We present iBERT (interpretable-BERT), an encoder to produce inherently interpretable and controllable embeddings - designed to modularize and expose the discriminative cues present in language, such as stylistic and semantic structure. Each input token is represented as a sparse, non-negative mixture over k context-independent sense vectors, which can be pooled into sentence embeddings or used directly at the token level. This enables modular control over representation, before any decoding or downstream use. To demonstrate our model’s interpretability, we evaluate it on a suite of style-focused tasks. On the STEL benchmark, it improves style representation effectiveness by  8 points over SBERT-style baselines, while maintaining competitive performance on authorship verification. Because each embedding is a structured composition of interpretable senses, we highlight how specific style attributes - such as emoji use, formality, or misspelling can be assigned to specific sense vectors. While our experiments center on style, iBERT is not limited to stylistic modeling. Its structural modularity is designed to interpretably decompose whichever discriminative signals are present in the data - enabling generalization even when supervision blends stylistic and semantic factors.

We propose a new approach for the authorship attribution task that leverages the various linguistic representations learned at different layers of pre-trained transformer-based models. We evaluate our approach on two popular authorship attribution models and three evaluation datasets, in in-domain and out-of-domain scenarios. We found that utilizing various transformer layers improves the robustness of authorship attribution models when tested on out-of-domain data, resulting in a much stronger performance. Our analysis gives further insights into how our model’s different layers get specialized in representing certain linguistic aspects that we believe benefit the model when tested out of the domain.

Distributed word embeddings have become ubiquitous in natural language processing as they have been shown to improve performance in many semantic and syntactic tasks. Popular models for learning cross-lingual word embeddings do not consider the morphology of words. We propose an approach to learn bilingual embeddings using parallel data and subword information that is expressed in various forms, i.e. character n-grams, morphemes obtained by unsupervised morphological segmentation and byte pair encoding. We report results for three low resource morphologically rich languages (Swahili, Tagalog, and Somali) and a high resource language (German) in a simulated a low-resource scenario. Our results show that our method that leverages subword information outperforms the model without subword information, both in intrinsic and extrinsic evaluations of the learned embeddings. Specifically, analogy reasoning results show that using subwords helps capture syntactic characteristics. Semantically, word similarity results and intrinsically, word translation scores demonstrate superior performance over existing methods. Finally, qualitative analysis also shows better-quality cross-lingual embeddings particularly for morphological variants in both languages.

We present iBERT (interpretable-BERT), an encoder to produce inherently interpretable and controllable embeddings - designed to modularize and expose the discriminative cues present in language, such as stylistic and semantic structure. Each input token is represented as a sparse, non-negative mixture over k context-independent sense vectors, which can be pooled into sentence embeddings or used directly at the token level. This enables modular control over representation, before any decoding or downstream use. To demonstrate our model’s interpretability, we evaluate it on a suite of style-focused tasks. On the STEL benchmark, it improves style representation effectiveness by  8 points over SBERT-style baselines, while maintaining competitive performance on authorship verification. Because each embedding is a structured composition of interpretable senses, we highlight how specific style attributes - such as emoji use, formality, or misspelling can be assigned to specific sense vectors. While our experiments center on style, iBERT is not limited to stylistic modeling. Its structural modularity is designed to interpretably decompose whichever discriminative signals are present in the data - enabling generalization even when supervision blends stylistic and semantic factors.

The disclosure herein describes training a document recommendation model using loss data generated from a linear score difference vector. A training data entry is provided including a query and a set of candidate documents. A document recommendation model generates a set of document prediction scores indicative of a likelihood that the candidate documents are responses to the query and a pairwise score difference matrix is generated using the set of document prediction scores. The pairwise score difference matrix is transformed into a score difference vector using a correct document vector that indicates a correct document among the set of candidate documents. Loss data of the document recommendation model is generated using the score difference vector and the document recommendation model is adjusted using the calculated loss data. Training the document recommendation model based on the linear score difference vector reduces resource usage when compared to training with a difference matrix.

We propose a new approach for the authorship attribution task that leverages the various linguistic representations learned at different layers of pre-trained transformer-based models. We evaluate our approach on two popular authorship attribution models and three evaluation datasets, in in-domain and out-of-domain scenarios. We found that utilizing various transformer layers improves the robustness of authorship attribution models when tested on out-of-domain data, resulting in a much stronger performance. Our analysis gives further insights into how our model’s different layers get specialized in representing certain linguistic aspects that we believe benefit the model when tested out of the domain.

Systems and methods are directed to training and utilizing a generative language model that is constrained by a predetermined template that is used to train the generative language model. Once trained, customer data is accessed and transmitted to an evaluation component associated with the generative language model. The generative language model generates one or more sentences based on a feedback input of the plurality of feedback inputs, whereby the one or more sentences each include a sentiment, a target, and a reason for the sentiment in a format defined by the predetermined template. The evaluation component then identifies the sentiment, the target, and the reason from a sentence of the one or more sentences. A communication is then presented, on a device of a user, based on at least the sentiment and the reason identified from the sentence. The communication can be an alert or a report.

The core challenge in numerous real-world applications is to match an inquiry to the best document from a mutable and finite set of candidates. Existing industry solutions, especially latency-constrained services, often rely on similarity algorithms that sacrifice quality for speed. In this paper we introduce a generic semantic learning-to-rank framework, Self-training Semantic Cross-attention Ranking (sRank). This transformer-based framework uses linear pairwise loss with mutable training batch sizes and achieves quality gains and high efficiency, and has been applied effectively to show gains on two industry tasks at Microsoft over real-world large-scale data sets: Smart Reply (SR) and Ambient Clinical Intelligence (ACI). In Smart Reply, sRank assists live customers with technical support by selecting the best reply from predefined solutions based on consumer and support agent messages. It achieves 11.7% gain in offline top-one accuracy on the SR task over the previous system, and has enabled 38.7% time reduction in composing messages in telemetry recorded since its general release in January 2021. In the ACI task, sRank selects relevant historical physician templates that serve as guidance for a text summarization model to generate higher quality medical notes. It achieves 35.5% top-one accuracy gain, along with 46% relative ROUGE-L gain in generated medical notes.

In the intricate problem of understanding long-form multi-modal inputs, few key-aspects in scene-understanding and dialogue-and-discourse are often overlooked. In this paper, we investigate two such key-aspects for better semantic and relational understanding - (i). head-object-tracking in addition to usual face-tracking, and (ii). fusing scene-to-text representation with external common-sense knowledge-base for effective mapping to sub-tasks of interest. The usage of head-tracking especially helps with enriching sparse entity mapping to inter-entity conversation interactions. These methods are guided by natural language supervision on visual models, and perform well for interaction and sentiment understanding tasks.

The machine learning community has recently had increased interest in the climate and disaster damage domain due to a marked increase in occurrences of natural hazards (e.g., hurricanes, forest fires, floods, earthquakes). However, not enough attention has been devoted to mitigating probable destruction from impending natural hazards. We explore this crucial space by predicting building-level damages on a before-the-fact basis that would allow state actors and non-governmental organizations to be best equipped with resource distribution to minimize or preempt losses. We introduce PREDISM that employs an ensemble of ResNets and fully connected layers over decision trees to capture image-level and meta-level information to accurately estimate weakness of man-made structures to disaster-occurrences. Our model performs well and is responsive to tuning across types of disasters and highlights the space of preemptive hazard damage modelling.

The natural language processing community has had a major interest in auto-regressive [4, 13] and span-prediction based language models [7] recently, while knowledge graphs are often referenced for common-sense based reasoning and fact-checking models. In this paper, we present an equivalence representation of span-prediction based language models and knowledge-graphs to better leverage recent developments of language modelling for multi-modal problem statements. Our method performed well, especially with sentiment understanding for multi-modal inputs, and discovered potential bias in naturally occurring videos when compared with movie-data interaction-understanding. We also release a dataset of an auto-generated questionnaire with ground-truths consisting of labels spanning across 120 relationships, 99 sentiments, and 116 interactions, among other labels for finer-grained analysis of model comparisons in the community.

We create multi-modal fusion models to predict relational classes within entities in free-form inputs such as unseen movies. Our approach identifies information rich features within individual sources – emotion, text-attention, age, gender, and contextual background object tracking. These information are absorbed and contrasted from baseline fusion architectures. These five models then showcase future research areas on this challenging problem of relational knowledge extraction from movies and free-form multi-modal input sources. We find that, generally, the Kinetics model added with Attributes and Objects beat the baseline models.

We consider the problem of building semantic relationship of unseen entities from free-form multi-modal sources. This intelligent agent understands semantic properties by creating (1) logical segments from sources, (2) finds interacting objects, (3) infers their interaction actions using (4) extracted textual, auditory, visual, and tonal information. The conversational dialogue discourses are automatically mapped to interacting co-located objects, and fused with their Kinetic action embeddings at each scene of occurrence. This generates a combined probability distribution representation for interacting entities spanning over every semantic relation class. Using these probabilities, we create knowledge graphs capable of answering semantic queries and infer missing properties in a given context.

Distributed word embeddings have become ubiquitous in natural language processing as they have been shown to improve performance in many semantic and syntactic tasks. Popular models for learning cross-lingual word embeddings do not consider the morphology of words. We propose an approach to learn bilingual embeddings using parallel data and subword information that is expressed in various forms, i.e. character n-grams, morphemes obtained by unsupervised morphological segmentation and byte pair encoding. We report results for three low resource morphologically rich languages (Swahili, Tagalog, and Somali) and a high resource language (German) in a simulated a low-resource scenario. Our results show that our method that leverages subword information outperforms the model without subword information, both in intrinsic and extrinsic evaluations of the learned embeddings. Specifically, analogy reasoning results show that using subwords helps capture syntactic characteristics. Semantically, word similarity results and intrinsically, word translation scores demonstrate superior performance over existing methods. Finally, qualitative analysis also shows better-quality cross-lingual embeddings particularly for morphological variants in both languages.

A computer - implemented method of fraud detection comprising receiving a user identification, a standard authentication key , and an alternative authentication key associated with a user. The method includes storing the standard and alternative authentication keys in a user profile associated with the user identification, and storing a contingent action corresponding to the alternative authentication key. The method includes receiving an authorization request including the user identification and an authentication input, and comparing the authentication input with the standard authentication key and the alternative authentication key in the user profile. The method includes determining that the authentication input matches the alternative authentication key. Based on the determination that the authentication input matches the alternative authentication key, the method includes initiating the contingent action stored in the user profile corresponding to the alternative authentication key. The method may include determining if the authorization request matches a third party fraud alert.

One embodiment of the invention is directed to a computer-implemented method comprising, receiving an indication that an avatar of a user has initiated a transaction in a virtual reality environment. The method further comprises obtaining a first biometric sample from the user interacting with the virtual reality hardware. The method further comprises generating a partial biometric template based at least in part on the first biometric sample. The method further comprises providing the partial biometric template and personal authentication information for the avatar to an authentication computer where the personal authentication information and the partial biometric template are used to authenticate the avatar.

Medical textual data present us with huge volumes of useful raw data, which can be pre-processed with the use of proper learning techniques. The information that is generated using the techniques can be leveraged to extract knowledge and predict varied medical correlations and conditions. Some examples of the predictions can include deciding the course of a given person’s condition in the context of the related data from the medical corpus, the correlation of varied diseases with each other, among others. To generate these analytical data, we have utilized the corpus of textual medical data to train artificial neural networks(ANNs). The ANNs have been loaded with the words from the corpus in an unsupervised fashion to train itself. The method employed utilizes the concepts of backpropagation using hidden layers. With the word vector representation generated, it can be used in varied tasks such as Parts-of-Speech tagging, Chunking, Named Entity Recognition, Semantic Role Labelling in a corpus’ data. In this part of the thesis, I have focussed on the window-based model to analyze and generate the word-vector representation so as to extend it to a multi-task framework once the vectors have been trained.