| { | |
| "paper_id": "I13-1039", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:14:05.869165Z" | |
| }, | |
| "title": "Detecting deceptive opinions with profile compatibility", | |
| "authors": [ | |
| { | |
| "first": "Vanessa", | |
| "middle": [ | |
| "Wei" | |
| ], | |
| "last": "Feng", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Toronto Toronto", | |
| "location": { | |
| "postCode": "M5S 3G4", | |
| "region": "ON", | |
| "country": "Canada" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Graeme", | |
| "middle": [], | |
| "last": "Hirst", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Toronto Toronto", | |
| "location": { | |
| "postCode": "M5S 3G4", | |
| "region": "ON", | |
| "country": "Canada" | |
| } | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "We propose using profile compatibility to differentiate genuine and fake product reviews. For each product, a collective profile is derived from a separate collection of reviews. Such a profile contains a number of aspects of the product, together with their descriptions. For a given unseen review about the same product, we build a test profile using the same approach. We then perform a bidirectional alignment between the test and the collective profile, to compute a list of aspect-wise compatible features. We adopt Ott et al. (2011)'s op spam v1.3 dataset for identifying truthful vs. deceptive reviews. We extend the recently proposed N-GRAM+SYN model of Feng et al. (2012a) by incorporating profile compatibility features, showing such an addition significantly improves upon their state-ofart classification performance.", | |
| "pdf_parse": { | |
| "paper_id": "I13-1039", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "We propose using profile compatibility to differentiate genuine and fake product reviews. For each product, a collective profile is derived from a separate collection of reviews. Such a profile contains a number of aspects of the product, together with their descriptions. For a given unseen review about the same product, we build a test profile using the same approach. We then perform a bidirectional alignment between the test and the collective profile, to compute a list of aspect-wise compatible features. We adopt Ott et al. (2011)'s op spam v1.3 dataset for identifying truthful vs. deceptive reviews. We extend the recently proposed N-GRAM+SYN model of Feng et al. (2012a) by incorporating profile compatibility features, showing such an addition significantly improves upon their state-ofart classification performance.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "With the rapid development of e-commerce and the increasing popularity of various product review websites, people are more and more used to making purchase decisions based on the reported experience of other customers. A product rated positively by its previous users is able to attract potential new customers, while a poorly rated product is certainly not a good option for most new customers. Given this influential power of product reviews, there comes a huge potential for deceptive opinion spam to distort the true evaluation of a product. The promoters of a product may post false complimentary reviews, and competitors may post false derogatory reviews.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Although the task of detecting deceptive opinion spam can be formulated as a traditional bi-nary classification problem with two classes, deceptive and truthful, where supervised learning can be applied, it is essentially very challenging. Its major difficulty arises from the lack of reliably labeled data, because it is extremely difficult for humans to identify through reading (see discussion in Section 4). Therefore, much previous work on detecting deceptive opinions usually relies on some meta-information, such as the IP address of the reviewer or the average rating of the product, rather than the actual content of the review Liu (2012) . However, thanks to the release of the op spam v1.3 dataset by , a gold-standard dataset composed of 400 truthful and 400 deceptive hotel reviews (see Section 3), we are now at an appropriate stage for conducting supervised learning and reliable evaluation of the task.", | |
| "cite_spans": [ | |
| { | |
| "start": 637, | |
| "end": 647, | |
| "text": "Liu (2012)", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In their work, Ott et al. proposed to use n-grams as features, and trained an SVM classifier to classify whether a given review is deceptive or truthful, achieving nearly 90% accuracy. More recently, Feng et al. (2012a) extended Ott et al.'s n-gram feature set by incorporating deep syntax features, i.e., syntactic production rules derived from Probabilistic Context Free Grammar (PCFG) parse trees, and obtained 91.2% accuracy on the same dataset.", | |
| "cite_spans": [ | |
| { | |
| "start": 200, | |
| "end": 219, | |
| "text": "Feng et al. (2012a)", | |
| "ref_id": "BIBREF2" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "While both Ott et al. and Feng et al. 's cues of deception come purely from the surface realization in the reviews, we propose the use of additional signals of truthfulness by characterizing the degree of compatibility between the personal experience described in a test review and a product profile derived from a collection of reference reviews about the same product. Our intuition comes from the hypothesis that since the writer of a deceptive review usually does not have any actual experience with that product, the resulting review might contain some contradictions with facts about the product, or the writer might fail to mention those as-pects of the product that are commonly mentioned in truthful reviews. Conversely, a writer of a truthful review is more likely to make similar comments about some particular aspects of the product as other truthful reviewers. In other words, we postulate that by aligning a the profile of a product and the description of the writer's personal experience, some useful clues can be revealed to help identify possible deception.", | |
| "cite_spans": [ | |
| { | |
| "start": 11, | |
| "end": 37, | |
| "text": "Ott et al. and Feng et al.", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In line with the intuition above, we want to capture two types of compatibility: (1) Compatibility with the existence of some distinct aspect of the product, such as the mention of the famous art museum nearby the hotel; (2) Compatibility with the description of some general aspect of the product, such as commenting that the breakfast at the hotel is charged for.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "We show that, by incorporating our computed profile alignment features with Feng et al.'s ngrams and deep syntax features, we significantly improve on the state-of-art performance presented by their N-GRAM+SYN model.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Before the existence of gold-standard datasets which include both truthful and deceptive product reviews, several attempts were made to detect deceptive opinions.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "With regard to unsupervised approaches, previous work was primarily based on various patterns of atypical behaviors. proposed several behavior models to identify unusual reviewer patterns; e.g., spammers may contribute a fake review soon after product launch to maximally affect subsequent reviewers. Each of these individual models assigns a numeric score indicating the extent to which the reviewer has engaged in apparent spam behaviors, and these scores are then combined to product a final spam score. employed data-mining techniques to discover unexpected class association rules; e.g., reviewers who give all high ratings to products of a brand while most other reviews are generally negative about the brand. Wu et al. (2010) proposed to use the distortion of popularity rankings as an indicator of opinion spamming, in the sense that deleting a set of random reviews should not overly disrupt the popularity ranking of a list of entities, while deleting fake reviews should significantly distort the overall rankings, under the assumption that deceptive reviews usually express a sentiment at odds with legitimate reviews. Feng et al. (2012b) postulated that for a given domain, there exists a set of representative distributions of review rating scores, and fake reviews written by hired spammers will distort such natural distributions. Experimenting with a range of pseudo-goldstandard datasets, they provided quantitative insights into the characteristics of natural distributions of opinions in various product review domains.", | |
| "cite_spans": [ | |
| { | |
| "start": 717, | |
| "end": 733, | |
| "text": "Wu et al. (2010)", | |
| "ref_id": "BIBREF19" | |
| }, | |
| { | |
| "start": 1132, | |
| "end": 1151, | |
| "text": "Feng et al. (2012b)", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "With respect to supervised approaches, Jindal and Liu (2008) first conducted a tentative study of detecting deceptive opinions. In the absence of gold-standard datasets, they trained models using features from the review texts, as well as from meta-information on the reviewer and the product, to distinguish between duplicate reviews (regarded as deceptive), i.e., reviews whose major content appears more than once in the corpus, and non-duplicate reviews (regarded as truthful). However, these (near-)duplicate reviews are often not sophisticated in their composition, and therefore are relatively easy to identify, even by offthe-shelf plagiarism detection software . Yoo and Gretzel (2009) constructed a small gold-standard dataset with 40 truthful and 42 deceptive hotel reviews, and manually inspected the statistical differences of psychologically relevant linguistic features for truthful and deceptive reviews. released the op spam v1.3 dataset (see Section 3), a much larger dataset with 400 truthful and 400 deceptive hotel reviews with positive comments, which allows the application of machine learning techniques for training models to automatically detect deceptive opinions. In their work, Ott et al. focused on the textual content of the reviews, and approached the task of detecting deceptive opinions by finding any stretch of imagination in the text -they treated fake reviews as imaginative writing, and used standard computational approaches of genre identification, psycholinguistic deception detection, and text categorization to identify them. Among the set of features explored in their classification experiments, they discovered that the features traditionally employed in either psychological studies of deception or genre identification were both significantly outperformed by a much simpler N-GRAM model with n-grams only as features, which achieved nearly 90% accuracy. Later, Feng et al. (2012a) proposed a strengthened model, N-GRAM+SYN, by incorporating syntactic production rules derived from Probabilistic Context Free Grammar (PCFG) parse trees. They obtained 91.2% accuracy on the op spam v1.3 dataset, which is a 14% error reduction.", | |
| "cite_spans": [ | |
| { | |
| "start": 39, | |
| "end": 60, | |
| "text": "Jindal and Liu (2008)", | |
| "ref_id": "BIBREF5" | |
| }, | |
| { | |
| "start": 672, | |
| "end": 694, | |
| "text": "Yoo and Gretzel (2009)", | |
| "ref_id": "BIBREF20" | |
| }, | |
| { | |
| "start": 1910, | |
| "end": 1929, | |
| "text": "Feng et al. (2012a)", | |
| "ref_id": "BIBREF2" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "More recently, Ott et al. released a new version, the op spam v1.4 dataset, with gold-standard negative reviews included as well (Ott et al., 2013) , which offers an opportunity to more extensively study the problem of detecting deceptive reviews. However, the work described in this paper is focused on positive review spam, and based directly on the work of Feng et al. We extend their N-GRAM+SYN model by incorporating potential signals of truthfulness, derived from aligning the description in a test review with the product profile constructed from a large collection of reference reviews about the same product. Somewhat orthogonal to Feng et al.'s n-grams and deep syntax features, which are focused on the surface realization of a given product review, our alignment features are able to employ useful information from the product itself.", | |
| "cite_spans": [ | |
| { | |
| "start": 129, | |
| "end": 147, | |
| "text": "(Ott et al., 2013)", | |
| "ref_id": "BIBREF14" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Due to the difficulty for humans to identify deceptive opinions, traditional approaches to constructing annotated corpora by recruiting human judges to label a given set of texts does not apply to the task of deception detection. Consequently, crowdsourcing services such as Amazon Mechanical Turk 1 (AMT) have been adopted as a better solution Rubin and Vashchilko, 2012) . By asking paid subjects on AMT to compose deceptive and/or truthful texts, corpora with reliable labels can be constructed.", | |
| "cite_spans": [ | |
| { | |
| "start": 345, | |
| "end": 372, | |
| "text": "Rubin and Vashchilko, 2012)", | |
| "ref_id": "BIBREF17" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The op spam v1.Dataset", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "In this work, we use Ott et al.'s op spam v1.3 dataset 2 , which contains positive reviews for the 20 most-rated hotels on TripAdvisor 3 (TA) in Chicago. For each hotel, Ott et al. selected 20 deceptive reviews from submissions on AMT, and 20 truthful reviews from 5-star reviews on TA, resulting in 800 reviews in total for 20 hotels. The average length of deceptive reviews is 115.75 words, while the truthful reviews are chosen to have roughly the same average length.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The op spam v1.Dataset", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "It has been shown that humans can differentiate truthful reviews from deceptive ones with only a modest accuracy. With respect to the op spam v1.3 dataset used in our work, human judges were only able to achieve 60% accuracy, and the inter-annotator agreement was low: 0.11 computed by Fleiss's kappa . Since as humans, we do not have a tangible intuition of what signals deception and what characterizes truthfulness, we have great difficulty designing features for automatic identifying deceptive opinions.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Difficulty of the Task", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "In addition, there is difficulty with regard to the construction of gold-standard datasets for the task of deception detection. On one hand, although crowdsourcing provides a relatively cheap and reliable approach, gathering deceptive data is still nevertheless laborious. On the other hand, since for truthful reviews, we are able to select a small subset for which we have high confidence only by using a combination of various heuristics, the constructed dataset is inherently biased. Therefore, given the limited number and size of currently available gold-standard datasets, in the process of developing more sophisticated models involving a larger number of features, we may inevitably face the issue of overfitting on a relatively small set of data.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Difficulty of the Task", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "We adopt Feng et al.'s N-GRAM+SYN model as the baseline system, which employs two distinct sets of features.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Baseline features", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "N-GRAM features: As shown by Ott et al., the bag-of-words model is effective for identifying deceptive reviews. However, the optimal choice of n is not consistent in Ott et al.'s original implementation and Feng et al.'s strengthened model: Ott et al. used the union of unigrams and bigrams, while Feng et al. obtained their best performance using unigrams alone, together with deep syntax features. Therefore, for a fair comparison, we consider using unigrams, bigrams, and the union of both, and choose the best combination with deep syntax features as our baseline system. SYN features: Following Feng et al., deep syntax features are encoded as production rules derived from PCFG parse trees. These production rules include lexicalized ones, i.e., POS-tag-toword rules, and are combined with the grandparent node.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Baseline features", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "All baseline features are encoded as their tf-idf weights, with features appearing only once in the dataset eliminated.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Baseline features", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "As introduced in Section 1, we postulate that by aligning the profile of a product with the description of the writer's personal experience, we can characterize the degree of compatibility between them. Such alignment features could serve as useful signals to differentiate between deception and truthfulness.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspects and compatibility", | |
| "sec_num": "5.2.1" | |
| }, | |
| { | |
| "text": "In particular, we define compatibility for two types of aspects of a product -distinct aspects and general aspects. Distinct aspects are special features of the product. For hotels, those distinct aspects are usually realized as proper noun phrases in the text, typically the names of the landmarks nearby the hotel, including museums, parks, restaurants, bars, and shopping malls. On the other hand, general aspects are common features which typically appear in any product of this particular kind. For hotels, location, service, and breakfast are typical general aspects. The method we use to identify and extract distinct and general aspects from reviews is described in Section 5.2.4.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspects and compatibility", | |
| "sec_num": "5.2.1" | |
| }, | |
| { | |
| "text": "Compatibility for distinct and general aspects is defined as follows:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspects and compatibility", | |
| "sec_num": "5.2.1" | |
| }, | |
| { | |
| "text": "1. Compatibility with the existence of some distinct aspect of the product, e.g., truthful reviews of this Chicago hotel often mention the famous nearby Field Museum, and the test review also mentions this museum.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspects and compatibility", | |
| "sec_num": "5.2.1" | |
| }, | |
| { | |
| "text": "2. Compatibility with the description of some general aspect of the product. For example, breakfast at most Chicago hotels is complimentary; however, the breakfast at this hotel is charged for, and the test review describes it as \"the breakfast is kinda expensive\".", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspects and compatibility", | |
| "sec_num": "5.2.1" | |
| }, | |
| { | |
| "text": "A product's profile P is composed of a set of its aspects A = {a 1 , . . . , a N }, where each a i is an individual aspect of the product. Each aspect a i is associated with a description to describe the aspect, along with the weight w j assigned to that word.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Definition of product profile", | |
| "sec_num": "5.2.2" | |
| }, | |
| { | |
| "text": "D i = {p 1 : w 1 , . . . , p m : w m }, in", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Definition of product profile", | |
| "sec_num": "5.2.2" | |
| }, | |
| { | |
| "text": "Distinct aspects, a i,d , and general aspects, a i,g , are treated differently when it comes to their descriptions. For a distinct aspect, its description can contain only one word, existence, because we only care about whether this aspect is mentioned in the text, not the particular words used to describe this aspect. In fact, most of these distinct aspects do not occur with any associated adjectives or adverbs.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Definition of product profile", | |
| "sec_num": "5.2.2" | |
| }, | |
| { | |
| "text": "An example fragment of a hotel's profile is shown in Table 1 , in which the last aspect Michigan Ave. is a distinct aspect, and thus only one word, existence, appears in its description.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 53, | |
| "end": 60, | |
| "text": "Table 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Definition of product profile", | |
| "sec_num": "5.2.2" | |
| }, | |
| { | |
| "text": "To establish a profile P for each target hotel in the op spam v1.3 dataset, we first gather all its reviews on TripAdvisor, from which we choose up to 200 reviews, subject to the following criteria:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data for profile construction", | |
| "sec_num": "5.2.3" | |
| }, | |
| { | |
| "text": "1. It is not present in the op spam v1.3 dataset.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data for profile construction", | |
| "sec_num": "5.2.3" | |
| }, | |
| { | |
| "text": "2. Its rating is five stars (the highest).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data for profile construction", | |
| "sec_num": "5.2.3" | |
| }, | |
| { | |
| "text": "3. Its language is English.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data for profile construction", | |
| "sec_num": "5.2.3" | |
| }, | |
| { | |
| "text": "4. It contains at least 150 characters.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data for profile construction", | |
| "sec_num": "5.2.3" | |
| }, | |
| { | |
| "text": "5. The author has written at least 10 reviews.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data for profile construction", | |
| "sec_num": "5.2.3" | |
| }, | |
| { | |
| "text": "6. The author has written reviews for at least 5 different hotels.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data for profile construction", | |
| "sec_num": "5.2.3" | |
| }, | |
| { | |
| "text": "7. The author has received at least 5 helpfulness votes from other users.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data for profile construction", | |
| "sec_num": "5.2.3" | |
| }, | |
| { | |
| "text": "If there are more than 200 reviews satisfying the above criteria, we choose the 200 with the highest number of helpfulness votes received by their authors. Most of the above criteria are consistent with the original setup of collecting the truthful reviews in the op spam v1.3 dataset. However, we add the last three strong criteria to ensure the quality of the collected reviews. Table 2 lists the statistics of the reviews used in our profile construction. The statistics are categorized into (1) the information about the reviews themselves, including the total number of qualified reviews for each hotel, and the length (in characters) of each review; and (2) the information about the distinct authors of these reviews, including the total number of reviews each author has written on TripAdvisor, the total number of hotels each author has reviewed, and the total number of helpfulness votes that each author has received from other users.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 381, | |
| "end": 388, | |
| "text": "Table 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Data for profile construction", | |
| "sec_num": "5.2.3" | |
| }, | |
| { | |
| "text": "Raw aspect extraction For each particular hotel h, from the set of its corresponding reviews R h , we extract the aspects A h = {a 1,t 1 ,h , . . . , a M,t M ,h } (e.g., room, service, and Michigan Ave. in Table 1 ) to be included into h's profile P h , where t i is the type, distinct or general, of each aspect. For the sake of notational clarity, from now on, when talking about a specific hotel h, we will omit the subscript h. We first extract all noun phrases present in R, and use these noun phrases as raw aspects. For each raw aspect extracted, we identify whether it is a distinct or general aspect: if the aspect is extracted as a proper noun phrase (as tagged by the Stanford parser (Klein and Manning, 2003) ), then the aspect is labeled as distinct, and general otherwise.", | |
| "cite_spans": [ | |
| { | |
| "start": 695, | |
| "end": 720, | |
| "text": "(Klein and Manning, 2003)", | |
| "ref_id": "BIBREF8" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 206, | |
| "end": 213, | |
| "text": "Table 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Aspect extraction", | |
| "sec_num": "5.2.4" | |
| }, | |
| { | |
| "text": "Noun phrase extraction is performed solely on the syntactic parse trees. We do not attempt to resolve coreferences for the following two reasons: (1) since most coreference resolution tools are trained on news texts, their performance on product reviews might not be reliable;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspect extraction", | |
| "sec_num": "5.2.4" | |
| }, | |
| { | |
| "text": "(2) we observe that authors of these product reviews rarely employ coreference in their writing, and therefore the impact of resolving coreference might be minor after all.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspect extraction", | |
| "sec_num": "5.2.4" | |
| }, | |
| { | |
| "text": "Aspect clustering Due to the flexibility of natural languages, it is common to see people express the same concept through different lexical usage. For example, both \"price\" and \"rate\" can refer to the price of the hotel. In addition to synonyms, semantically related words can also be adopted to describe the same aspect of the product, for example, \"lighting\" and \"painting\" can both describe the decoration of the hotel. In order to deal with the use of these (near-) synonyms in text, we perform separate clustering for the sets of distinct and general aspects.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspect extraction", | |
| "sec_num": "5.2.4" | |
| }, | |
| { | |
| "text": "For distinct aspects {a i,d }, the distance between a pair of aspects is defined by the length of their longest common substring divided by the average length of the two aspects. The motivation for this definition is to bring distinct aspects such as Michigan Ave. and Michigan Avenue into the same cluster.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspect extraction", | |
| "sec_num": "5.2.4" | |
| }, | |
| { | |
| "text": "For general aspects {a i,g }, the distance between a pair of aspects is defined as their Lin Similarity (Lin, 1998) , implemented by NLTK's similarity package (Bird et al., 2009) .", | |
| "cite_spans": [ | |
| { | |
| "start": 104, | |
| "end": 115, | |
| "text": "(Lin, 1998)", | |
| "ref_id": "BIBREF10" | |
| }, | |
| { | |
| "start": 159, | |
| "end": 178, | |
| "text": "(Bird et al., 2009)", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspect extraction", | |
| "sec_num": "5.2.4" | |
| }, | |
| { | |
| "text": "For both types of aspects, we perform hierarchical agglomerative clustering with average linkage. Cluster merging is terminated once the distance between all pairs of clusters is greater than 0.3.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspect extraction", | |
| "sec_num": "5.2.4" | |
| }, | |
| { | |
| "text": "After clustering distinct and general aspects separately, we keep track of the total occurrences of each aspect across the entire set R. Then, we let A be the set of the most common N aspects 4 . N is a parameter identical for all hotels, indicating how many aspects are included in each profile. We expect that the choice of N is affected by the particular type of product, and we experiment with several values of N.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Sorting aspects by occurrences", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "For a hotel h, each of its aspects a i is then associated with a description", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Description extraction", | |
| "sec_num": "5.2.5" | |
| }, | |
| { | |
| "text": "D i = {p i,1 : w i,1 , . . . , p i,m : w i,m },", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Description extraction", | |
| "sec_num": "5.2.5" | |
| }, | |
| { | |
| "text": "which has two components: description words p i, j and their weights w i, j , for 1 \u2264 j \u2264 m. The list {p i,1 , . . . , p i,m } is composed of the description words used to modify the particular aspect a i , e.g., {wonderful, deluxe, and huge} for the aspect room in Table 1 .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 93, | |
| "end": 112, | |
| "text": "The list {p i,1 , .", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 266, | |
| "end": 273, | |
| "text": "Table 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Description extraction", | |
| "sec_num": "5.2.5" | |
| }, | |
| { | |
| "text": "For each distinct aspect a i,d , as discussed in Section 5.2.2, there can be only one description word, existence, indicating the presence of this aspect in the profile.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Description words", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "For each general aspect a i,g , its description words are automatically extracted from the dependencies in the set of reviews R. From the dependencies output by the Stanford dependency parser (de Marneffe et al., 2006) , for a particular aspect a i , we first locate all relevant dependency relations which have a i (or the head noun of a i if a i is a noun phrase) as the governors (or dependents, depending on the particular dependency types). Then from these relevant dependency relations, we extract the words that appear in the dependent (or governor) slots to be the description words, provided that those words are tagged as adjectives or adverbs.", | |
| "cite_spans": [ | |
| { | |
| "start": 192, | |
| "end": 218, | |
| "text": "(de Marneffe et al., 2006)", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Description words", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Weight assignment Each description word p i, j is assigned a weight w i, j to indicate the reliability of that particular description word. There are many possibilities for these weight assignment functions. Here, as a preliminary study, for distinct aspects {a i,d }, we simply use the number of occurrences of the aspect in the review collection R as the assigned weight w i, j , and for general aspects {a i,d }, we use the number of occurrences of p i, j in R as w i, j .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Description words", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Once we establish a profile P for each hotel h from the set of relevant reviews R, then given an unseen test review r about h, we can perform a bidirectional alignment with P, and compute a list of compatibility features.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Computing profile compatibility", | |
| "sec_num": "5.3" | |
| }, | |
| { | |
| "text": "Using the same approach as described in Sections 5.2.4 and 5.2.5, we construct a profile Q from this single review. For the sake of clarity, we call Q the test profile, in contrast to P, the collective profile, which is constructed from a collection of reviews.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Computing profile compatibility", | |
| "sec_num": "5.3" | |
| }, | |
| { | |
| "text": "We perform a bidirectional alignment between the test profile P and the collective profile Q. The direction of an alignment is defined as the following.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Bidirectional alignment", | |
| "sec_num": "5.3.1" | |
| }, | |
| { | |
| "text": "When aligning the test profile Q with the collective profile P (Q \u2192 P), we compute aspect-wise compatibility features (to be defined in Section 5.3.2) for each aspect a i in P. Similarly, when aligning the collective profile P with the test profile Q (P \u2192 Q), we compute for each aspect a i in Q.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Bidirectional alignment", | |
| "sec_num": "5.3.1" | |
| }, | |
| { | |
| "text": "The Q \u2192 P alignment captures whether the aspects that most reviewers would describe in their reviews are mentioned in the test review, and whether there exist any similarity or conflicts between the common opinions represented in the collective profile P and the test review.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Bidirectional alignment", | |
| "sec_num": "5.3.1" | |
| }, | |
| { | |
| "text": "The P \u2192 Q alignment captures whether the test review falsely includes some imaginary extra aspects about the product, such as an indoor pool in the hotel, while the collective profile P does not include this aspect.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Bidirectional alignment", | |
| "sec_num": "5.3.1" | |
| }, | |
| { | |
| "text": "For each direction of the bidirectional alignment, we compute a list of aspect-wise compatibility features. Without loss of generality, assume that the direction of the alignment is Q \u2192 P. Thus, for each aspect a i in P, we include three features to indicate the compatibility between P and Q:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspect-wise compatibility features", | |
| "sec_num": "5.3.2" | |
| }, | |
| { | |
| "text": "1. m a i : a boolean value indicating whether a i is mentioned in Q or not. If a i does appear in Q, we then compute the following two values, based on a i 's description D i in P, and its description D i in Q.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspect-wise compatibility features", | |
| "sec_num": "5.3.2" | |
| }, | |
| { | |
| "text": "2. s a i \u2208 [0, 1]: a numeric value indicating the similarity between P and Q with respect to a i . S a i is computed as the cosine similarity between D i and D i . We treat D i and D i as two vectors following the bag-of-words model, where the description words are the words, and the weights are their corresponding frequencies.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspect-wise compatibility features", | |
| "sec_num": "5.3.2" | |
| }, | |
| { | |
| "text": "3. c a i \u2208 [0, 1]: a numeric value indicating the conflicts between P and Q with respect to a i . For each description word p i in D i , we determine whether D i includes its antonym.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspect-wise compatibility features", | |
| "sec_num": "5.3.2" | |
| }, | |
| { | |
| "text": "Antonyms are determined using the lexical relations defined in WordNet 5 . c a i is computed as the fraction of description words in D i that have corresponding antonyms in D i .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspect-wise compatibility features", | |
| "sec_num": "5.3.2" | |
| }, | |
| { | |
| "text": "Clearly, when computing aspect-wise similarity score s a i , we do not consider any synonyms or semantic similarity between two description words -the similarity is based solely on the use of exact words. Our strategy of exact word matching is a simplification, but since similarity of adjectives and adverbs is not as well-studied as that of nouns (various similarity metrics for nouns are available for WordNet, but none of those work on adjectives and adverbs), clustering description words is not as trivial as clustering aspects 6 .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Aspect-wise compatibility features", | |
| "sec_num": "5.3.2" | |
| }, | |
| { | |
| "text": "To test whether a given review r about the hotel h is deceptive or not, we extend Feng et al.' s N-GRAM+SYN model by incorporating our aspectwise compatibility features into it. We call this extension the C+N-GRAM+SYN model, which includes the following two categories of features: 2. Baseline features: 3,009 unigrams, 10,538 bigrams, and 6,571 syntactic production rules, as described in Section 5.1.", | |
| "cite_spans": [ | |
| { | |
| "start": 82, | |
| "end": 94, | |
| "text": "Feng et al.'", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "C+N-GRAM+SYN model", | |
| "sec_num": "5.4" | |
| }, | |
| { | |
| "text": "The resulting dimension of our full feature set is 20,118 + 2N, where N ranges from 10 to 100 in our experiments (see Section 6.2). For all experiments, we use SVM perf (Joachims, 2006) to train all the linear SVM classifiers.", | |
| "cite_spans": [ | |
| { | |
| "start": 169, | |
| "end": 185, | |
| "text": "(Joachims, 2006)", | |
| "ref_id": "BIBREF7" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "C+N-GRAM+SYN model", | |
| "sec_num": "5.4" | |
| }, | |
| { | |
| "text": "6 Results and discussion", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "C+N-GRAM+SYN model", | |
| "sec_num": "5.4" | |
| }, | |
| { | |
| "text": "We first demonstrate the performance of our reimplemented baseline models. We conduct 5-fold Table 3 : Results (in %) of our reimplemented baseline models. Performance is reported using 5-fold cross-validation over the op spam v1.3 dataset (800 reviews of 20 hotels).", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 93, | |
| "end": 100, | |
| "text": "Table 3", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Reimplementing baseline models", | |
| "sec_num": "6.1" | |
| }, | |
| { | |
| "text": "cross-validation experiments over the 800 reviews of 20 hotels, in which each fold contains all reviews from four hotels, such that the learned models are always tested against unseen hotels. Performance is reported by accuracy and the microaveraged precision, recall, and F-score of retrieving deceptive reviews (see Ott et al. (2013) for details). We experiment with five different baseline feature sets: 1 is noticeably lower (88.9%). Though its improvement over the N-GRAM model is not significant, the N-GRAM+SYN model performs the best (90.1%) among all baseline models, and thus is chosen to be the baseline in our later experiments.", | |
| "cite_spans": [ | |
| { | |
| "start": 318, | |
| "end": 335, | |
| "text": "Ott et al. (2013)", | |
| "ref_id": "BIBREF14" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Reimplementing baseline models", | |
| "sec_num": "6.1" | |
| }, | |
| { | |
| "text": "We now demonstrate the performance achieved by our profile alignment compatibility features. We experiment with using our profile alignment features in combination with baseline features, i.e., the C+N-GRAM+SYN model. We first study the 7 Actually, the best performance reported by Ott et al. was obtained using their LIWC + N-GRAM model, which is the combination of N-GRAM features and 80 features output by the Linguistic Inquiry and Word Count (LIWC) software (Pennebaker et al., 2007) . However, since the improvement over n-gram features alone was small and insignificant (0.2% absolute improvement in accuracy), and we were not able to successfully reproduce the improvement, the comparison with our model is based on using N-GRAM features alone. effect of the number of reviews used in constructing collective profiles (see Section 5.2.3). We experiment with using the top 50, 100, 150, and 200 reviews written by authors with the highest helpfulness votes (for some hotels, the total number of available reviews is less than 200). For each set of constructed profiles, we tried different values of N, the number of aspects included in each profile (both collective and test profiles), ranging from 10 to 100, and determine the optimal number by conducting 4-fold cross-validation experiments using 80% of the training data.", | |
| "cite_spans": [ | |
| { | |
| "start": 463, | |
| "end": 488, | |
| "text": "(Pennebaker et al., 2007)", | |
| "ref_id": "BIBREF15" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Comparison with baseline models", | |
| "sec_num": "6.2" | |
| }, | |
| { | |
| "text": "The results are shown in Table 4 . First, we see that the best overall performance, in terms of F 1 score and accuracy, is achieved using the top 50 or 100 reviews for profile construction. This suggests that using a fairly small amount of reliable data is sufficient to recognize compatibility between collective profiles and test reviews. In fact, using a larger number of reviews is slightly detrimental to the overall performance, since more noise might be included in the process.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 25, | |
| "end": 32, | |
| "text": "Table 4", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Comparison with baseline models", | |
| "sec_num": "6.2" | |
| }, | |
| { | |
| "text": "Moreover, we see that our best model, achieved using 50 or 100 reviews for profile construction, outperforms the best baseline model, N-GRAM+SYN, on all four metrics, by over 1%, which is a 12.1% error reduction rate. The improvement is confirmed to be statistically significant (p < .05) using the Wilcoxon sign-rank test. In addition, we inspect each model's predictions on individual test instances, and discover that our C+N-GRAM+SYN model correctly classifies some instances on which the baseline model makes an error, and at the same time does not make any additional errors that the baseline model does not make.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Comparison with baseline models", | |
| "sec_num": "6.2" | |
| }, | |
| { | |
| "text": "Finally, with respect to the value of N, which is the number of aspects included in each profile, we discover that smaller values of N, i.e., 10 or 20, consistently give the most satisfying results.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Comparison with baseline models", | |
| "sec_num": "6.2" | |
| }, | |
| { | |
| "text": "In this paper we proposed the use of profile alignment compatibility as an indicator of truthfulness in product reviews. We defined two types of compatibility between product profiles, and designed a methodology to tackle them by extracting aspects and associated descriptions from reviews. We adopted Ott et al.'s op spam v1.3 dataset of hotel reviews, and improved the N-GRAM+SYN model of Feng et al. (2012a) Table 4 : Results (in %) of our profile alignment compatibility models. Performance is reported using 5-fold cross-validation over the op spam v1.3 dataset (800 reviews of 20 hotels). The best result of each metric is shown in bold face. All numbers are significantly better (denoted by *) than the baseline in Table 3 , using the Wilcoxon sign-rank test (p < .05).", | |
| "cite_spans": [ | |
| { | |
| "start": 391, | |
| "end": 410, | |
| "text": "Feng et al. (2012a)", | |
| "ref_id": "BIBREF2" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 411, | |
| "end": 418, | |
| "text": "Table 4", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 722, | |
| "end": 729, | |
| "text": "Table 3", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Conclusions and perspectives", | |
| "sec_num": "7" | |
| }, | |
| { | |
| "text": "shown to significantly improve the performance of identifying deceptive reviews.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusions and perspectives", | |
| "sec_num": "7" | |
| }, | |
| { | |
| "text": "In future work, it is critical to know how well our methodology of extracting aspects and their descriptions from hotel reviews generalizes on other kinds of reviews. We suspect that the approach should work equally well on other domains, as long as the aspects are realized mainly by noun phrases, especially that distinct aspects are realized by proper noun phrases.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusions and perspectives", | |
| "sec_num": "7" | |
| }, | |
| { | |
| "text": "Another particularly interesting direction is to explore how our C+N-GRAM+SYN model performs on identifying fake negative reviews, as recently released by Ott et al. (2013) , rather than the positive reviews used in this work. While negative opinion spam is more hazardous to a brand's fame compared to positive ones, and thus identifying fake negative reviews might be more crucial, one potential difficulty of our approach is that genuine extremely negative reviews written by renowned reviewers are much more sparse than extremely positive ones, especially for famous products, such as the most popular Chicago hotels in the op spam v1.3 dataset.", | |
| "cite_spans": [ | |
| { | |
| "start": 155, | |
| "end": 172, | |
| "text": "Ott et al. (2013)", | |
| "ref_id": "BIBREF14" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusions and perspectives", | |
| "sec_num": "7" | |
| }, | |
| { | |
| "text": "http://mturk.com. 2 http://www.cs.cornell.edu/\u02dcmyleott 3 http://tripadvisor.com.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "N is usually much smaller than M, the total number of raw aspects.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "http://wordnet.princeton.edu/wordnet/ 6 In fact, we have experimented with using LSA-based word vectors to cluster description words, but the performance is slightly lower than using exact word matching.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "This work was financially supported by the Natural Sciences and Engineering Research Council of Canada and by the University of Toronto. We would like to thank Myle Ott from Cornell University for kindly providing us with the op spam v1.3 dataset.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Acknowledgments", | |
| "sec_num": null | |
| } | |
| ], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Natural Language Processing with Python", | |
| "authors": [ | |
| { | |
| "first": "Steven", | |
| "middle": [], | |
| "last": "Bird", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ewan", | |
| "middle": [], | |
| "last": "Klein", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Edward", | |
| "middle": [], | |
| "last": "Loper", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural Language Processing with Python. O'Reilly Media Inc.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Generating typed dependency parses from phrase structure parses", | |
| "authors": [ | |
| { | |
| "first": "Marie-Catherine", | |
| "middle": [], | |
| "last": "De Marneffe", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bill", | |
| "middle": [], | |
| "last": "Maccartney", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Christopher", | |
| "middle": [ | |
| "D" | |
| ], | |
| "last": "Manning", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Proceedings of the 5th International Conference on Language Resources and Evaluation", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Marie-Catherine de Marneffe, Bill MacCartney, and Christopher D. Manning. 2006. Generating typed dependency parses from phrase structure parses. In Proceedings of the 5th International Conference on Language Resources and Evaluation (LREC 2006).", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Syntactic stylometry for deception detection", | |
| "authors": [ | |
| { | |
| "first": "Song", | |
| "middle": [], | |
| "last": "Feng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ritwik", | |
| "middle": [], | |
| "last": "Banerjee", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yejin", | |
| "middle": [], | |
| "last": "Choi", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics", | |
| "volume": "2", | |
| "issue": "", | |
| "pages": "171--175", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Song Feng, Ritwik Banerjee, and Yejin Choi. 2012a. Syntactic stylometry for deception detection. In Proceedings of the 50th Annual Meeting of the As- sociation for Computational Linguistics (Volume 2: Short Papers), pages 171-175, Jeju Island, Korea, July. Association for Computational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Distributional footprints of deceptive product reviews", | |
| "authors": [ | |
| { | |
| "first": "Song", | |
| "middle": [], | |
| "last": "Feng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Longfei", | |
| "middle": [], | |
| "last": "Xing", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Anupam", | |
| "middle": [], | |
| "last": "Gogar", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yejin", | |
| "middle": [], | |
| "last": "Choi", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proceedings of International AAAI Conference on Weblogs and Social Media", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Song Feng, Longfei Xing, Anupam Gogar, and Yejin Choi. 2012b. Distributional footprints of decep- tive product reviews. In Proceedings of Interna- tional AAAI Conference on Weblogs and Social Me- dia (ICWSM 2012), June.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "In search of a gold standard in studies of deception", | |
| "authors": [ | |
| { | |
| "first": "Stephanie", | |
| "middle": [], | |
| "last": "Gokhman", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jeff", | |
| "middle": [], | |
| "last": "Hancock", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Poornima", | |
| "middle": [], | |
| "last": "Prabhu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Myle", | |
| "middle": [], | |
| "last": "Ott", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Claire", | |
| "middle": [], | |
| "last": "Cardie", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proceedings of the Workshop on Computational Approaches to Deception Detection", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "23--30", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Stephanie Gokhman, Jeff Hancock, Poornima Prabhu, Myle Ott, and Claire Cardie. 2012. In search of a gold standard in studies of deception. In Pro- ceedings of the Workshop on Computational Ap- proaches to Deception Detection, pages 23-30, Avi- gnon, France, April. Association for Computational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Opinion spam and analysis", | |
| "authors": [ | |
| { | |
| "first": "Nitin", | |
| "middle": [], | |
| "last": "Jindal", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bing", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "Proceedings of First ACM International Conference on Web Search and Data Mining (WSDM-2008)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "219--230", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nitin Jindal and Bing Liu. 2008. Opinion spam and analysis. In Proceedings of First ACM Interna- tional Conference on Web Search and Data Mining (WSDM-2008), pages 219-230.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Finding unusual review patterns using unexpected rules", | |
| "authors": [ | |
| { | |
| "first": "Nitin", | |
| "middle": [], | |
| "last": "Jindal", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bing", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ee-Peng", | |
| "middle": [], | |
| "last": "Lim", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Proceedings of the 19th ACM International Conference on Information and Knowledge Management (CIKM 2010)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1549--1552", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nitin Jindal, Bing Liu, and Ee-Peng Lim. 2010. Find- ing unusual review patterns using unexpected rules. In Proceedings of the 19th ACM International Con- ference on Information and Knowledge Manage- ment (CIKM 2010), pages 1549-1552.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Training linear SVMs in linear time", | |
| "authors": [ | |
| { | |
| "first": "Thorsten", | |
| "middle": [], | |
| "last": "Joachims", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, KDD '06", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "217--226", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Thorsten Joachims. 2006. Training linear SVMs in lin- ear time. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, KDD '06, pages 217-226, New York, NY, USA. ACM.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Accurate unlexicalized parsing", | |
| "authors": [ | |
| { | |
| "first": "Dan", | |
| "middle": [], | |
| "last": "Klein", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "D", | |
| "middle": [], | |
| "last": "Christopher", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Manning", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics (ACL 2003)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "423--430", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Dan Klein and Christopher D. Manning. 2003. Ac- curate unlexicalized parsing. In Proceedings of the 41st Annual Meeting of the Association for Com- putational Linguistics (ACL 2003), pages 423-430, Sapporo, Japan, July.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Detecting product review spammers using rating behaviors", | |
| "authors": [ | |
| { | |
| "first": "Ee-Peng", | |
| "middle": [], | |
| "last": "Lim", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Viet-An", | |
| "middle": [], | |
| "last": "Nguyen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Nitin", | |
| "middle": [], | |
| "last": "Jindal", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bing", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hady Wirawan", | |
| "middle": [], | |
| "last": "Lauw", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Proceedings of the 19th ACM International Conference on Information and Knowledge Management (CIKM 2010)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "939--948", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ee-Peng Lim, Viet-An Nguyen, Nitin Jindal, Bing Liu, and Hady Wirawan Lauw. 2010. Detecting prod- uct review spammers using rating behaviors. In Pro- ceedings of the 19th ACM International Conference on Information and Knowledge Management (CIKM 2010), pages 939-948.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "An information-theoretic definition of similarity", | |
| "authors": [ | |
| { | |
| "first": "Dekang", | |
| "middle": [], | |
| "last": "Lin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1998, | |
| "venue": "Proceedings of International Conference on Machine Learning", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Dekang Lin. 1998. An information-theoretic defini- tion of similarity. In Proceedings of International Conference on Machine Learning, Madison, Wis- consin, USA, July.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Sentiment Analysis and Opinion Mining", | |
| "authors": [ | |
| { | |
| "first": "Bing", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Synthesis Lectures on Human Language Technologies. Morgan & Claypool Publishers", | |
| "volume": "5", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Bing Liu. 2012. Sentiment Analysis and Opinion Min- ing, volume 5 of Synthesis Lectures on Human Lan- guage Technologies. Morgan & Claypool Publish- ers, May.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Multi-aspect sentiment analysis with topic models", | |
| "authors": [ | |
| { | |
| "first": "Bin", | |
| "middle": [], | |
| "last": "Lu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Myle", | |
| "middle": [], | |
| "last": "Ott", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Claire", | |
| "middle": [], | |
| "last": "Cardie", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Benjamin", | |
| "middle": [ | |
| "K" | |
| ], | |
| "last": "Tsou", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "ICDM Workshops", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "81--88", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Bin Lu, Myle Ott, Claire Cardie, and Benjamin K. Tsou. 2011. Multi-aspect sentiment analysis with topic models. In ICDM Workshops, pages 81-88.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Finding deceptive opinion spam by any stretch of the imagination", | |
| "authors": [ | |
| { | |
| "first": "Myle", | |
| "middle": [], | |
| "last": "Ott", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yejin", | |
| "middle": [], | |
| "last": "Choi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Claire", | |
| "middle": [], | |
| "last": "Cardie", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jeffrey", | |
| "middle": [ | |
| "T" | |
| ], | |
| "last": "Hancock", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "309--319", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Myle Ott, Yejin Choi, Claire Cardie, and Jeffrey T. Hancock. 2011. Finding deceptive opinion spam by any stretch of the imagination. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Tech- nologies, pages 309-319, Portland, Oregon, USA, June.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "Negative deceptive opinion spam", | |
| "authors": [ | |
| { | |
| "first": "Myle", | |
| "middle": [], | |
| "last": "Ott", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Claire", | |
| "middle": [], | |
| "last": "Cardie", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jeffrey", | |
| "middle": [ | |
| "T" | |
| ], | |
| "last": "Hancock", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Short Papers, Atlanta", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Myle Ott, Claire Cardie, and Jeffrey T. Hancock. 2013. Negative deceptive opinion spam. In Proceedings of the 2013 Conference of the North American Chap- ter of the Association for Computational Linguistics: Human Language Technologies, Short Papers, At- lanta, Georgia, USA, June. Association for Compu- tational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "The development and psychometric properties of LIWC2007", | |
| "authors": [ | |
| { | |
| "first": "James", | |
| "middle": [ | |
| "W" | |
| ], | |
| "last": "Pennebaker", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Cindy", | |
| "middle": [ | |
| "K" | |
| ], | |
| "last": "Chung", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Molly", | |
| "middle": [], | |
| "last": "Ireland", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Amy", | |
| "middle": [], | |
| "last": "Gonzales", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Roger", | |
| "middle": [ | |
| "J" | |
| ], | |
| "last": "Booth", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "James W. Pennebaker, Cindy K. Chung, Molly Ireland, Amy Gonzales, and Roger J. Booth. 2007. The development and psychometric proper- ties of LIWC2007. Technical report, Austin, TX, LIWC.Net.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "Improved inference for unlexicalized parsing", | |
| "authors": [ | |
| { | |
| "first": "Slav", | |
| "middle": [], | |
| "last": "Petrov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Dan", | |
| "middle": [], | |
| "last": "Klein", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2007, | |
| "venue": "Proceedings of Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics (NAACL-HLT 2007)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "404--411", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Slav Petrov and Dan Klein. 2007. Improved inference for unlexicalized parsing. In Proceedings of Hu- man Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics (NAACL-HLT 2007), pages 404-411, Rochester, New York, April. Asso- ciation for Computational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "Identification of truth and deception in text: Application of vector space model to rhetorical structure theory", | |
| "authors": [ | |
| { | |
| "first": "L", | |
| "middle": [], | |
| "last": "Victoria", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tatiana", | |
| "middle": [], | |
| "last": "Rubin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Vashchilko", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proceedings of the Workshop on Computational Approaches to Deception Detection", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "97--106", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Victoria L. Rubin and Tatiana Vashchilko. 2012. Iden- tification of truth and deception in text: Application of vector space model to rhetorical structure theory. In Proceedings of the Workshop on Computational Approaches to Deception Detection, pages 97-106, Avignon, France, April.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "A joint model of text and aspect ratings for sentiment summarization", | |
| "authors": [ | |
| { | |
| "first": "Ivan", | |
| "middle": [], | |
| "last": "Titov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ryan", | |
| "middle": [], | |
| "last": "Mcdonald", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2008, | |
| "venue": "Proceedings of the 46th Annual Meeting of the Association for Computational Linguistcs: Human Language Technologies (ACL-08: HLT)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "308--316", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ivan Titov and Ryan McDonald. 2008. A joint model of text and aspect ratings for sentiment summariza- tion. In Proceedings of the 46th Annual Meeting of the Association for Computational Linguistcs: Hu- man Language Technologies (ACL-08: HLT), pages 308-316, Columbus, Ohio, June. Association for Computational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "Distortion as a validation criterion in the identification of suspicious reviews", | |
| "authors": [ | |
| { | |
| "first": "Guangyu", | |
| "middle": [], | |
| "last": "Wu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Derek", | |
| "middle": [], | |
| "last": "Greene", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Barry", | |
| "middle": [], | |
| "last": "Smyth", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "P\u00e1draig", | |
| "middle": [], | |
| "last": "Cunningham", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2010, | |
| "venue": "Proceedings of the First Workshop on Social Media Analytics (SOMA 2010)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "10--13", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Guangyu Wu, Derek Greene, Barry Smyth, and P\u00e1draig Cunningham. 2010. Distortion as a validation cri- terion in the identification of suspicious reviews. In Proceedings of the First Workshop on Social Media Analytics (SOMA 2010), pages 10-13, New York, NY, USA.", | |
| "links": null | |
| }, | |
| "BIBREF20": { | |
| "ref_id": "b20", | |
| "title": "Comparison of deceptive and truthful travel reviews. In Information and Communication Technologies in Tourism", | |
| "authors": [ | |
| { | |
| "first": "Kyung-Hyan", | |
| "middle": [], | |
| "last": "Yoo", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ulrike", | |
| "middle": [], | |
| "last": "Gretzel", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "37--47", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Kyung-Hyan Yoo and Ulrike Gretzel. 2009. Com- parison of deceptive and truthful travel reviews. In Information and Communication Technologies in Tourism 2009, pages 37-47. Springer Vienna.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "uris": null, | |
| "type_str": "figure", | |
| "text": "Alignment compatibility features (a) Compatibility features for alignment Q \u2192 P, including m a i , s a i , and c a i for each of the N aspects {a 1 , . . . , a N } with the highest occurrences in P. (b) Compatibility features for alignment P \u2192 Q, including m a i , s a i , and c a i for each of the N aspects {a 1 , . . . , a N } with the highest occurrences in Q.", | |
| "num": null | |
| }, | |
| "FIGREF1": { | |
| "uris": null, | |
| "type_str": "figure", | |
| "text": "SYN: syntax features, (2) N-GRAM: unigram and bigram features, (3) 1-GRAM+SYN: unigram and syntax features, (4) 2-GRAM+SYN: bigram and syntax features, and (5) N-GRAM+SYN: unigram, bigram and syntax features, where the second is Ott et al.'s model 7 , and the third is Feng et al. (2012a)'s model. The results are shown in Table 3. Our reimplementation of Ott et al.'s model (N-GRAM) obtains 89.6% accuracy, while the accuracy of our reimplemented Feng et al. model (1-GRAM+SYN)", | |
| "num": null | |
| }, | |
| "TABREF3": { | |
| "num": null, | |
| "type_str": "table", | |
| "content": "<table><tr><td>Reviews</td><td>N Prec</td><td>Rec</td><td>F 1</td></tr><tr><td>used</td><td/><td/><td/></tr></table>", | |
| "text": ". Our approach was Acc 50 10 90.2* 92.5* 91.4* 91.3* 100 10 90.0* 92.8* 91.4* 91.3* 150 20 90.6* 92.0* 91.3* 91.3* 200 20 90.4* 91.8* 91.1* 91.0*", | |
| "html": null | |
| } | |
| } | |
| } | |
| } |