# Research Statement

(Last updated October 2016)

Most of human knowledge is encoded in natural language. A longstanding goal of artificial intelligence has been to automate the understanding of natural language. Formulating an appropriate representation for the meaning of language has proved problematic. Approaches that employ complex semantic representations, like first order predicate logic, are difficult or impossible to scale to cover the broad range of expressions used in real language. I approach natural language understanding using pairs of English phrases as the basic unit of representation, and automatically labeling them with a small number of semantic relationships that allows a subset of automated reasoning to be applied. This design decision allows us to scale to open domains and to implement data-driven algorithms for acquiring semantic knowledge about language.

My approach to language understanding is inspired by my past research in machine translation. The advent of data-driven, statistical models has resulted in dramatically improved quality for machine translation. Commercial systems like Google Translate, or state-of-the-art research software that I have helped develop (the Moses system (Koehn et al, 2007) and the Joshua decoder (Li et al, 2009, Li et al, 2010, Weese et al, 2011, Ganitkevitch et al, 2012, Post et al, 2013)) use pairs of English and foreign phrases as their underlying representation. These phrase pairs are automatically acquired from a large volume of translated documents, and are treated as meaning-equivalent without having an explicit semantic representation. Vast quantities of bilingual training data allow us to extract a huge number of phrase pairs and to estimate associated probabilities. I assembled the largest publicly available bilingual training data for statistical machine translation, consisting of 22 million sentence pairs with 1.5 billion French + English words (Callison-Burch et al, 2009). This encompasses a huge range of language use from scientific abstracts to movie dialog slang, and thus allows the system to translate a wide variety of input sentences.

Currently, I have three areas of research. My primary research focus is to automate the understanding of English via paraphrases. I adapt the data, representations, and algorithms from statistical machine translation to facilitate natural language understanding. In addition, I have two other research directions. One attempts to extend machine translation so that it may be applied to a wider range of languages by doing away with the necessity for bilingual parallel corpora. Instead, my research focuses on learning translations from monolingual texts in two languages. My third research focus is on using crowdsourcing to explore new areas of natural language processing. (This crowdsourcing work has even extended beyond NLP and now includes social justice issues, including workers’ rights and gun violence in the United States.)

## Natural Language Understanding via Paraphrasing

Figure 1: The German festgenommen links the English phrase thrown into jail to its paraphrase imprisoned.

I developed a method that extracts paraphrases from bilingual parallel corpora by identifying equivalent English expressions using a shared foreign phrase. This ensures that their meaning is similar. Figure 1 illustrates the method (Bannard and Callison-Burch (2005), Callison-Burch (2007)). Thrown into jail occurs many times in the training data, aligning with several different foreign phrases. Each of these may align with a variety of other English paraphrases. Thus, thrown into jail not only paraphrases as imprisoned, but also as arrested, detained, incarcerated, jailed, locked up, taken into custody, and thrown into prison. However, not all the paraphrases are uniformly good. The baseline method also extracts candidate paraphrases that mean the same thing, but do not share the same syntactic category as the original phrase, such as be thrown in prison, been thrown into jail, being arrested, in jail, in prison, put in prison for, were thrown into jail, and who are held in detention. It is also prone to generating many bad paraphrases, such as maltreated, thrown, cases, custody, arrest, owners, and protection, because of noisy/inaccurate word alignments and other problems. Separating good paraphrases from bad presents fascinating research challenges (Pavlick et al (2015)).

Figure 2: We learn the English possessive rule by pivoting over SCFG translation rules.

I have extended the bilingual pivoting methodology to syntactic representations of translation rules. This builds on my research group’s work into adding syntactic information into statistical machine translation rules. We have adopted a synchronous context free grammar (SCFG) representation for our Joshua decoder, and we demonstrated that it is useful for translating between languages with different word orders like Urdu’s subject-object-verb order and English’s subject-verb-object order (Baker et al (2010)). Instead of pivoting over foreign phrases, we can pivot over foreign SCFG rules, as shown in Figure 2. This allows us to automatically acquire meaning-preserving syntactic transformations like the English possessive rule. This rule is a general transformation that can apply to most noun phrases in English. It allows us to recognize that the laptop’s screen can be rewritten as the screen of the laptop. Table 1 shows a variety of other meaning-preserving structural transformations that we learn in this way (Ganitkevitch et al (2011)).

Table 1: We are able to automatically acquire a variety of meaning-preserving structural translations in English by pivoting over SCFG translation rules.

We used my bilingual pivoting technique to create the paraphrase database, called PPDB for short (Ganitkevitch et al, 2013). PPDB contains 8 million synonyms, 68 million phrasal paraphrases, and 94 million meaning-preserving syntactic transformations. PPDB is freely available from our web site paraphrase.org. It is a much larger resource than the manually-constructed WordNet resource that is heavily used in NLP research. PPDB has made immediate impact and was widely adopted by other researchers. It has been cited 185 times in the three years since its publication, and it was central to the research described in the NAACL 2015 best paper (Faruqui et al, 2015) on retrofitting word vectors to semantic lexicons. The NAACL paper shows that PPDB has enormous potential for improving deep learning of word embeddings. I plan to explore this connection further.

Table 2: Examples of different types of entailment relations appearing in PPDB.

Over the past two years we made several advances to PPDB that improve its usefulness for understanding natural language:

• Semantics: In Pavlick et al (2015), we add an interpretable semantics to PPDB. The relationship between the phrase pairs in the database had been defined as approximately equivalent. Our new research allows these pairs to be encoded with more nuanced semantic relations, including directed entailment (little girl/girl) and exclusion (nobody/someone). We automatically assign semantic entailment relations to all 100 million entries in PPDB using features derived from past work on discovering inference rules from text and semantic taxonomy induction. Examples are given in Table 2.

Figure 3: We partition paraphrases of an input word like bug into clusters representing its distinct senses.
• Domain adaptation: Language is used differently in different domains. In Pavlick et al (2015) we demonstrate an algorithm that is able to automatically adapt paraphrases to suit a particular domain. For instance, paraphrase of the word divide when used in biology should include division, break, split, dispense, multiply, cleave, fracture, separate, mitotic division, partition since it refers to cellular division/multiplication. In a parliamentary domain it more commonly refers to the divide between rich and poor, and should be paraphrased as gap, division, gulf, separate, distinction, rift, difference.
• Natural language generation: Paraphrases are useful in the generation components of dialog systems like Apple’s Siri, question answering, and automatic summarization. We are investigating using paraphrases for text to text generation. Given an input text, rewrite it subject to constraints: for summarization make it shorter; for simplification use words that are easier to understand; for poetry generation conform to a meter and a rhyming scheme. In Xu et al (2015) and Xu et al (2016), we show how paraphrasing and machine translation techniques can be used for the problem of text simplification.
• Word sense clustering: The original release of PPDB grouped all senses of polysemous words together, and did not partition paraphrases into groups like WordNet does with its synsets. Thus a search for paraphrases of the noun bug would yield a single list of paraphrases that includes insect, glitch, beetle, error, microbe, wire, cockroach, malfunction, microphone, mosquito, virus, tracker, pest, informer, snitch, parasite, bacterium, fault, mistake, failure and many others. In Cocos and Callison-Burch (2016), we systematically explore different clustering algorithms, and ways of defining the similarity matrix that they use as input (Figure 3). We produce high quality sense clusters that represent a substantial improvement to PPDB. We are currently exploring using our sense clusters to perform word sense disambiguation.

The goal of the paraphrasing line of my research is to advance the longstanding AI goal of language understanding data-driven methods and statistical models. If successful, it has the potential to impact a wide variety of NLP tasks including information retrieval, question answering, and machine translation. My research into this area has been sponsored by two NSF EAGER awards, multiple grants from the Allen Institute for Artificial Intelligence (AI2) and its predecessor Vulcan, and a 1.6 million DARPA DEFT award. I have submitted an NSF CAREER proposal on this topic. It is currently under review. ## Statistical Machine Translation Without Parallel Corpora Figure 4: Example of projecting contextual vectors over a seed bilingual lexicon. Statistical machine translation has long been purported to be “language independent” since linguistic information is not typically used when training systems. This has been touted as a strength of the paradigm, because the only requirement for building a new system is a sentence-aligned bilingual parallel corpus. However, this language independence does not mean that statistical machine translation works equally well for every language. Translation quality depends on many factors, including the amount of training data, morphological complexity, and divergences in word order. Since very large parallel corpora are required to achieve good translation quality, statistical machine translation can, in reality, only be applied to a very limited number of languages. My research agenda for the next decade is to transform statistical machine translation so that (1) models can be trained without the use bilingual parallel corpora, (2) linguistic information can be integrated directly into the models and as priors for learning translations and re-ordering patterns, and (3) large groups of Internet users can collaborate to improve translation quality. If successful, these efforts will radically change the field and make statistical machine translation applicable to (nearly) all of the world’s languages. Figure 5: The temporal histograms are collected from monolingual texts spanning several years and show the number of occurrences of each word (on the y-axes) across time. While the correct translation has a good temporal match, the non-translations are less temporally similar. Like other statistical NLP systems and machine learning applications, the performance of statistical machine translation improves as more training data is used. For a few language pairs, we have tremendous amounts of training data – I created a French-English parallel corpus with nearly 1 billion words on each side, the DARPA GALE program produced Arabic-English and Chinese-English parallel corpora with 250 million words in each language, and we have somewhere on the order of 50–100 million words worth of parallel data for the official languages of the European Union. However, for most language pairs, we have comparatively tiny amounts of bilingual training data, which means that current statistical machine translation techniques will not work. To build statistical machine translation systems without parallel corpora, I have revived research started by Rapp (1999), who investigated inducing bilingual lexicons from monolingual texts. The method uses vector-space semantic models to build a context vector representing words whose translations are unknown. The elements in an unknown word’s vector are projected into the vector space of the other language using the known translations from a small seed bilingual dictionary. This sparse projected vector is compared to the vectors for all words in the target language. The word whose vector is most similar to the projected vector is considered to be the best translation of the unknown word. This process is illustrated in Figure 4. I have successfully replicated the results of Rapp (1999), and used the method to estimate the parameters of phrase-based statistical machine translation systems (Klementiev et al (2012), (Irvine and Callison-Burch (2015)). The advantage of this paradigm is that it only requires a small bilingual dictionary and large monolingual corpora, rather than bilingual parallel data. My students and I have examined combining a diverse set of monolingually-derived signals of translation equivalence (Irvine and Callison-Burch (2013)). In addition to vector space models, we have incorporated a diverse set of signals including temporal similarity (Figure 5), orthographic similarity, and topical similarity. Table 3 shows examples of the highest ranking English translations of 5 Spanish words for several of our signals of translation equivalence. Each signal produces different types of errors.(For instance, using topic similarity, montana, miley, and hannah are ranked highly as candidate translations of the Spanish word montana. The TV character Hannah Montana is played by actress Miley Cyrus, so the topic similarity between these words makes sense.) Table 3: Examples of translation candidates ranked using contextual similarity, temporal similarity, orthographic similarity and topic similarity. The correct English translations, when found, are bolded. My group has conducted a study of bilingual lexicon induction on a wide range of languages and data sizes (Irvine and Callison-Burch (accepted)). We examine translation into English from 25 foreign languages: Albanian, Azeri, Bengali, Bosnian, Bulgarian, Cebuano, Gujarati, Hindi, Hungarian, Indonesian, Latvian, Nepali, Romanian, Serbian, Slovak, Somali, Spanish, Swedish, Tamil, Telugu, Turkish, Ukrainian, Uzbek, Vietnamese and Welsh. Rather than testing solely on high frequency words, as previous research has done, we test on low frequency as well, so that our results are more relevant to statistical machine translation, where systems typically lack translations of rare words that fall outside of their training data. We systematically explore a wide range of features and phenomena that affect the quality of the translations discovered by bilingual lexicon induction. We give illustrative examples of the highest ranking translations for orthogonal signals of translation equivalence like contextual similarity and temporal similarity. We analyze the effects of frequency and burstiness, and the sizes of the seed bilingual dictionaries and the monolingual training corpora. Our model performs better than the previous state-of-the-art matching canonical correlation analysis (MCCA) algorithm, achieving an accuracy of 42% versus MCCA’s 15%. My goal is to go beyond simply expanding bilingual dictionaries so that we can use bilingual lexicon induction techniques in order to produce full translations systems. I am extending the model to translations of phrases (instead of just words), to calculate translation probabilities, and to predict word order of translations. My research has been supported by two DARPA grants – a500k follow-on grant to my participation in the DARPA Computer Science Study Group, and a new $450k multi-year DARPA LORELEI project. ## Crowdsourcing My third research focus is crowdsourcing. The idea of using crowdsourcing to create annotated data for natural language processing applications is a relatively new topic, and it raises a number of scientific challenges. Rather than treating annotated training data as a gold standard created by experts whose labels are authoritative, we must cope with the fact that we have anonymous, non-expert annotators whose labels are noisy and who may not even be doing the task conscientiously. We build models of the annotators themselves, and use those models to create high quality labeled training data by soliciting redundant labels and making predictions about which labels and which annotators are most likely to be correct. The ability to accurately model the annotators has direct implications for the cost of creating a labeled training data set, since we can decide how much to trust a given annotator’s label and whether soliciting a redundant label would be likely to improve the accuracy. Figure 6: A comparison of the translation quality (approximated by Bleu score) for professionals against different ways of selecting the Turker translations from among 4 redundant translations. One of my first successes with crowdsourcing for NLP was to show that the quality of Urdu-English translations produced by non-professional translators can be made to approach the quality of professional translation at a fraction of the cost (Zaidan and Callison-Burch (2011)) Figure 6 highlights the main findings of the study. It shows that through judicious application of quality control techniques, crowdsourced translations can fall in the range that we would expect of professional translators. This, combined with the fact that crowdsourced translations are more than an order of magnitude cheaper, opens the real possibility of creating training data for SMT, previously thought to be too expensive to create from scratch. I have established myself as one of the foremost experts in crowdsourcing as applied to language (Callison-Burch and Dredze (2010)) . I have used crowdsourcing to create a wide range of new NLP data sets, which approach the size and quality of ones produced by the Linguistics Data Consortium (LDC). Here’s a sample of what I have produced: • A large scale evaluation of machine translation output by crowd workers with a comprehensive comparison against expert annotators and recommendations for quality control (Callison-Burch et al (2010)). • An Arabic dialect corpus with manual classification of the dialect for 150k sentences harvested from the online comments of Arabic newspapers from different countries. We used this data to train a system that could automatically identify Arabic dialects (Zaidan and Callison-Burch (2011)). • Translations of 1.5 million words worth of Egyptian Arabic and Levine Arabic data. This allowed us to train a statistical machine translation system to translate Arabic dialect that was better than a system trained on 100 times more MSA data (Zbib et al (2012)). • Parallel corpora for six Indian languages. These data sets are valuable for evaluating syntactic models of translations, since Indian languages are verb final and require a lot of long-distance reordering (Post et al (2012)). • Translations of 1.8 million words of spoken Spanish and of 143k words of spoke Egyptian Arabic. This data facilitates research into spoken language translation (Kumar et al (2014), Post et al (2013)). • A demographic study of the languages spoken by workers on the Mechanical Turk crowdsourcing platform, which resulted in bilingual dictionaries for 100 languages with 10,000 words translated in each of the languages (Pavlick et al (2014)). My research into crowdsourcing for NLP has been facilitated by a$500,000 DARPA grant, two Google faculty research awards totaling $195,000, and approximately$100,000 towards a Mechanical Turk annotation budget from the Human Language Technology Center of Excellence. I have a new 2 year DARPA LORELEI grant under the LDC for creating data for low resource languages using crowdsourcing.

My interests in crowdsourcing have expanded beyond natural language processing. I have designed a semester-long course on Crowdsourcing and Human Computation (described in my teaching statement). I have two research projects related to social justice that use crowdsourcing. In one project, I am designing tools to help crowd workers find better higher paying work (Callison-Burch (2014)) (see crowd-workers.com). In another, I am using crowdsourcing to create a structured database of all shootings in the United States to facilitate public health research (which the Republican congress has blocked the CDC and NIH from doing). See my teaching statement for a description of the gun violence database project (Pavlick and Callison-Burch (2016)).

## Bibliography

Optimizing Statistical Machine Translation for Text Simplification. Wei Xu, Courtney Napoles, Ellie Pavlick, Jim Chen, and Chris Callison-Burch. TACL-2016. Abstract BibTex
Clustering Paraphrases by Word Sense. Anne Cocos and Chris Callison-Burch. NAACL-2016. Abstract BibTex
End-to-End Statistical Machine Translation with Zero or Small Parallel Texts. Ann Irvine and Chris Callison-Burch. Journal of Natural Language Engineering-2016. Abstract BibTex
A Comprehensive Analysis of Bilingual Lexicon Induction. Ann Irvine and Chris Callison-Burch. Computational Linguistics-2016. Abstract
The Gun Violence Database: A new task and data set for NLP. Ellie Pavlick, Heng Ji, Xiaoman Pan and Chris Callison-Burch. EMNLP-2016. Abstract BibTex
PPDB 2.0: Better paraphrase ranking, fine-grained entailment relations, word embeddings, and style classification. Ellie Pavlick, Pushpendre Rastogi, Juri Ganitkevich, Ben Van Durme, Chris Callison-Burch. ACL-2015. Abstract BibTex
Adding Semantics to Data-Driven Paraphrasing. Ellie Pavlick, Johan Bos, Malvina Nissim, Charley Beller, Benjamin Van Durme, and Chris Callison-Burch. ACL-2015. Abstract Figures BibTex
Domain-Specific Paraphrase Extraction. Ellie Pavlick, Juri Ganitkevich, Tsz Ping Chan, Xuchen Yao, Ben Van Durme, Chris Callison-Burch. ACL-2015. Abstract BibTex
Problems in Current Text Simplification Research: New Data Can Help. Wei Xu, Chris Callison-Burch, and Courtney Napoles. TACL-2015. Abstract Figures BibTex
Translations of the CALLHOME Egyptian Arabic corpus for conversational speech translation. Gaurav Kumar, Yuan Cao, Ryan Cotterell, Chris Callison-Burch, Daniel Povey, and Sanjeev Khudanpur. IWSLT-2014. Abstract Figures BibTex
The Language Demographics of Amazon Mechanical Turk. Ellie Pavlick, Matt Post, Ann Irvine, Dmitry Kachaev, and Chris Callison-Burch. TACL-2014. Abstract Figures BibTex
Crowd-Workers: Aggregating Information Across Turkers To Help Them Find Higher Paying Work. Chris Callison-Burch. HCOMP Poster-2014. Abstract Figures BibTex
Joshua 5.0: Sparser, better, faster, server. Matt Post, Juri Ganitkevitch, Luke Orland, Jonathan Weese, Yuan Cao, and Chris Callison-Burch. WMT-2013. Abstract Figures BibTex
PPDB: The Paraphrase Database. Juri Ganitkevitch, Benjamin Van Durme, and Chris Callison-Burch. NAACL-2013. Abstract Figures BibTex
Supervised Bilingual Lexicon Induction with Multiple Monolingual Signals. Ann Irvine and Chris Callison-Burch. NAACL-2013. Abstract Figures BibTex
Improved Speech-to-Text Translation with the Fisher and Callhome Spanish–English Speech Translation Corpus. Matt Post, Gaurav Kumar, Adam Lopez, Damianos Karakos, Chris Callison-Burch and Sanjeev Khudanpur. IWSLT-2013. Abstract Figures BibTex
Joshua 4.0: Packing, PRO, and Paraphrases. Juri Ganitkevitch, Yuan Cao, Jonathan Weese, Matt Post, and Chris Callison-Burch. WMT-2012. Abstract Figures BibTex
Toward Statistical Machine Translation without Parallel Corpora. Alex Klementiev, Ann Irvine, Chris Callison-Burch, and David Yarowsky. EACL-2012. Abstract Figures BibTex
Machine Translation of Arabic Dialects. Rabih Zbib, Erika Malchiodi, Jacob Devlin, David Stallard, Spyros Matsoukas, Richard Schwartz, John Makhoul, Omar F. Zaidan and Chris Callison-Burch. NAACL-2012. Abstract Figures BibTex
Constructing Parallel Corpora for Six Indian Languages via Crowdsourcing. Matt Post, Chris Callison-Burch, and Miles Osborne. WMT-2012. Abstract Figures BibTex
Joshua 3.0: Syntax-based Machine Translation with the Thrax Grammar Extractor. Jonathan Weese, Juri Ganitkevitch, Chris Callison-Burch, Matt Post and Adam Lopez. WMT-2011. Abstract Figures BibTex
Learning Sentential Paraphrases from Bilingual Parallel Corpora for Text-to-Text Generation. Juri Ganitkevitch, Chris Callison-Burch, Courtney Napoles, and Benjamin Van Durme. EMNLP-2011. Abstract Figures BibTex
Crowdsourcing Translation: Professional Quality from Non-Professionals. Omar Zaidan and Chris Callison-Burch. ACL-2011. Abstract Figures BibTex
The Arabic Online Commentary Dataset: An Annotated Dataset of Informal Arabic with High Dialectal Content. Omar Zaidan and Chris Callison-Burch. ACL-2011. Abstract Figures BibTex
Joshua 2.0: A Toolkit for Parsing-Based Machine Translationwith Syntax, Semirings, Discriminative Training and Other Goodies. Zhifei Li, Chris Callison-Burch, Chris Dyer, Juri Ganitkevitch, Ann Irvine, Lane Schwartz, Wren N. G. Thornton, Ziyuan Wang, Jonathan Weese and Omar F. Zaidan. WMT-2010. Abstract BibTex
Semantically-Informed Syntactic Machine Translation: A Tree-Grafting Approach. Kathryn Baker, Michael Bloodgood, Chris Callison-Burch, Bonnie Dorr, Scott Miller, Christine Piatko, Nathaniel W. Filardo, and Lori Levin. AMTA-2010. Abstract Figures BibTex
Creating Speech and Language Data With Amazon’s Mechanical Turk. Chris Callison-Burch and Mark Dredze. NAACL Workshop on Creating Speech and Language Data With Amazon’s Mechanical Turk-2010. Abstract Figures BibTex
Findings of the 2010 Joint Workshop on Statistical Machine Translation and Metrics for Machine Translation. Chris Callison-Burch, Philipp Koehn, Christof Monz, Kay Peterson, Mark Przybocki, Omar Zaidan. WMT-2010. Abstract Figures BibTex
Joshua: An Open Source Toolkit for Parsing-based Machine Translation. Zhifei Li, Chris Callison-Burch, Chris Dyer, Juri Ganitkevitch, Sanjeev Khudanpur, Lane Schwartz, Wren Thornton, Jonathan Weese and Omar Zaidan. WMT-2009. Abstract Figures BibTex
Findings of the 2009 Workshop on Statistical Machine Translation. Chris Callison-Burch, Philipp Koehn, Christof Monz and Josh Schroeder. WMT-2009. Abstract Figures BibTex
Moses: Open source toolkit for statistical machine translation. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondřej Bojar, Alexandra Constantin, and Evan Herbst. ACL-2007. Abstract Figures BibTex
Paraphrasing and Translation. Chris Callison-Burch. PhD Thesis, University of Edinburgh-2007. Abstract Figures BibTex
Paraphrasing with Bilingual Parallel Corpora. Colin Bannard and Chris Callison-Burch. ACL-2005. Abstract Figures BibTex