Abstract: Systems and methods are provided for improving language models for speech recognition by personalizing knowledge sources utilized by the language models to specific users or user-population characteristics. A knowledge source, such as a knowledge graph, is personalized for a particular user by mapping entities or user actions from usage history for the user, such as query logs, to the knowledge source. The personalized knowledge source may be used to build a personal language model by training a language model with queries corresponding to entities or entity pairs that appear in usage history. In some embodiments, a personalized knowledge source for a specific user can be extended based on personalized knowledge sources of similar users.

Abstract: Systems and methods are provided for training language models using in-domain-like data collected automatically from one or more data sources. The data sources (such as text data or user-interactional data) are mined for specific types of data, including data related to style, content, and probability of relevance, which are then used for language model training. In one embodiment, a language model is trained from features extracted from a knowledge graph modified into a probabilistic graph, where entity popularities are represented and the popularity information is obtained from data sources related to the knowledge. Embodiments of language models trained from this data are particularly suitable for domain-specific conversational understanding tasks where natural language is used, such as user interaction with a game console or a personal assistant application on personal device.

Abstract: The customization of language modeling components for speech recognition is provided. A list of language modeling components may be made available by a computing device. A hint may then be sent to a recognition service provider for combining the multiple language modeling components from the list. The hint may be based on a number of different domains. A customized combination of the language modeling components based on the hint may then be received from the recognition service provider.

Abstract: An incremental speech recognition system. The incremental speech recognition system incrementally decodes a spoken utterance using an additional utterance decoder only when the additional utterance decoder is likely to add significant benefit to the combined result. The available utterance decoders are ordered in a series based on accuracy, performance, diversity, and other factors. A recognition management engine coordinates decoding of the spoken utterance by the series of utterance decoders, combines the decoded utterances, and determines whether additional processing is likely to significantly improve the recognition result. If so, the recognition management engine engages the next utterance decoder and the cycle continues. If the accuracy cannot be significantly improved, the result is accepted and decoding stops.

Abstract: Embodiment described herein are directed to an enhanced search engine with multiple feedback loops for providing optimal search results that are responsive a user's search query. The user's search query is parsed, and based on the underlying terms, different linguistic models and refinement techniques generate alternative candidate search queries that may yield better results. Searches are performed for the original search query and the candidate search queries, and different scores are used to select the best search results to present to the user. Results making it onto the list, as well as the underlying candidate search query, linguistic model, or refinement technique for generating that search query, will have their corresponding scores updated to reflect their success of generating a search result. Scores are stored and used by future searches to come up with better results.

Abstract: One or more techniques and/or systems are provided for maintaining user tagged content. For example, a user may experience content (e.g., watch a scene of a movie, create a photo, create a social network post, read an email, etc.), which the user may desire to save and/or organize for later retrieval. Accordingly, a personalization tag for the content may be received from the user (e.g., “Paris vacation photo”). The content may be indexed with the personalization tag within a personalization index (e.g., a cloud-based index for the user that may be accessible to any device associated with the user). In this way, the user may retrieve the content at a later point in time from any device. For example, a search query “Paris photos” may be received from the user. The personalization index may be queried using the search query to identify content that may be provided to the user.

Abstract: The present invention is directed towards systems and methods for providing search assistance technologies based on a user's search self-efficacy and search frustration. The method according to one embodiment of the present invention comprises receiving a search query from a user. The method then calculates the user's search self-efficacy and calculates the user's frustration with the current information task. The method then identifies a plurality of search assistance technologies based on the user's search self-efficacy and frustration. Finally, the method provides a search engine results page comprising a plurality of search results and the identified plurality of search assistance technologies.

Abstract: Content is proactively presented to a user, to enable the user to more efficiently access such content. A user context is correlated to content that is likely to be subsequently accessed. One such a correlation is specific to a given user, while another such correlation is general to a collection, or class, of users. Correlations between a current user context and content subsequently accessed are based on historical data and are defined in terms of mathematical functions or semantic relationships. Such correlations are then utilized to identify content that is likely to be subsequently accessed, and such content is proactively presented to a user. A user interface provides a defined area within which proactive presentations of content are made, including while the user is utilizing other application programs.

Abstract: Embodiment described herein are directed to an enhanced search engine with multiple feedback loops for providing optimal search results that are responsive a user's search query. The user's search query is parsed, and based on the underlying terms, different linguistic models and refinement techniques generate alternative candidate search queries that may yield better results. Searches are performed for the original search query and the candidate search queries, and different scores are used to select the best search results to present to the user. Results making it onto the list, as well as the underlying candidate search query, linguistic model, or refinement technique for generating that search query, will have their corresponding scores updated to reflect their success of generating a search result. Scores are stored and used by future searches to come up with better results.

Abstract: One embodiment selects from a set of query-suggestion pairs a first query and a subset of query-suggestion pairs that each has the first query as its query; computes a Log Likelihood Ratio (LLR) value for each query-suggestion pair from the subset of query-suggestion pairs; ranks the subset of query-suggestion pairs according to their respective LLR values; removes from the subset of query-suggestion pairs all query-suggestion pairs whose LLR values are below a predetermined LLR threshold; computes a Pointwise Mutual Information (PMI) value for each remaining query suggestion pair from the subset of query-suggestion pairs; removes from the subset of query-suggestion pairs all query-suggestion pairs whose PMI values are below a predetermine PMI threshold; and constructs a ranked set of suggestions for the first query, wherein the ranked set of suggestions comprises one or more suggestions of the remaining query-suggestion pairs from the subset of query-suggestion pairs.

Abstract: The present invention is directed towards systems and methods for providing search assistance technologies based on a user's search self-efficacy and search frustration. The method according to one embodiment of the present invention comprises receiving a search query from a user. The method then calculates the user's search self-efficacy and calculates the user's frustration with the current information task. The method then identifies a plurality of search assistance technologies based on the user's search self-efficacy and frustration. Finally, the method provides a search engine results page comprising a plurality of search results and the identified plurality of search assistance technologies.

Abstract: A method for handling abbreviations in web queries includes building a dictionary of possible word expansions for potential abbreviations related to query terms received and anticipated to be received by a search engine; accepting a query including an abbreviation from a searching user, where a probability of finding a most probably-correct expansion in the dictionary is a first probability, and a probability that the expansion is the abbreviation itself is a second probability; determining a ratio between the first and second probabilities; expanding the abbreviation in accordance with the most probably-correct expansion when the ratio is above a first threshold value; and highlighting the abbreviation with a suggested expansion of the most probably-correct expansion for the user so that the user may accept the suggested expansion when the ratio is between a second, lower threshold value and the first threshold value.

Abstract: One embodiment selects from a set of query-suggestion pairs a first query and a subset of query-suggestion pairs that each has the first query as its query; computes a Log Likelihood Ratio (LLR) value for each query-suggestion pair from the subset of query-suggestion pairs; ranks the subset of query-suggestion pairs according to their respective LLR values; removes from the subset of query-suggestion pairs all query-suggestion pairs whose LLR values are below a predetermined LLR threshold; computes a Pointwise Mutual Information (PMI) value for each remaining query suggestion pair from the subset of query-suggestion pairs; removes from the subset of query-suggestion pairs all query-suggestion pairs whose PMI values are below a predetermine PMI threshold; and constructs a ranked set of suggestions for the first query, wherein the ranked set of suggestions comprises one or more suggestions of the remaining query-suggestion pairs from the subset of query-suggestion pairs.

Abstract: In one embodiment, access a search query comprising one or more query words, at least one of the query words representing one or more query concepts; access a network document identified for a search query by a search engine, the network document comprising one or more document words, at least one of the document words representing one or more document concepts; semantic-text match the search query and the network document to determine one or more negative semantic-text matches; and construct one or more negative features based on the negative semantic-text matches.

Abstract: A method for normalizing query words in web search includes populating a dictionary with join and split candidates and corresponding joined and split words from an aggregate of query logs; determining a confidence score for join and split candidates, a highest confidence score for each being characterized in the dictionary as must-join and must-split, respectively; accepting queries with words amenable to being split or joined, or amenable to an addition or deletion of a hyphen or an apostrophe; generating, based on the accepted queries, split candidates obtained from the dictionary, and candidates of join, hyphen, or apostrophe algorithmically; and submitting to a search engine the generated possible candidates characterized as must-join or must-split in the dictionary, to improve search results returned in response to the queries; applying a language dictionary to generated candidates not characterized as must-split or must-join, to rank them, and submitting those highest-ranked to the search engine.

Abstract: One particular embodiment accesses a first set of search queries comprising one or more first search queries; extracts one or more features based on the first set of search queries, trains a search-query classifier using the features; accesses a second search query provided by a user; determines whether the second search query has implicit and general local intent using the search-query classifier; if the second search query has implicit and general local intent, then determines a location associated with the user; and identifies a search result in response to the second search query based at least in part on the location associated with the user; and presents the search result to the user.

Abstract: In one embodiment, access a search query comprising one or more query words, at least one of the query words representing one or more query concepts; access a network document identified for a search query by a search engine, the network document comprising one or more document words, at least one of the document words representing one or more document concepts; semantic-text match the search query and the network document to determine one or more negative semantic-text matches; and construct one or more negative features based on the negative semantic-text matches.

Abstract: A method for handling abbreviations in web queries includes building a dictionary of possible word expansions for potential abbreviations related to query terms received and anticipated to be received by a search engine; accepting a query including an abbreviation from a searching user, where a probability of finding a most probably-correct expansion in the dictionary is a first probability, and a probability that the expansion is the abbreviation itself is a second probability; determining a ratio between the first and second probabilities; expanding the abbreviation in accordance with the most probably-correct expansion when the ratio is above a first threshold value; and highlighting the abbreviation with a suggested expansion of the most probably-correct expansion for the user so that the user may accept the suggested expansion when the ratio is between a second, lower threshold value and the first threshold value.

Abstract: Techniques for automatically scoring submissions to an online question-and-answer submission system are disclosed. According to one such technique, an initial set of user submissions are scored by human operators and/or automated algorithmic mechanisms. The submissions and their accompanying scores are provided as training data to an automated machine learning mechanism. The machine learning mechanism processes the training data and automatically detects patterns in the provided submissions. The machine learning mechanism automatically correlates these patterns with the scores assigned to the submissions that match those patterns. As a result, the machine learning mechanism is trained. Thereafter, the machine learning mechanism processes unscored submissions. The machine learning mechanism automatically identifies, from among the patterns that the machine learning mechanism has already detected, one or more patterns that these submissions match.

Abstract: Techniques for determining when and which name variant candidates to use to re-write a search query that includes a person's name in order to provide the most relevant search results are provided. A determination is made whether a person name is present in a search query request entered by a user. Name variant candidates are generated for each person name. Then, the name variant candidates are ranked for each person name based upon one or more models that calculate a probability value for each name variant candidate. Based upon these rankings, the query may be re-written to include the original person name and a specified number of top ranked name variant candidates to present the user with the most relevant search results.