Base_JSON/prefixU/json/U17/U17-1016.json

{
    "paper_id": "U17-1016",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T03:11:29.383591Z"
    },
    "title": "SuperOCR for ALTA 2017 Shared Task",
    "authors": [
        {
            "first": "Yufei",
            "middle": [],
            "last": "Wang",
            "suffix": "",
            "affiliation": {},
            "email": "yufei.wang@student.unsw.edu.au"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "This paper describes the SuperOCR system submitted for the ALTA 2017 shared task, which aims at correcting noisy OCR output for the Trove database. We used heuristic rules and patterns in submitted system and we apply language model to further improve our system. Experiment shows that language model plays an vital role in performance. Surprisingly, a trigram language model outperforms LSTM language model in this task.",
    "pdf_parse": {
        "paper_id": "U17-1016",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "This paper describes the SuperOCR system submitted for the ALTA 2017 shared task, which aims at correcting noisy OCR output for the Trove database. We used heuristic rules and patterns in submitted system and we apply language model to further improve our system. Experiment shows that language model plays an vital role in performance. Surprisingly, a trigram language model outperforms LSTM language model in this task.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "acters being dis-recognized during OCR process, and (b) delete frequent noisy text pattern. For (a), we first filtered out normal words using vocabulary list; then we applied correction to those errorlike tokens. For (b), we extracted the frequent corrected patterns from the aligned documents. The experiment result shows that language model and high-quality vocabulary list are vital to boost performance. Surprisingly, a simple tri-gram language model outperforms a state-of-the-art LSTM language model in selecting candidate words for correction.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "This paper is organized as follows: Section 2 profiles data set and OCR errors. Section 3 introduces submitted system and improved system. Section 4 shows the experiment result. We summarize our finding in Section 5.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "In this section, we first summarize some basic statistics of the OCR documents from this shared task in Table 1. #Docs 6000 #avg. Words 571.9 #avg. Errors 39.6 #Error Ratio 6.93% We are also interested in the types of correction made by annotators in this data set. To extract the correction, we first align document pairs and unaligned words in raw documents are the corrections. We characterize them largely based on the vocabulary list used in our final system. We refer words in the list to \"in-vocabulary\" (IV), otherwise \"Out-of-Vocabulary\" (OOV). We categorize these corrections into the following types:",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 104,
                        "end": 112,
                        "text": "Table 1.",
                        "ref_id": "TABREF0"
                    }
                ],
                "eq_spans": [],
                "section": "OCR Documents and Errors",
                "sec_num": "2"
            },
            {
                "text": "\u2022 Single Word Split: Split an IV into two OOVs. 21.89%",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "OCR Documents and Errors",
                "sec_num": "2"
            },
            {
                "text": "\u2022 Single Word Correction: Change an OOV to an IV. 18.16%",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "OCR Documents and Errors",
                "sec_num": "2"
            },
            {
                "text": "\u2022 Words Deletion: Delete words from OCRed Text. 12.79%",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "OCR Documents and Errors",
                "sec_num": "2"
            },
            {
                "text": "\u2022 Multiple Words Merge: Merge multiple words into a single IV. 8.05%",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "OCR Documents and Errors",
                "sec_num": "2"
            },
            {
                "text": "\u2022 Punctuation Transform: Change a punctuation to another punctuation. 4.52%",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "OCR Documents and Errors",
                "sec_num": "2"
            },
            {
                "text": "\u2022 Known Word Correction: Change an IV to another IV. 4.18%",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "OCR Documents and Errors",
                "sec_num": "2"
            },
            {
                "text": "\u2022 Unknown Word Modify: Change an OOV to another OOV. 3.95%",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "OCR Documents and Errors",
                "sec_num": "2"
            },
            {
                "text": "\u2022 Character Case: Change the character case of a word. 1.44%",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "OCR Documents and Errors",
                "sec_num": "2"
            },
            {
                "text": "\u2022 Other: Other Corrections 25%",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "OCR Documents and Errors",
                "sec_num": "2"
            },
            {
                "text": "It should be noted that, the correction Single Word Split often splits valid words into two OOVs randomly, for example, randomly modifying \"yesterday\" to \"yes\" and \"terday\"; modifying \"Australian\" to \"Austra\" and \"lian\" etc. We suspect that this is caused by a text processing errors. The correction distribution also shows the difficulties of this shared task as only 6.93% of words are modified while majority of words remain unchanged. Even worse, around 25% of errors are multiple words correction, which cannot be solved by checking single words.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "OCR Documents and Errors",
                "sec_num": "2"
            },
            {
                "text": "Lastly, we analysis the length of the continuous corrected word sequence in the data set. As shown in Table 2 , around 75% errors are length 1. Therefore, most of OCR errors are not continuous and separated by context words.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 102,
                        "end": 109,
                        "text": "Table 2",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "OCR Documents and Errors",
                "sec_num": "2"
            },
            {
                "text": "Len. 1 2 3 4 5+ % 74.9% 13.2% 5.7% 3.9% 2.3% ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "OCR Documents and Errors",
                "sec_num": "2"
            },
            {
                "text": "There are two important factors to consider when designing system:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System Description",
                "sec_num": "3"
            },
            {
                "text": "\u2022 Candidate Filtering In our data, only 6.9% of words are corrected by annotators, which means that we are facing an imbalanced situation. If we apply correction to every word in document, it will generate a lot of false positive examples. Our intuition is that, a valid word is unlikely to be an error in OCRed text.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System Description",
                "sec_num": "3"
            },
            {
                "text": "\u2022 Independent Correction As shown previously, most of errors are isolated by their unchanged context. Correction using sequence modeling may not be helpful as the dependency between errors are weak. Therefore, we correct words individually in our system.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System Description",
                "sec_num": "3"
            },
            {
                "text": "Therefore, we design our system as shown in Alg 1. The system includes following postcorrection components:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System Description",
                "sec_num": "3"
            },
            {
                "text": "\u2022 ProcRawText: Correct frequent errors and split text into tokens.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System Description",
                "sec_num": "3"
            },
            {
                "text": "\u2022 Word Filtering: Filtering out most of correct words in OCRed text.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System Description",
                "sec_num": "3"
            },
            {
                "text": "\u2022 Correction: Correct an OOV word into a non-OOV word.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System Description",
                "sec_num": "3"
            },
            {
                "text": "\u2022 ProcessKnownWord: Correct a non-OOV word to another non-OOV word.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System Description",
                "sec_num": "3"
            },
            {
                "text": "Both of our systems follow the above framework and they only differ in strategies used in each component.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System Description",
                "sec_num": "3"
            },
            {
                "text": "Data: OCRed Text W ordList = ProcRawText(OCR T ext); Create CorrectedList ; foreach w 2 W ordList do if w should be Corrected then w = Correction(w); else w = ProcessKnownWord(w);",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Algorithm 1: System Framework",
                "sec_num": null
            },
            {
                "text": "Add w To CorrectedList; return TextJoin(CorrectedList) ;",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Algorithm 1: System Framework",
                "sec_num": null
            },
            {
                "text": "In our submitted system, we mainly used heuristic rules and patterns obtained from training data to correct text.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Submitted System",
                "sec_num": "3.1"
            },
            {
                "text": "In P rocRawT ext part, we used three strategies:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Submitted System",
                "sec_num": "3.1"
            },
            {
                "text": "1. Deleting a set of errors patterns with format \"-* \" and \"-* \" where \"*\" stands for one of \"< \u21e4 > i j : ? ! 1 ; l\". These patterns are the most frequent deleted error sequences in training pairs. Interestingly, these patterns often result in valid words being splitted and their character shapes are similar to each other. So these errors may be caused by similar noise in image input.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Submitted System",
                "sec_num": "3.1"
            },
            {
                "text": "2. Splitting text based on white space and removing length one tokens except for \"a\" and \"A\". These length one tokens tend to be noise in the data set. Although \"I\" is indeed a valid word, our experiment result shows that there are much more noisy \"I\" than the valid one. So, we remove it as well.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Submitted System",
                "sec_num": "3.1"
            },
            {
                "text": "3. the leading and following non-alphabet characters are removed from each word except for the following punctuation \".\" and \",\". We skip numbers and punctuation in this step.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Submitted System",
                "sec_num": "3.1"
            },
            {
                "text": "In W ord F iltering part, we constructed a vocabulary list by merging the most frequent 15000 words from corrected OCR documents in training data and most frequent words 10000 from 1 Billion Word Language Model Benchmark 3 . Given the vocabulary list, OOVs or words with most three non-alphabet characters (words with four or more non-alphabet characters are too noisy to be corrected) are selected as correction target. All other words remain unchanged.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Submitted System",
                "sec_num": "3.1"
            },
            {
                "text": "In Correction part, we extracted frequent single correction pairs in training data (e.g. \"tne\" ) \"the\"). If a candidate matched one of the pairs, we would correct it. Otherwise, we exhaustively searched for words in vocabulary list that are k edit distance from the correction target. We refer these two methods as word-level correction and exhaustive correction respectively. Finally, we will correct the character case based on the original word shape.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Submitted System",
                "sec_num": "3.1"
            },
            {
                "text": "We skipped P rocessKnownW ord stage in our submitted system.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Submitted System",
                "sec_num": "3.1"
            },
            {
                "text": "This system was submitted after the shared task. Following (Tursun and Cakici, 2017) , we used trigram KenLM (Heafield et al., 2013) 4 language model. Experiment result shows that tri-gram language model is sufficient for this task. To train the language model, we used the corrected documents in training data and lower-cased all words before the training. Intuitively, language model would capture the context information and therefore, it is helpful when we rank the correction candidate.",
                "cite_spans": [
                    {
                        "start": 59,
                        "end": 84,
                        "text": "(Tursun and Cakici, 2017)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 109,
                        "end": 132,
                        "text": "(Heafield et al., 2013)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 133,
                        "end": 134,
                        "text": "4",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Language model enhanced System",
                "sec_num": "3.2"
            },
            {
                "text": "In P rocRawT ext part, we still used strategy 1 and 2 in submitted system. However, we changed the tokenization method to the one used in the provided evaluation script. This method splits punctuation from words and provides better boundaries between words and punctuation, but, this could potentially lead to inappropriate punctuation split in noisy text. For example, given noisy word \"-Mr.\" whose ground truth of is \"Mr.\", our method splits it into \"-Mr\" and \".\", making it impossible to merge the punctuation back. To tackle this issue, we collected the frequent cases from training data to correct the errors before tokenization.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Language model enhanced System",
                "sec_num": "3.2"
            },
            {
                "text": "In W ord F iltering part, we constructed the vocabulary list by combining non-singleton words that are no shorter than 5 in training corrected text and words that are no longer than 4 in 1 Billion Language model benchmark.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Language model enhanced System",
                "sec_num": "3.2"
            },
            {
                "text": "In Correction part, we additionally applied character error information to suggest correction candidates. We first extracted all Single Word Correction pairs (18.16% of all errors) and then aligned each of them in character level. We refer this as character-level correction. Besides, we corrected an OOV by merging or splitting if we can obtain IVs.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Language model enhanced System",
                "sec_num": "3.2"
            },
            {
                "text": "To combine both language model and word/char transformation information, we applied \"Noisy Channel Model\" (Mays et al., 1991) to select optimal candidates. Formally, we tried to find the optimal word c for correction target w such that it maximize P (c|w), as shown in 1:",
                "cite_spans": [
                    {
                        "start": 106,
                        "end": 125,
                        "text": "(Mays et al., 1991)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Language model enhanced System",
                "sec_num": "3.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "arg min c P r(w|c) \u21e4 P r(c)",
                        "eq_num": "(1)"
                    }
                ],
                "section": "Language model enhanced System",
                "sec_num": "3.2"
            },
            {
                "text": "where P r(c) is the language model score indicating how likely it should be there given the context; while P r(w|c) is the error model indicating how like w is an error of c. To unify our character-level, word-level and exhaustive correction, we grouped error pairs with same correction target together and normalize their count as weight. We always assigned exhaustive correction a constant weight. Note that, we only applied character transformation once to each word, the combination of trans-formation have not been considered here. Given a word, its context window and candidates list, (a) we calculated the language model score for all candidates and original word, with context, which is a window with size of 5, which includes itself and its previous and following two context words. Candidates words that receive higher score than original word are chosen for comparison using above model. If no candidate words get higher score than the original word, we remained original word unchanged.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Language model enhanced System",
                "sec_num": "3.2"
            },
            {
                "text": "In P rocessKnownW ord, our improved model applied known word transformation correction by using the frequent patterns in training data. If a known word was found in the patterns, we used the above Noisy Channel Model to decide if we should make correction.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Language model enhanced System",
                "sec_num": "3.2"
            },
            {
                "text": "In this section, we show two performance measure, the randomly sampled development set (Dev.) and final test set (Test.) performance for our systems. During system development, we split provided documents into 5500 documents for training and 500 documents for validation. The performance is shown in Table 3 .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 300,
                        "end": 307,
                        "text": "Table 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Experiment Result and Discussion",
                "sec_num": "4"
            },
            {
                "text": "Dev.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System",
                "sec_num": null
            },
            {
                "text": "Test. Sub Sys.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System",
                "sec_num": null
            },
            {
                "text": "17.72% 16.82% LM-Imprved Sys.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "System",
                "sec_num": null
            },
            {
                "text": "20.68% 20.72% Table 3 : Performance for both System Table 3 shows that the language model boosts system performance by around 4%. Additionally, our new system no longer suffered from overfitting as the submitted system did. This indicates that the context information provided by language model surpassed hand-crafted rules in earlier system.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 14,
                        "end": 21,
                        "text": "Table 3",
                        "ref_id": null
                    },
                    {
                        "start": 52,
                        "end": 59,
                        "text": "Table 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "System",
                "sec_num": null
            },
            {
                "text": "To show the effectiveness of each components, an ablation study is conducted for our final system in this section. Note that all reported performance is based on development data set. Table 4 shows that language model is the most vital component in the system. This shows the importance of context modeling components for spelling correction task. In addition, the contribution of the vocabulary list cannot be neglected. We manually investigated the vocabulary of corrected OCR text and found that low frequent words tend to be noisy. Many of these low frequent words should have been corrected during the annotation process. This indicates that the quality of training data needs to be improved.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 184,
                        "end": 191,
                        "text": "Table 4",
                        "ref_id": "TABREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Ablation Study",
                "sec_num": "4.1"
            },
            {
                "text": "In our final system, we used a tri-gram traditional language model. Will higher order language model or advanced neural model continuously improve the performance? We conduct an experiment regarding the order of language model in this section.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Optimal Language Model",
                "sec_num": "4.2"
            },
            {
                "text": "In this experiment, we applied a language model based on LSTM (Hochreiter and Schmidhuber, 1997) . Comparing with transitional language model, LSTM language model can be viewed as 1-order because LSTM can capture long-range dependent information using the cell and hidden information (Hochreiter and Schmidhuber, 1997) . In our experiment, we used the tensorflow implementation 5 of (Kim et al., 2016) . We did not change the default parameter setting in the source code as they are optimized based on English Penn Treebank (PTB) (Marcus et al., 1993) . We used the corrected documents with the same text prepossess technologies as we train KenLM. We also experimented with uni-, bi-, 4-, 5-, 6gram KenLM for comparsion. Note that, we used LSTM language model in the same way as we use KenLM.",
                "cite_spans": [
                    {
                        "start": 62,
                        "end": 96,
                        "text": "(Hochreiter and Schmidhuber, 1997)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 284,
                        "end": 318,
                        "text": "(Hochreiter and Schmidhuber, 1997)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 383,
                        "end": 401,
                        "text": "(Kim et al., 2016)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 530,
                        "end": 551,
                        "text": "(Marcus et al., 1993)",
                        "ref_id": "BIBREF5"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Optimal Language Model",
                "sec_num": "4.2"
            },
            {
                "text": "The performance for different language models are shown in Fig. 4.2 . As the result show, biand tri-gram language models have already provided satisfying performance. Higher order language model even slightly decrease performance by 0.1%. Surprisingly, the LSTM-based language model dramatically decrease the performance by over 5%. During the LSTM model training, we monitored average perplexity over development set. The final model performance is around 80 perplexity which is a reasonable performance compared with (Kim et al., 2016) , showing no overfitting in model training. We argue that two possible reasons could explain this:",
                "cite_spans": [
                    {
                        "start": 519,
                        "end": 537,
                        "text": "(Kim et al., 2016)",
                        "ref_id": "BIBREF3"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 59,
                        "end": 67,
                        "text": "Fig. 4.2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Optimal Language Model",
                "sec_num": "4.2"
            },
            {
                "text": "1. The training data for neural networks is too noisy. It has been shown that neural networks cannot work well when training on noisy data. (Natarajan et al., 2013) 2. In the task of OCR post-correction, correcting errors only require nearby words, rather than long-dependency information would provide noisy information.",
                "cite_spans": [
                    {
                        "start": 140,
                        "end": 164,
                        "text": "(Natarajan et al., 2013)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Optimal Language Model",
                "sec_num": "4.2"
            },
            {
                "text": "We can conclude that bi-and tri-gram language model are the optimal choice for OCR postcorrection task. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Optimal Language Model",
                "sec_num": "4.2"
            },
            {
                "text": "We applied heuristic rule and patterns to the task of OCR post-correction. We further apply language model to boost the performance. Our system finally achieve average F1 score 20.68%, a 2 nd score in all submitted systems. Experiment result suggest that 3-order language model is more capable in modeling context information than state-of-the-art LSTM-based language model when ranking correction candidates.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "5"
            },
            {
                "text": "http://www.alta.asn.au/events/sharedtask2017/description.html 2 http://trove.nla.gov.au/",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "http://www.statmt.org/lm-benchmark/ 4 https://github.com/kpu/kenlm",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "https://github.com/dhyeon/character-aware-neurallanguage-models",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "We would like to thank Prof. Wei Wang for discussion, Dr. Stephen Wan and Dr. Diego Moll-Aliod for their time in proofreading this paper. We would also like to thank the organizers of the shared task for their support.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgments",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Scalable modified Kneser-Ney language model estimation",
                "authors": [
                    {
                        "first": "Kenneth",
                        "middle": [],
                        "last": "Heafield",
                        "suffix": ""
                    },
                    {
                        "first": "Ivan",
                        "middle": [],
                        "last": "Pouzyrevsky",
                        "suffix": ""
                    },
                    {
                        "first": "Jonathan",
                        "middle": [
                            "H"
                        ],
                        "last": "Clark",
                        "suffix": ""
                    },
                    {
                        "first": "Philipp",
                        "middle": [],
                        "last": "Koehn",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kenneth Heafield, Ivan Pouzyrevsky, Jonathan H. Clark, and Philipp Koehn. 2013. Scalable modi- fied Kneser-Ney language model estimation. In Pro- ceedings of the 51st Annual Meeting of the Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Long short-term memory",
                "authors": [
                    {
                        "first": "Sepp",
                        "middle": [],
                        "last": "Hochreiter",
                        "suffix": ""
                    },
                    {
                        "first": "J\u00fcrgen",
                        "middle": [],
                        "last": "Schmidhuber",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Neural computation",
                "volume": "9",
                "issue": "8",
                "pages": "1735--1780",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sepp Hochreiter and J\u00fcrgen Schmidhuber. 1997. Long short-term memory. Neural computation 9(8):1735-1780.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Trove: Innovation in access to information in australia",
                "authors": [
                    {
                        "first": "Rose",
                        "middle": [],
                        "last": "Holley",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Ariadne",
                "volume": "",
                "issue": "64",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rose Holley. 2010. Trove: Innovation in access to in- formation in australia. Ariadne (64).",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Character-aware neural language models",
                "authors": [
                    {
                        "first": "Yoon",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    },
                    {
                        "first": "Yacine",
                        "middle": [],
                        "last": "Jernite",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Sontag",
                        "suffix": ""
                    },
                    {
                        "first": "Alexander M",
                        "middle": [],
                        "last": "Rush",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Thirtieth AAAI Conference on Artificial Intelligence",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yoon Kim, Yacine Jernite, David Sontag, and Alexan- der M Rush. 2016. Character-aware neural language models. In Thirtieth AAAI Conference on Artificial Intelligence.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Finding names in trove: Named entity recognition for australian historical newspapers",
                "authors": [
                    {
                        "first": "Mac",
                        "middle": [],
                        "last": "Sunghwan",
                        "suffix": ""
                    },
                    {
                        "first": "Steve",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Cassidy",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the Australasian Language Technology Association Workshop",
                "volume": "",
                "issue": "",
                "pages": "57--65",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sunghwan Mac Kim and Steve Cassidy. 2015. Find- ing names in trove: Named entity recognition for australian historical newspapers. In Proceedings of the Australasian Language Technology Association Workshop 2015. pages 57-65.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Building a large annotated corpus of english: The penn treebank",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Mitchell",
                        "suffix": ""
                    },
                    {
                        "first": "Mary",
                        "middle": [
                            "Ann"
                        ],
                        "last": "Marcus",
                        "suffix": ""
                    },
                    {
                        "first": "Beatrice",
                        "middle": [],
                        "last": "Marcinkiewicz",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Santorini",
                        "suffix": ""
                    }
                ],
                "year": 1993,
                "venue": "Computational linguistics",
                "volume": "19",
                "issue": "2",
                "pages": "313--330",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. 1993. Building a large annotated corpus of english: The penn treebank. Computa- tional linguistics 19(2):313-330.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Context based spelling correction",
                "authors": [
                    {
                        "first": "Eric",
                        "middle": [],
                        "last": "Mays",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Fred",
                        "suffix": ""
                    },
                    {
                        "first": "Robert L",
                        "middle": [],
                        "last": "Damerau",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Mercer",
                        "suffix": ""
                    }
                ],
                "year": 1991,
                "venue": "Information Processing & Management",
                "volume": "27",
                "issue": "5",
                "pages": "517--522",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Eric Mays, Fred J Damerau, and Robert L Mercer. 1991. Context based spelling correction. Informa- tion Processing & Management 27(5):517-522.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Learning with noisy labels",
                "authors": [
                    {
                        "first": "Nagarajan",
                        "middle": [],
                        "last": "Natarajan",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Inderjit",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Dhillon",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Pradeep",
                        "suffix": ""
                    },
                    {
                        "first": "Ambuj",
                        "middle": [],
                        "last": "Ravikumar",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Tewari",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Advances in neural information processing systems",
                "volume": "",
                "issue": "",
                "pages": "1196--1204",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nagarajan Natarajan, Inderjit S Dhillon, Pradeep K Ravikumar, and Ambuj Tewari. 2013. Learning with noisy labels. In Advances in neural information pro- cessing systems. pages 1196-1204.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "The effects of ocr error on the extraction of private information",
                "authors": [
                    {
                        "first": "Kazem",
                        "middle": [],
                        "last": "Taghva",
                        "suffix": ""
                    },
                    {
                        "first": "Russell",
                        "middle": [],
                        "last": "Beckley",
                        "suffix": ""
                    },
                    {
                        "first": "Jeffrey",
                        "middle": [],
                        "last": "Coombs",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Document Analysis Systems",
                "volume": "3872",
                "issue": "",
                "pages": "348--357",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kazem Taghva, Russell Beckley, and Jeffrey Coombs. 2006. The effects of ocr error on the extraction of private information. In Document Analysis Systems. Springer, volume 3872, pages 348-357.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Noisy uyghur text normalization",
                "authors": [
                    {
                        "first": "Osman",
                        "middle": [],
                        "last": "Tursun",
                        "suffix": ""
                    },
                    {
                        "first": "Ruket",
                        "middle": [],
                        "last": "Cakici",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 3rd Workshop on Noisy User-generated Text",
                "volume": "",
                "issue": "",
                "pages": "85--93",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Osman Tursun and Ruket Cakici. 2017. Noisy uyghur text normalization. In Proceedings of the 3rd Work- shop on Noisy User-generated Text. pages 85-93.",
                "links": null
            }
        },
        "ref_entries": {
            "TABREF0": {
                "text": "Statistics of Share Task Data",
                "type_str": "table",
                "html": null,
                "num": null,
                "content": "<table/>"
            },
            "TABREF1": {
                "text": "Error Length Distribution",
                "type_str": "table",
                "html": null,
                "num": null,
                "content": "<table/>"
            },
            "TABREF3": {
                "text": "",
                "type_str": "table",
                "html": null,
                "num": null,
                "content": "<table><tr><td>: Ablation Performance. LM: Set lan-guage model score to be 1.0; Vocabulary: Only using words in corrected training data; Prepos-sess: Disable patterns and word cleaning in begin-ning; Multi Word &amp; Trans: No word merge split and known word trans correction; Wegt. Error:</td></tr><tr><td>Removing the Error Model, all weights for correct</td></tr><tr><td>candidates are 1.0.</td></tr></table>"
            }
        }
    }
}