| { |
| "paper_id": "N06-1016", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T14:45:54.050532Z" |
| }, |
| "title": "An Empirical Study of the Behavior of Active Learning for Word Sense Disambiguation", |
| "authors": [ |
| { |
| "first": "Jinying", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "University of Pennsylvania Philadelphia", |
| "location": { |
| "postCode": "19104", |
| "region": "PA" |
| } |
| }, |
| "email": "jinying@cis.upenn.edu" |
| }, |
| { |
| "first": "Andrew", |
| "middle": [], |
| "last": "Schein", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "University of Pennsylvania Philadelphia", |
| "location": { |
| "postCode": "19104", |
| "region": "PA" |
| } |
| }, |
| "email": "" |
| }, |
| { |
| "first": "Lyle", |
| "middle": [], |
| "last": "Ungar", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "University of Pennsylvania Philadelphia", |
| "location": { |
| "postCode": "19104", |
| "region": "PA" |
| } |
| }, |
| "email": "ungar@cis.upenn.edu" |
| }, |
| { |
| "first": "Martha", |
| "middle": [], |
| "last": "Palmer", |
| "suffix": "", |
| "affiliation": {}, |
| "email": "martha.palmer@colorado.edu" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "This paper shows that two uncertaintybased active learning methods, combined with a maximum entropy model, work well on learning English verb senses. Data analysis on the learning process, based on both instance and feature levels, suggests that a careful treatment of feature extraction is important for the active learning to be useful for WSD. The overfitting phenomena that occurred during the active learning process are identified as classic overfitting in machine learning based on the data analysis.", |
| "pdf_parse": { |
| "paper_id": "N06-1016", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "This paper shows that two uncertaintybased active learning methods, combined with a maximum entropy model, work well on learning English verb senses. Data analysis on the learning process, based on both instance and feature levels, suggests that a careful treatment of feature extraction is important for the active learning to be useful for WSD. The overfitting phenomena that occurred during the active learning process are identified as classic overfitting in machine learning based on the data analysis.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "Corpus-based methods for word sense disambiguation (WSD) have gained popularity in recent years. As evidenced by the SENSEVAL exercises (http://www.senseval.org), machine learning models supervised by sense-tagged training corpora tend to perform better on the lexical sample tasks than unsupervised methods. However, WSD tasks typically have very limited amounts of training data due to the fact that creating large-scale high-quality sense-tagged corpora is difficult and time-consuming. Therefore, the lack of sufficient labeled training data has become a major hurdle to improving the performance of supervised WSD.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "A promising method for solving this problem could be the use of active learning. Researchers use active learning methods to minimize the labeling of examples by human annotators. A decrease in overall labeling occurs because active learners (the machine learning models used in active learning) pick more informative examples for the target word (a word whose senses need to be learned) than those that would be picked randomly. Active learning requires human labeling of the newly selected training data to ensure high quality.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "We focus here on pool-based active learning where there is an abundant supply of unlabeled data, but where the labeling process is expensive. In NLP problems such as text classification (Lewis and Gale, 1994; McCallum and Nigam, 1998) , statistical parsing (Tang et al., 2002) , information extraction (Thompson et al., 1999) , and named entity recognition (Shen et al., 2004) , pool-based active learning has produced promising results.", |
| "cite_spans": [ |
| { |
| "start": 186, |
| "end": 208, |
| "text": "(Lewis and Gale, 1994;", |
| "ref_id": "BIBREF5" |
| }, |
| { |
| "start": 209, |
| "end": 234, |
| "text": "McCallum and Nigam, 1998)", |
| "ref_id": "BIBREF8" |
| }, |
| { |
| "start": 257, |
| "end": 276, |
| "text": "(Tang et al., 2002)", |
| "ref_id": "BIBREF12" |
| }, |
| { |
| "start": 302, |
| "end": 325, |
| "text": "(Thompson et al., 1999)", |
| "ref_id": "BIBREF13" |
| }, |
| { |
| "start": 357, |
| "end": 376, |
| "text": "(Shen et al., 2004)", |
| "ref_id": "BIBREF11" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "This paper presents our experiments in applying two active learning methods, a min-margin based method and a Shannon-entropy based one, to the task of the disambiguation of English verb senses. The contribution of our work is not only in demonstrating that these methods work well for the active learning of coarse-grained verb senses, but also analyzing the behavior of the active learning process on two levels: the instance level and the feature level. The analysis suggests that a careful treatment of feature design and feature generation is important for a successful application of active learning to WSD. We also accounted for the overfitting phenomena that occurred in the learning process based on our data analysis.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "The rest of the paper is organized as follows. In Section 2, we introduce two uncertainty sampling methods used in our active learning experiments and review related work in using active learning for WSD. We then present our active learning experiments on coarse-grained English verb senses in Section 3 and analyze the active learning process in Section 4. Section 5 presents conclusions of our study.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "The methods evaluated in this work fit into a common framework described by Algorithm 1 (see Table 1 ). The key difference between alternative active learning methods is how they assess the value of labeling individual examples, i.e., the methods they use for ranking and selecting the candidate examples for labeling. The framework is wide open to the type of ranking rule employed. Usually, the ranking rule incorporates the model trained on the currently labeled data. This is the reason for the requirement of a partial training set when the algorithm begins. In our experiments we look at two variants of the uncertainty sampling heuristic: entropy sampling and margin sampling. Uncertainty sampling is a term invented by Lewis and Gale (Lewis and Gale, 1994) to describe a heuristic where a probabilistic classifier picks examples for which the model's current predictions are least certain. The intuitive justification for this approach is that regions where the model is uncertain indicate a decision boundary, and clarifying the position of decision boundaries is the goal of learning classifiers. Schein (2005) demonstrates the two methods run quickly and compete favorably against alternatives when combined with the logistic regression classifier.", |
| "cite_spans": [ |
| { |
| "start": 727, |
| "end": 764, |
| "text": "Lewis and Gale (Lewis and Gale, 1994)", |
| "ref_id": "BIBREF5" |
| }, |
| { |
| "start": 1107, |
| "end": 1120, |
| "text": "Schein (2005)", |
| "ref_id": "BIBREF10" |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 93, |
| "end": 100, |
| "text": "Table 1", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Active Learning Algorithms", |
| "sec_num": "2" |
| }, |
| { |
| "text": "A key question is how to measure uncertainty. Different methods of measuring uncertainty will lead to different variants of uncertainty sampling. We will look at two such measures. As a convenient notation we use q (a vector) to represent the trained model's predictions, with c q equal to the predicted probability of class c . One method is to pick the example whose prediction vector q displays the greatest Shannon entropy:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Entropy Sampling", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "\u2211 \u2212 c c c q q log", |
| "eq_num": "(1)" |
| } |
| ], |
| "section": "Entropy Sampling", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "Such a rule means ranking candidate examples in Algorithm 1 by Equation 1.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Entropy Sampling", |
| "sec_num": "2.1" |
| }, |
| { |
| "text": "An alternative method picks the example with the smallest margin: the difference between the largest two values in the vector q (Abe and Mamitsuka, 1998) . In other words, if c and ' c are the two most likely categories for example n", |
| "cite_spans": [ |
| { |
| "start": 128, |
| "end": 153, |
| "text": "(Abe and Mamitsuka, 1998)", |
| "ref_id": "BIBREF0" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Margin Sampling", |
| "sec_num": "2.2" |
| }, |
| { |
| "text": "x , the margin is measured as follows:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Margin Sampling", |
| "sec_num": "2.2" |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": ") | ' Pr( ) | Pr( n n n x c x c M \u2212 =", |
| "eq_num": "(2)" |
| } |
| ], |
| "section": "Margin Sampling", |
| "sec_num": "2.2" |
| }, |
| { |
| "text": "In this case Algorithm 1 would rank examples by increasing values of margin, with the smallest value at the top of the ranking.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Margin Sampling", |
| "sec_num": "2.2" |
| }, |
| { |
| "text": "Using either method of uncertainty sampling, the computational cost of picking an example from T candidates is: O(TD) where D is the number of model parameters.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Margin Sampling", |
| "sec_num": "2.2" |
| }, |
| { |
| "text": "To our best knowledge, there have been very few attempts to apply active learning to WSD in the literature (Fujii and Inui, 1999; Chklovski and Mihalcea, 2002; Dang, 2004) . Fujii and Inui (1999) developed an example sampling method for their example-based WSD system in the active learning of verb senses in a pool-based setting. Unlike the uncertainty sampling methods (such as the two methods we used), their method did not select examples for which the system had the minimal certainty. Rather, it selected the examples such that after training using those examples the system would be most certain about its predictions on the rest of the unlabeled examples in the next iteration. This sample selection criterion was enforced by calculating a training utility function. The method performed well on the active learning of Japanese verb senses. However, the efficient computation of the training utility function relied on the nature of the example-based learning method, which made their example sampling method difficult to export to other types of machine learning models.", |
| "cite_spans": [ |
| { |
| "start": 107, |
| "end": 129, |
| "text": "(Fujii and Inui, 1999;", |
| "ref_id": null |
| }, |
| { |
| "start": 130, |
| "end": 159, |
| "text": "Chklovski and Mihalcea, 2002;", |
| "ref_id": "BIBREF2" |
| }, |
| { |
| "start": 160, |
| "end": 171, |
| "text": "Dang, 2004)", |
| "ref_id": "BIBREF3" |
| }, |
| { |
| "start": 174, |
| "end": 195, |
| "text": "Fujii and Inui (1999)", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Related Work", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "Open Mind Word Expert (Chklovski and Mihalcea, 2002) was a real application of active learning for WSD. It collected sense-annotated examples from the general public through the Web to create the training data for the SENSEVAL-3 lexical sample tasks. The system used the disagreement of two classifiers (which employed different sets of features) on sense labels to evaluate the difficulty of the unlabeled examples and ask the web users to tag the difficult examples it selected. There was no formal evaluation for this active learning system. Dang (2004) used an uncertainty sampling method to get additional training data for her WSD system. At each iteration the system selected a small set of examples for which it had the lowest confidence and asked the human annotators to tag these examples. The experimental results on 5 English verbs with fine-grained senses (from WordNet 1.7) were a little surprising in that active learning performed no better than random sampling. The proposed explanation was that the quality of the manually sense-tagged data was limited by an inconsistent or unclear sense inventory for the fine-grained senses.", |
| "cite_spans": [ |
| { |
| "start": 22, |
| "end": 52, |
| "text": "(Chklovski and Mihalcea, 2002)", |
| "ref_id": "BIBREF2" |
| }, |
| { |
| "start": 545, |
| "end": 556, |
| "text": "Dang (2004)", |
| "ref_id": "BIBREF3" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Related Work", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "We experimented with the two uncertainty sampling methods on 5 English verbs that had coarse-grained senses (see Table 2 ), as described below. By using coarse-grained senses, we limit the impact of noisy data due to unclear sense boundaries and therefore can get a clearer observation of the effects of the active learning methods themselves. Table 2 . The number of senses, the baseline accuracy, the number of instances used for active learning and for held-out evaluation for each verb", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 113, |
| "end": 120, |
| "text": "Table 2", |
| "ref_id": null |
| }, |
| { |
| "start": 344, |
| "end": 351, |
| "text": "Table 2", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Active Learning Experiments 3.1 Experimental Setting", |
| "sec_num": "3" |
| }, |
| { |
| "text": "The coarse-grained senses are produced by grouping together the original WordNet senses using syntactic and semantic criteria (Palmer et al., 2006) . Double-blind tagging is applied to 50 instances of the target word. If the ITA < 90%, the sense entry is revised by adding examples and explanations of distinguishing criteria. Table 2 summarizes the statistics of the data. The baseline accuracy was computed by using the \"most frequent sense\" heuristic to assign sense labels to verb instances (examples). The data used in active learning (Column 4 in Table 2 ) include two parts: an initial labeled training set and a pool of unlabeled training data. We experimented with sizes 20, 50 and 100 for the initial training set. The pool of unlabeled data had actually been annotated in advance, as in most pool-based active learning experiments. Each time an example was selected from the pool by the active learner, its label was returned to the learner. This simulates the process of asking human annotators to tag the selected unlabeled example at each time. The advantage of using such a simulation is that we can experiment with different settings (different sizes of the initial training set and different sampling methods).", |
| "cite_spans": [ |
| { |
| "start": 126, |
| "end": 147, |
| "text": "(Palmer et al., 2006)", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 327, |
| "end": 334, |
| "text": "Table 2", |
| "ref_id": null |
| }, |
| { |
| "start": 553, |
| "end": 560, |
| "text": "Table 2", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Active Learning Experiments 3.1 Experimental Setting", |
| "sec_num": "3" |
| }, |
| { |
| "text": "The data sets used for active learning and for held-out evaluation were randomly sampled from a large data pool for each round of the active learning experiment. We ran ten rounds of the experiments for each verb and averaged the learning curves for the ten rounds.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Active Learning Experiments 3.1 Experimental Setting", |
| "sec_num": "3" |
| }, |
| { |
| "text": "In the experiments, we used random sampling (picking up an unlabeled example randomly at each time) as a lower bound. Another control (ultimate-maxent) was the learner's performance on the test set when it was trained on a set of labeled data that were randomly sampled from a large data pool and equaled the amount of data used in the whole active learning process (e.g., 400 training data for the verb add).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Active Learning Experiments 3.1 Experimental Setting", |
| "sec_num": "3" |
| }, |
| { |
| "text": "The machine learning model we used for active learning was a regularized maximum entropy (MaxEnt) model (McCallum, 2002) . The features used for disambiguating the verb senses included topical, collocation, syntactic (e.g., the subject, object, and preposition phrases taken by a target verb), and semantic (e.g., the WordNet synsets and hypernyms of the head nouns of a verb's NP arguments) features (Chen and Palmer, 2005) .", |
| "cite_spans": [ |
| { |
| "start": 104, |
| "end": 120, |
| "text": "(McCallum, 2002)", |
| "ref_id": "BIBREF7" |
| }, |
| { |
| "start": 401, |
| "end": 424, |
| "text": "(Chen and Palmer, 2005)", |
| "ref_id": "BIBREF1" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Active Learning Experiments 3.1 Experimental Setting", |
| "sec_num": "3" |
| }, |
| { |
| "text": "Due to space limits, Figure 1 only shows the learning curves for 4 verbs do, feel, see, and work (size of the initial training set = 20). The curve for the verb add is similar to that for feel. These curves clearly show that the two uncertainty sampling methods, the entropy-based (called entropy-maxent in the figure) and the margin-based (called min_margin-maxent), work very well for active learning of the senses of these verbs.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 21, |
| "end": 29, |
| "text": "Figure 1", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Experimental Results", |
| "sec_num": "3.2" |
| }, |
| { |
| "text": "Both methods outperformed the random sampling method in that they reached the upperbound accuracy earlier and had smoother learning curves. For the four verbs add, do, feel and see, their learning curves reached the upper bound at about 200~300 iterations, which means 1/2 or 1/3 of the annotation effort can be saved for these verbs by using active learning, while still achieving the same level of performance as supervised WSD without using active learning. Given the largescale annotation effort currently underway in the OntoNotes project (Hovy et al., 2006) , this could provide considerable savings in annotation effort and speed up the process of providing sufficient data for a large vocabulary. The OntoNotes project has now provided coarse-grained entries for over 350 verbs, with corresponding double-blind annotation and adjudication in progress. As this adjudicated data becomes available, we will be able to train our system accordingly. Preliminary results for 22 of these coarse-grained verbs (with an average grouping polysemy of 4.5) give us an average accuracy of 86.3%. This will also provide opportunities for more experiments with active learning, where there are enough instances. Active learning could also be beneficial in porting these supervised taggers to new genres with different sense distributions.", |
| "cite_spans": [ |
| { |
| "start": 544, |
| "end": 563, |
| "text": "(Hovy et al., 2006)", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Figure 1 Active learning for four verbs", |
| "sec_num": null |
| }, |
| { |
| "text": "We also experimented with different sizes of the initial training set (20, 50 and 100) and found no significant differences in the performance at different settings. That means, for these 5 verbs, only 20 labeled training instances will be enough to initiate an efficient active learning process.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Figure 1 Active learning for four verbs", |
| "sec_num": null |
| }, |
| { |
| "text": "From Figure 1 , we can see that the two uncertainty sampling methods generally perform equally well except that for the verb do, the minmargin method is slightly better than the entropy method at the beginning of active learning. This may not be so surprising, considering that the two methods are equal for two-class classification tasks (see Equations 1 and 2 for their definition) and the verbs used in our experiments have coarse-grained senses and often have only 2 or 3 major senses.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 5, |
| "end": 13, |
| "text": "Figure 1", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Figure 1 Active learning for four verbs", |
| "sec_num": null |
| }, |
| { |
| "text": "An interesting phenomenon observed from these learning curves is that for the two verbs add and feel, the active learner reached the upper bound very soon (at about 100 iterations) and then even breached the upper bound. However, when the training set was extended, the learner's performance dropped and eventually returned to the same level of the upper bound. We discuss the phenomenon below.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Figure 1 Active learning for four verbs", |
| "sec_num": null |
| }, |
| { |
| "text": "In addition to verifying the usefulness of active learning for WSD, we are also interested in a deeper analysis of the learning process. For example, why does the active learner's performance drop sometimes during the learning process? What are the characteristics of beneficial features that help to boost the learner's accuracy? How do we account for the overfitting phenomena that occurred during the active learning for the verbs add and feel? We analyzed the effect of both instances and features throughout the course of active learning using min-margin-based sampling.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Analysis of the Learning Process", |
| "sec_num": "4" |
| }, |
| { |
| "text": "Intuitively, if the learner's performance drops after a new example is added to the training set, it is likely that something has gone wrong with the new example. To find out such bad examples, we define a measure credit_inst for instance i as:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Instance-level Analysis", |
| "sec_num": "4.1" |
| }, |
| { |
| "text": "\u2211\u2211 = + = \u2212 m r l l n l Acc Acc l i sel m 1 1 1 ) ( ) , ( 1 (3)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Instance-level Analysis", |
| "sec_num": "4.1" |
| }, |
| { |
| "text": "where Acc l and Acc l+1 are the classification accuracies of the active learner at the lth and (l+1)th iterations. n is the total number of iterations of active learning and m is the number of rounds of active learning (m=10 in our case).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Instance-level Analysis", |
| "sec_num": "4.1" |
| }, |
| { |
| "text": "is 1 iff instance i is selected by the active learner at the lth iteration and is 0 if otherwise.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": ") , ( l i sel", |
| "sec_num": null |
| }, |
| { |
| "text": "An example is a bad example if and only if it satisfies the following conditions: a) its credit_inst value is negative b) it increases the learner's performance, if it does, less often than it decreases the performance in the 10 rounds. We ranked the bad examples by their credit_inst values and their frequency of decreasing the learner's performance in the 10 rounds. Table 3 shows the top five bad examples for feel and work. There are several reasons why the bad examples may hurt the learner's performance. Column 3 of Table 3 proposes reasons for many of our bad examples. We categorized these reasons into three major types.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 370, |
| "end": 377, |
| "text": "Table 3", |
| "ref_id": null |
| }, |
| { |
| "start": 524, |
| "end": 531, |
| "text": "Table 3", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": ") , ( l i sel", |
| "sec_num": null |
| }, |
| { |
| "text": "I. The major senses of a target verb depend heavily on the semantic categories of its NP arguments but WordNet sometimes fails to provide the appropriate semantic categories (features) for the head nouns of these NP arguments. For example, feel in the board apparently felt no pressure has Sense 1 (experience). In Sense 1, feel typically takes an animate subject. However, board, the head word of the verb's subject in the above sentence has no animate meanings defined in WordNet. Even worse, the major meaning of board, i.e., artifact, is typical for the subject of feel in Sense 2 (touch, grope). Similar semantic type mismatches hold for the last four bad examples of the verb work in Table 3 .", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 690, |
| "end": 697, |
| "text": "Table 3", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": ") , ( l i sel", |
| "sec_num": null |
| }, |
| { |
| "text": "II. The contexts of the target verb are difficult for our feature exaction module to analyze. For example, the antecedent for the pronoun subject they in the first example of work in Table 3 should be ringers, an agent subject that is typical for Sense 1 (exert oneself in an activity). However, the feature exaction module found the wrong antecedent changes that is an unlikely fit for the intended verb sense. In the fourth example for feel, the feature extraction module cannot handle the expletive \"it\" (a dummy subject) in \"it was felt that\", therefore, it cannot identify the typical syntactic pattern for Sense 3 (find, conclude), i.e., subject+feel+relative clause.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 183, |
| "end": 190, |
| "text": "Table 3", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": ") , ( l i sel", |
| "sec_num": null |
| }, |
| { |
| "text": "III. Sometimes, deep semantic and discourse analyses are needed to get the correct meaning of the target verb. For example, in the third example of feel, \"\u2026, he or she feels age creeping up\", it is difficult to tell whether the verb has Sense 1 (experience) or Sense 3 (find) without an understanding of the meaning of the relative clause and without looking at a broader discourse context. The syntactic pattern identified by our feature extraction module, subject+feel+relative clause, favors Sense 3 (find), which leads to an inaccurate interpretation for this case.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": ") , ( l i sel", |
| "sec_num": null |
| }, |
| { |
| "text": "Recall that the motivation behind uncertainty samplers is to find examples near decision boundaries and use them to clarify the position of these boundaries. Active learning often does find informative examples, either ones from the less common senses or ones close to the boundary between the different senses. However, active learning also identifies example sentences that are difficult to analyze. The failure of our feature extraction module, the lack of appropriate semantic categories for certain NP arguments in WordNet, the lack of deep analysis (semantic and discourse analysis) of the context of the target verb can all Table 3 Data analysis of the top-ranked bad examples found for two verbs produce misleading features. Therefore, in order to make active learning useful for its applications, both identifying difficult examples and getting good features for these examples are equally important. In other words, a careful treatment of feature design and feature generation is necessary for a successful application of active learning.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 631, |
| "end": 638, |
| "text": "Table 3", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": ") , ( l i sel", |
| "sec_num": null |
| }, |
| { |
| "text": "There is a positive side to identifying such \"bad\" examples; one can have human annotators look at the features generated from the sentences (as we did above), and use this to improve the data or the classifier. Note that this is exactly what we did above: the identification of bad sentences was automatic, and they could then be reannotated or removed from the training set or the feature extraction module needs to be refined to generate informative features for these sentences.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": ") , ( l i sel", |
| "sec_num": null |
| }, |
| { |
| "text": "Not all sentences have obvious interpretations; hence the two question marks in Table 3 . An example can be bad for many reasons: conflicting features (indicative of different senses), misleading features (indicative of non-intended senses), or just containing random features that are incorrectly incorporated into the model. We will return to this point in our discussion of the overfitting phenomena for active learning in Section 4.3.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 80, |
| "end": 87, |
| "text": "Table 3", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": ") , ( l i sel", |
| "sec_num": null |
| }, |
| { |
| "text": "The purpose of our feature-level analysis is to identify informative features for verb senses. The learning curve of the active learner may provide some clues. The basic idea is, if the learner's performance increases after adding a new example, it is likely that the good example contains good features that contribute to the clarification of sense boundaries. However, the feature-level analysis is much less straightforward than the instance-level analysis since we cannot simply say the features that are active (present) in this good example are all good. Rather, an example often contains both good and bad features, and many other features that are somehow neutral or uninformative. The interaction or balance between these features determines the final outcome. On the other hand, a statistics based analysis may help us to find features that tend to be good or bad. For this analysis, we define a measure credit_feat for feature i as:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Feature-level Analysis", |
| "sec_num": "4.2" |
| }, |
| { |
| "text": "Proposed reasons for bad examples Senses Some days the coaches make you feel as though you are part of a large herd of animals .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "feel", |
| "sec_num": null |
| }, |
| { |
| "text": "? S1: experience And , with no other offers on the table , the board apparently felt no pressure to act on it. subject: board, no \"animate\" meaning in WordNet S1: experience Sometimes a burst of aggressiveness will sweep over a man --or his wife --because he or she feels age creeping up.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "feel", |
| "sec_num": null |
| }, |
| { |
| "text": "syntactic pattern: sbj+feel+relative clause headed by that, a typical pattern for Sense 3 (find) rather than Sense 1 (experience) S1: experience At this stage it was felt I was perhaps more pertinent as chief. executive .", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "feel", |
| "sec_num": null |
| }, |
| { |
| "text": "syntactic pattern: sbj+feel+relative clause, typical for Sense 3 (find) but has not been detected by the feature exaction module S3: find, conclude I felt better Tuesday evening when I woke up. ? S1: experience Work When their changes are completed, and after they have worked up a sweat, ringers often \u2026\u2026 subject: they, the feature exaction module found the wrong antecedent (changes rather than ringers) for they S1: exert oneself in an activity Others grab books, records , photo albums , sofas and chairs , working frantically in the fear that an aftershock will jolt the house again . subject: others (means people here), no definition in WordNet S1: exert oneself in an activity Security Pacific 's factoring business works with companies in the apparel, textile and food industries \u2026 subject: business, no \"animate\" meaning in WordNet S1: exert oneself in an activity \u2026 ; blacks could work there , but they had to leave at night . subject: blacks, no \"animate\" meaning in WordNet S1: exert oneself in an activity \u2026 has been replaced by alginates (gelatin-like material ) that work quickly and accurately and with least discomfort to a child . is 1 iff feature i is active in the example selected by the active learner at the lth iteration and is 0 if otherwise. act l is the total number of active features in the example selected at the lth iteration. n and m have the same definition as in Equation 3.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "feel", |
| "sec_num": null |
| }, |
| { |
| "text": "A feature is regarded as good if its credit_feat value is positive. We ranked the good features by their credit_feat values. By looking at the topranked good features for the verb work (due to space limitations, we omit the table data), we identify two types of typically good features.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "feel", |
| "sec_num": null |
| }, |
| { |
| "text": "The first type of good feature occurs frequently in the data and has a frequency distribution over the senses similar to the data distribution over the senses. Such features include those denoting that the target verb takes a subject (subj), is not used in a passive mode (morph_normal), does not take a direct object (intransitive), occurs in present tense (word_work, pos_vb, word_works, pos_vbz) , and semantic features denoting an abstract subject (subjsyn_16993 1 ) or an entity subject (subjsyn_ 1742), etc. We call such features background features. They help the machine learning model learn the appropriate sense distribution of the data. In other words, a learning model only using such features will be equal to the \"most frequent sense\" heuristic used in WSD.", |
| "cite_spans": [ |
| { |
| "start": 358, |
| "end": 398, |
| "text": "(word_work, pos_vb, word_works, pos_vbz)", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "feel", |
| "sec_num": null |
| }, |
| { |
| "text": "Another type of good feature occurs less frequently and has a frequency distribution over senses that mismatches with the sense distribution of the data. Such features include those denoting that the target verb takes an inanimate subject (subj_it), takes a particle out (prt_out), is followed directly by the word out (word+1_out), or occurs at the end of the sentence. Such features are indicative of less frequent verb senses that still occur fairly frequently in the data. For example, taking an inanimate subject (subj_it) is a strong clue for Sense 2 (perform, function, behave) of the verb work. Occurring at the end of the sentence is also indicative of Sense 2 since when work is used in Sense 1 (exert oneself in an activity), it tends to take adjuncts to modify the activity as in He is working hard to bring up his grade.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "feel", |
| "sec_num": null |
| }, |
| { |
| "text": "There are some features that don't fall into the above two categories, such as the topical feature tp_know and the collocation feature pos-2_nn. There are no obvious reasons why they are good for the learning process, although it is possible that the combination of two or more such features could make a clear sense distinction. However, this hypothesis cannot be verified by our current statistics-based analysis. It is also worth noting that our current feature analysis is post-experimental (i.e., based on the results). In the future, we will try automatic feature selection methods that can be used in the training phase to select useful features and/or their combinations.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "feel", |
| "sec_num": null |
| }, |
| { |
| "text": "We have similar results for the feature analysis of the other four verbs.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "feel", |
| "sec_num": null |
| }, |
| { |
| "text": "Recall that in the instance-level analysis in Section 4.1, we found that some examples hurt the learning performance during active learning but for no obvious reasons (the two examples marked by ? in Table 3 ). We found that these two examples occurred in the overfitting region for feel. By looking at the bad examples (using the same definition for bad example as in Section 4.1) that occurred in the overfitting region for both feel and add, we identified two major properties of these examples. First, most of them occurred only once as bad examples (19 out 23 for add and 40 out of 63 for feel). Second, many of the examples had no obvious reasons for their badness.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 200, |
| "end": 207, |
| "text": "Table 3", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Account for the Overfitting Phenomena", |
| "sec_num": "4.3" |
| }, |
| { |
| "text": "Based on the above observations, we believe that the overfitting phenomena that occurred for the two verbs during active learning is typical of classic overfitting, which is consistent with a \"death by a thousand mosquito bites\" of rare bad features, and consistent with there often being (to mix a metaphor) no \"smoking gun\" of a bad feature/instance that is added in, especially in the region far away from the starting point of active learning.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Account for the Overfitting Phenomena", |
| "sec_num": "4.3" |
| }, |
| { |
| "text": "We have shown that active learning can lead to substantial reductions (often by half) in the number of observations that need to be labeled to achieve a given accuracy in word sense disambiguation, compared to labeling randomly selected instances. In a follow-up experiment, we also compared a larger number of different active learning methods.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusions", |
| "sec_num": "5" |
| }, |
| { |
| "text": "The results suggest that for tasks like word sense disambiguation where maximum entropy methods are used as the base learning models, the minimum margin active criterion for active learning gives superior results to more comprehensive competitors including bagging and two variants of query by committee (Schein, 2005) . By also taking into account the high running efficiency of the min-margin method, it is a very promising active learning method for WSD.", |
| "cite_spans": [ |
| { |
| "start": 304, |
| "end": 318, |
| "text": "(Schein, 2005)", |
| "ref_id": "BIBREF10" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusions", |
| "sec_num": "5" |
| }, |
| { |
| "text": "We did an analysis on the learning process on two levels: instance-level and feature-level. The analysis suggests that a careful treatment of feature design and feature generation is very important for the active learner to take advantage of the difficult examples it finds during the learning process. The feature-level analysis identifies some characteristics of good features. It is worth noting that the good features identified are not particularly tied to active learning, and could also be obtained by a more standard feature selection method rather than by looking at how the features provide benefits as they are added in.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusions", |
| "sec_num": "5" |
| }, |
| { |
| "text": "For a couple of the verbs examined, we found that active learning gives higher prediction accuracy midway through the training than one gets after training on the entire corpus. Analysis suggests that this is not due to bad examples being added to the training set. It appears that the widely used maximum entropy model with Gaussian priors is overfitting: the model by including too many features and thus fitting noise as well as signal. Using different strengths of the Gaussian prior does not solve the problem. If a very strong prior is used, then poorer accuracy is obtained. We believe that using appropriate feature selection would cause the phenomenon to vanish.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusions", |
| "sec_num": "5" |
| }, |
| { |
| "text": "Those features are from the WordNet. The numbers are WordNet ids of synsets and hypernyms.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [ |
| { |
| "text": "This work was supported by National Science Foundation Grant NSF-0415923, Word Sense Disambiguation, the DTO-AQUAINT NBCHC-040036 grant under the University of Illinois subcontract to University of Pennsylvania 2003-07911-01 and the GALE program of the Defense Advanced Research Projects Agency, Contract No. HR0011-06-C-0022. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation, the DTO, or DARPA.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Acknowledgements", |
| "sec_num": null |
| } |
| ], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "Query learning strategies using boosting and bagging", |
| "authors": [ |
| { |
| "first": "Naoki", |
| "middle": [], |
| "last": "Abe", |
| "suffix": "" |
| }, |
| { |
| "first": "Hiroshi", |
| "middle": [], |
| "last": "Mamitsuka", |
| "suffix": "" |
| } |
| ], |
| "year": 1998, |
| "venue": "Proc. of ICML1998", |
| "volume": "", |
| "issue": "", |
| "pages": "1--10", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Naoki Abe and Hiroshi Mamitsuka. 1998. Query learning strategies using boosting and bagging. In Proc. of ICML1998, pages 1-10.", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "Towards Robust High Performance Word Sense Disambiguation of English Verbs Using Rich Linguistic Features", |
| "authors": [ |
| { |
| "first": "Jinying", |
| "middle": [], |
| "last": "Chen", |
| "suffix": "" |
| }, |
| { |
| "first": "Martha", |
| "middle": [], |
| "last": "Palmer", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proc. of IJCNLP2005, Oct", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Jinying Chen and Martha Palmer. 2005. Towards Robust High Performance Word Sense Disambiguation of English Verbs Using Rich Linguistic Features, In Proc. of IJCNLP2005, Oct., Jeju, Republic of Korea.", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "Building a Sense Tagged Corpus with Open Mind Word Expert", |
| "authors": [ |
| { |
| "first": "Tim", |
| "middle": [], |
| "last": "Chklovski", |
| "suffix": "" |
| }, |
| { |
| "first": "Rada", |
| "middle": [], |
| "last": "Mihalcea", |
| "suffix": "" |
| } |
| ], |
| "year": 2002, |
| "venue": "Proceedings of the ACL 2002 Workshop on \"Word Sense Disambiguation: Recent Successes and Future Directions", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Tim Chklovski and Rada Mihalcea, Building a Sense Tagged Corpus with Open Mind Word Expert, in Proceedings of the ACL 2002 Workshop on \"Word Sense Disambiguation: Recent Successes and Future Directions\", Philadelphia, July 2002.", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "Investigations into the role of lexical semantics in word sense disambiguation", |
| "authors": [ |
| { |
| "first": "Hoa", |
| "middle": [ |
| "T" |
| ], |
| "last": "Dang", |
| "suffix": "" |
| } |
| ], |
| "year": 2004, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Hoa T. Dang. 2004. Investigations into the role of lexical semantics in word sense disambiguation. PhD Thesis. University of Pennsylvania.", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "Selective sampling for example-based word sense disambiguation", |
| "authors": [ |
| { |
| "first": "Atsushi", |
| "middle": [], |
| "last": "Fujii", |
| "suffix": "" |
| }, |
| { |
| "first": "Takenobu", |
| "middle": [], |
| "last": "Tokunaga", |
| "suffix": "" |
| }, |
| { |
| "first": "Kentaro", |
| "middle": [], |
| "last": "Inui", |
| "suffix": "" |
| }, |
| { |
| "first": "Hozumi", |
| "middle": [], |
| "last": "Tanaka", |
| "suffix": "" |
| } |
| ], |
| "year": 1998, |
| "venue": "Computational Linguistics", |
| "volume": "4", |
| "issue": "", |
| "pages": "573--597", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Atsushi Fujii, Takenobu Tokunaga, Kentaro Inui, Hozumi Tanaka. 1998. Selective sampling for example-based word sense disambiguation, Computational Linguistics, v.24 n.4, p.573-597, Dec. Eduard Hovy, Mitchell Marcus, Martha Palmer, Lance Ramshaw and Ralph Weischedel. OntoNotes: The 90% Solution. Accepted by HLT-NAACL06. Short paper.", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "A sequential algorithm for training text classifiers", |
| "authors": [ |
| { |
| "first": "D", |
| "middle": [], |
| "last": "David", |
| "suffix": "" |
| }, |
| { |
| "first": "William", |
| "middle": [ |
| "A" |
| ], |
| "last": "Lewis", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Gale", |
| "suffix": "" |
| } |
| ], |
| "year": 1994, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "David D. Lewis and William A. Gale. 1994. A sequential algorithm for training text classifiers. In W.", |
| "links": null |
| }, |
| "BIBREF6": { |
| "ref_id": "b6", |
| "title": "Proceedings of SIGIR-94", |
| "authors": [ |
| { |
| "first": "Bruce", |
| "middle": [], |
| "last": "Croft", |
| "suffix": "" |
| }, |
| { |
| "first": "J", |
| "middle": [], |
| "last": "Cornelis", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Van Rijsbergen", |
| "suffix": "" |
| } |
| ], |
| "year": null, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Bruce Croft and Cornelis J. van Rijsbergen, editors, Proceedings of SIGIR-94, Dublin, IE.", |
| "links": null |
| }, |
| "BIBREF7": { |
| "ref_id": "b7", |
| "title": "MALLET: A Machine Learning for Language Toolkit", |
| "authors": [ |
| { |
| "first": "K", |
| "middle": [], |
| "last": "Andrew", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Mccallum", |
| "suffix": "" |
| } |
| ], |
| "year": 2002, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Andrew K. McCallum. 2002. MALLET: A Machine Learning for Language Toolkit. http://www.cs. umass.edu/~mccallum/mallet.", |
| "links": null |
| }, |
| "BIBREF8": { |
| "ref_id": "b8", |
| "title": "Employing EM in pool-based active learning for text classification", |
| "authors": [ |
| { |
| "first": "Andew", |
| "middle": [], |
| "last": "Mccallum", |
| "suffix": "" |
| }, |
| { |
| "first": "Kamal", |
| "middle": [], |
| "last": "Nigam", |
| "suffix": "" |
| } |
| ], |
| "year": 1998, |
| "venue": "Proc. of ICML '98", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Andew McCallum and Kamal Nigam. 1998. Employing EM in pool-based active learning for text classification. In Proc. of ICML '98.", |
| "links": null |
| }, |
| "BIBREF9": { |
| "ref_id": "b9", |
| "title": "Making fine-grained and coarse-grained sense distinctions, both manually and automatically", |
| "authors": [], |
| "year": 2006, |
| "venue": "Natural Language Engineering", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Martha Palmer, Hoa Trang Dang and Christiane Fellbaum. (to appear, 2006). Making fine-grained and coarse-grained sense distinctions, both manually and automatically. Natural Language Engineering.", |
| "links": null |
| }, |
| "BIBREF10": { |
| "ref_id": "b10", |
| "title": "Active Learning for Logistic Regression", |
| "authors": [ |
| { |
| "first": "Andrew", |
| "middle": [ |
| "I" |
| ], |
| "last": "Schein", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Andrew I. Schein. 2005. Active Learning for Logistic Regression. Ph.D. Thesis. Univ. of Pennsylvania.", |
| "links": null |
| }, |
| "BIBREF11": { |
| "ref_id": "b11", |
| "title": "Multi-criteria-based active learning for named entity recognition", |
| "authors": [ |
| { |
| "first": "Dan", |
| "middle": [], |
| "last": "Shen", |
| "suffix": "" |
| }, |
| { |
| "first": "Jie", |
| "middle": [], |
| "last": "Zhang", |
| "suffix": "" |
| }, |
| { |
| "first": "Jian", |
| "middle": [], |
| "last": "Su", |
| "suffix": "" |
| }, |
| { |
| "first": "Guodong", |
| "middle": [], |
| "last": "Zhou", |
| "suffix": "" |
| }, |
| { |
| "first": "Chew", |
| "middle": [], |
| "last": "Lim Tan", |
| "suffix": "" |
| } |
| ], |
| "year": 2004, |
| "venue": "Proc. of ACL04", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Dan Shen, Jie Zhang, Jian Su, Guodong Zhou and Chew Lim Tan. 2004 Multi-criteria-based active learning for named entity recognition, In Proc. of ACL04, Barcelona, Spain.", |
| "links": null |
| }, |
| "BIBREF12": { |
| "ref_id": "b12", |
| "title": "Active learning for statistical natural language parsing", |
| "authors": [ |
| { |
| "first": "Min", |
| "middle": [], |
| "last": "Tang", |
| "suffix": "" |
| }, |
| { |
| "first": "Xiaoqiang", |
| "middle": [], |
| "last": "Luo", |
| "suffix": "" |
| }, |
| { |
| "first": "Salim", |
| "middle": [], |
| "last": "Roukos", |
| "suffix": "" |
| } |
| ], |
| "year": 2002, |
| "venue": "Proc. of ACL", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Min Tang, Xiaoqiang Luo, and Salim Roukos. 2002. Active learning for statistical natural language parsing. In Proc. of ACL 2002.", |
| "links": null |
| }, |
| "BIBREF13": { |
| "ref_id": "b13", |
| "title": "Active learning for natural language parsing and information extraction", |
| "authors": [ |
| { |
| "first": "Cynthia", |
| "middle": [ |
| "A" |
| ], |
| "last": "Thompson", |
| "suffix": "" |
| }, |
| { |
| "first": "Mary", |
| "middle": [ |
| "Elaine" |
| ], |
| "last": "Califf", |
| "suffix": "" |
| }, |
| { |
| "first": "Raymond", |
| "middle": [ |
| "J" |
| ], |
| "last": "Mooney", |
| "suffix": "" |
| } |
| ], |
| "year": 1999, |
| "venue": "Proc. of ICML-99", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Cynthia A. Thompson, Mary Elaine Califf, and Raymond J. Mooney. 1999. Active learning for natural language parsing and information extraction. In Proc. of ICML-99.", |
| "links": null |
| } |
| }, |
| "ref_entries": {} |
| } |
| } |
