Rukmini Iyer

Modeling long distance dependence in language: topic mixtures versus dynamic cache models

Standard statistical language models use n-grams to capture local dependencies, or use dynamic modeling techniques to track dependencies within an article. In this paper, we investigate a new statistical language model that captures topic-related dependencies of words within and across sentences. First, we develop a topic-dependent, sentence-level mixture language model which takes advantage of the topic constraints in a sentence or article. Second, we introduce topic-dependent dynamic adaptation techniques in the framework of the mixture model, using n-gram caches and content word unigram caches. Experiments with the static (or unadapted) mixture model on the North American Business (NAB) task show a 21% reduction in perplexity and a 3-4% improvement in recognition accuracy over a general n-gram model, giving a larger gain than that obtained with supervised dynamic cache modeling. Further experiments on the Switchboard corpus also showed a small improvement in performance with the sentence-level mixture model. Cache modeling techniques introduced in the mixture framework contributed a further 14% reduction in perplexity and a small improvement in recognition accuracy on the NAB task for both supervised and unsupervised adaptation.

Modeling long distance dependence in language: topic mixtures vs. dynamic cache models

We investigate a new statistical language model which captures topic-related dependencies of words within and across sentences. First, we develop a sentence-level mixture language model that takes advantage of the topic constraints in a sentence or article. Second, we introduce topic-dependent dynamic cache adaptation techniques in the framework of the mixture model. Experiments with the static (or unadapted) mixture model on the 1994 WSJ task indicated a 21% reduction in perplexity and a 3-4% improvement in recognition accuracy over a general n-gram model. The static mixture model also improved recognition performance over an adapted n-gram model. Mixture adaptation techniques contributed a further 14% reduction in perplexity and a small improvement in recognition accuracy.

Using out-of-domain data to improve in-domain language models

Standard statistical language modeling techniques suffer from sparse data problems when applied to real tasks in speech recognition, where large amounts of domain-dependent text are not available. We investigate new approaches to improve sparse application-specific language models by combining domain dependent and out-of-domain data, including a back-off scheme that effectively leads to context-dependent multiple interpolation weights, and a likelihood-based similarity weighting scheme to discriminatively use data to train a task-specific language model. Experiments with both approaches on a spontaneous speech recognition task (switchboard), lead to reduced word error rate over a domain-specific n-gram language model, giving a larger gain than that obtained with previous brute-force data combination approaches.

Language modeling with sentence-level mixtures

This paper introduces a simple mixture language model that attempts to capture long distance constraints in a sentence or paragraph. The model is an m-component mixture of trigram models. The models were constructed using a 5K vocabulary and trained using a 76 million word Wall Street Journal text corpus. Using the BU recognition system, experiments show a 7% improvement in recognition accuracy with the mixture trigram models as compared to using a trigram model.

Speech recognition in multiple languages and domains: the 2003 BBN/LIMSI EARS system

We report on the results of the first evaluations for the BBN/LIMSI system under the new DARPA EARS program. The evaluations were carried out for conversational telephone speech (CTS) and broadcast news (BN) for three languages: English, Mandarin, and Arabic. In addition to providing system descriptions and evaluation results, the paper highlights methods that worked well across the two domains and those few that worked well on one domain but not the other. For the BN evaluations, which had to be run under 10 times real-time, we demonstrated that a joint BBN/LIMSI system with a time constraint achieved better results than either system alone.

Analyzing and predicting language model improvements

Statistical n-gram language models are traditionally developed using perplexity as a measure of goodness. However, perplexity often demonstrates a poor correlation with recognition improvements, mainly because it fails to account for the acoustic confusability between words and for search errors in a recognizer. In this paper, we study alternatives to perplexity for predicting language model performance, including other global features as well as a new approach that predicts, with a high correlation (0.96), performance differences associated with localized changes in language models, given a recognition system. Experiments focus on the problem of augmenting in-domain Switchboard text with out-of-domain text from the Wall Street Journal and broadcast news that differ in both style and content from the in-domain data.

Employing web search query click logs for multi-domain spoken language understanding

Logs of user queries from a search engine (such as Bing or Google) together with the links clicked provide valuable implicit feedback to improve statistical spoken language understanding (SLU) models. In this work, we propose to enrich the existing classification feature set for domain detection with features computed using the click distribution over a set of clicked URLs from search query click logs (QCLs) of user utterances. Since the form of natural language utterances differs stylistically from that of keyword search queries, to be able to match natural language utterances with related search queries, we perform a syntax-based transformation of the original utterances, after filtering out domain-independent salient phrases. This approach results in significant improvements for domain detection, especially when detecting the domains of web-related user utterances.

Probabilistic first pass retrieval for search advertising: from theory to practice

Information retrieval in search advertising, as in other ad-hoc retrieval tasks, aims to find the most appropriate ranking of the ad documents of a corpus for a given query. In addition to ranking the ad documents, we also need to filter or threshold irrelevant ads from participating in the auction to be displayed alongside search results. In this work, we describe our experience in implementing a successful ad retrieval system for a commercial search engine based on the Language Modeling (LM) framework for retrieval. The LM demonstrates significant performance improvements over the baseline vector space model (TF-IDF) system that was in production at the time. From a modeling perspective, we propose a novel approach to incorporate query segmentation and phrases in the LM framework, discuss impact of score normalization for relevance filtering, and present preliminary results of incorporating query expansions using query rewriting techniques. From an implementation perspective, we also discuss real-time latency constraints of a production search engine and how we overcome them by adapting the WAND algorithm to work with language models. In sum, our LM formulation is considerably better in terms of accuracy metrics such as Precision-Recall (10% improvement in AUC) and nDCG (8% improvement in nDCG@5) on editorial data and also demonstrates significant improvements in clicks in live user tests (0.787% improvement in Click Yield, with 8% coverage increase). Finally, we hope that this paper provides the reader with adequate insights into the challenges of building a system that serves millions of users every day.

Translating natural language utterances to search queries for SLU domain detection using query click logs

Logs of user queries from a search engine (such as Bing or Google) together with the links clicked provide valuable implicit feedback to improve statistical spoken language understanding (SLU) models. However, the form of natural language utterances occurring in spoken interactions with a computer differs stylistically from that of keyword search queries. In this paper, we propose a machine translation approach to learn a mapping from natural language utterances to search queries. We train statistical translation models, using task and domain independent semantically equivalent natural language and keyword search query pairs mined from the search query click logs. We then extend our previous work on enriching the existing classification feature sets for input utterance domain detection with features computed using the click distribution over a set of clicked URLs from search engine query click logs of user utterances with automatically translated queries. This approach results in significant improvements for domain detection, especially when detecting the domains of user utterances that are formulated as natural language queries and effectively complements to the earlier work using syntactic transformations.