Stephen Soderland’s research while affiliated with University of Washington and other places

Most approaches to relation extraction, the task of extracting ground facts from natural language text, are based on machine learning and thus starved by scarce training data. Manual annotation is too expensive to scale to a comprehensive set of relations. Distant supervision, which automatically creates training data, only works with relations that already populate a knowledge base (KB). Unfortunately, KBs such as FreeBase rarely cover event relations ( e.g. “person travels to location”). Thus, the problem of extracting a wide range of events — e.g., from news streams — is an important, open challenge. This paper introduces NewsSpike-RE, a novel, unsupervised algorithm that discovers event relations and then learns to extract them. NewsSpike-RE uses a novel probabilistic graphical model to cluster sentences describing similar events from parallel news streams. These clusters then comprise training data for the extractor. Our evaluation shows that NewsSpike-RE generates high quality training sentences and learns extractors that perform much better than rival approaches, more than doubling the area under a precision-recall curve compared to Universal Schemas.

A translation graph is created using a plurality of reference sources that include translations between a plurality of different languages. Each entry in a source is used to create a wordsense entry, and each new word in a source is used to create a wordnode entry. A pair of wordnode and wordsense entries corresponds to a translation. In addition, a probability is determined for each wordsense entry and is decreased for each translation entry that includes more than a predefined number of translations into the same language. Bilingual translation entries are removed if subsumed by a multilingual translation entry. Triangulation is employed to identify pairs of common wordsense translations between a first, second, and third language. Translations not found in reference sources can also be inferred from the data comprising the translation graph. The translation graph can then be used for searches of a data collection in different languages.

Chambers and Jurafsky (2009) demonstrated that event schemas can be automatically induced from text corpora. However, our analysis of their schemas identifies several weaknesses, e.g., some schemas lack a common topic and distinct roles are incorrectly mixed into a single actor. It is due in part to their pair-wise representation that treats subjectverb independently from verb-object. This often leads to subject-verb-object triples that are not meaningful in the real-world. We present a novel approach to inducing open-domain event schemas that overcomes these limitations. Our approach uses cooccurrence statistics of semantically typed relational triples, which we call Rel-grams (relational n-grams). In a human evaluation, our schemas outperform Chambers’s schemas by wide margins on several evaluation criteria. Both Rel-grams and event schemas are freely available to the research community.

Both lexical translation and knowledge-based translation systems require sense-distinguished translation lexicons, yet such lexicons are expensive to create manually. However, the abundance of untagged monolingual corpora and the availability of bilingual, machine-readable dictionaries (MRDs) suggest an opportunity. Our PanLexicon system takes advantage of these resources to automatically construct a sense-distinguished multilingual lexicon. The challenge for PanLexicon is that free, bilingual MRDs do not make sense distinctions, and often have spotty coverage. PanLexicon uses word contexts from monolingual corpora to guide it in finding translation sets – sets of words that share the same word sense across multiple languages. By maintaining word sense distinctions, PanLexicon finds translations between language pairs that are not supported by any of its bilingual source dictionaries. PanLexicon runs in time linear in the size of its input, and thus scales readily to large numbers of languages. We built a prototype of PanLexicon with inputs from Spanish-English and Chinese-English dictionaries. Our initial experimental results show that PanLexicon is able to find high-quality translation sets despite the limitations of its inputs.

The need for translation among the world's thousands of natural languages makes information access and communication costly. One possible solution is lemmatic communication: A human sender encodes a message into sequences of lemmata (dictionary words), a massively multilingual lexical translation engine translates them into lemma sequences in a target language, and a human receiver interprets them to infer the sender's intended meanings. Using a 13-million-lemma, 1300-language translation engine, we conducted an experiment in lemmatic communication with Spanish-and Hungarian-speaking subjects. Translingual communication was less successful than intralingual communication, and intralingual communication was less successful when the lemma sequences were artificially randomized before the receiver saw them (simulating word-order differences among languages). In all conditions, however, meanings were transmitted with high or moderate fidelity in at least 40% of the cases. The results suggest interface and translation-algorithm improvements that could increase the efficacy of lemmatic communication.

Open Information Extraction extracts relations from text without requiring a pre-specified domain or vocabulary. While existing techniques have used only shallow syntactic features, we investigate the use of semantic role labeling techniques for the task of Open IE. Semantic role labeling (SRL) and Open IE, although developed mostly in isolation, are quite related. We compare SRL-based open extractors, which perform computationally expensive, deep syntactic analysis, with TextRunner, an open extractor, which uses shallow syntactic analysis but is able to analyze many more sentences in a fixed amount of time and thus exploit corpus-level statistics. Our evaluation answers questions regarding these systems, including, can SRL extractors, which are trained on PropBank, cope with heterogeneous text found on the Web? Which extractor attains better precision, recall, f-measure, or running time? How does extractor performance vary for binary, n-ary and nested relations? How much do we gain by running multiple extractors? How do we select the optimal extractor given amount of data, available time, types of extractions desired?