Because warfighters need to translate and comprehend more than just, “Where’s the train station?”
Devices and dictionaries for converting basic phrases from one language to another are common tools for travelers in foreign countries. But when understanding context, dialect and personalities offers the chance to stop an attack or catch a terrorist, official personnel need more sophisticated technology. The U.S. Defense Department’s premier research organization has embarked on several projects designed to give troops in any situation the tools they need to secure the spoken word.
These efforts span a range of problems for the military, from communicating with locals in theater to sifting through intelligence in an analysis unit at higher headquarters. Bonnie Dorr, the Defense Advanced Research Projects Agency (DARPA) program manager for nearly all the work, explains that each research area has its own focus. She emphasizes that more than one tool is necessary to meet the different natural-language translation needs of the military; no one-size-fits-all solution exists. Not only could troops in the field need different capabilities, but what they need also varies from what the analysts working in secure stateside locations require. In addition, the range of projects enables the agency to create different teams of experts from industry, academia and other fields to address specific issues. The programs vary in size depending on the effort necessary to move forward in a particular field of study.
Dorr oversees four programs: two are underway and one recently finished; and work on a final one begins this fall. They are Robust Automatic Transcription of Speech (RATS); Multilingual Automatic Document Classification, Analysis and Translation (MADCAT); Global Autonomous Language Exploitation (GALE); and Broad Operational Language Translation (BOLT), respectively. Another project—the Spoken Language Communication and Translation System for Tactical Use, or TRANSTAC, is an ongoing research project at DARPA, but it falls under a different program manager.
During its active period, GALE examined computer software technologies to absorb, translate, analyze and interpret large volumes of speech and text in multiple languages. The program converted and distilled usable information from various sources, then presented it to military personnel and analysts in simple formats. In the program, researchers focused on Arabic and Chinese from formal genres such as print and broadcast news outlets. Studies included language-specific and language-independent algorithms. By its end, GALE achieved significant results. Translation error rates in Arabic dropped from 35 percent to 10 percent for unstructured audio. Similar numbers were reached in Chinese. According to Dorr, GALE systems can perform better than humans or Google when handling many queries. The technology now has transitioned to operational systems across the military, including those in Iraq.
BOLT picks up where GALE left off. The success of the completed program in accurately translating formal news stories will be leveraged, but BOLT advances to the entirely new and much more difficult task of translating informal conversational speech, including that used in media such as emails and messaging. The goal of the five-year program is to analyze the information sources and pull out the pieces that could impact national security or deployed troops. It also will help those stationed overseas to communicate better both with local populaces and foreign allies.
The effort will focus mainly on dialects, translation and retrieval—including English-language queries that retrieve targeted information from multilanguage sources and yield the most relevant results. However, project personnel also will undertake the demanding task of creating electronics that acquire deep semantic language concepts using robotic visual and tactile information. Semantics encompass the true meaning of words and sentences as they are used. Dialects and slang, in which a literal translated meaning does not convey the true intent of a speaker or writer, will be explored through BOLT as well. By the end of the five-year program, personnel plan to develop new metrics for items such as automatic measures of semantic closeness.
RATS is a fairly new project begun earlier this year. The timeline for the work spans three and a half years, at the end of which personnel plan to have better capabilities for dealing with noisy signals. Researchers are developing methods to determine what actually is speech in a noisy transmission; to identify the language spoken and the speaker; and to pick out important words. Doing so will enable users to extract reliable, relevant data to inform decisions.
Dorr explains that through one of RATS’ most innovative features, it will analyze speech in a way similar to human ears. The project also should yield new techniques to detect automatically regions of the spectrum that are more reliable than others. “Currently, staff linguists undertaking the difficult and exhausting task of listening to a noisy channel have a probability of 0.5 percent to 2.5 percent that they will find relevant material,” Dorr says. “By the end of three years, RATS aims to increase that chance to 90 percent.” The technology should be ready for field testing in two years.
Dorr’s final project is in its fourth of five years of effort trying to develop technology to classify, analyze and interpret Arabic text images from hard-copy documents. These include not only electronic images such as faxes or PDFs, but also handwritten documents and even graffiti. MADCAT detects words based on optical character recognition, a technology that enables users to convert written images into files that can be edited or searched. Dorr explains that the program has become a type of application for GALE in which GALE’s machine translation tools are used by MADCAT’s component technologies.
In the last four years, MADCAT has advanced from 40 percent accuracy per character, which is less than 5 percent per word, to 60 to 65 percent accuracy. The ultimate goal is to reach 90 percent accuracy. Initial insertion of the technology is complete and has proved successful, according to the agency.
TRANSTAC, which is managed by Dr. Mari Maeda, takes a slightly different approach to translation. This technology enables robust, spontaneous, two-way tactical speech between U.S. personnel and native speakers. It seeks to fill the gaps in current translation capabilities that prevent warfighters from communicating directly with locals when no human translator is available to assist.
Dorr says the variety of programs underway at DARPA cover the major areas of natural-language processing important to defense: translation, spoken communication, speech signal processing, optical character recognition and information extraction. These efforts go beyond converting words or phrases between tongues and help fill other needs, such as finding the right information in large volumes of materials; finding ways to use the information; and enabling effective communication regardless of language or medium. She believes anyone in the defense and intelligence communities who needs to understand, exploit or convey information in foreign languages could benefit from DARPA’s research.
Each initiative comes with its own challenges, and all such computer translation efforts are complex. “Translation in these contexts ... is not substituting words from one language to another,” Dorr explains. “People underestimate the challenge it poses to computers.” Dialects, slang and other informal language pieces that humans understand easily require advanced computing in machines. Yet, the need for such capabilities in the defense realm is high. “[Military] operations take place all over the world, and the world speaks many different languages,” Dorr states. “Warfighters need translators, but human translators are expensive and limited in number and speed. This technology has the potential to bridge the language gap.”
An area of interest to researchers is reaching a new level of understanding with translations that could fill in missing pieces in dialogue. Dorr explains that in some languages, the subject of a sentence is never mentioned. “If you’re going to translate into English, you need to know the subject,” she says. In addition, these tools can help follow a conversation, understanding that a person mentioned in one part of a conversation could be referenced later by another name or title. The various projects aim to perform tasks at an in-depth level to pull meaning and relevance out of language sources.
Though many companies in the private sector have language-translation offerings, Dorr says the DARPA efforts move beyond the commercial. The agency wants to understand deep semantic meanings in many vernaculars. These include more obscure languages that may be unprofitable for industry to invest in but could offer benefits for national security and defense operations. In addition, accuracy is crucial for the military to perform its missions and keep people safe. Traditional search engines may return findings that are quick or good enough but not up to the level of robustness or accuracy required in life-or-death situations.