| { | |
| "paper_id": "S07-1034", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T15:23:14.923206Z" | |
| }, | |
| "title": "HIT-IR-WSD: A WSD System for English Lexical Sample Task", | |
| "authors": [ | |
| { | |
| "first": "Yuhang", | |
| "middle": [], | |
| "last": "Guo", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Information Retrieval Lab Harbin Institute of technology Harbin", | |
| "location": { | |
| "postCode": "150001", | |
| "country": "China" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Wanxiang", | |
| "middle": [], | |
| "last": "Che", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Information Retrieval Lab Harbin Institute of technology Harbin", | |
| "location": { | |
| "postCode": "150001", | |
| "country": "China" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Yuxuan", | |
| "middle": [], | |
| "last": "Hu", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Information Retrieval Lab Harbin Institute of technology Harbin", | |
| "location": { | |
| "postCode": "150001", | |
| "country": "China" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Wei", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Information Retrieval Lab Harbin Institute of technology Harbin", | |
| "location": { | |
| "postCode": "150001", | |
| "country": "China" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Ting", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Information Retrieval Lab Harbin Institute of technology Harbin", | |
| "location": { | |
| "postCode": "150001", | |
| "country": "China" | |
| } | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "HIT-IR-WSD is a word sense disambiguation (WSD) system developed for English lexical sample task (Task 11) of Semeval 2007 by Information Retrieval Lab, Harbin Institute of Technology. The system is based on a supervised method using an SVM classifier. Multi-resources including words in the surrounding context, the partof-speech of neighboring words, collocations and syntactic relations are used. The final micro-avg raw score achieves 81.9% on the test set, the best one among participating runs.", | |
| "pdf_parse": { | |
| "paper_id": "S07-1034", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "HIT-IR-WSD is a word sense disambiguation (WSD) system developed for English lexical sample task (Task 11) of Semeval 2007 by Information Retrieval Lab, Harbin Institute of Technology. The system is based on a supervised method using an SVM classifier. Multi-resources including words in the surrounding context, the partof-speech of neighboring words, collocations and syntactic relations are used. The final micro-avg raw score achieves 81.9% on the test set, the best one among participating runs.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "Lexical sample task is a kind of WSD evaluation task providing training and test data in which a small pre-selected set of target words is chosen and the target words are marked up. In the training data the target words' senses are given, but in the test data are not and need to be predicted by task participants.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "HIT-IR-WSD regards the lexical sample task as a classification problem, and devotes to extract effective features from the instances. We didn't use any additional training data besides the official ones the task organizers provided. Section 2 gives the architecture of this system. As the task provides correct word sense for each instance, a supervised learning approach is used. In this system, we choose Support Vector Machine (SVM) as classifier. SVM is introduced in section 3. Knowledge sources are presented in section 4. The last section discusses the experimental results and present the main conclusion of the work performed.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "HIT-IR-WSD system consists of 2 parts: feature extraction and classification. Figure 1 portrays the architecture of the system. Features are extracted from original instances and are made into digitized features to feed the SVM classifier. The classifier gets the features of training data to make a model of the target word. Then it uses the model to predict the sense of target word in the test data.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 78, | |
| "end": 86, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "The Architecture of the System", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "SVM is an effective learning algorithm to WSD (Lee and Ng, 2002) . The SVM tries to find a hyperplane with the largest margin separating the training samples into two classes. The instances in the same side of the hyperplane have the same class label. A test instance's feature decides the position where the sample is in the feature space and which side of the hyperplane it is. In this way, it leads to get a prediction. SVM could be extended to tackle multi-classes problems by using oneagainst-one or one-against-rest strategy.", | |
| "cite_spans": [ | |
| { | |
| "start": 46, | |
| "end": 64, | |
| "text": "(Lee and Ng, 2002)", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Learning Algorithm", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "In the WSD problem, input of SVM is the feature vector of the instance. Features that appear in all the training samples are arranged as a vector space. Every instance is mapped to a feature vector. If the feature of a certain dimension exists in a sample, assign this dimension 1 to this sample, else assign it 0. For example, assume the feature vector space is <x1, x2, x3, x4, x5, x6, x7> ; the instance is \"x2 x6 x5 x7\". The feature vector of this sample should be <0, 1, 0, 0, 1, 1, 1>.", | |
| "cite_spans": [ | |
| { | |
| "start": 368, | |
| "end": 391, | |
| "text": "x2, x3, x4, x5, x6, x7>", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Learning Algorithm", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "The implementation of SVM here is libsvm 1 (Chang and Lin, 2001 ) for multi-classes.", | |
| "cite_spans": [ | |
| { | |
| "start": 43, | |
| "end": 63, | |
| "text": "(Chang and Lin, 2001", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Learning Algorithm", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "We used 4 kinds of features of the target word and its context as shown in Table 1 .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 75, | |
| "end": 82, | |
| "text": "Table 1", | |
| "ref_id": "TABREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Knowledge Sources", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "Part of the original text of an example is \"\u2026 This is the <head>age</head> of new media , the era of \u2026\". ", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Knowledge Sources", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "We take the neighboring words in the context of the target word as a kind of features ignoring their exact position information, which is called bag-ofwords approach.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Words in the Surrounding Context", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "Mostly, a certain sense of a word is tend to appear in a certain kind of context, so the context words could contain some helpful information to disambiguate the sense of the target word.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Words in the Surrounding Context", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "Because there would be too many context words to be added into the feature vector space, data sparseness problem is inevitable. We need to reduce the sparseness as possible as we can. A simple way is to use the words' morphological root forms. In addition, we filter the tokens which contain no alphabet character (including punctuation symbols) and stop words. The stop words are tested separately, and only the effective ones would be added into the stop words list. All remaining words in the instance are gathered, converted to lower case and replaced by their morphological root forms. The implementation for getting the morphological root forms is WordNet (morph).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Words in the Surrounding Context", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "As mentioned above, the data sparseness is a serious problem in WSD. Besides changing tokens to their morphological root forms, part-of-speech is a good choice too. The size of POS tag set is much smaller than the size of surrounding words set. And the neighboring words' part-of-speeches also contain useful information for WSD. In this part, we use a POS tagger (Gim\u00e9nez and M\u00e1rquez, 2004) to assign POS tags to those tokens.", | |
| "cite_spans": [ | |
| { | |
| "start": 364, | |
| "end": 391, | |
| "text": "(Gim\u00e9nez and M\u00e1rquez, 2004)", | |
| "ref_id": "BIBREF2" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Part-of-Speechs of Neighboring Words", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "We get the left and right 3 words' POS tags together with their position information in the target words' sentence.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Part-of-Speechs of Neighboring Words", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "For example, the word age is to be disambiguated in the sentence of \"\u2026 This is the <head>age</head> of new media , the era of \u2026\". The features then will be added to the feature vector are \"DT_0, VBZ_0, DT_0, NN_t, IN_1, JJ_1, NNS_1\", in which _0/_1 stands for the word with current POS tag is in the left/right side of the target word. The POS tag set in use here is Penn Treebank Tagset 5 .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Part-of-Speechs of Neighboring Words", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "Different from bag-of-words, collocation feature contains the position information of the target words' neighboring words. To make this feature in the same form with the bag-of-words, we appended a symbol to each of the neighboring words' morphological root forms to mark whether this word is in the left or in the right of the target word. Like POS feature, collocation was extracted in the sentence where the target word belongs to. The window size of this feature is 5 to the left and 5 to the right of the target word, which is attained by empirical value. In this part, punctuation symbol and stop words are not removed.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Collocations", | |
| "sec_num": "4.3" | |
| }, | |
| { | |
| "text": "Take the same instance last subsection has mentioned as example. The features we extracted are \"this_0, be_0, the_0, age_t, of_1, new_1, me-dium_1\". Like POS, _0/_1 stands for the word is in the left/right side of the target word. Then the features were added to the feature vector space.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Collocations", | |
| "sec_num": "4.3" | |
| }, | |
| { | |
| "text": "Many effective context words are not in a short distance to the target word, but we shouldn't enlarge the window size too much in case of including too many noises. A solution to this problem is to use the syntactic relations of the target word and its parent head word.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Syntactic Relations", | |
| "sec_num": "4.4" | |
| }, | |
| { | |
| "text": "We use Nivre et al., (2006) 's dependency parser. In this part, we get 4 features from every instance: head word of the target word, the head word's POS, the head word's dependency relation with the target word and the relative position of the head word to the target word.", | |
| "cite_spans": [ | |
| { | |
| "start": 7, | |
| "end": 27, | |
| "text": "Nivre et al., (2006)", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Syntactic Relations", | |
| "sec_num": "4.4" | |
| }, | |
| { | |
| "text": "Still take the same instance which has been mentioned in the las subsection as example. The features we extracted are \"SYN_HEAD_is, SYN_HEADPOS_VBZ, SYN_RELATION_PRD, SYN_HEADRIGHT\", in which SYN_HEAD_is stands for is is the head word of age; SYN_HEADPOS_VBZ stands for the POS of the 5 http://www.lsi.upc.es/~nlp/SVMTool/PennTreebank.html head word is is VBZ; SYN_RELATION_PRD stands for the relationship between the head word is and target word age is PRD; and SYN_HEADRIGHT stands for the target word age is in the right side of the head word is.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Syntactic Relations", | |
| "sec_num": "4.4" | |
| }, | |
| { | |
| "text": "This English lexical sample task: Semeval 2007 task 11 6 provides two tracks of the data set for participants. The first one is from LDC and the second from web.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Set and Results", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "We took part in this evaluation in the second track. The corpus is from web. In this track the task organizers provide a training data and test data set for 20 nouns and 20 adjectives.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Set and Results", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "In order to develop our system, we divided the training data into 2 parts: training and development sets. The size of the training set is about 2 times of the development set. The development set contains 1,781 instances.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Set and Results", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "4 kinds of features were merged into 15 combinations. Here we use a vector (V) to express which features are used. The four dimensions stand for syntactic relations, POS, surrounding words and collocations, respectively. For example, 1010 means that the syntactic relations feature and the surrounding words feature are used. Table 2 , we can conclude that the surrounding words feature is the most useful kind of features. It obtains much better performance than other kinds of features individually. In other words, without it, the performance drops a lot. Among these features, syntactic relations feature is the most unstable one (the improvement with it is unstable), partly because the performance of the dependency parser is not good enough. As the ones with the vector 0111 and 1111 get the best perfor-mance, we chose all of these kinds of features for our final system.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 326, | |
| "end": 333, | |
| "text": "Table 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Data Set and Results", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "V Precision V", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Set and Results", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "A trade-off parameter C in SVM is tuned, and the result is shown in Figure 2 . We have also tried 4 types of kernels of the SVM classifier (parameters are set by default). The experimental results show that the linear kernel is the most effective as Table 3 shows. Accuracy 82.9% 68.3% 68.3% 68.3% Table 3 : Accuracy with different kernel function types", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 68, | |
| "end": 76, | |
| "text": "Figure 2", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 250, | |
| "end": 257, | |
| "text": "Table 3", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 298, | |
| "end": 305, | |
| "text": "Table 3", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Data Set and Results", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Another experiment (as shown in Figure 3 ) also validate that the linear kernel is the most suitable one. We tried using polynomial function. Unlike the parameters set by default above (g=1/k, d=3), here we set its Gama parameter as 1 (g=1) but other parameters excepting degree parameter are still set by default. The performance gets better when the degree parameter is tuned towards 1. That means the closer the kernel function to linear function the better the system performs. In order to get the relation between the system performance and the size of training data, we made several groups of training-test data set from the training data the organizers provided. Each of them has the same test data but different size of training data which are 2, 3, 4 and 5 times of the test data respectively. Figure 4 shows the performance curve with the training data size. Indicated in Figure 4 , the accuracy increases as the size of training data enlarge, from which we can infer that we could raise the performance by using more training data potentially. Feature extraction is the most time-consuming part of the system, especially POS tagging and parsing which take 2 hours approximately on the training and test data. The classification part (using libsvm) takes no more than 5 minutes on the training and test data. We did our experiment on a PC with 2.0GHz CPU and 960 MB system memory.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 32, | |
| "end": 40, | |
| "text": "Figure 3", | |
| "ref_id": "FIGREF2" | |
| }, | |
| { | |
| "start": 803, | |
| "end": 811, | |
| "text": "Figure 4", | |
| "ref_id": "FIGREF3" | |
| }, | |
| { | |
| "start": 882, | |
| "end": 890, | |
| "text": "Figure 4", | |
| "ref_id": "FIGREF3" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Data Set and Results", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Our official result of HIT-IR-WSD is: microavg raw score 81.9% on the test set, the top one among the participating runs.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data Set and Results", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "http://w3.msi.vxu.se/~nivre/research/MaltParser.html", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "http://nlp.cs.swarthmore.edu/semeval/tasks/task11/descript ion.shtml", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "We gratefully acknowledge the support for this study provided by the National Natural Science Foundation of China (NSFC) via grant 60435020, 60575042, 60575042 and 60675034.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Acknowledgement", | |
| "sec_num": null | |
| } | |
| ], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "An empirical evaluation of knowledge sources and learning algorithms for word sense disambiguation", | |
| "authors": [ | |
| { | |
| "first": "Y", | |
| "middle": [ | |
| "K" | |
| ], | |
| "last": "Lee", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [ | |
| "T" | |
| ], | |
| "last": "Ng", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2002, | |
| "venue": "Proceedings of EMNLP02", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "41--48", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Lee, Y. K., and Ng, H. T. 2002. An empirical evaluation of knowledge sources and learning algorithms for word sense disambiguation. In Proceedings of EMNLP02, 41-48.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "LIBSVM: a library for support vector machines", | |
| "authors": [ | |
| { | |
| "first": "Chih-Chung", | |
| "middle": [], | |
| "last": "Chang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Chih-Jen", | |
| "middle": [], | |
| "last": "Lin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2001, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Chih-Chung Chang and Chih-Jen Lin, 2001. LIBSVM: a library for support vector machines.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "SVMTool: A general POS tagger generator based on Support Vector Machines", | |
| "authors": [ | |
| { | |
| "first": "Jes\u00fas", | |
| "middle": [], | |
| "last": "Gim\u00e9nez", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Llu\u00eds", | |
| "middle": [], | |
| "last": "M\u00e1rquez", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "Proceedings of the 4th International Conference on Language Resources and Evaluation (LREC'04)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Jes\u00fas Gim\u00e9nez and Llu\u00eds M\u00e1rquez. 2004. SVMTool: A general POS tagger generator based on Support Vec- tor Machines. Proceedings of the 4th International Conference on Language Resources and Evaluation (LREC'04). Lisbon, Portugal.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Labeled Pseudo-Projective Dependency Parsing with Support Vector Machines", | |
| "authors": [ | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Nivre", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Hall", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Nilsson", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "G", | |
| "middle": [], | |
| "last": "Eryigit", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Marinov", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Proceedings of the Tenth Conference on Computational Natural Language Learning (CoNLL)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nivre, J., Hall, J., Nilsson, J., Eryigit, G. and Marinov, S. 2006. Labeled Pseudo-Projective Dependency Pars- ing with Support Vector Machines. In Proceedings of the Tenth Conference on Computational Natural Language Learning (CoNLL).", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "type_str": "figure", | |
| "uris": null, | |
| "num": null, | |
| "text": "The architecture of HIT-IR-WSD" | |
| }, | |
| "FIGREF1": { | |
| "type_str": "figure", | |
| "uris": null, | |
| "num": null, | |
| "text": "Figure 2: Accuracy with different C parameters Kernel Function Type Linear Polynomial RBF Sigmoid" | |
| }, | |
| "FIGREF2": { | |
| "type_str": "figure", | |
| "uris": null, | |
| "num": null, | |
| "text": "Accuracy with different degree in polynomial function" | |
| }, | |
| "FIGREF3": { | |
| "type_str": "figure", | |
| "uris": null, | |
| "num": null, | |
| "text": "Accuracy's trend with the training data size" | |
| }, | |
| "TABREF0": { | |
| "html": null, | |
| "text": "Features the system extracted The next 4 subsections elaborate these features.", | |
| "type_str": "table", | |
| "num": null, | |
| "content": "<table><tr><td/><td>Extraction Tools</td><td>Example</td></tr><tr><td>Surrounding</td><td>WordNet</td><td>\u2026, this, be, age, new,</td></tr><tr><td>words</td><td>(morph) 2</td><td>medium, ,, era, \u2026</td></tr><tr><td>Part-of-speech</td><td>SVMTool 3</td><td>DT_0, VBZ_0, DT_0, NN_t, IN_1, JJ_1, NNS_1</td></tr></table>" | |
| } | |
| } | |
| } | |
| } |