQA at same levels of accuracy across 3 languages - English, Hindi, and Turkish.

Healthcare systems have increased patients’ exposure to their own health materials to enhance patients’ health levels, but this has been impeded by patients’ lack of understanding of their health material. We address potential barriers to their comprehension by developing a context-aware text simplification system for health material. Given the scarcity of annotated parallel corpora in healthcare domains, we design our system to be independent of a parallel corpus, complementing the availability of data-driven neural methods when such corpora are available. Our system compensates for the lack of direct supervision using a biomedical lexical database: Unified Medical Language System (UMLS). Compared to a competitive prior approach that uses a tool for identifying biomedical concepts and a consumer-directed vocabulary list, we empirically show the enhanced accuracy of our system due to improved handling of ambiguous terms. We also show the enhanced accuracy of our system over directly-supervised neural methods in this low-resource setting. Finally, we show the direct impact of our system on laypeople’s comprehension of health material via a human subjects’ study (n=160).

Acquiring labeled speech for low-resource languages is a difficult task in the absence of native speakers of the language. One solution to this problem involves collecting speech transcriptions from crowd workers who are foreign or non-native speakers of a given target language. From these mismatched transcriptions, one can derive probabilistic phone transcriptions that are defined over the set of all target language phones using a noisy channel model. This paper extends prior work on deriving probabilistic transcriptions (PTs) from mismatched transcriptions by 1) modelling multilingual channels and 2) introducing a clustering-based phonetic mapping technique to improve the quality of PTs. Mismatched crowdsourcing for multilingual channels has certain properties of projection mapping, e.g., it can be interpreted as a clustering based on singular value decomposition of the segment alignments. To this end, we explore the use of distinctive feature weights, lexical tone confusions, and a two-step clustering algorithm to learn projections of phoneme segments from mismatched multilingual transcriber languages to the target language. We evaluate our techniques using mismatched transcriptions for Cantonese speech acquired from native English and Mandarin speakers. We observe a 5-9% relative reduction in phone error rate for the predicted Cantonese phone transcriptions using our proposed techniques compared with the previous PT method.

A major problem with dialectal Arabic speech recognition is due to the sparsity of speech resources. In this paper, a transfer learning framework is proposed to jointly use a large amount of Modern Standard Arabic (MSA) data and little amount of dialectal Arabic data to improve acoustic and language modeling. The Qatari Arabic (QA) dialect has been chosen as a typical example for an under-resourced Arabic dialect. A wide-band speech corpus has been collected and transcribed from several Qatari TV series and talk-show programs. A large vocabulary speech recognition baseline system was built using the QA corpus. The proposed MSA-based transfer learning technique was performed by applying orthographic normalization, phone mapping, data pooling, acoustic model adaptation, and system combination. The proposed approach can achieve more than 28% relative reduction in WER.

In this paper, a framework for long audio alignment for conversational Arabic speech is proposed. Accurate alignments help in many speech processing tasks such as audio indexing, speech recognizer acoustic model (AM) training, audio summarizing and retrieving, etc. We have collected more than 1,400 hours of conversational Arabic besides the corresponding human generated non-aligned transcriptions. Automatic audio segmentation is performed using a split and merge approach. A biased language model (LM) is trained using the corresponding text after a pre-processing stage. Because of the dominance of non-standard Arabic in conversational speech, a graphemic pronunciation model (PM) is utilized. The proposed alignment approach is performed in two passes. Firstly, a generic standard Arabic AM is used along with the biased LM and the graphemic PM in a fast speech recognition pass. In a second pass, a more restricted LM is generated for each audio segment, and unsupervised acoustic model adaptation is applied. The recognizer output is aligned with the processed transcriptions using Levenshtein algorithm. The proposed approach resulted in an initial alignment accuracy of 97.8-99.0% depending on the amount of disfluencies. A confidence scoring metric is proposed to accept/reject aligner output. Using confidence scores, it was possible to reject the majority of mis-aligned segments resulting in alignment accuracy of 99.0-99.8% depending on the speech domain and the amount of disfluencies.