DOCSLIB.ORG

Blank Language Model: Flexible Sequence Modeling by Any-Order Generation by Victor Quach B.S

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Blank Language Model: Flexible Sequence Modeling by Any-Order Generation by Victor Quach B.S Blank Language Model: Flexible Sequence Modeling by Any-Order Generation by Victor Quach B.S. in Engineering., École polytechnique (2016) M.S. in Computer Science and Mathematics, École polytechnique (2017) Submitted to the Department of Electrical Engineering and Computer Science in partial fulfillment of the requirements for the degree of Master of Science in Electrical Engineering and Computer Science at the MASSACHUSETTS INSTITUTE OF TECHNOLOGY May 2020 ○c Massachusetts Institute of Technology 2020. All rights reserved. Author................................................................ Department of Electrical Engineering and Computer Science May 15, 2020 Certified by. Regina Barzilay Professor of Electrical Engineering and Computer Science Thesis Supervisor Accepted by . Leslie A. Kolodziejski Professor of Electrical Engineering and Computer Science Chair, Department Committee on Graduate Students 2 Blank Language Model: Flexible Sequence Modeling by Any-Order Generation by Victor Quach Submitted to the Department of Electrical Engineering and Computer Science on May 15, 2020, in partial fulfillment of the requirements for the degree of Master of Science in Electrical Engineering and Computer Science Abstract We propose Blank Language Model (BLM), a model that generates sequences by dy- namically creating and filling in blanks. Unlike previous masked language models [7] or the Insertion Transformer [26], BLM uses blanks to control which part of the se- quence to expand. This fine-grained control of generation is ideal for a variety of text editing and rewriting tasks. The model can start from a single blank or partially completed text with blanks at specified locations. It iteratively determines which word to place in a blank and whether to insert new blanks, and stops generating when no blanks are left to fill. BLM can be efficiently trained using a lower bound of the marginal data likelihood, and achieves perplexity comparable to traditional left-to-right language models on the Penn Treebank and WikiText datasets. On the task of filling missing text snippets, BLM significantly outperforms all other baselines in terms of both accuracy and fluency. Experiments on style transfer and damaged ancient text restoration demonstrate the potential of this framework for a wide range of applications. Thesis Supervisor: Regina Barzilay Title: Professor of Electrical Engineering and Computer Science 3 4 Acknowledgments First, I want to express my deepest gratitude to my advisor, Regina Barzilay. Over the years, she has never failed to provide me with the academic guidance or life advice I needed. Regina has pushed me to think critically and creatively, and to push the boundaries of what I can achieve. I really appreciate her undying patience and support in this journey. I would like to thank Tianxiao Shen and Tommi Jaakola, co-authors of the work on which this thesis is based. I am grateful to have had the opportunity to collaborate with such brilliant minds. I would also like to thank Yujia Bao, Adam Fisch, Benson Chen, Adam Yala, Tal Schuster, Yujie Qian, Jiang Guo, Darsh Shah, Jiaming Luo, Wengong Jin and all other members of MIT NLP group. In these times of self-quarantine, I more than ever measure the value of our research discussions (planned or impromptu) who have always left me more curious and intellectually stimulated. I am thankful to be part of a group composed of such smart, talented, and generous people. I want to thank Yuening Zhang, my girlfriend, dance partner, life partner, and friend. This work would not have been possible without her undiminished support. Finally, I would like to thank my family to which I am deeply indebted. To my sister Hélène who has been a lifelong confidant and friend. To my parents for their sacrifical love, constant encouragement, and support of my studies. I would not be where I am today without all you. Thank you. 5 6 Bibliographic Note This thesis is based on our previous work available as a preprint [25]. 7 8 Contents 1 Introduction 15 2 Related Work 17 3 Blank Language Models 19 4 Experiments 25 4.1 Language Modeling . 26 4.2 Text Infilling . 27 4.3 Ancient Text Restoration . 30 4.4 Sentiment Transfer . 32 5 Conclusion 37 9 10 List of Figures 1-1 BLM fills in blanks of arbitrary length. 16 1-2 An example trajectory that generates the sentence “customer service is awesome”. Each action is a tuple (b; w; l; r), indicating the blank location b selected for expansion, the word w to fill in, whether to create a left blank l, and whether to create a right blank r....... 16 3-1 Architecture of the Blank Language Model. In the first stage, an in- dex is chosen among all current blank positions. For that location, a word is selected in the second stage. In the final stage, the blank representation is concatenated with the chosen word’s embedding and fed into a multilayer perceptron (MLP) to determine the creation of the following blanks. 20 4-1 Examples of inputs and outputs for the three rewriting tasks (text infilling, ancient test restoration and style transfer). We contrast text infilling, where blanks can cover an arbitrary number of words, with ancient text restoration, where the number of characters to recover is indicated by the number of ‘?’ symbols in the input. 25 4-2 Failure rate, BLEU score and perplexity of generated documents for the text infilling task. The “No infill” line reports the BLEU scoreof the blanked document. The “Data PPL” dotted line serves as reference for the perplexity of the original documents . 27 11 4-3 Example generations for the text infilling task, with mask ratios 0.1 and 0.5. Completions are in italic. Invalid completions are in red. For the seq2seq-fill baseline, we represent the outputs of the model along with the merged document. In this example, the insertion transformer produces invalid completions by failing to generate tokens in the “? the” blank. At mask ratio 0.5, the seq2seq-fill baseline also generates an invalid document by producing too many ‘|’ tokens, i.e. filling to many blanks. 29 4-4 Example generations for the style transfer task using attention-based masking mechanism. Masked words are in bold. 34 12 List of Tables 4.1 Perplexity on the Penn Treebank and WikiText datasets. 26 4.2 Character error rate for the ancient text restoration task in both single- slot and multi-slot settings. 31 4.3 Accuracy and BLEU scores for the Yelp sentiment transfer task. Ac- curacy measures the percentage of sentences labeled as the target sen- timent by the classifier. BLEU is evaluated against human reference generations. For reference, we also report accuracy and BLEU scores of the canvas (i.e. the original masked sentence). 33 13 14 Chapter 1 Introduction Neural language models have been successfully applied to many sequence genera- tion tasks, including machine translation [3], summarization [23], and image caption- ing [32]. Typically, sequences are modeled autoregressively from left to right, making the log-likelihood tractable and allowing efficient training and inference. While left- to-right models are effective, they are not well-suited for text completion or editing. In these tasks, we are given a partial draft of the text and the goal is to add new text to complete it. Models such as Masked Language Model [7, MLM] and Insertion Transformer [26] are able to fill in words to complete partially written text. However, neither ofthemis tailored to rewriting/editing. MLM assumes that the length of the text to be inserted is known in advance. Insertion Transformer, on the other hand, does not explicitly control where insertions can take place. In this paper, we introduce Blank Language Model (BLM). The model exploits a special “ ” symbol to control where tokens can be placed. In each stage of generation, a blank can be replaced by any word, and potentially accompanied by a new blank on the left, right or both sides of the word to continue writing. As shown in Fig. 1-1, such models can be used to fill in missing words in incomplete sentences, generate a new sentence in between two given sentences, and so on. BLM can start with a single blank or partial text with blanks in specified locations. The model iterates through generation steps, replacing blanks with words and possibly adjoining blanks, until no 15 They also have which . They also have ice cream which is really good . Figure 1-1: BLM fills in blanks of arbitrary length. Canvas Action Step Location b Word w (Left l, Right r) 0. #1 #1 is Y Y 1. #1 is #2 #1 customer N Y 2. customer #1 is #2 #2 awesome N N 3. customer #1 is awesome #1 service N N 4. customer service is awesome -End- Figure 1-2: An example trajectory that generates the sentence “customer service is awesome”. Each action is a tuple (b; w; l; r), indicating the blank location b selected for expansion, the word w to fill in, whether to create a left blank l, and whether to create a right blank r. blanks remain. Our BLM is based on a Transformer encoder that maps the input text contain- ing blanks into a sequence of vector representations. The representations at blank locations are further processed to select a blank, word to fill in it, and whether to generate adjoining blanks. Since there are multiple trajectories through the actions in the BLM that all result in the same final text, we train the model by maximizing the marginal likelihood. To make training more efficient, and to introduce an inductive bias towards order independence, we maximize instead a lower bound on the marginal likelihood. At test time, BLM can in principle fill in any amount of text in any ofthe given blank positions.
Load more

Recommended publications
  • How Do BERT Embeddings Organize Linguistic Knowledge?
    How Do BERT Embeddings Organize Linguistic Knowledge? Giovanni Puccettiy , Alessio Miaschi? , Felice Dell’Orletta y Scuola Normale Superiore, Pisa ?Department of Computer Science, University of Pisa Istituto di Linguistica Computazionale “Antonio Zampolli”, Pisa ItaliaNLP Lab – www.italianlp.it [email protected], [email protected], [email protected] Abstract et al., 2019), we proposed an in-depth investigation Several studies investigated the linguistic in- aimed at understanding how the information en- formation implicitly encoded in Neural Lan- coded by BERT is arranged within its internal rep- guage Models. Most of these works focused resentation. In particular, we defined two research on quantifying the amount and type of in- questions, aimed at: (i) investigating the relation- formation available within their internal rep- ship between the sentence-level linguistic knowl- resentations and across their layers. In line edge encoded in a pre-trained version of BERT and with this scenario, we proposed a different the number of individual units involved in the en- study, based on Lasso regression, aimed at understanding how the information encoded coding of such knowledge; (ii) understanding how by BERT sentence-level representations is ar- these sentence-level properties are organized within ranged within its hidden units. Using a suite of the internal representations of BERT, identifying several probing tasks, we showed the existence groups of units more relevant for specific linguistic of a relationship between the implicit knowl- tasks. We defined a suite of probing tasks based on edge learned by the model and the number of a variable selection approach, in order to identify individual units involved in the encodings of which units in the internal representations of BERT this competence.
    [Show full text]
  • Treebanks, Linguistic Theories and Applications Introduction to Treebanks
    Treebanks, Linguistic Theories and Applications Introduction to Treebanks Lecture One Petya Osenova and Kiril Simov Sofia University “St. Kliment Ohridski”, Bulgaria Bulgarian Academy of Sciences, Bulgaria ESSLLI 2018 30th European Summer School in Logic, Language and Information (6 August – 17 August 2018) Plan of the Lecture • Definition of a treebank • The place of the treebank in the language modeling • Related terms: parsebank, dynamic treebank • Prerequisites for the creation of a treebank • Treebank lifecycle • Theory (in)dependency • Language (in)dependency • Tendences in the treebank development 30th European Summer School in Logic, Language and Information (6 August – 17 August 2018) Treebank Definition A corpus annotated with syntactic information • The information in the annotation is added/checked by a trained annotator - manual annotation • The annotation is complete - no unannotated fragments of the text • The annotation is consistent - similar fragments are analysed in the same way • The primary format of annotation - syntactic tree/graph 30th European Summer School in Logic, Language and Information (6 August – 17 August 2018) Example Syntactic Trees from Wikipedia The two main approaches to modeling the syntactic information 30th European Summer School in Logic, Language and Information (6 August – 17 August 2018) Pros vs. Cons (Handbook of NLP, p. 171) Constituency • Easy to read • Correspond to common grammatical knowledge (phrases) • Introduce arbitrary complexity Dependency • Flexible • Also correspond to common grammatical
    [Show full text]
  • Senserelate::Allwords - a Broad Coverage Word Sense Tagger That Maximizes Semantic Relatedness
    WordNet::SenseRelate::AllWords - A Broad Coverage Word Sense Tagger that Maximizes Semantic Relatedness Ted Pedersen and Varada Kolhatkar Department of Computer Science University of Minnesota Duluth, MN 55812 USA tpederse,kolha002 @d.umn.edu http://senserelate.sourceforge.net{ } Abstract Despite these difficulties, word sense disambigua- tion is often a necessary step in NLP and can’t sim- WordNet::SenseRelate::AllWords is a freely ply be ignored. The question arises as to how to de- available open source Perl package that as- velop broad coverage sense disambiguation modules signs a sense to every content word (known that can be deployed in a practical setting without in- to WordNet) in a text. It finds the sense of vesting huge sums in manual annotation efforts. Our each word that is most related to the senses answer is WordNet::SenseRelate::AllWords (SR- of surrounding words, based on measures found in WordNet::Similarity. This method is AW), a method that uses knowledge already avail- shown to be competitive with results from re- able in the lexical database WordNet to assign senses cent evaluations including SENSEVAL-2 and to every content word in text, and as such offers SENSEVAL-3. broad coverage and requires no manual annotation of training data. SR-AW finds the sense of each word that is most 1 Introduction related or most similar to those of its neighbors in the Word sense disambiguation is the task of assigning sentence, according to any of the ten measures avail- a sense to a word based on the context in which it able in WordNet::Similarity (Pedersen et al., 2004).
    [Show full text]
  • Deep Linguistic Analysis for the Accurate Identification of Predicate
    Deep Linguistic Analysis for the Accurate Identification of Predicate-Argument Relations Yusuke Miyao Jun'ichi Tsujii Department of Computer Science Department of Computer Science University of Tokyo University of Tokyo [email protected] CREST, JST [email protected] Abstract obtained by deep linguistic analysis (Gildea and This paper evaluates the accuracy of HPSG Hockenmaier, 2003; Chen and Rambow, 2003). parsing in terms of the identification of They employed a CCG (Steedman, 2000) or LTAG predicate-argument relations. We could directly (Schabes et al., 1988) parser to acquire syntac- compare the output of HPSG parsing with Prop- tic/semantic structures, which would be passed to Bank annotations, by assuming a unique map- statistical classifier as features. That is, they used ping from HPSG semantic representation into deep analysis as a preprocessor to obtain useful fea- PropBank annotation. Even though PropBank tures for training a probabilistic model or statistical was not used for the training of a disambigua- tion model, an HPSG parser achieved the ac- classifier of a semantic argument identifier. These curacy competitive with existing studies on the results imply the superiority of deep linguistic anal- task of identifying PropBank annotations. ysis for this task. Although the statistical approach seems a reason- 1 Introduction able way for developing an accurate identifier of Recently, deep linguistic analysis has successfully PropBank annotations, this study aims at establish- been applied to real-world texts. Several parsers ing a method of directly comparing the outputs of have been implemented in various grammar for- HPSG parsing with the PropBank annotation in or- malisms and empirical evaluation has been re- der to explicitly demonstrate the availability of deep ported: LFG (Riezler et al., 2002; Cahill et al., parsers.
    [Show full text]
  • The Procedure of Lexico-Semantic Annotation of Składnica Treebank
    The Procedure of Lexico-Semantic Annotation of Składnica Treebank Elzbieta˙ Hajnicz Institute of Computer Science, Polish Academy of Sciences ul. Ordona 21, 01-237 Warsaw, Poland [email protected] Abstract In this paper, the procedure of lexico-semantic annotation of Składnica Treebank using Polish WordNet is presented. Other semantically annotated corpora, in particular treebanks, are outlined first. Resources involved in annotation as well as a tool called Semantikon used for it are described. The main part of the paper is the analysis of the applied procedure. It consists of the basic and correction phases. During basic phase all nouns, verbs and adjectives are annotated with wordnet senses. The annotation is performed independently by two linguists. Multi-word units obtain special tags, synonyms and hypernyms are used for senses absent in Polish WordNet. Additionally, each sentence receives its general assessment. During the correction phase, conflicts are resolved by the linguist supervising the process. Finally, some statistics of the results of annotation are given, including inter-annotator agreement. The final resource is represented in XML files preserving the structure of Składnica. Keywords: treebanks, wordnets, semantic annotation, Polish 1. Introduction 2. Semantically Annotated Corpora It is widely acknowledged that linguistically annotated cor- Semantic annotation of text corpora seems to be the last pora play a crucial role in NLP. There is even a tendency phase in the process of corpus annotation, less popular than towards their ever-deeper annotation. In particular, seman- morphosyntactic and (shallow or deep) syntactic annota- tically annotated corpora become more and more popular tion. However, there exist semantically annotated subcor- as they have a number of applications in word sense disam- pora for many languages.
    [Show full text]
  • Unified Language Model Pre-Training for Natural
    Unified Language Model Pre-training for Natural Language Understanding and Generation Li Dong∗ Nan Yang∗ Wenhui Wang∗ Furu Wei∗† Xiaodong Liu Yu Wang Jianfeng Gao Ming Zhou Hsiao-Wuen Hon Microsoft Research {lidong1,nanya,wenwan,fuwei}@microsoft.com {xiaodl,yuwan,jfgao,mingzhou,hon}@microsoft.com Abstract This paper presents a new UNIfied pre-trained Language Model (UNILM) that can be fine-tuned for both natural language understanding and generation tasks. The model is pre-trained using three types of language modeling tasks: unidirec- tional, bidirectional, and sequence-to-sequence prediction. The unified modeling is achieved by employing a shared Transformer network and utilizing specific self-attention masks to control what context the prediction conditions on. UNILM compares favorably with BERT on the GLUE benchmark, and the SQuAD 2.0 and CoQA question answering tasks. Moreover, UNILM achieves new state-of- the-art results on five natural language generation datasets, including improving the CNN/DailyMail abstractive summarization ROUGE-L to 40.51 (2.04 absolute improvement), the Gigaword abstractive summarization ROUGE-L to 35.75 (0.86 absolute improvement), the CoQA generative question answering F1 score to 82.5 (37.1 absolute improvement), the SQuAD question generation BLEU-4 to 22.12 (3.75 absolute improvement), and the DSTC7 document-grounded dialog response generation NIST-4 to 2.67 (human performance is 2.65). The code and pre-trained models are available at https://github.com/microsoft/unilm. 1 Introduction Language model (LM) pre-training has substantially advanced the state of the art across a variety of natural language processing tasks [8, 29, 19, 31, 9, 1].
    [Show full text]
  • Building a Treebank for French
    Building a treebank for French £ £¥ Anne Abeillé£ , Lionel Clément , Alexandra Kinyon ¥ £ TALaNa, Université Paris 7 University of Pennsylvania 75251 Paris cedex 05 Philadelphia FRANCE USA abeille, clement, [email protected] Abstract Very few gold standard annotated corpora are currently available for French. We present an ongoing project to build a reference treebank for French starting with a tagged newspaper corpus of 1 Million words (Abeillé et al., 1998), (Abeillé and Clément, 1999). Similarly to the Penn TreeBank (Marcus et al., 1993), we distinguish an automatic parsing phase followed by a second phase of systematic manual validation and correction. Similarly to the Prague treebank (Hajicova et al., 1998), we rely on several types of morphosyntactic and syntactic annotations for which we define extensive guidelines. Our goal is to provide a theory neutral, surface oriented, error free treebank for French. Similarly to the Negra project (Brants et al., 1999), we annotate both constituents and functional relations. 1. The tagged corpus pronoun (= him ) or a weak clitic pronoun (= to him or to As reported in (Abeillé and Clément, 1999), we present her), plus can either be a negative adverb (= not any more) the general methodology, the automatic tagging phase, the or a simple adverb (= more). Inflectional morphology also human validation phase and the final state of the tagged has to be annotated since morphological endings are impor- corpus. tant for gathering constituants (based on agreement marks) and also because lots of forms in French are ambiguous 1.1. Methodology with respect to mode, person, number or gender. For exam- 1.1.1.
    [Show full text]
  • Converting an HPSG-Based Treebank Into Its Parallel Dependency-Based Treebank
    Converting an HPSG-based Treebank into its Parallel Dependency-based Treebank Masood Ghayoomiy z Jonas Kuhnz yDepartment of Mathematics and Computer Science, Freie Universität Berlin z Institute for Natural Language Processing, University of Stuttgart [email protected] [email protected] Abstract A treebank is an important language resource for supervised statistical parsers. The parser induces the grammatical properties of a language from this language resource and uses the model to parse unseen data automatically. Since developing such a resource is very time-consuming and tedious, one can take advantage of already extant resources by adapting them to a particular application. This reduces the amount of human effort required to develop a new language resource. In this paper, we introduce an algorithm to convert an HPSG-based treebank into its parallel dependency-based treebank. With this converter, we can automatically create a new language resource from an existing treebank developed based on a grammar formalism. Our proposed algorithm is able to create both projective and non-projective dependency trees. Keywords: Treebank Conversion, the Persian Language, HPSG-based Treebank, Dependency-based Treebank 1. Introduction treebank is introduced in Section 4. The properties Supervised statistical parsers require a set of annotated of the available dependency treebanks for Persian and data, called a treebank, to learn the grammar of the their pros and cons are explained in Section 5. The target language and create a wide coverage grammar tools and the experimental setup as well as the ob- model to parse unseen data. Developing such a data tained results are described and discussed in Section source is a tedious and time-consuming task.
    [Show full text]
  • Corpus Based Evaluation of Stemmers
    View metadata, citation and similar papers at core.ac.uk brought to you by CORE provided by Repository of the Academy's Library Corpus based evaluation of stemmers Istvan´ Endredy´ Pazm´ any´ Peter´ Catholic University, Faculty of Information Technology and Bionics MTA-PPKE Hungarian Language Technology Research Group 50/a Prater´ Street, 1083 Budapest, Hungary [email protected] Abstract There are many available stemmer modules, and the typical questions are usually: which one should be used, which will help our system and what is its numerical benefit? This article is based on the idea that any corpora with lemmas can be used for stemmer evaluation, not only for lemma quality measurement but for information retrieval quality as well. The sentences of the corpus act as documents (hit items of the retrieval), and its words are the testing queries. This article describes this evaluation method, and gives an up-to-date overview about 10 stemmers for 3 languages with different levels of inflectional morphology. An open source, language independent python tool and a Lucene based stemmer evaluator in java are also provided which can be used for evaluating further languages and stemmers. 1. Introduction 2. Related work Stemmers can be evaluated in several ways. First, their Information retrieval (IR) systems have a common crit- impact on IR performance is measurable by test collec- ical point. It is when the root form of an inflected word tions. The earlier stemmer evaluations (Hull, 1996; Tordai form is determined. This is called stemming. There are and De Rijke, 2006; Halacsy´ and Tron,´ 2007) were based two types of the common base word form: lemma or stem.
    [Show full text]
  • Building a User-Generated Content North-African Arabizi Treebank: Tackling Hell
    Building a User-Generated Content North-African Arabizi Treebank: Tackling Hell Djamé Seddah1 Farah Essaidi1 Amal Fethi1 Matthieu Futeral1 Benjamin Muller1;2 Pedro Javier Ortiz Suárez1;2 Benoît Sagot1 Abhishek Srivastava1 1Inria, Paris, France 2Sorbonne Université, Paris, France [email protected] Abstract sourced, namely Haitian Creole and Algerian di- alectal Arabic in these cases. No readily avail- We introduce the first treebank for a romanized able parsing and machine translations systems are user-generated content variety of Algerian, a available for such languages. Taking as an ex- North-African Arabic dialect known for its fre- quent usage of code-switching. Made of 1500 ample the Arabic dialects spoken in North-Africa, sentences, fully annotated in morpho-syntax mostly from Morocco to Tunisia, sometimes called and Universal Dependency syntax, with full Maghribi, sometimes Darija, these idioms notori- translation at both the word and the sentence ously contain various degrees of code-switching levels, this treebank is made freely available. with languages of former colonial powers such as It is supplemented with 50k unlabeled sen- French, Spanish, and, to a much lesser extent, Ital- tences collected from Common Crawl and web- ian, depending on the area of usage (Habash, 2010; crawled data using intensive data-mining tech- Cotterell et al., 2014; Saadane and Habash, 2015). niques. Preliminary experiments demonstrate its usefulness for POS tagging and dependency They share Modern Standard Arabic (MSA) as parsing. We believe that what we present in their matrix language (Myers-Scotton, 1993), and this paper is useful beyond the low-resource of course present a rich morphology.
    [Show full text]
  • Lecture 5: Part-Of-Speech Tagging
    CS447: Natural Language Processing http://courses.engr.illinois.edu/cs447 Lecture 5: Part-of-Speech Tagging Julia Hockenmaier [email protected] 3324 Siebel Center POS tagging Tagset: NNP: proper noun CD: numeral, POS tagger JJ: adjective, ... Raw text Tagged text Pierre_NNP Vinken_NNP ,_, 61_CD Pierre Vinken , 61 years old years_NNS old_JJ ,_, will_MD join_VB , will join the board as a the_DT board_NN as_IN a_DT nonexecutive director Nov. nonexecutive_JJ director_NN Nov._NNP 29 . 29_CD ._. CS447: Natural Language Processing (J. Hockenmaier) !2 Why POS tagging? POS tagging is a prerequisite for further analysis: –Speech synthesis: How to pronounce “lead”? INsult or inSULT, OBject or obJECT, OVERflow or overFLOW," DIScount or disCOUNT, CONtent or conTENT –Parsing: What words are in the sentence? –Information extraction: Finding names, relations, etc. –Machine Translation: The noun “content” may have a different translation from the adjective. CS447: Natural Language Processing (J. Hockenmaier) !3 POS Tagging Words often have more than one POS: " -The back door (adjective) -On my back (noun) -Win the voters back (particle) -Promised to back the bill (verb)" The POS tagging task is to determine the POS tag " for a particular instance of a word. " Since there is ambiguity, we cannot simply look up the correct POS in a dictionary. These examples from Dekang Lin CS447: Natural Language Processing (J. Hockenmaier) !4 Defining a tagset CS447: Natural Language Processing (J. Hockenmaier) !5 Defining a tag set We have to define an inventory of labels for the word classes (i.e. the tag set)" -Most taggers rely on models that have to be trained on annotated (tagged) corpora.
    [Show full text]
  • Merging Propbank, Nombank, Timebank, Penn Discourse Treebank and Coreference James Pustejovsky, Adam Meyers, Martha Palmer, Massimo Poesio
    Merging PropBank, NomBank, TimeBank, Penn Discourse Treebank and Coreference James Pustejovsky, Adam Meyers, Martha Palmer, Massimo Poesio Abstract annotates the temporal features of propositions Many recent annotation efforts for English and the temporal relations between propositions. have focused on pieces of the larger problem The Penn Discourse Treebank (Miltsakaki et al of semantic annotation, rather than initially 2004a/b) treats discourse connectives as producing a single unified representation. predicates and the sentences being joined as This paper discusses the issues involved in arguments. Researchers at Essex were merging four of these efforts into a unified responsible for the coreference markup scheme linguistic structure: PropBank, NomBank, the developed in MATE (Poesio et al, 1999; Poesio, Discourse Treebank and Coreference 2004a) and have annotated corpora using this Annotation undertaken at the University of scheme including a subset of the Penn Treebank Essex. We discuss resolving overlapping and (Poesio and Vieira, 1998), and the GNOME conflicting annotation as well as how the corpus (Poesio, 2004a). This paper discusses the various annotation schemes can reinforce issues involved in creating a Unified Linguistic each other to produce a representation that is Annotation (ULA) by merging annotation of greater than the sum of its parts. examples using the schemata from these efforts. Crucially, all individual annotations can be kept 1. Introduction separate in order to make it easy to produce alternative annotations of a specific type of The creation of the Penn Treebank (Marcus et al, semantic information without need to modify the 1993) and the word sense-annotated SEMCOR annotation at the other levels. Embarking on (Fellbaum, 1997) have shown how even limited separate annotation efforts has the advantage of amounts of annotated data can result in major allowing researchers to focus on the difficult improvements in complex natural language issues in each area of semantic annotation and understanding systems.
    [Show full text]