Abstract: Speech models are trained using one or more of three different training systems. They include competitive training which reduces a distance between a recognized result and a true result, data boosting which divides and weights training data, and asymmetric training which trains different model components differently.

Abstract: An alteration candidate for a query can be scored. The scoring may include computing one or more query-dependent feature scores and/or one or more intra-candidate dependent feature scores. The computation of the query-dependent feature score(s) can be based on dependencies to multiple query terms from each of one or more alteration terms (i.e., for each of the one or more alteration terms, there can be dependencies to multiple query terms that form at least a portion of the basis for the query-dependent feature score(s)). The computation of the intra-candidate dependent feature score(s) can be based on dependencies between different terms in the alteration candidate. A candidate score can be computed using the query dependent feature score(s) and/or the intra-candidate dependent feature score(s). Additionally, the candidate score can be used in determining whether to select the candidate to expand the query. If selected, the candidate can be used to expand the query.

Abstract: A supervised technique uses relevance judgments to train a dependency parser such that it approximately optimizes Normalized Discounted Cumulative Gain (NDCG) in information retrieval. A weighted tree edit distance between the parse tree for a query and the parse tree for a document is added to a ranking function, where the edit distance weights are parameters from the parser. Using parser parameters in the ranking function enables approximate optimization of the parser's parameters for NDCG by adding some constraints to the objective function.

Abstract: Architecture that formulates speech translation as a unified log-linear model with a plurality of feature functions, some of which are derived from generative models. The architecture employs discriminative training for the generative features based on an optimization technique referred to as growth transformation. A discriminative training objective function is formulated for speech translation as well as a growth transformation-based model training method that includes an iterative training formula. This architecture is used to design and perform the global end-to-end optimization of speech translation, which when compared with conventional methods for speech translation provides not only a learning method with faster convergence but also improves speech translation accuracy.

Abstract: A method and apparatus for training an acoustic model are disclosed. A training corpus is accessed and converted into an initial acoustic model. Scores are calculated for a correct class and competitive classes, respectively, for each token given the initial acoustic model. Also, a sample-adaptive window bandwidth is calculated for each training token. From the calculated scores and the sample-adaptive window bandwidth values, loss values are calculated based on a loss function. The loss function, which may be derived from a Bayesian risk minimization viewpoint, can include a margin value that moves a decision boundary such that token-to-boundary distances for correct tokens that are near the decision boundary are maximized. The margin can either be a fixed margin or can vary monotonically as a function of algorithm iterations. The acoustic model is updated based on the calculated loss values. This process can be repeated until an empirical convergence is met.

Abstract: Architecture that provides the integration of automatic speech recognition (ASR) and machine translation (MT) components of a full speech translation system. The architecture is an integrative and discriminative approach that employs an end-to-end objective function (the conditional probability of the translated sentence (target) given the source language's acoustic signal, as well as the associated BLEU score in the translation, as a goal in the integrated system. This goal defines the theoretically correct variables to determine the speech translation system output using a Bayesian decision rule. These theoretically correct variables are modified in practical use due to known imperfections of the various models used in building the full speech translation system. The disclosed approach also employs automatic training of these variables using minimum classification error (MCE) criterion.

Abstract: An expected dialog-turn (ED) value is estimated for evaluating a speech application. Parameters such as a confidence threshold setting can be adjusted based on the expected dialog-turn value. In a particular example, recognition results and corresponding confidence scores are used to estimate the expected dialog-turn value. The recognition results can be associated with a possible outcome for the speech application and a cost for the possible outcome can be used to estimate the expected dialog-turn value.

Abstract: Hidden Markov Model (HMM) parameters are updated using update equations based on growth transformation optimization of a minimum classification error objective function. Using the list of N-best competitor word sequences obtained by decoding the training data with the current-iteration HMM parameters, the current HMM parameters are updated iteratively. The updating procedure involves using weights for each competitor word sequence that can take any positive real value. The updating procedure is further extended to the case where a decoded lattice of competitors is used. In this case, updating the model parameters relies on determining the probability for a state at a time point based on the word that spans the time point instead of the entire word sequence. This word-bound span of time is shorter than the duration of the entire word sequence and thus reduces the computing time.

Abstract: Systems, methods, and computer media for identifying query rewriting replacement terms are provided. A list of related string pairs each comprising a first string and second string is received. The first string of each related string pair is a user search query extracted from user click log data. For one or more of the related string pairs, the string pair is provided as inputs to a statistical machine translation model. The model identifies one or more pairs of corresponding terms, each pair of corresponding terms including a first term from the first string and a second term from the second string. The model also calculates a probability of relatedness for each of the one or more pairs of corresponding terms. Term pairs whose calculated probability of relatedness exceeds a threshold are characterized as query term replacements and incorporated, along with the probability of relatedness, into a query rewriting candidate database.

Abstract: Systems, methods, and computer media for identifying related strings for search query rewriting are provided. Session data for a user search query session in an accessed click log data is identified. It is determined whether a first additional search query in the session data is related to a first user search query based on at least one of: dwell time; a number of search result links clicked on; and similarity between web page titles or uniform resource locators (URLs). When related, the first additional search query is incorporated into a list of strings related to the first user search query. One or more supplemental strings that are related to the first user search query are also identified. The identified supplemental strings are also included in the list of strings related to the first user search query.

Abstract: Architecture that provides the capability to subselect the most relevant data from an out-domain corpus to use either in isolation or in combination conjunction with in-domain data. The architecture is a domain adaptation for machine translation that selects the most relevant sentences from a larger general-domain corpus of parallel translated sentences. The methods for selecting the data include monolingual cross-entropy measure, monolingual cross-entropy difference, bilingual cross entropy, and bilingual cross-entropy difference. A translation model is trained on both the in-domain data and an out-domain subset, and the models can be interpolated together to boost performance on in-domain translation tasks.

Abstract: A supervised technique uses relevance judgments to train a dependency parser such that it approximately optimizes Normalized Discounted Cumulative Gain (NDCG) in information retrieval. A weighted tree edit distance between the parse tree for a query and the parse tree for a document is added to a ranking function, where the edit distance weights are parameters from the parser. Using parser parameters in the ranking function enables approximate optimization of the parser's parameters for NDCG by adding some constraints to the objective function.

Abstract: Architecture that provides the integration of automatic speech recognition (ASR) and machine translation (MT) components of a full speech translation system. The architecture is an integrative and discriminative approach that employs an end-to-end objective function (the conditional probability of the translated sentence (target) given the source language's acoustic signal, as well as the associated BLEU score in the translation, as a goal in the integrated system. This goal defines the theoretically correct variables to determine the speech translation system output using a Bayesian decision rule. These theoretically correct variables are modified in practical use due to known imperfections of the various models used in building the full speech translation system. The disclosed approach also employs automatic training of these variables using minimum classification error (MCE) criterion.

Abstract: An alteration candidate for a query can be scored. The scoring may include computing one or more query-dependent feature scores and/or one or more intra-candidate dependent feature scores. The computation of the query-dependent feature score(s) can be based on dependencies to multiple query terms from each of one or more alteration terms (i.e., for each of the one or more alteration terms, there can be dependencies to multiple query terms that form at least a portion of the basis for the query-dependent feature score(s)). The computation of the intra-candidate dependent feature score(s) can be based on dependencies between different terms in the alteration candidate. A candidate score can be computed using the query dependent feature score(s) and/or the intra-candidate dependent feature score(s). Additionally, the candidate score can be used in determining whether to select the candidate to expand the query. If selected, the candidate can be used to expand the query.

Abstract: The language of origin of a word is determined by analyzing non-uniform letter sequence portions of the word.

Abstract: A joint optimization strategy is employed for combining translation hypotheses from multiple machine-translation systems. Decisions on word alignment, between the hypotheses, ordering, and selection of a combined translation output are made jointly in accordance with a set of features. Additional features that model alignment and ordering behavior are also provided and utilized.

Abstract: A computing system configured to produce an optimized translation hypothesis of text input into the computing system. The computing system includes a plurality of translation machines. Each of the translation machines is configured to produce their own translation hypothesis from the same text. An optimization machine is connected to the plurality of translation machines. The optimization machine is configured to receive the translation hypotheses from the translation machines. The optimization machine is further configured to align, word-to-word, the hypotheses in the plurality of hypotheses by using a hidden Markov model.

Abstract: A word alignment modeler uses probabilistic learning techniques to train “word-dependent transition models” for use in constructing phrase level Hidden Markov Model (HMM) based word alignment models. As defined herein, “word-dependent transition models” provide a probabilistic model wherein for each source word in training data, a self-transition probability is modeled in combination with a probability of jumping from that particular word to a different word, thereby providing a full transition model for each word in a source phrase. HMM based word alignment models are then used for various word alignment and machine translation tasks. In additional embodiments sparse data problems (i.e., rarely used words) are addressed by using probabilistic learning techniques to estimate word-dependent transition model parameters by maximum a posteriori (MAP) training.

Abstract: Pattern model parameters are updated using update equations based on competing patterns that are identical to a reference pattern except for one segment at a time that is replaced with a competing segment. This allows pattern recognition parameters to be tuned one segment at a time, rather than have to try to model distinguishing features of the correct pattern model as a whole, according to an illustrative embodiment. A reference pattern and competing patterns are divided into pattern segments. A set of training patterns are generated by replacing one of the pattern segments in the reference pattern with a corresponding competing pattern segment. For each of the training patterns, a pattern recognition model is applied to evaluate a relative degree of correspondence of the reference pattern with the pattern signal compared to a degree of correspondence of the training patterns with the pattern signal.

Abstract: Described is a technology for learning a foreign language or other subject. Answers (e.g., translations) to questions (e.g., sentences to translate) received from learners are combined into a combined answer that serves as a representative model answer for those learners. The questions also may be provided to machine subsystems to generate machine answers, e.g., machine translators, with those machine answers used in the combined answer. The combined answer is used to evaluate each learner's individual answer. The evaluation may be used to compute profile information that is then fed back for use in selecting further questions, e.g., more difficult sentences as the learners progress. Also described is integrating the platform/technology into a web service.