An entity linking system links named entities in a text to their corresponding entries in a knowledge base. In recent years, building an entity linking system that leverages the transformer architecture has gained lots of attention. However, deploying a transformer-based neural entity linking system in industrial production environments in a limited resource setting is a challenging task. In this work, we present an entity linking system that leverages a transformer-based BERT encoder (the BLINK model) to connect the product and organization type entities in business phone conversations to their corresponding Wikipedia entries. We propose a dimensionality reduction technique via utilizing an auto encoder that can effectively compress the dimension of the pre-trained BERT embeddings to 256 from the original size of 1024. This allows our entity linking system to significantly optimize the space requirement when deployed in a resource limited cloud machine while reducing the inference time along with retaining high accuracy.

We describe a method to define and use subwebs, user-defined neighborhoods of the Internet. Subwebs help improve search performance by inducing a topic-specific page relevance bias over a collection of documents. Subwebs may be automatically identified using a simple algorithm we describe, and used to provide highly-relevant topic-specific information retrieval. Using subwebs in a Help and Support topic, we see marked improvements in precision compared to generic search engine results.

We approach to correcting features in transferred linguistic representations in machine translation. The hybrid approach combines decision trees and transformation-based learning. Decision trees serve as a filter on the intractably large search space of possible interrelations among features. Transformation-based learning results in a simple set of ordered rules that can be compiled and executed after transfer and before sentence realization in the target language. We measure the reduction in noise in the linguistic representations and the results of human evaluations of end-to-end English-German machine translation.

We show that it is possible to learn the contexts for linguistic operations which map a semantic representation to a surface syntactic tree in sentence realization with high accuracy. We cast the problem of learning the contexts for the linguistic operations as classification tasks, and apply straightforward machine learning techniques, such as decision tree learning. The training data consist of linguistic features extracted from syntactic and semantic representations produced by a linguistic analysis system. The target features are extracted from links to surface syntax trees. Our evidence consists of four examples from the German sentence realization system code-named Amalgam: case assignment, assignment of verb position features, extraposition, and syntactic aggregation.

We present a machine learning approach to evaluating the well-formedness of output of a machine translation system, using classifiers that learn to distinguish human reference translations from machine translations. This approach can be used to evaluate an MT system, tracking improvements over time; to aid in the kind of failure analysis that can help guide system development; and to select among alternative output strings. The method presented is fully automated and independent of source language, target language and domain.

We present an automated, system-internal evaluation technique for linguistic representations in a large-scale, multilingual MT system. We use machine-learned classifiers to recognize the differences between linguistic representations generated from transfer in an MT context from representations that are produced by “native” analysis of the target language. In the MT scenario, convergence of the two is the desired result. Holding the feature set and the learning algorithm constant, the accuracy of the classifiers provides a measure of the overall difference between the two sets of linguistic representations: classifiers with higher accuracy correspond to more pronounced differences between representations. More importantly, the classifiers yield the basis for error-analysis by providing a ranking of the importance of linguistic features. The more salient a linguistic criterion is in discriminating transferred representations from “native” representations, the more work will be needed in order to get closer to the goal of producing native-like MT. We present results from using this approach on the Microsoft MT system and discuss its advantages and possible extensions.

We perform a linguistic analysis of documents during indexing for information retrieval. By eliminating index terms that occur only in subordinate clauses, index size is reduced by approximately 30% without adversely affecting precision or recall. These results hold for two corpora: a sample of the world wide web and an electronic encyclopedia.