Word error rate (WER) is a standard metric for the evaluation of Automated Speech Recognition (ASR) systems. However, WER fails to provide a fair evaluation of human perceived quality in presence of spelling variations, abbreviations, or compound words arising out of agglutination. Multiple spelling variations might be acceptable based on locale/geography, alternative abbreviations, borrowed words, and transliteration of code-mixed words from a foreign language to the target language script. Similarly, in case of agglutination, often times the agglutinated, as well as the split forms, are acceptable. Previous work handled this problem by using manually identified normalization pairs and applying them to both the transcription and the hypothesis before computing WER. In this paper, we propose an automatic WER normalization system consisting of two modules: spelling normalization and segmentation normalization. The proposed system is unsupervised and language agnostic, and therefore scalable. Experiments with ASR on 35K utterances across four languages yielded an average WER reduction of 13.28%. Human judgements of these automatically identified normalization pairs show that our WER-normalized evaluation is highly consistent with the perceived quality of ASR output.
Syntactic parsing is the task of assigning a syntactic structure to a sentence. There are two popular syntactic parsing methods: constituency and dependency parsing. Recent works have used syntactic embeddings based on constituency trees, incremental top-down parsing, and other word syntactic features for brain activity prediction given the text stimuli to study how the syntax structure is represented in the brain's language network. However, the effectiveness of dependency parse trees or the relative predictive power of the various syntax parsers across brain areas, especially for the listening task, is yet unexplored. In this study, we investigate the predictive power of the brain encoding models in three settings: (i) individual performance of the constituency and dependency syntactic parsing based embedding methods, (ii) efficacy of these syntactic parsing based embedding methods when controlling for basic syntactic signals, (iii) relative effectiveness of each of the syntactic embedding methods when controlling for the other. Further, we explore the relative importance of syntactic information (from these syntactic embedding methods) versus semantic information using BERT embeddings. We find that constituency parsers help explain activations in the temporal lobe and middle-frontal gyrus, while dependency parsers better encode syntactic structure in the angular gyrus and posterior cingulate cortex. Although semantic signals from BERT are more effective compared to any of the syntactic features or embedding methods, syntactic embedding methods explain additional variance for a few brain regions.
Multiple studies have focused on predicting the prospective popularity of an online document as a whole, without paying attention to the contributions of its individual parts. We introduce the task of proactively forecasting popularities of sentences within online news documents solely utilizing their natural language content. We model sentence-specific popularity forecasting as a sequence regression task. For training our models, we curate InfoPop, the first dataset containing popularity labels for over 1.7 million sentences from over 50,000 online news documents. To the best of our knowledge, this is the first dataset automatically created using streams of incoming search engine queries to generate sentence-level popularity annotations. We propose a novel transfer learning approach involving sentence salience prediction as an auxiliary task. Our proposed technique coupled with a BERT-based neural model exceeds nDCG values of 0.8 for proactive sentence-specific popularity forecasting. Notably, our study presents a non-trivial takeaway: though popularity and salience are different concepts, transfer learning from salience prediction enhances popularity forecasting. We release InfoPop and make our code publicly available: https://github.com/sayarghoshroy/InfoPopularity
Language models have been shown to be very effective in predicting brain recordings of subjects experiencing complex language stimuli. For a deeper understanding of this alignment, it is important to understand the alignment between the detailed processing of linguistic information by the human brain versus language models. In NLP, linguistic probing tasks have revealed a hierarchy of information processing in neural language models that progresses from simple to complex with an increase in depth. On the other hand, in neuroscience, the strongest alignment with high-level language brain regions has consistently been observed in the middle layers. These findings leave an open question as to what linguistic information actually underlies the observed alignment between brains and language models. We investigate this question via a direct approach, in which we eliminate information related to specific linguistic properties in the language model representations and observe how this intervention affects the alignment with fMRI brain recordings obtained while participants listened to a story. We investigate a range of linguistic properties (surface, syntactic and semantic) and find that the elimination of each one results in a significant decrease in brain alignment across all layers of a language model. These findings provide direct evidence for the role of specific linguistic information in the alignment between brain and language models, and opens new avenues for mapping the joint information processing in both systems.
Multiple business scenarios require an automated generation of descriptive human-readable text from structured input data. Hence, fact-to-text generation systems have been developed for various downstream tasks like generating soccer reports, weather and financial reports, medical reports, person biographies, etc. Unfortunately, previous work on fact-to-text (F2T) generation has focused primarily on English mainly due to the high availability of relevant datasets. Only recently, the problem of cross-lingual fact-to-text (XF2T) was proposed for generation across multiple languages alongwith a dataset, XALIGN for eight languages. However, there has been no rigorous work on the actual XF2T generation problem. We extend XALIGN dataset with annotated data for four more languages: Punjabi, Malayalam, Assamese and Oriya. We conduct an extensive study using popular Transformer-based text generation models on our extended multi-lingual dataset, which we call XALIGNV2. Further, we investigate the performance of different text generation strategies: multiple variations of pretraining, fact-aware embeddings and structure-aware input encoding. Our extensive experiments show that a multi-lingual mT5 model which uses fact-aware embeddings with structure-aware input encoding leads to best results on average across the twelve languages. We make our code, dataset and model publicly available, and hope that this will help advance further research in this critical area.
Natural Language Generation (NLG) represents a large collection of tasks in the field of NLP. While many of these tasks have been tackled well by the cross-entropy (CE) loss, the task of dialog generation poses a few unique challenges for this loss function. First, CE loss assumes that for any given input, the only possible output is the one available as the ground truth in the training dataset. In general, this is not true for any task, as there can be multiple semantically equivalent sentences, each with a different surface form. This problem gets exaggerated further for the dialog generation task, as there can be multiple valid responses (for a given context) that not only have different surface forms but are also not semantically equivalent. Second, CE loss does not take the context into consideration while processing the response and, hence, it treats all ground truths with equal importance irrespective of the context. But, we may want our final agent to avoid certain classes of responses (e.g. bland, non-informative or biased responses) and give relatively higher weightage for more context-specific responses. To circumvent these shortcomings of the CE loss, in this paper, we propose a novel loss function, CORAL, that directly optimizes recently proposed estimates of human preference for generated responses. Using CORAL, we can train dialog generation models without assuming non-existence of response other than the ground-truth. Also, the CORAL loss is computed based on both the context and the response. Extensive comparisons on two benchmark datasets show that the proposed methods outperform strong state-of-the-art baseline models of different sizes.
Several popular Transformer based language models have been found to be successful for text-driven brain encoding. However, existing literature leverages only pretrained text Transformer models and has not explored the efficacy of task-specific learned Transformer representations. In this work, we explore transfer learning from representations learned for ten popular natural language processing tasks (two syntactic and eight semantic) for predicting brain responses from two diverse datasets: Pereira (subjects reading sentences from paragraphs) and Narratives (subjects listening to the spoken stories). Encoding models based on task features are used to predict activity in different regions across the whole brain. Features from coreference resolution, NER, and shallow syntax parsing explain greater variance for the reading activity. On the other hand, for the listening activity, tasks such as paraphrase generation, summarization, and natural language inference show better encoding performance. Experiments across all 10 task representations provide the following cognitive insights: (i) language left hemisphere has higher predictive brain activity versus language right hemisphere, (ii) posterior medial cortex, temporo-parieto-occipital junction, dorsal frontal lobe have higher correlation versus early auditory and auditory association cortex, (iii) syntactic and semantic tasks display a good predictive performance across brain regions for reading and listening stimuli resp.
Enabling effective brain-computer interfaces requires understanding how the human brain encodes stimuli across modalities such as visual, language (or text), etc. Brain encoding aims at constructing fMRI brain activity given a stimulus. There exists a plethora of neural encoding models which study brain encoding for single mode stimuli: visual (pretrained CNNs) or text (pretrained language models). Few recent papers have also obtained separate visual and text representation models and performed late-fusion using simple heuristics. However, previous work has failed to explore: (a) the effectiveness of image Transformer models for encoding visual stimuli, and (b) co-attentive multi-modal modeling for visual and text reasoning. In this paper, we systematically explore the efficacy of image Transformers (ViT, DEiT, and BEiT) and multi-modal Transformers (VisualBERT, LXMERT, and CLIP) for brain encoding. Extensive experiments on two popular datasets, BOLD5000 and Pereira, provide the following insights. (1) To the best of our knowledge, we are the first to investigate the effectiveness of image and multi-modal Transformers for brain encoding. (2) We find that VisualBERT, a multi-modal Transformer, significantly outperforms previously proposed single-mode CNNs, image Transformers as well as other previously proposed multi-modal models, thereby establishing new state-of-the-art. The supremacy of visio-linguistic models raises the question of whether the responses elicited in the visual regions are affected implicitly by linguistic processing even when passively viewing images. Future fMRI tasks can verify this computational insight in an appropriate experimental setting.
How the brain captures the meaning of linguistic stimuli across multiple views is still a critical open question in neuroscience. Consider three different views of the concept apartment: (1) picture (WP) presented with the target word label, (2) sentence (S) using the target word, and (3) word cloud (WC) containing the target word along with other semantically related words. Unlike previous efforts, which focus only on single view analysis, in this paper, we study the effectiveness of brain decoding in a zero-shot cross-view learning setup. Further, we propose brain decoding in the novel context of cross-view-translation tasks like image captioning (IC), image tagging (IT), keyword extraction (KE), and sentence formation (SF). Using extensive experiments, we demonstrate that cross-view zero-shot brain decoding is practical leading to ~0.68 average pairwise accuracy across view pairs. Also, the decoded representations are sufficiently detailed to enable high accuracy for cross-view-translation tasks with following pairwise accuracy: IC (78.0), IT (83.0), KE (83.7) and SF (74.5). Analysis of the contribution of different brain networks reveals exciting cognitive insights: (1) A high percentage of visual voxels are involved in image captioning and image tagging tasks, and a high percentage of language voxels are involved in the sentence formation and keyword extraction tasks. (2) Zero-shot accuracy of the model trained on S view and tested on WC view is better than same-view accuracy of the model trained and tested on WC view.
While transformers demonstrate impressive performance on many knowledge intensive (KI) tasks, their ability to serve as implicit knowledge bases (KBs) remains limited, as shown on several slot-filling, question-answering (QA), fact verification, and entity-linking tasks. In this paper, we implement an efficient, data-programming technique that enriches training data with KB-derived context and improves transformer utilization of encoded knowledge when fine-tuning for a particular QA task, namely answer sentence selection (AS2). Our method outperforms state of the art transformer approach on WikiQA and TrecQA, two widely studied AS2 benchmarks, increasing by 2.0% p@1, 1.3% MAP, 1.1% MRR, and 4.4% p@1, 0.9% MAP, 2.4% MRR, respectively. To demonstrate our improvements in an industry setting, we additionally evaluate our approach on a proprietary dataset of Alexa QA pairs, and show increase of 2.3% F1 and 2.0% MAP. We additionally find that these improvements remain even when KB context is omitted at inference time, allowing for the use of our models within existing transformer workflows without additional latency or deployment costs.