INTRODUCTION
The importance of text mining applications is growing proportionally with the exponential growth of electronic text. Along with the growth of internet many other sources of electronic text have become really popular. With increasing penetration of internet, many forms of communication and interaction such as email, chat, newsgroups, blogs, discussion groups, scraps etc. have become increasingly popular. These generate huge amount of noisy text data everyday. Apart from these the other big contributors in the pool of electronic text documents are call centres and customer relationship management organizations in the form of call logs, call transcriptions, problem tickets, complaint emails etc., electronic text generated by Optical Character Recognition (OCR) process from hand written and printed documents and mobile text such as Short Message Service (SMS). Though the nature of each of these documents is different but there is a common thread between all of these—presence of noise.
An example of information extraction is the extraction of instances of corporate mergers, more formally MergerBetween(company1,company2,date), from an online news sentence such as: “Yesterday New-York based Foo Inc. announced their acquisition of Bar Corp. ” Opinion (product 1,good), from a blog post such as: “I absolutely liked the texture ofSheetK quilts.”
At superficial level, there are two ways for information extraction from noisy text. The first one is cleaning text by removing noise and then applying existing state of the art techniques for information extraction. There in lies the importance of techniques for automatically correcting noisy text. In this chapter, first we will review some work in the area of noisy text correction. The second approach is to devise extraction techniques which are robust with respect to noise. Later in this chapter, we will see how the task of information extraction is affected by noise.
NOISY TEXT CORRECTION
Before moving on to techniques for processing noisy text we will briefly introduce methods for correcting noisy text. One of the most common forms of noise in text is wrong spelling. Kukich provides a comprehensive survey of techniques pertaining to detecting and correcting spelling errors (Kukich, 1992). According to this survey, three types of nonword misspellings are typically found viz. typographic such as teh, speel, cognitive such as recieve, conspeeracy and phonetic such as abiss, nacherly. A distinction must be made between automatically detecting such errors and automatically correcting those errors. The latter is a much harder problem. Most of the recent work in this area is about correcting spelling mistakes automatically. Golding and Roth (Golding & Roth, 1999) proposed a combination of a variant of Winnow, a multiplicative weight-update algorithm and weighted majority voting for context sensitive spelling correction. Mangu and Brill (Mangu & Brill, 1997) have shown that a small set of human understandable rules is more meaningful than a large set of opaque features and weights. Hybrid methods capturing the context using trigrams of the parts-of-speech tags and a feature based method have also been proposed to handle context sensitive spelling correction (Golding & Schabes, 1996). There is a lot of work related to automatic correction of spelling errors (Agirre et. al., 1998), (Zamora et. al., 1983), (Golding, 1995). A complete bibliography of all the work related to spelling error detection and correction can be found in (Beebe, 2005). On a related note, automatic spelling error correction techniques have been applied for other applications such as semantic role labelling (Sang et. al., 2005).
There is also recent work on correcting the output of SMS text (Aw et. al., 2006) (Choudhury et. al., 2007), OCR errors (Nartker et. al., 2003) and ASR errors (Sarma & Palmer, 2004).
INFORMATION EXTRACTION FROM NOISY TEXT
The goal of Information Extraction (IE) is to automatically extract structured information from the unstructured documents. The extracted structured information has to be contextually and semantically well-defined data from a given domain. A typical application of IE is to scan a set of documents written in natural language and populate a database with the information extracted. The MUC (Message Understanding Conference) conference was one effort at codifying the IE task and expanding it (Chinchor, 1998).
There are two basic approaches to the design of IE systems. One comprises the knowledge engineering approach where a domain expert writes a set of rules to extract the sought after information. Typically the process of building the system is iterative whereby a set of rules is written, the system is run and the output examined to see how the system is performing. The domain expert then modifies the rules to overcome any under- or over-generation in the output. The second is the automatic training approach. This approach is similar to classification where the texts are appropriately annotated with the information being extracted. For example, if we would like to build a city name extractor, then the training set would include documents with all the city names marked. An IE system would be trained on this annotated corpus to learn the patterns that would help in extracting the necessary entities.
An information extraction system typically consists of natural language processing steps such as morphological processing, lexical processing and syntactic analysis. These include stemming to reduce inflected forms of words to their stem, parts of speech tagging to assign labels such as noun, verb, etc. to each word and parsing to determine the grammatical structure of sentences.
Named Entity Annotation of Web Posts
Extraction of named entities is a key IE task. It seeks to locate and classify atomic elements in the text into predefined categories such as the names of persons, organizations, locations, expressions of times, quantities, monetary values, percentages, etc. Entity recognition systems either use rule based techniques or statistical models. Typically a parser or a parts of speech tagger identifies elements such as nouns, noun phrases, or pronouns. These elements along with surface forms of the text are used to define templates for extracting the named entities. For example, to tag company names it would be desirable to look at noun phrases that contain the words company or incorporated in them. These rules can be automatically learnt using a tagged corpus or could be defined manually. Most known approaches do this on clean well formed text. However, named entity annotation of web posts such as online classifieds, product listings etc. is harder because these texts are not grammatical or well written. In such cases reference sets have been used to annotate parts of the posts (Michelson & Knoblock, 2005). The reference set is thought of as a relational set of data with a defined schema and consistent attribute values. Posts are now matched to their nearest records in the reference set. In the biological domain gene name annotation, even though it is performed on well written scientific articles, can be thought of in the context of noise, because many gene names overlap with common English words or biomedical terms. There have been studies on the performance of the gene name annotator when trained on noisy data (Vlachos, 2006).
Information Extraction from OCRed Documents
Documents obtained from OCR may have not only unknown words and compound words, but also incorrect words due to OCR errors. In their work Miller et. al. (Miller et. al., 2000) have measured the effect of OCR noise on IE performance. Many IE methods work directly on the document image to avoid errors resulting from converting to text. They adopt keyword matching by searching for string patterns and then use global document models consisting of keyword models and their logical relationships to achieve robustness in matching (Lu & Tan, 2004). The presence of OCR errors has a detrimental effect on information access from these documents (Taghva et. al., 2004). However, post processing of these documents to correct these errors exist and have been shown to give large improvements.
Information Extraction from ASRed Documents
The output of an ASR system does not contain case information and punctuations. It has been shown that in the absence of punctuations extraction of different syntactic entities like parts of speech and noun phrases is not accurate (Nasukawa et. al., 2007). So IE from ASRed documents becomes harder. Miller et. al. (Miller et. al., 2000) have shown how IE performance varies with ASR noise. It has been shown that it is possible to build aggregate models from ASR data (Roy & Subramaniam, 2006). In this work topical models are constructed by utilizing inter document redundancy to overcome the noise. In this work only a few natural language processing steps have been used. Phrases have been aggregated over the noisy collection to get to the clean underlying text.
FUTURE TRENDS
More and more data from sources like chat, conversations, blogs, discussion groups need to be mined to capture opinions, trends, issues and opportunities. These forms of communication encourage informal language which can be considered noisy due to spelling errors, grammatical errors and informal writing styles. Companies are interested in mining such data to observe customer preferences and improve customer satisfaction. Online agents need to be able to understand web posts to take actions and communicate with other agents. Customers are interested in collated product reviews from web posts of other users. The nature of the noisy text warrants moving beyond traditional text analytics techniques. There is need for developing natural language processing techniques that are robust to noise. Also techniques that implicitly and explicitly tackle textual noise need to be developed.
CONCLUSION
In this chapter we have looked at information extraction from noisy text. This topic is gaining in importance as more and more noisy data gets generated and useful information needs to be obtained from this. We have presented a survey of existing techniques information extraction techniques. We have also presented some of the future trends in noisy text analytics.
KEY TERMS
Automatic Speech Recognition: Machine recognition and conversion of spoken words into text.
Data Mining: The application of analytical methods and tools to data for the purpose of identifying patterns, relationships or obtaining systems that perform useful tasks such as classification, prediction, estimation, or affinity grouping.
Information Extraction: Automatic extraction of structured knowledge from unstructured documents.
Knowledge Extraction: Explicitation of the internal knowledge of a system or set of data in a way that is easily interpretable by the user.
Noisy Text: Text with any kind of difference in the surface form, from the intended, correct or original text.
Optical Character Recognition: Translation of images of handwritten or typewritten text (usually captured by a scanner) into machine-editable text.
Rule Induction: Process of learning, from cases or instances, if-then rule relationships that consist of an antecedent (if-part, defining the preconditions or coverage of the rule) and a consequent (then-part, stating a classification, prediction, or other expression of a property that holds for cases defined in the antecedent).
TextAnalytics: The process of extracting useful and structured knowledge from unstructured documents to find useful associations and insights.
