Semi-Markov Phrase-Based Monolingual Alignment

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 590–600, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Semi-Markov Phrase-based Monolingual Alignment Xuchen Yao and Benjamin Van Durme Johns Hopkins University Baltimore, MD, USA Chris Callison-Burch∗ University of Pennsylvania Philadelphia, PA, USA Peter Clark Allen Institute for Artificial Intelligence Seattle, WA, USA Abstract We introduce a novel discriminative model for phrase-based monolingual alignment using a semi-Markov CRF. Our model achieves state- of-the-art alignment accuracy on two phrase- based alignment datasets (RTE and para- phrase), while doing significantly better than other strong baselines in both non-identical alignment and phrase-only alignment. Addi- tional experiments highlight the potential ben- efit of our alignment model to RTE, para- phrase identification and question answering, where even a naive application of our model’s alignment score approaches the state of the art. 1 Introduction Various NLP tasks can be treated as an alignment problem: machine translation (aligning words in one language with words in another language), ques- tion answering (aligning question words with the an- swer phrase), textual entailment recognition (align- ing premise with hypothesis), paraphrase detection (aligning semantically equivalent words), etc. Even though most of these tasks involve only a single lan- guage, alignment research has primarily focused on the bilingual setting (i.e., machine translation) rather than monolingual. Moreover, most work has con- sidered token-based approaches over phrase-based.1 Here we seek to address this imbalance by proposing better phrase-based models for monolingual word alignment. ∗Performed while faculty at Johns Hopkins University. 1In this paper we use the term token-based alignment for one-to-one alignment and phrase-based for non one-to-one alignment, and word alignment in general for both. Most token-based alignment models can extrin- sically handle phrase-based alignment to some ex- tent. For instance, in the case of NYC align- ing to New York City, the single source word NYC may align three times separately to the tar- get words: NYC↔New, NYC↔York, NYC↔City. Or in the case of identical alignment, New York City aligning to New York City is simply New↔New, York↔York, City↔City. How- ever, it is not as clear how to token-align New York (as a city) with New York City. The problem is more prominent when aligning phrasal paraphrases or multiword expressions, such as pass away and kick the bucket. This suggests an intrinsi- cally phrase-based alignment model. The token aligner jacana-align (Yao et al., 2013a) has achieved state-of-the-art result on the task of monolingual alignment, based on previous work of Blunsom and Cohn (2006). It employs a Conditional Random Field (Lafferty et al., 2001) to align tokens from the source sentence to tokens in the target sen- tence, by treating source tokens as “observation” and target tokens as “hidden states”. However, it is not designed to handle phrase-based alignment, largely due to the Markov nature of the underlying model: a state can only span one token each time, making it unable to align multiple consecutive tokens (i.e. a phrase). We extend this model by introducing semi- Markov states …

Content truncated to 500 words