Seed-GuidedTopicModelforDocumentFilteringandClassification资源-CSDN文库

研究论文

178 浏览量 2021-02-07 13:14:13 上传评论收藏 2.58MB PDF 举报

资源推荐

资源详情

资源评论

Seed-Guided Topic Model for Document Filtering

and Classification

CHENLIANG LI and SHIQIAN CHEN, Wuhan University, China

JIAN XING, Hithink RoyalFlush Information Network Co., Ltd, China

AIXIN SUN, Nanyang technological University, Singapore

ZONGYANG MA, Microsoft (China) Co., Ltd, China

One important necessity is to lter out the irrelevant information and organize the relevant information into

meaningful categories. However, developing text classiers often requires a large number of labeled docu-

ments as training examples. Manually labeling documents is costly and time-consuming. More importantly, it

becomes unrealistic to know all the categories covered by the documents beforehand. Recently, a few methods

have been proposed to label documents by using a small set of relevant keywords for each category, known as

dataless text classication. In this article, we propose a seed-guided topic model for the dataless text ltering

and classication (named DFC). Given a collection of unlabeled documents, and for each specied category

a small set of seed words that are relevant to the semantic meaning of the category, DFC lters out the ir-

relevant documents and classies the relevant documents into the corresponding categories through topic

inuence. DFC models two kinds of topics: category-topics and general-topics. Also, there are two kinds of

category-topics: relevant-topics and irrelevant-topics. Each relevant-topic is associated with one specic cat-

egory, representing its semantic meaning. The irrelevant-topics represent the semantics of the unknown cat-

egories covered by the document collection. And the general-topics capture the global semantic information.

DFC assumes that each document is associated with a single category-topic and a mixture of general-topics.

A novelty of the model is that DFC learns the topics by exploiting the explicit word co-occurrence patterns

between the seed words and regular words (i.e., non-seed words) in the document collection. A document is

then ltered, or classied, based on its posterior category-topic assignment. Experiments on two widely used

datasets show that DFC consistently outperforms the state-of-the-art dataless text classiers for both classi-

cation with ltering and classication without ltering. In many tasks, DFC can also achieve comparable or

even better classication accuracy than the state-of-the-art supervised learning solutions. Our experimental

results further show that DFC is insensitive to the tuning parameters. Moreover, we conduct a thorough study

about the impact of seed words for existing dataless text classication techniques. The results reveal that it

This article is an extended version of Reference [28], a paper presented at the 25th International ACM CIKM Conference

(Indianapolis, IN, Oct. 24–28, 2016).

This research was supported by National Natural Science Foundation of China (Grant No. 61502344), Natural Science

Foundation of Hubei Province (Grant No. 2017CFB502), Natural Scientic Research Program of Wuhan University (Grants

No. 2042017kf0225 and No. 2042016kf0190), Academic Team Building Plan for Young Scholars from Wuhan University

(Grant No. Whu2016012) and Singapore Ministry of Education Academic Research Fund Tier 2 (Grant No. MOE2014-T2-

2-066).

Authors’ addresses: C. Li (corresponding author) and S. Chen, Wuhan University, School of Cyber Science and Engineering,

Bayi Road, Wuhan, Hubei, 430075, China; emails: {cllee, sqchen}@whu.edu.cn; J. Xing, Hithink RoyalFlush Information

Network Co., Ltd, Hangzhou, Zhejiang, 310023, China; email: xingjian@myhexin.com; A. Sun, Nanyang Technological

University, School of Computer Science and Engineering, 639798, Singapore; email: axsun@ntu.edu.sg; Z. Ma, Microsoft

(China) Co., Ltd, Soochow, 215123, China; email: mzyone@gmail.com.

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee

provided that copies are not made or distributed for prot or commercial advantage and that copies bear this notice and

the full citation on the rst page. Copyrights for components of this work owned by others than ACM must be honored.

Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires

prior specic permission and/or a fee. Request permissions from permissions@acm.org.

https://doi.org/10.1145/3238250

ACM Transactions on Information Systems, Vol. 37, No. 1, Article 9. Publication date: December 2018.

9:2 C. Li et al.

is not using more seed words but the document coverage of the seed words for the corresponding category

that aects the dataless classication performance.

CCS Concepts: • Information systems → Document topic models; Clustering and classication;

Additional Key Words and Phrases: Topic model, dataless classication, document ltering

ACM Reference format:

Chenliang Li, Shiqian Chen, Jian Xing, Aixin Sun, and Zongyang Ma. 2018. Seed-Guided Topic Model for

Document Filtering and Classication. ACM Trans. Inf. Syst. 37, 1, Article 9 (December 2018), 37 pages.

https://doi.org/10.1145/3238250

1 INTRODUCTION

With the advance of Information Technology, the tremendous amounts of textual information gen-

erated everyday is far beyond the scope that people can manage manually. The recent prevalence of

social media further exacerbates this information overload, because rich information about various

kinds of events, user opinions, and daily life activities are generated in an unprecedented speed.

This timely information breeds the new and dynamic information needs everywhere. For exam-

ple, a data analyst might need to track an emerging event by using a few relevant keywords [36].

A data-driven company often needs to conduct a focused and deep analysis on the documents of

the specied categories. Within these semantic applications, one fundamental task is to lter out

irrelevant information and organize relevant information into meaningful topical categories.

During the past decade, text classiers have become important tools in managing and analyz-

ing large document collections. Text classication refers to the task of assigning category labels

to documents based on their semantics. Due to its wide usage, text classication has been studied

intensively for many years. Existing solutions are mainly based on supervised learning techniques

that require tremendous human eort in annotating documents as labeled examples, as shown in

the upper part of Figure 1. To reduce the labeling eort, many semi-supervised algorithms have

been proposed for text classication [7, 34]. Considering the diversity of the documents in many

applications, constructing a relatively small training set required by the semi-supervised algo-

rithms remains very expensive. Recently, a number of dataless text classication methods have

been proposed [7, 9, 13, 14, 18, 21, 22, 28, 29, 38]. Instead of using labeled documents as training

examples, dataless methods only require a small set of relevant words for each category or label-

ing the topics learned from a standard LDA model [3], to build text classiers. As illustrated in the

lower part of Figure 1, dataless classiers do not require labeled documents, which saves a lot of

human eorts. It has been reported that a speed-up of up to ve times can be achieved to build a

dataless text classier with indistinguishable performance to a supervised classier, by assuming

that labeling a word is ve times faster than labeling a document [14]. These promising results sug-

gest that dataless text classication is a practical alternative to the supervised approaches, when

constructing the training documents is not an easy task. More importantly, the labeled documents

produced by a dataless classier can also be used as training examples to learn supervised text

classiers if necessary [28]. However, these existing dataless classication techniques do not con-

sider document ltering. That is, we like to retrieve all the documents relevant to a specied set

of categories from a given document collection, and organize these relevant documents into the

corresponding categories. With the existing dataless classiers, we need to provide all the cate-

gories and the corresponding seed words covered by the document collection. Unfortunately, it is

often unrealistic to foresee all possible categories covered by a document collection, since the doc-

uments streamed in are likely to cover dynamic topics. In an extreme case, the number of possible

categories covered by documents could be potentially limitless.

ACM Transactions on Information Systems, Vol. 37, No. 1, Article 9. Publication date: December 2018.

9:4 C. Li et al.

of the standard LDA [3], a probabilistic topic model (PTM) that implicitly infers the hidden topics

from the documents based on the higher-order word co-occurrence patterns [40]. However, con-

ventional PTMs like PLSA and LDA are unsupervised techniques that implicitly infer the hidden

topics based on word co-occurrences [3, 23]. It is dicult or even infeasible to lter and classify

documents in such a purely unsupervised manner.

Inspired by the recent success of the PTM-based dataless text classication techniques [9, 21, 22,

28], in this article, we propose a seed-guided topic model for dataless text ltering and classica-

tion (DFC). Given a collection of unlabeled documents, DFC is able to achieve the goal of ltering

and classication by taking only a few semantically relevant words for each category of interest

(called “seed words”). To enable document ltering, we model two sets of category-topics: relevant-

topics and irrelevant-topics. A one-to-one correspondence between relevant-topics and categories

of interest is made. That is, each relevant-topic is associated with one specic category of interest,

and vice versa. The relevant-topic is assumed to represent the meaning of that category.

Since

the documents relevant to the categories of interest could comprise a limited proportion of the

whole collection, irrelevant-topics are expected to model other categories covered by the irrele-

vant documents. In our earlier work [28], a topic model based dataless classication technique

(named STM) is proposed by using a set of general-topics to model the general semantics of the

whole document collection. Although the modeling of general and specic aspects of documents

was studied previously for information retrieval [8], it had been overlooked for dataless text clas-

sication in previous studies [9, 21, 22]. Our earlier work has proven that this model setting is

benecial for dataless classication performance. Following this modeling strategy, in DFC, we

also utilize a set of general-topics to represent the general semantic information. The task of l-

tering and classication is achieved by associating each document with a single category-topic

and a mixture of general-topics. The posterior category-topic assignment is then used to label the

document as a category of interest, or an irrelevant one. DFC also subsumes STM under some par-

ticular parameter settings. This means that DFC is also able to conduct dataless text classication

without ltering.

Seeking useful supervision from the seed words to precisely infer category-topics and general-

topics is vital to the ecacy of DFC. In other words, precise relevance estimation between a word

and a category-topic via a small set of seed words is crucial for DFC. It is noteworthy to underline

that no seed word can be provided for any irrelevant-topic, because the possible semantic cate-

gories covered by a document collection are unknown beforehand. However, the precise inference

for the irrelevant-topics is an essential factor for classication performance guarantee. Without

any supervision from the corresponding seed words, the model is hard to identify the irrelevant-

topics successfully, leading to inferior classication performance. Here, we devsie a simple but

eective mechanism to identify a set of pseudo seed words for each irrelevant-topic. Specically,

we resort to using standard LDA to extract the hidden topics for the document collection in an

unsupervised manner. Then, a relevance measure is proposed to calculate the distance between

each LDA hidden topic and all the seed words provided for the relevant-topics. After a heuristic

procedure to lter out noisy LDA hidden topics, the top topical words from the least relevant LDA

hidden topics are then considered as the pseudo seed words for the irrelevant-topics.

In contrast to the existing dataless classication methods that simply exploit the semantic guid-

ance provided by the seed words in an implicit way (i.e., the word co-occurrence information), we

adopt an explicit strategy to estimate the relevance between a word and a category-topic, and also

Category and relevant-topic are considered equivalent and exchangeable in this work when the context has no ambiguity.

A category-topic refers to either a relevant-topic or an irrelevant-topic in this work when the context is for DFC.

ACM Transactions on Information Systems, Vol. 37, No. 1, Article 9. Publication date: December 2018.

Seed-Guided Topic Model for Document Filtering and Classification 9:5

the initial category-topic distribution for each document. The estimated relevance is then utilized

to supervise the topic learning process of DFC. In particular, we investigate two mechanisms (i.e.,

Doc-Rel and Topic-Rel) to estimate the probability of a word being generated by a category-topic,

by measuring its correlations to the (pseudo) seed words of that category-topic based on either

document-level word co-occurrence or topical-level word co-occurrence information. We call the

words that are generated by a category-topic category words.

In summary, DFC conducts the document ltering and classication in a weakly supervised

manner, just as what humans do in learning to classify documents with just few words: (i) rst

to identify the highly relevant documents based on the given seed words of a category; (ii) then

based on these highly relevant documents, to collectively identify the category words in addition

to the seed words; (iii) next to use both the seed words and category words to nd new relevant

documents and new category words; the last step repeats until a global equilibrium is optimized.

We conduct extensive experiments on two datasets Reuters-10 and -20 Newsgroup, and compare

DFC with state-of-the-art dataless text classiers and supervised learning solutions. In terms of

classication accuracy measured by F

, our experimental results show that DFC outperforms all

the dataless competitors in almost all the tasks and performs better than the supervised classiers

sLDA and SVM in many tasks for both classication with ltering and classication without lter-

ing. We also conduct a comprehensive performance evaluation to analyze the impact of parameter

settings in DFC. The results show that the proposed DFC is reliable to a broad range of parameter

values, indicating its superiority in real scenarios.

We need to emphasize that all existing dataless text classication techniques rely solely on the

weak supervision provided by the seed words. That is, the quality of seed words plays a crucial

factor regarding the classication performance. It is intuitive that fewer seed words could carry

less semantic information for the classication. However, the study on seed words in the paradise

of dataless text classication is still missing. Here, an open question naturally arises: Will using

more seed words lead to better classication accuracy? If the answer to this question is no,then,

What criterion should we use to build a set of seed words for a category? In this work, we conduct a

thorough study with the aim of answering these two questions. We nd that using more seed words

may not produce better classication accuracy. Also, we empirically observe that more relevant

documents covered by a set of seed words under a category, better classication accuarcy can be

obtained. In an extreme case, given two seed words for a category, no performance gain could be

obtained by using both seed words over using either one, if the two words always appear together

in documents. That is, it is not the number of seed words that matters, but the document coverage

for that category. In summary, the main contributions of this article are listed as follows:

(1) We propose and formalize a new task of dataless text ltering and classifcation. To the

best of our knowledge, this is the rst work to classify documents into relevant categories

of interest, and lter out irrelevant documents in a dataless manner. To enable precise

inference of irrelevant-topics, we propose a novel mechanism to identify the pseudo seed

words for irrelevant topics in an unsupervised manner.

(2) DFC does not solely rely on the implicit word co-occurrence patterns to guide the category

inference process. Instead, we introduce two mechanisms to estimate the probability of a

word being generated by a category-topic. The estimation is based on the explicit word

co-occurrence patterns derived from the document collection.

(3) We empirically study the impact of seed words for dataless text classication techniques.

Our results suggest that using more seed words may not lead to better classication accu-

racy. Instead, the document coverage of the selected seed words correlates positively with

the classication accuracy.

ACM Transactions on Information Systems, Vol. 37, No. 1, Article 9. Publication date: December 2018.

剩余36页未读，继续阅读

评论收藏

内容反馈

weixin_38623080

粉丝: 5
资源: 1002

Seed-Guided Topic Model for Document Filtering and Classificatio...

最新资源

Seed-Guided Topic Model for Document Filtering and Classificatio...

topicmodels

guided filter

Self-Guided and Cross-Guided Learning for Few-Shot Seg.pptx

Person Search by Separated Modeling and A Mask-Guided Two-Stream CNN Model.pdf

Attention-guided CNN for image denoising.pptx

Unsupervised-Attention-guided-Image-to-Image-Translation-master.zip

A Confidence-Guided Automated System for Non-emergency Calls.pdf

Feature-guided Image Stippling

Multi-scale self-guided attention for medical image segmentation.pdf

Partition-Guided GANs.pdf

fast-guided-filter-code-v1_边缘提取_导向滤波_图像边缘提取_纹理_图像特征提取

2014-History-Guided conversational recommendation-WWW.pdf

Laravel开发-guided-image

Orientation-Guided Similarity Learning for Person Re-identification

何恺明导向滤波器MATLAB源码--guidedfilter.m

Module 2-11 - Guided Consolidation.PPT

顺序引导的分离表征学习在溃疡性结肠炎分类中的应用_Order-Guided Disentangled Representatio

PGDMPose guided deep model for pedestrian attribute recognition in surveillance

General coupled-mode analysis for gain-guided, index-antiguided fibers, and index-guided fibers

attribute-guided_GaN_

B-GAP-Behavior-Guided-Action-Prediction-and-Navigation-for-Autonomous-Driving:这是我们的论文“ B-GAP

Self-Guided-Network-for-Fast-Image-Denoising:ICCV 2019论文SGN的PyTorch实施

drools-wb-guided-template-editor-api-6.1.0.Beta3.zip

行人重识别-用于被遮挡行人重识别优化的姿态引导pose-guided特征对齐-附项目源码+流程教程-优质项目实战.zip

【ICLR2022】图神经网络复杂时间序列建模

Single-Image-Based Rain and Snow Removal Using Multi-guided Filter（CCF C类会议）

最新资源