Improving QA Model Performance with Cartographic Inoculation

QA models are faced with complex and open-ended contextual reasoning problems, but can often learn well-performing solution heuristics by exploiting dataset-specific patterns in their training data. These patterns, or "dataset artifacts", reduce the model's ability to generalize to real-world QA problems. Utilizing an ElectraSmallDiscriminator model trained for QA, we analyze the impacts and incidence of dataset artifacts using an adversarial challenge set designed to confuse models reliant on artifacts for prediction. Extending existing work on methods for mitigating artifact impacts, we propose cartographic inoculation, a novel method that fine-tunes models on an optimized subset of the challenge data to reduce model reliance on dataset artifacts. We show that by selectively fine-tuning a model on ambiguous adversarial examples from a challenge set, significant performance improvements can be made on the full challenge dataset with minimal loss of model generalizability to other challenging environments and QA datasets.

High-precision Voice Search Query Correction via Retrievable Speech-text Embedings

Automatic speech recognition (ASR) systems can suffer from poor recall for various reasons, such as noisy audio, lack of sufficient training data, etc. Previous work has shown that recall can be improved by retrieving rewrite candidates from a large database of likely, contextually-relevant alternatives to the hypothesis text using nearest-neighbors search over embeddings of the ASR hypothesis text to correct and candidate corrections. However, ASR-hypothesis-based retrieval can yield poor precision if the textual hypotheses are too phonetically dissimilar to the transcript truth. In this paper, we eliminate the hypothesis-audio mismatch problem by querying the correction database directly using embeddings derived from the utterance audio; the embeddings of the utterance audio and candidate corrections are produced by multimodal speech-text embedding networks trained to place the embedding of the audio of an utterance and the embedding of its corresponding textual transcript close together. After locating an appropriate correction candidate using nearest-neighbor search, we score the candidate with its speech-text embedding distance before adding the candidate to the original n-best list. We show a relative word error rate (WER) reduction of 6% on utterances whose transcripts appear in the candidate set, without increasing WER on general utterances.