Effectively selecting data from subgroups where a model performs poorly is crucial for improving its performance. Traditional methods for identifying these subgroups often rely on sensitive information, raising privacy issues. Additionally, gathering such information at runtime might be impractical. This paper introduces a cost-effective strategy that addresses these concerns. We identify underperforming subgroups and train a model to predict if an utterance belongs to these subgroups without needing sensitive information. This model helps mitigate bias by selecting and adding new data, which is labeled as challenging, for re-training the speech model. Experimental results on intent classification and automatic speech recognition tasks show the effectiveness of our approach in reducing biases and enhancing performance, with improvements in reducing error rates of up to 39% for FSC, 16% for ITALIC, and 22% for LibriSpeech.

Video summarization aims to generate a condensed textual version of an original video. Summaries may consist of either plain text or a shortlist of salient events, possibly including temporal or spatial references. Video Large Language Models (VLLMs) exhibit impressive zero-shot capabilities in video analysis. However, their performance varies significantly according to the LLM prompt, the characteristics of the video, and the properties of the training data and LLM architecture.In this work, we thoroughly evaluate the zero-shot summarization performance of four state-of-the-art open-source VLLMs specifically designed to address spatial and temporal reasoning. In light of the detected summarization issues, we propose different cost-effective mitigation strategies, based on Chain-of-Thought prompting, that involve the injection of knowledge extracted by external, lightweight models. To perform the VLLM evaluation, we design a new video summarization benchmark consisting of 100 videos with varying characteristics in terms of domain, duration, and spatio-temporal properties. Videos are manually annotated by three independent human experts with plain text, event-based, and spatio-temporal summaries. The experimental evaluation shows that VLLMs significantly benefit from prompting a list of recognized actions, whereas injecting automatically recognized objects and scene changes respectively improve spatially contextualized and event-based summaries in specific cases.

The context of modern smart voice assistants is often multi-modal, where images, audio and video content are consumed by users simultaneously. In such a setup, co-reference resolution is especially challenging, and runs across modalities and dialogue turns. We explore the problem of multi-modal co-reference resolution in multi-turn dialogues and quantify the performance of multi-modal LLMs on a specially curated dataset of long, image-interleaved conversations between a voice assistant and human in a shopping use case. We propose a custom architecture for multi-modal embedding alignment using a novel parameter augmentation technique. Our proposed Parameter Augmented LLM approach shows a 4.9% absolute F1 improvement above a cross-attention baseline while reducing the number of parameters being trained by 4x.

Leveraging representations from pre-trained transformer-based encoders achieves state-of-the-art performance on numerous NLP tasks. Larger encoders can improve accuracy for spoken language understanding (SLU) but are challenging to use given the inference latency constraints of online systems (especially on CPU machines).We evaluate using a larger 170M parameter BERT encoder that shares representations across languages, domains and tasks for SLU compared to using smaller 17M parameter BERT encoders with language-, domain- and task-decoupled finetuning.Running inference with a larger shared encoder on GPU is latency neutral and reduces infrastructure cost compared to running inference for decoupled smaller encoders on CPU machines. The larger shared encoder reduces semantic error rates by 4.62% for test sets representing user requests to voice-controlled devices and 5.79% on the tail of the test sets on average across four languages.

Teacher-student knowledge distillation is a popular technique for compressing today’s prevailing large language models into manageable sizes that fit low-latency downstream applications. Both the teacher and the choice of transfer set used for distillation are crucial ingredients in creating a high quality student. Yet, the generic corpora used to pretrain the teacher and the corpora associated with the downstream target domain are often significantly different, which raises a natural question: should the student be distilled over the generic corpora, so as to learn from high-quality teacher predictions, or over the downstream task corpora to align with finetuning? Our study investigates this trade-off using Domain Classification (DC) and Intent Classification/Named Entity Recognition (ICNER) as downstream tasks. We distill several multilingual students from a larger multilingual LM with varying proportions of generic and task-specific datasets, and report their performance after finetuning on DC and ICNER. We observe significant improvements across tasks and test sets when only task-specific corpora is used. We also report on how the impact of adding task-specific data to the transfer set correlates with the similarity between generic and task-specific data. Our results clearly indicate that, while distillation from a generic LM benefits downstream tasks, students learn better using target domain data even if it comes at the price of noisier teacher predictions. In other words, target domain data still trumps teacher knowledge.

Unsupervised word alignments offer a lightweight and interpretable method to transfer labels from high- to low-resource languages, as long as semantically related words have the same label across languages. But such an assumption is often not true in industrial NLP pipelines, where multilingual annotation guidelines are complex and deviate from semantic consistency due to various factors (such as annotation difficulty, conflicting ontology, upcoming feature launches etc.);We address this difficulty by constraining the alignments models to remain consistent with both source and target annotation guidelines , leveraging posterior regularization and labeled examples. We illustrate the overall approach using IBM 2 (fast_align) as a base model, and report results on both internal and external annotated datasets. We measure consistent accuracy improvements on the MultiATIS++ dataset over AWESoME, a popular transformer-based alignment model, in the label projection task (+2.7% at word-level and +15% at sentence-level), and show how even a small amount of target language annotations help substantially.

True-casing, the task of restoring proper case to (generally) lower case input, is important in downstream tasks and for screen display. In this paper, we investigate truecasing as an in- trinsic task and present several experiments on noisy user queries to a voice-controlled dia- log system. In particular, we compare a rule- based, an n-gram language model (LM) and a recurrent neural network (RNN) approaches, evaluating the results on a German Q&A cor- pus and reporting accuracy for different case categories. We show that while RNNs reach higher accuracy especially on large datasets, character n-gram models with interpolation are still competitive, in particular on mixed- case words where their fall-back mechanisms come into play.