Butler Hill

“It’s a hard problem.” Or was that a “heart problem?” - Challenges and Opportunities for Speech Technologies

Attending SpeechTEK 2010, an annual conference and exhibition for speech technology companies and professionals, provided us at Butler Hill with great new insights about the challenges, opportunities, and trends in the industry. It was especially exciting to hear how strong of a role linguistics play in the evolution of speech technology and applications. Read on for our take-aways on what's buzzing in the speech technology industry.

While there are many similarities between speech technology and text based language technology development, there are also a lot of unique challenges to processing and producing speech data. The ambiguity introduced by homophones (words that are spelled differently but sound the same, making a “hard problem” very different from a “heart problem” in a medical record transcription context) and the speech data variety across different speakers are just the tip of the iceberg. People also tend to use different vocabulary when they talk compared to when they write; e.g., a customer might use more idiosyncratic words when they call a customer care center than when they write an email or comment on a company blog, since they are often biased by the immediate context of the company website and the “official” product terminology in the latter scenario. Also, speakers tend to adapt to their dialog partner, which is when cultural issues come into play like variations when talking to someone with a perceived age difference, or someone with a regional accent – characteristics that are much less apparent when entering a written dialog. All of these factors make speech a particularly intriguing focal point for new technologies and applications.

There is lots of energy around consumer focused mobile voice applications, driven by continued strong growth in smart phones worldwide and the fact that network capacity, database and storage capacities have finally caught up with the requirements of speech recognition and text-to-speech technology. As users become more and more attached to their mobile phones and devices, they expect them to really “work” for them and help them accomplish tasks in new ways and in new situations: Up to half of some mobile application’s usage occurs while driving, which calls for “eye free” and hands free modalities. Voice enabled applications are the answer, and need to move towards truly conversational interfaces. Interesting challenges are presented through the fact that users might use several voice enabled applications at the same time, so the ideal speech recognition engine can process voice input across different applications and domains rather than requiring the user to provide clunky voice commands for switching between constrained domain apps. We anticipate seeing increased efforts to fully integrate speech into the OS and grounding it in a broad coverage syntactic and semantic representation layer.

Enterprises, meanwhile, are looking to balance cost saving demands with customer experience requirements for their IVRs (interactive voice response systems, as in phone banking or call center routing systems.) Many large global corporations, from airlines to car rental agencies to health care providers, attended the conference to survey new solutions for efficiently handling incoming customer calls. Such a solution might include deep natural language processing (NLP), which is especially valuable for first-time caller and long tail scenarios. Beyond the general challenges of implementing a high quality automated system, linguistic and cultural challenges abound once you adapt the system for new markets. For example, a voice navigation system in the Japanese market might use a voice that sounds friendly, animated, and eager, demonstrating “helpfulness” – a style that would not work in the German market, where the same system might use a voice conveying “confidence” instead.

Another topic with a lot of buzz at SpeechTEK was Speech Analytics. While parts of these solution overlap with Text Analytics (mining unstructured text data and turning it into actionable insights), speech analytics has additional information to work with, from prosody (intonation, stress, and rhythm) and pitch to pauses in the utterance to how loud the speaker gets, which helps analyze and categorize data. One might think of this as “sentiment analysis with additional dimensions.” There are lots of valuable applications for speech analytics in the enterprise, e.g. mining customer data from call centers, and elsewhere as in the legal field and in forensics.

We at Butler Hill look forward to continued partnering with companies developing consumer and enterprise focused speech technologies – and we’ll see you at SpeechTEK 2011!

The Meaning of Meaning: Semantic Technologies

Butler Hill attended SemTech 2010, a large conference on semantic technologies. Read on for our take on how semantic technologies will help enterprises, organizations and consumers make sense of the massive amounts of data that lives on the web and in internal data stores.

When we at Butler Hill talk about semantics, more often than not we use the term in its linguistic sense, referring to the meaning of words or sentences in text and speech. There is, however, another “semantic universe” out there, in which semantics refers to techniques for incorporating meaning in the expression, storage, analysis and retrieval of structured data. At last week’s Semantic Technologies conference in San Francisco, SemTech 2010, both of these senses were represented.

Much of the conference focused on the challenges of and opportunities for dealing with structured data: data in enterprise knowledge bases, product catalogs, government databases. Data that is mostly hidden from current search engines like Google and Bing – and there is lots of it: some estimate that there is 500 times more structured data than unstructured data, with billions of data tables out on the web. Much of the “semantic web” or “web 3.0” discussions are about creating and utilizing formats for structuring, linking, and surfacing this data to enterprises and consumers. Ontologies – formal representations of concepts, terms, and relationships that model a specific domain – will play a big part in this effort. Domain specific, culture specific, and linguistic knowledge will be needed as existing ontologies are integrated into the semantic web, and/or domain or application specific ontologies are created through automatic or human assisted approaches.

Other interesting challenges arise for dealing with the massive quantities of unstructured data (including enterprise documents, traditional media content, user generated content in social media). The sheer explosion of unstructured text data demands sophisticated tools and approaches for evaluating, extracting, classifying and analyzing this data. We see lots of opportunities for helping companies and organizations tackle these challenges, from assisting with machine learning text mining techniques through the creation of “gold standard” by qualified annotators, to using our linguistic and cultural know-how to support the creation of domain specific grammars and the extraction of entities.

Data isn’t as neatly “siloed” as the distinction between structured and unstructured data suggests. In reality, structured and unstructured sources often exist side by side, and need to be integrated in order to truly mine and share information across the web. We are curious and excited to see how the semantic web unfolds!

How far will machine translation take sentiment analysis technology?

There has been a lot of buzz around sentiment analysis, a text analytics application that aims to extract and classify emotions, attitudes and opinions from unstructured text sources. To respond to the need for capturing “global sentiment”, some sentiment analysis providers are looking to machine translation as a cheap and fast way to extend their offerings into new languages.

We at Butler Hill wanted to explore how machine translation holds up in this context. We designed an experiment to investigate the impact of machine translation on precision and recall of sentiment analysis. Our findings provide valuable insights to anyone in the sentiment analysis industry with global customers.

Text Analytics and the Global Challenge

Part of Butler Hill’s leadership team just came back from the 6th Annual Text Analytics Summit in Boston, where we connected with text analytics companies and their customers. Read on for more on the challenges faced by the text analytics industry and how Butler Hill helps overcome those.

Butler Hill consultant co-authors book on Automated Grammatical Error Detection for Language Learners

Congratulations to Butler Hill consultant Claudia Leacock, whose book “Automated Grammatical Error Detection for Language Learners” just got published by Morgan & Claypool Publishers.

The book, co-authored with City University of New York Psychology Professor Martin Chodorow, Computational Linguist at Microsoft Michael Gamon, and Computational Linguistics Research Scientist at Educational Testing Service Joel Tetreault, explores the unique grammar challenges non-native speakers of English face and how automated systems help detect and correct such grammar errors.

Butler Hill Joins Microsoft in Haitian Relief Efforts

When aid workers reached out for a translation tool, Butler Hill helped its colleagues at the Microsoft Machine Translation group to add English/Haitian Creole to Microsoft Translator, accomplishing in five days what typically takes several months.

Microsoft Designates Butler Hill a Preferred Vendor

Microsoft’s Global Procurement Group has chosen Butler Hill to participate in the select Microsoft Vendor Program. The MSVP program identifies outstanding suppliers that have proven to be the best performers in their category while providing exceptional value.