Wei-Yun Ma

Introduction to CKIP Chinese Word Segmentation System for the First International Chinese Word Segmentation Bakeoff

In this paper, we roughly described the procedures of our segmentation system, including the methods for resolving segmentation ambiguities and identifying unknown words. The CKIP group of Academia Sinica participated in testing on open and closed tracks of Beijing University (PK) and Hong Kong Cityu (HK). The evaluation results show our system performs very well in either HK open track or HK closed track and just acceptable in PK tracks. Some explanations and analysis are presented in this paper.

Unknown word extraction for Chinese documents

There is no blank to mark word boundaries in Chinese text. As a result, identifying words is difficult, because of segmentation ambiguities and occurrences of unknown words. Conventionally unknown words were extracted by statistical methods because statistical methods are simple and efficient. However the statistical methods without using linguistic knowledge suffer the drawbacks of low precision and low recall, since character strings with statistical significance might be phrases or partial phrases instead of words and low frequency new words are hardly identifiable by statistical methods. In addition to statistical information, we try to use as much information as possible, such as morphology, syntax, semantics, and world knowledge. The identification system fully utilizes the context and content information of unknown words in the steps of detection process, extraction process, and verification process. A practical unknown word extraction system was implemented which online identifies new words, including low frequency new words, with high precision and high recall rates.

A Bottom-up Merging Algorithm for Chinese Unknown Word Extraction

Statistical methods for extracting Chinese unknown words usually suffer a problem that superfluous character strings with strong statistical associations are extracted as well. To solve this problem, this paper proposes to use a set of general morphological rules to broaden the coverage and on the other hand, the rules are appended with different linguistic and statistical constraints to increase the precision of the representation. To disambiguate rule applications and reduce the complexity of the rule matching, a bottom-up merging algorithm for extraction is proposed, which merges possible morphemes recursively by consulting above the general rules and dynamically decides which rule should be applied first according to the priorities of the rules. Effects of different priority strategies are compared in our experiment, and experimental results show that the performance of proposed method is very promising.

Who, What, When, Where, Why? Comparing Multiple Approaches to the Cross-Lingual 5W Task

Cross-lingual tasks are especially difficult due to the compounding effect of errors in language processing and errors in machine translation (MT). In this paper, we present an error analysis of a new cross-lingual task: the 5W task, a sentence-level understanding task which seeks to return the English 5Ws (Who, What, When, Where and Why) corresponding to a Chinese sentence. We analyze systems that we developed, identifying specific problems in language processing and MT that cause errors. The best cross-lingual 5W system was still 19% worse than the best monolingual 5W system, which shows that MT significantly degrades sentence-level understanding. Neither source-language nor target-language analysis was able to circumvent problems in MT, although each approach had advantages relative to the other. A detailed error analysis across multiple systems suggests directions for future research on the problem.

Where's the Verb? Correcting Machine Translation During Question Answering

When a multi-lingual question-answering (QA) system provides an answer that has been incorrectly translated, it is very likely to be regarded as irrelevant. In this paper, we propose a novel method for correcting a deletion error that affects overall understanding of the sentence. Our post-editing technique uses information available at query time: examples drawn from related documents determined to be relevant to the query. Our results show that 4%-7% of MT sentences are missing the main verb and on average, 79% of the modified sentences are judged to be more comprehensible. The QA performance also benefits from the improved MT: 7% of irrelevant response sentences become relevant.

Phrase-level System Combination for Machine Translation Based on Target-to-Target Decoding

In this paper, we propose a novel latticebased MT combination methodology that we call Target-to-Target Decoding (TTD). The combination process is carried out as a “translation” from backbone to the combination result. This perspective suggests the use of existing phrase-based MT techniques in the combination framework. We show how phrase extraction rules and confidence estimations inspired from machine translation improve results. We also propose system-specific LMs for estimating N-gram consensus. Our results show that our approach yields a strong improvement over the best single MT system and competes with other stateof-the-art combination systems.

System Combination for Machine Translation through Paraphrasing

In this paper, we propose a paraphrasing model to address the task of system combination for machine translation. We dynamically learn hierarchical paraphrases from target hypotheses and form a synchronous context-free grammar to guide a series of transformations of target hypotheses into fused translations. The model is able to exploit phrasal and structural system-weighted consensus and also to utilize existing information about word ordering present in the target hypotheses. In addition, to consider a diverse set of plausible fused translations, we develop a hybrid combination architecture, where we paraphrase every target hypothesis using different fusing techniques to obtain fused translations for each target, and then make the final selection among all fused translations. Our experimental results show that our approach can achieve a significant improvement over combination baselines.

Detecting and Correcting Syntactic Errors in Machine Translation Using Feature-Based Lexicalized Tree Adjoining Grammars

Statistical machine translation has made tremendous progress over the past ten years. The output of even the best systems, however, is often ungrammatical because of the lack of sufficient linguistic knowledge. Even when systems incorporate syntax in the translation process, syntactic errors still result. To address this issue, we present a novel approach for detecting and correcting ungrammatical translations. In order to simultaneously detect multiple errors and their corresponding words in a formal framework, we use feature-based lexicalized tree adjoining grammars, where each lexical item is associated with a syntactic elementary tree, in which each node is associated with a set of feature-value pairs to define the lexical item’s syntactic usage. Our syntactic error detection works by checking the feature values of all lexical items within a sentence using a unification framework. In order to simultaneously detect multiple error types and track their corresponding words, we propose a new unification method which allows the unification procedure to continue when unification fails and also to propagate the failure information to relevant words. Once error types and their corresponding words are detected, one is able to correct errors based on a unified consideration of all related words under the same error types. In this paper, we present some simple mechanism to handle part of the detected situations. We use our approach to detect and correct translations of six single statistical machine translation systems. The results show that most of the corrected translations are improved.

A Deep Learning Framework for Coreference Resolution Based on Convolutional Neural Network

Recently many researches have shown that word embeddings are able to represent information from word related contexts or its nearest neighborhood words, and thus are applied in many NLP tasks successfully. In this paper, we propose convolutional neural network model to extent word embeddings to mention/antecedent representation. These representations are obtained through convoluting neighboring word embeddings and other contextual information for coreference resolution. We evaluate our system on the English portion of the CoNLL 2012 Shared Task dataset and show that the proposed system achieves a competitive performance compared with the state-of-the-art approaches. We also show that our proposed model especially improves the coreference resolution of long spans significantly.

Leveraging Linguistic Structures for Named Entity Recognition with Bidirectional Recursive Neural Networks

In this paper, we utilize the linguistic structures of texts to improve named entity recognition by BRNN-CNN, a special bidirectional recursive network attached with a convolutional network. Motivated by the observation that named entities are highly related to linguistic constituents, we propose a constituent-based BRNN-CNN for named entity recognition. In contrast to classical sequential labeling methods, the system first identifies which text chunks are possible named entities by whether they are linguistic constituents. Then it classifies these chunks with a constituency tree structure by recursively propagating syntactic and semantic information to each constituent node. This method surpasses current state-of-the-art on OntoNotes 5.0 with automatically generated parses.