chunk_id stringlengths 3 7 | chunk stringlengths 1 823 | source_url stringclasses 416
values | title stringclasses 416
values | chunk_idx int64 0 294 | chunk_start_char int64 0 139k | chunk_end_char int64 303 139k |
|---|---|---|---|---|---|---|
1_29 | This defines a Markov chain on the candidate entities, that we will use to propagate the local evidence.
Classifying entities in context
BIBREF22 first combine the local features into a local evidence score, and then spread this local evidence using the Markov chain:
$$
G(d) = (\alpha I + (1 - \alpha ) M_d)^k \c... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 29 | 13,999 | 14,485 |
1_30 | Let $F$ be the matrix of all local features for each candidate entity: $F = (F(e_1,d[s_1]),
\dots , F(e_n, d[s_n]))$ . After $k$ iterations in the Markov chain, this defines features $M_d^k F$ . | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 30 | 14,485 | 14,680 |
1_31 | Rather than relying on these features for a fixed number of steps $k$ , we record the features at each step, which defines the vector $(F, M_d \cdot F, M_d^2 \cdot F, \dots , M_d^k \cdot F)$
This alleviates the need for an $\alpha $ parameter while keeping the number of features small. We train a linear support vect... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 31 | 14,680 | 15,208 |
1_32 |
Experimental setup
Most entity linking datasets are annotated against DBpedia or YAGO. Wikidata contains items which do not have any corresponding Wikipedia article (in any language), so these items do not have any DBpedia or YAGO URI either. Therefore, converting an entity linking dataset from DBpedia to Wikidata r... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 32 | 15,208 | 15,795 |
1_33 | We first translated all DBpedia URIs to Wikidata items. Then, we used OpenRefine BIBREF24 to extract the entities marked not covered by DBpedia and matched them against Wikidata. After human review, this added 63 new links to the 524 converted from DBpedia (out of 476 out-of-KB entities).
We also annotated a new data... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 33 | 15,795 | 16,290 |
1_34 | In this dataset, only 64 of the 2,624 Wikidata mentions do not have a corresponding DBpedia URI.
We use the Wikidata JSON dump of 2018-02-24 for our experiments, indexed with Solr (Lucene). We restrict the index to humans, organizations and locations, by selecting only items whose type was a subclass of (P279) human ... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 34 | 16,290 | 16,671 |
1_35 | Labels and aliases in all languages are added to a case-sensitive FST index.
We trained our classifier and its hyper-parameters by five-fold cross-validation on the training sets of the ISTEX and RSS datasets. We used GERBIL BIBREF23 to evaluate OpenTapioca against other approaches. We report the InKB micro and macro... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 35 | 16,671 | 17,233 |
1_36 | Our restriction to people, locations and organizations probably helps in this regard and we anticipate worse performance for broader domains. Our approach works best for scientific affiliations, where spelling is more canonical than in newswire. The availability of Twitter identifiers directly in Wikidata helps us to ... | https://arxiv.org/abs/1904.09131 | OpenTapioca: Lightweight Entity Linking for Wikidata | 36 | 17,233 | 17,886 |
2_0 | Spotting Rumors via Novelty Detection
Rumour detection is hard because the most accurate systems operate retrospectively, only recognising rumours once they have collected repeated signals. By then the rumours might have already spread and caused harm. We introduce a new category of features based on novelty, tailored... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 0 | 0 | 563 |
2_1 | Additionally we introduce pseudo feedback, which assumes that documents that are similar to previous rumours, are more likely to also be a rumour. Comparison with other real-time approaches shows that novelty based features in conjunction with pseudo feedback perform significantly better, when detecting rumours instan... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 1 | 563 | 1,168 |
2_2 | In contrast to traditional news wire, information on social media is spread without intensive investigation, fact and background checking. The combination of ease and fast pace of sharing information provides a fertile breeding ground for rumours, false- and disinformation. Social media users tend to share controversi... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 2 | 1,168 | 1,760 |
2_3 |
The cases in Germany and Austria in 2016, show how misleading and false information about crimes committed by refugees negatively influenced the opinion of citizens.
Detecting these rumours allows debunking them to prevent them from further spreading and causing harm. The further a rumour has spread, the more likely... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 3 | 1,760 | 2,327 |
2_4 |
Rumour detection on social media is challenging due to the short texts, creative lexical variations and high volume of the streams. The task becomes even harder if we attempt to perform rumour detection on-the-fly, without looking into the future. We provide an effective and highly scalable approach to detect rumours... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 4 | 2,327 | 3,049 |
2_5 | In a nutshell, documents that are similar to previously detected rumours are considered to be more likely to also be a rumour. The proposed features can be computed in constant time and space allowing us to process high-volume streams in real-time (Muthukrishnan, 2005). Our experiments reveal that novelty based featur... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 5 | 3,049 | 3,636 |
2_6 | Unconfirmed (novel) information with respect to trusted resources is considered as an indication of rumours.
Pseudo Feedback for Detection/Classification
Pseudo feedback increases detection accuracy by harnessing repeated signals, without the need of retrospective operation.
Related Work
Before rumour detection, s... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 6 | 3,636 | 4,258 |
2_7 | The most successful methods proposed focus on classification harnessing lexical, user-centric, propagation-based (Wu et. al, 2015) and cluster-based (Cai et. al, 2014; Liu et. al, 2015; Zhao et. al, 2015) features.
Many of these context based features originate from a study by Castillo et. al (2011), which pioneered ... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 7 | 4,258 | 4,659 |
2_8 | They observed a significant correlation between the trustworthiness of a tweet with context-based characteristics including hashtags, punctuation characters and sentiment polarity. When assessing the credibility of a tweet, they also assessed the source of its information by constructing features based on provided URL... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 8 | 4,659 | 5,380 |
2_9 | al (2010) found that the topology of a distrustful tweet's propagation pattern differs from those of news and normal tweets. These findings along with the fact that rumours tend to more likely be questioned by responses than news paved the way for future research examining propagation graphs and clustering methods (Ca... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 9 | 5,380 | 5,886 |
2_10 | al, 2015), sentiment analysis (Qazvinian et. al, 2011; Wu et. al, 2015) as well as propagation graphs (Kwon, et. al, 2013; Wang et. al, 2015).
Recent research mainly focuses on further improving the quality of rumour detection while neglecting the increasing delay between the publication and detection of a rumour. Th... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 10 | 5,886 | 6,311 |
2_11 | Unfortunately, state-of-the-art systems operate in a retrospective manner, meaning they detect rumours long after they have spread. The most accurate systems rely on features based on propagation graphs and clustering techniques. These features can only detect rumours after the rumours have spread and already caused h... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 11 | 6,311 | 6,816 |
2_12 | Their focus on latency aware rumour detection makes their approaches conceptually related to ours. Zhao et. al (1015) found clustering tweets containing enquiry patterns as an indication of rumours. Also clustering tweets by keywords and subsequently judging rumours using an ensemble model that combine user, propagati... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 12 | 6,816 | 7,400 |
2_13 | The approach with the lowest latency banks on the 'wisdom of the crowd' (Liu et. al, 2015). In addition to traditional context and user based features they also rely on clustering micro-blogs by their topicality to identify conflicting claims, which indicate increased likelihood of rumours. Although they claim to oper... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 13 | 7,400 | 7,970 |
2_14 |
Rumour Detection
Rumour detection is a challenging task, as it requires determining the truth of information (Zhao et. al, 2015). The Cambridge dictionary, defines a rumour as information of doubtful or unconfirmed truth. We rely on classification using an SVM, which is the state-of-the-art approach for novelty dete... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 14 | 7,970 | 8,472 |
2_15 | Unfortunately, the currently most successful features rely on information based on graph propagation and clustering, which can only be computed retrospectively. This renders them close to useless when detecting rumours early on. We introduce two new classes of features, one based on novelty, the other on pseudo feedba... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 15 | 8,472 | 9,160 |
2_16 |
More formally, we denote by $d_t$ the document that arrives from stream $S:\lbrace d_0, d_1, . . . d_n\rbrace $ at time $t$ . Upon arrival of document $d_t$ we compute its corresponding feature vector $f_{d,t}$ . Given $f_{d,t}$ and the previously obtained weigh vector $w$ we compute the rumour score $RS_{d,t} = w^T ... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 16 | 9,160 | 9,589 |
2_17 | We predict that message $d_t$ is likely to become a rumour if its rumour score exceeds the detection threshold $S:\lbrace d_0, d_1, . . . d_n\rbrace $0 . The optimal parameter setting for weight vector $S:\lbrace d_0, d_1, . . . d_n\rbrace $1 and detection threshold $S:\lbrace d_0, d_1, . . . d_n\rbrace $2 are learned... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 17 | 9,589 | 9,952 |
2_18 |
Novelty-based Features
To increase instantaneous detection performance, we compensate for the absence of future information by consulting additional data sources. In particular, we make use of news wire articles, which are considered to be of high credibility. This is reasonable as according to Petrovic et. al (2013... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 18 | 9,952 | 10,505 |
2_19 | In a nutshell, the presence of information unconfirmed by the official media is construed as an indication of being a rumour. Note that this closely resembles the definition of what a rumour is.
Novelty Feature Construction
High volume streams demand highly efficient feature computation. This applies in particular t... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 19 | 10,505 | 11,186 |
2_20 | To make vector proximity applicable, we slide a term-level based window, whose length resembles the average social media message length, through each of the news articles. This results in sub-documents whose length resembles those of social media messages. Novelty is computed using term weighted tf-idf dot products be... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 20 | 11,186 | 11,737 |
2_21 | al, 2015), a recent advance in novelty detection that improved the efficiency by an order of magnitude without sacrificing effectiveness. Kterm hashing computes novelty non-comparatively. Instead of measuring similarity between documents, a single representation of previously seen information is constructed. For each ... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 21 | 11,737 | 12,415 |
2_22 | This filter then captures the combined information of all trusted resources. A single representation allows computing novelty with a single step, instead of comparing each social media message individually with all trusted resources.
When kterm hashing was introduced by Wurzer et. al (2015) for novelty detection on E... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 22 | 12,415 | 13,024 |
2_23 | This allows us to compute a dedicated weight for kterms based on these top 10 keywords. The distinction in weights between kterms based on all versus keyword yields superior rumour detection quality, as described in section "Feature analysis" . This leaves us with a total of 6 novelty based features for kterm hashing ... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 23 | 13,024 | 13,537 |
2_24 | The focus lies on features that can be computed instantly based only on the text of a message to keep the latency of our approach to a minimum. Most of these 51 features overlap with previous studies (Castillo et. al, 2011; Liu et. al, 2015; Qazvinian et. al, 2011; Yang et. al, 2012; Zhao et. al, 2015). | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 24 | 13,537 | 13,842 |
2_25 | This includes features based on the presence or number of URLs, hash-tags and user-names, POS tags, punctuation characters as well as 8 different categories of sentiment and emotions.
On the arrival of a new message from a stream, all its features are computed and linearly combined using weights obtained from an SVM ... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 25 | 13,842 | 14,453 |
2_26 | The feature is conceptually related to pseudo relevance feedback found in retrieval and ranking tasks in IR. The concept builds upon the idea that documents, which reveal similar characteristics as previously detected rumours are also likely to be a rumour. During detection, feedback about which of the previous docume... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 26 | 14,453 | 15,122 |
2_27 | Conceptually, PF passes on evidence to repeated signals by increasing the rumour score of future documents if they are similar to a recently detected rumour. Note that this allows harnessing information from repeated signals without the need of operating retrospectively.
Training Pseudo Feedback Features
The trainin... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 27 | 15,122 | 15,756 |
2_28 | This provides a model for all but the PF features. Then the trainings set is processed to computing rumour scores based on the model obtained from our initial trainings round. This time, we additionally compute the PF feature value by measuring the minimum distance in term space between the current document vector and... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 28 | 15,756 | 16,393 |
2_29 | Once we obtained the value for the PF feature, we compute its weight using the SVM. The combination of the weight for the PF feature with the weights for all other features, obtained in the initial trainings round, resembles the final model.
Experiments
The previous sections introduced two new categories of features... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 29 | 16,393 | 17,012 |
2_30 | Messages with high rumour scores are considered likely being rumours. The classification decision is based on an optimal thresholding strategy based on the trainings set.
Evaluation metrics
We report accuracy to evaluate effectiveness, as is usual in the literature (Zhou et. al, 2015). Additionally we use the standa... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 30 | 17,012 | 17,472 |
2_31 | This procedure evaluates detection tasks using Detection Error Trade-off (DET) curves, which show the trade-off between miss and false alarm probability. By visualizing the full range of thresholds, DET plots provide a more comprehensive illustration of effectiveness than single value metrics (Allan et. al, 2000). We ... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 31 | 17,472 | 18,148 |
2_32 | We therefore followed previous researchers like Liu et. al (2015) and Yang et. al (2012) and created our own dataset.
trusted resources: We randomly collected 200 news articles about broad topics commonly reported by news wires over our target time period. These range from news about celebrities and disasters to fina... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 32 | 18,148 | 18,641 |
2_33 | To ensure a fair evaluation, we collected the news articles before judging rumours, not knowing which rumours we would find later on. We also only consider news articles published before the timestamps of the social media messages.
For our social media stream, we chose Sina Weibo, a Chinese social media service with ... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 33 | 18,641 | 19,198 |
2_34 | al (2015), we use this service to obtain a high quality set of 202 confirmed rumours.
non-rumours: We additionally gathered 202 non-rumours using the public Sina Weibo API. Three human annotators judged these weibos based on unanimous decision making to ensure that they don't contain rumours.
Since we operate in a s... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 34 | 19,198 | 19,667 |
2_35 |
We ordered the rumours and non-rumours chronologically and divided them in half, forming a training and test set. We ensured that each of the sets consists of 50% rumours and non-rumours. This is important when effectiveness is measured by accuracy. All training and optimization use the trainings set. Performance is ... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 35 | 19,667 | 20,229 |
2_36 | al (2012), which we re-implemented. We chose the algorithm by Yang et. al (2012), dubbed Yang, because they proposed a feature set for early detection tailored to Sina Weibo and were used as a state-of-the-art baseline before by Liu et. al (2015). The algorithm by Liu et. al (2015), dubbed Liu, is said to operate in r... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 36 | 20,229 | 20,633 |
2_37 | Both apply various message-, user-, topic- and propagation-based features and rely on an SVM classifier which they also found to perform best. The approaches advertise themselves as suitable for early or real-time detection and performed rumour detection with the smallest latency across all published methods. Yang per... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 37 | 20,633 | 21,304 |
2_38 |
Table 2 compares the performance of our features with the two classifiers on the 101 rumours and 101 non-rumours of the test set, when detecting rumour instantly after their publication. The table reveals comparable accuracy for Yang and Liu at around 60%. Our observed performance of Yang matches those by Liu et. al ... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 38 | 21,304 | 21,813 |
2_39 | al (2015) report performance based on the first 5 messages which clearly outperforms Yang for early rumour detection. However, we find that when reducing the set from 5 to 1, their superiority is only marginal. In contrast, the combination of novelty and pseudo relevance based features performs significantly better (s... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 39 | 21,813 | 22,336 |
2_40 | In particular for instantaneous rumour detection, where information can only be obtained from a single message, the use of external data proves to perform superior. Note that accuracy is a single value metric describing performance at an optimal threshold. Figure 1 compares the effectiveness of the three algorithms fo... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 40 | 22,336 | 23,043 |
2_41 |
When increasing the detection delay to 12 and 24 hours, all three algorithms reach comparable performance with no statistically significant difference, as seen in table 4. For our approach, none of the features are computed retrospectively, which explains why the performance does not change when increasing the detect... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 41 | 23,043 | 23,685 |
2_42 |
Feature analysis
We group our 57 features into 7 categories shown in Table 6 and analyse their contribution using feature ablation, as seen in Table 5 . Feature ablation illustrates the importance of a feature by measuring performance, when removing it from the set of features. Novelty related features based on kter... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 42 | 23,685 | 24,090 |
2_43 | 'Sentence char' features, which include punctuation, hashtags, user-symbols and URLs, contributed the most of the traditional features, followed by Part of Speech ('POS') and 'extreme word' features. Our experiments found 'sentiment' and 'emotion' based features to contribute the least. Since excluding them both resul... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 43 | 24,090 | 24,718 |
2_44 | This is interesting, as when kterm hashing was introduced (Wurzer et. al, 2015), all kterms were considered as equally important. We found that prioritising certain kterms yields increased performance.
Interestingly, novelty based features computed by the vector similarity between weibos and news sub-documents perfor... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 44 | 24,718 | 25,278 |
2_45 | Pulling all information into a single representation bridges the gab between documents and allows finding information matches within documents. We hypothesize that this causes increased detection performance.
Pseudo Feedbaack
Features ablation revealed that pseudo feedback (PF) increased detection performance by 5.3... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 45 | 25,278 | 25,985 |
2_46 | Beside causing a detection delay these methods are also blind to less popular rumours that don't go viral. In contrast, novelty based feature require only a single message enabling them to detect even the smallest rumours. Examples for such small rumours are shown in table 3 .
Efficiency and Scalability
To demonstra... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 46 | 25,985 | 26,567 |
2_47 | We measure the throughput on an idle machine and average the observed performance over 5 runs. Figure 2 presents performance when processing more and more weibos. The average throughput of our system is around 7,000 weibos per second, which clearly exceeds the average volume of the full Twitter (5,700 tweets/sec.) and... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 47 | 26,567 | 27,107 |
2_48 | Figure 2 also illustrates that our proposed features can be computed in constant time with respect to the number of messages processed. This is crucial to keep operation in a true streaming environment feasible. Approaches, whose runtime depend on the number of documents processed become progressively slower, which is... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 48 | 27,107 | 27,756 |
2_49 | Novelty based features consider the increased presence of unconfirmed information within a message with respect to trusted sources as an indication of being a rumour. Pseudo feedback features consider messages that are similar to previously detected rumours as more likely to also be a rumour. Pseudo feedback and its v... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 49 | 27,756 | 28,476 |
2_50 | We also showed that the proposed features can be computed efficiently enough to operate on the average Twitter and Sina Weibo stream while keeping time and space requirements constant. | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 50 | 28,476 | 28,661 |
2_51 |
Table 1: Excerpt of topics with synopsis of corresponding rumours
Figure 1: DET plot, revealing superior effectiveness of our approach for instant rumour detection for the full range of thresholds
Table 4: Detection accuracy at different levels of delay; Asterisk indicates significance (p < 0.05)
Table 2: Effectiv... | https://arxiv.org/abs/1611.06322 | Spotting Rumors via Novelty Detection | 51 | 28,661 | 29,243 |
3_0 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves
We propose Sentence Level Recurrent Topic Model (SLRTM), a new topic model that assumes the generation of each word within a sentence to depend on both the topic of the sentence and the whole history of its preceding words in the sentence. Diffe... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 0 | 0 | 631 |
3_1 | Furthermore, SLRTM can automatically generate sentences given a topic (i.e., topics to sentences), which is a key technology for real world applications such as personalized short text conversation.
Introduction
Statistic topic models such as Latent Dirichlet Allocation (LDA) and its variants BIBREF0 , BIBREF1 , BIB... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 1 | 631 | 1,039 |
3_2 | In these models, a word token in a document is assumed to be generated by a hidden mixture model, where the hidden variables are the topic indexes for each word and the topic assignments for words are related to document level topic weights. Due to the effectiveness and efficiency in modeling the document generation p... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 2 | 1,039 | 1,557 |
3_3 |
Most topic models take the bag-of-words assumption, in which every document is treated as an unordered set of words and the word tokens in such a document are sampled independently with each other. The bag-of-words assumption brings computational convenience, however, it sacrifices the characterization of sequential ... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 3 | 1,557 | 2,258 |
3_4 |
There have been some works trying to solve the aforementioned problems, although still insufficiently. For example, several sentence level topic models BIBREF9 , BIBREF10 , BIBREF11 tackle the topic coherence problem by assuming all the words in a sentence to share the same topic (i.e., every sentence has only one to... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 4 | 2,258 | 2,850 |
3_5 | For another example, in BIBREF12 , the embedding based neural language model BIBREF13 , BIBREF14 , BIBREF15 and topic model are integrated. They assume the generation of a given word in a sentence to depend on its local context (including its preceding words within a fixed window) as well as the topics of the sentence... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 5 | 2,850 | 3,349 |
3_6 | Furthermore, there is no explicit coherence constraints on the word topics and sentence topics, since every word can have its own topics in their model.
We propose Sentence Level Recurrent Topic Model (SLRTM) to tackle the limitations of the aforementioned works. In the new model, we assume the words in the same sent... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 6 | 3,349 | 3,883 |
3_7 | Specifically, for a particular word INLINEFORM0 within a sentence INLINEFORM1 , we assume its generation depends on two factors: the first is the whole set of its historical words in the sentence and the second is the sentence topic, which we regard as a pseudo word and has its own distributed representations. We use ... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 7 | 3,883 | 4,378 |
3_8 |
With the proposed SLRTM, we can not only model the document generation process more accurately, but also construct new natural sentences that are coherent with a given topic (we call it topic2sentence, similar to image2sentece BIBREF19 ). Topic2sentence has its huge potential for many real world tasks. For example, i... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 8 | 4,378 | 4,953 |
3_9 |
We have conducted experiments to compare SLRTM with several strong topic model baselines on two tasks: generative model evaluation (i.e. test set perplexity) and document classification. The results on several benchmark datasets quantitatively demonstrate SLRTM's advantages in modeling documents. We further provide s... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 9 | 4,953 | 5,633 |
3_10 | Several variants of LDA have been developed such as hierarchical topic models BIBREF22 and supervised topic models BIBREF3 . With the recent development of deep learning, there are also neural network based topic models such as BIBREF23 , BIBREF24 , BIBREF25 , BIBREF26 , which use distributed representations of words ... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 10 | 5,633 | 6,210 |
3_11 |
There are some efforts trying to address the limitations of the bag-of-words assumption. For example, in BIBREF27 , both semantic (i.e., related with topics) and syntactic properties of words were modeled. After that, a hidden Markov transition model for topics was proposed BIBREF9 , in which all the words in a sente... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 11 | 6,210 | 6,682 |
3_12 | Although these works have made some meaningful attempts on topic coherence and sequential dependency across sentences, they have not sufficiently model the sequential dependency of words within a sentence. To address this problem, the authors of BIBREF12 adopted the neural language model technology BIBREF13 to enhance... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 12 | 6,682 | 7,368 |
3_13 | The embedding vectors sampled from the GMM are further used to generate words in a sentence according to a feedforward neural network. To be specific, the preceding words in a fixed sized window, together with the sentence and document, act as the context to generate the next word by a softmax conditional distribution... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 13 | 7,368 | 7,875 |
3_14 |
Another line of research related to our model is Recurrent Neural Network (RNN), especially some recently developed effective RNN models such as Long Short Term Memory BIBREF17 and Gated Recurrent Unit BIBREF18 . These new RNN models characterize long range dependencies for a sequence, and has been widely adopted in ... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 14 | 7,875 | 8,294 |
3_15 | In particular, for language modeling tasks, it has been shown that RNN (and its variants such as LSTM) is much more effective than simple feedforward neural networks with fixed window size BIBREF16 given that it can model dependencies with nearly arbitrary length.
Sentence Level Recurrent Topic Model
In this section... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 15 | 8,294 | 8,758 |
3_16 |
With the three points in mind, let us introduce the detailed generative process of SLRTM, as well as the stochastic variational inference and learning algorithm for SLRTM in the following subsections.
The generative process
Suppose we have INLINEFORM0 topics, INLINEFORM1 words contained in dictionary INLINEFORM2 , ... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 16 | 8,758 | 9,263 |
3_17 | Similar to LDA, we assume there is a INLINEFORM10 -dimensional Dirichlet prior distribution INLINEFORM11 for topic mixture weights of each document. | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 17 | 9,263 | 9,412 |
3_18 | With these notations, the generative process for document INLINEFORM12 can be written as below:
Sample the multinomial parameter INLINEFORM0 from INLINEFORM1 ;
For the INLINEFORM0 th sentence of document INLINEFORM1 INLINEFORM2 , INLINEFORM3 , where INLINEFORM4 is the INLINEFORM5 th word for INLINEFORM6 :
Draw the... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 18 | 9,412 | 9,979 |
3_19 | the corresponding items. For example, INLINEFORM0 and INLINEFORM1 denote the embeddings for word INLINEFORM2 and topic INLINEFORM3 , respectively. INLINEFORM4 is a zero vector and INLINEFORM5 is a fake starting word. Function INLINEFORM6 is the LSTM unit to generate hidden states, for which we omit the details due to ... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 19 | 9,979 | 10,554 |
3_20 |
Then the probability of observing document INLINEFORM0 can be written as: DISPLAYFORM0
where INLINEFORM0 is the probability of generating sentence INLINEFORM1 under topic INLINEFORM2 , and it is decomposed through the probability chain rule; INLINEFORM3 is specified in equation ( EQREF11 ) and ( EQREF12 ); INLINEFO... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 20 | 10,554 | 11,029 |
3_21 |
To sum up, we use Figure FIGREF14 to illustrate the generative process of SLRTM, from which we can see that in SLRTM, the historical words and topic of the sentence jointly affect the LSTM hidden state and the next word.
Stochastic Variational Inference and Learning
As the computation of the true posterior of hidde... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 21 | 11,029 | 11,461 |
3_22 | Particularly, we use multinomial distribution INLINEFORM0 and Dirichlet distribution INLINEFORM1 as the variational distribution for the hidden variables INLINEFORM2 and INLINEFORM3 , and we denote the variational parameters for document INLINEFORM4 as INLINEFORM5 , with the subscript INLINEFORM6 omitted. Then the var... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 22 | 11,461 | 11,938 |
3_23 |
The introduction of LSTM-RNN makes the optimization of ( EQREF16 ) computationally expensive, since we need to update both the model parameters INLINEFORM0 and variational parameters INLINEFORM1 after scanning the whole corpus. Considering that mini-batch (containing several sentences) inference and training are nece... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 23 | 11,938 | 12,529 |
3_24 | The execution of the whole inference and learning process includes several epochs of iteration over all documents INLINEFORM2 with Algorithm SECREF15 (starting with INLINEFORM3 ).
[ht] Stochastic Variational EM for SLRTM Input: document INLINEFORM0 , variation parameters INLINEFORM1 , and model weights INLINEFORM2 . ... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 24 | 12,529 | 13,097 |
3_25 | M-Step: Compute the gradient INLINEFORM1 by LSTM backward pass. Use INLINEFORM2 to obtain INLINEFORM3 by stochastic gradient descent methods such as Adagrad BIBREF30 . In Algorithm SECREF15 , INLINEFORM4 is the digamma function. Equation ( EQREF18 ) guarantees the estimate of INLINEFORM5 is unbiased. In equation (), I... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 25 | 13,097 | 13,520 |
3_26 | Due to space limit, we omit the derivation details for the updating equations in Algorithm SECREF15 , as well as the forward/backward pass details for LSTM BIBREF17 .
Experiments
We report our experimental results in this section. Our experiments include two parts: (1) quantitative experiments, including a generativ... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 26 | 13,520 | 14,091 |
3_27 |
Quantitative Results
We compare SLRTM with several state-of-the-art topic models on two tasks: generative document evaluation and document classification. The former task is to investigate the generation capability of the models, while the latter is to show the representation ability of the models.
We base our expe... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 27 | 14,091 | 14,668 |
3_28 |
Wiki10+ BIBREF31 , which contains Web documents from Wikipedia, each of which is associated with several tags such as philosophy, software, and music. Following BIBREF25 , we kept the most frequent 25 tags and removed those documents without any of these tags, forming a training set and a test set with 11,164 and 6,1... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 28 | 14,668 | 15,114 |
3_29 | Wiki10+ contains much more words per document (i.e., 1,704) than 20Newsgroup (i.e., 135).
We followed the practice in many previous works and removed infrequent words. After that, the dictionary contains about INLINEFORM0 unique words for 20Newsgroup and INLINEFORM1 for Wiki10+. We adopted the NLTK sentence tokenizer... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 29 | 15,114 | 15,576 |
3_30 | LDA is the classic topic model, and we used GibbsLDA++ for its implementation.
Doc-NADE BIBREF24 . Doc-NADE is a representative neural network based topic model. We used the open-source code provided by the authors.
HTMM BIBREF9 . HTMM models consider the sentence level Markov transitions. Similar to Doc-NADE, the i... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 30 | 15,576 | 16,125 |
3_31 |
For SLRTM, we implemented it in C++ using Eigen and Intel MKL. For the sake of fairness, similar to BIBREF12 , we set the word embedding size, topic embedding size, and LSTM hidden layer size to be 128, 128, and 600 respectively. In the experiment, we tested the performances of SLRTM and the baselines with respect to... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 31 | 16,125 | 16,506 |
3_32 | In initialization (values of INLINEFORM2 and INLINEFORM3 ), the LSTM weight matrices were initialized as orthogonal matrices, the word/topic embeddings were randomly sampled from the uniform distribution INLINEFORM4 and are fined-tuned through the training process, INLINEFORM5 and INLINEFORM6 were both set to INLINEFO... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 32 | 16,506 | 17,068 |
3_33 | Gradient clipping with a clip value of 20 was used during the optimization of LSTM weights. Asynchronous stochastic gradient descent BIBREF32 with Adagrad was used to perform multi-thread parallel training.
We measure the performances of different topic models according to the perplexity per word on the test set, def... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 33 | 17,068 | 17,536 |
3_34 | Based on the table, we have the following discussions:
Our proposed SLRTM consistently outperforms the baseline models by significant margins, showing its outstanding ability in modelling the generative process of documents. In fact, as tested in our further verifications, the perplexity of SLRTM is close to that of ... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 34 | 17,536 | 18,077 |
3_35 |
Models that consider sequential property within sentences (i.e., GMNTM and SLRTM) are generally better than other models, which verifies the importance of words' sequential information. Furthermore, LSTM-RNN is much better in modelling such a sequential dependency than standard feed-forward networks with fixed words ... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 35 | 18,077 | 18,667 |
3_36 | For 20Newsgroup, we used the multi-class logistic regression classifier and used accuracy as the evaluation criterion. For Wiki10+, since multiple labels (tags) might be associated with each document, we used logistic regression for each label and the classification result is measured by Micro- INLINEFORM1 score BIBRE... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 36 | 18,667 | 19,262 |
3_37 | We can further find that the embedding based methods (Doc-NADE, GMNTM and SLRTM) generate better document representations than other models, demonstrating the representative power of neural networks based on distributed representations. In addition, when the training data is larger (i.e., with more sentences per docum... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 37 | 19,262 | 19,833 |
3_38 |
Qualitative Results
In this subsection, we demonstrate the capability of SLRTM in generating reasonable and understandable sentences given particular topics. In the experiment, we trained a larger SLRTM with 128 topics on a randomly sampled INLINEFORM0 Wikipedia documents in the year of 2010 with average 275 words p... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 38 | 19,833 | 20,323 |
3_39 | The size of word embedding, topic embedding and RNN hidden layer are set to 512, 1024 and 1024, respectively.
We used two different mechanisms in sentence generating. The first mechanism is random sampling new word INLINEFORM0 at every time step INLINEFORM1 from the probability distribution defined in equation ( EQRE... | https://arxiv.org/abs/1604.02038 | Sentence Level Recurrent Topic Model: Letting Topics Speak for Themselves | 39 | 20,323 | 20,808 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.