Spreadsheet as Syntax

If we look at a spreadsheet as a document we can develop a parser that locates the user-supplied information in that document. Like any language, we'll have to treat some elements of the spreadsheet as pure syntax, little more than punctuation. In the case of embedded formulas, these are elaborate punctuation that is an important part of overall usability.

We'll look at parsing an XML version of a spreadsheet, since that is a delightfully simple exercise in Python. Looking back at "The Problem with Spreadsheets", we're looking at simplifying some of the business processes, and saving the spreadsheets as XML files so that we can write application programs to extract meaningful information from the spreadsheets.

Generally, parsing happens at two levels: lexical analysis and syntax analysis. A lexical scanner will break the stream of input characters into discrete tokens. The syntax analysis looks at the stream of tokens to create a usable representation, often called an abstract syntax tree (AST).

Layers of Meaning.

Compounding the essential parsing problem are the number of layers of syntax that we have to work with:
⁃ XML. XML lexical analysis locates tags, and punctuation, building up a document object model of Nodes, Elements and Attributes.
⁃ Spreadsheet. The spreadsheet analysis interprets the XML nodes are parts of a Workbook, with Worksheets, Rows and Cells.
⁃ Business Process Objects. What we observe in the spreadsheet is really an implementation of some essential business process. There are business objects which have been written down in the cells of the spreadsheet. The mapping from business entity to spreadsheet representation is sometimes haphazard. The objects we find in the spreadsheet rows and cells are something we'll call Business Process Objects. They aren't the real essence of the processing, they're just one possible implementation.
⁃ Essential Business Entities. These are the fundamental business entities that we are really interested in. These are plans for selling or making product, commitments from vendors, prospective customers, financial arrangements, etc. These get encoded in a number of forms, one of which is Business Process Objects in a spreadsheet.

Since there are multiple levels of meaning, there are multiple abstract syntax trees that we will be developing. In essence we work from the bottom up, accreting information to create the essential business entities. Our SAX parser will examine the XML syntax. We can use a generic Document Object Model (DOM) for the XML structure.

We can traverse the DOM to uncover spreadsheet objects. We can traverse the spreadsheet objects to uncover the Business Process Objects (BPO). Typically, we can optimize some of these traversal steps and locate the BPO's while traversing the DOM. The essential business entities, however, can present a more complex problem. Business Entities often exist as dimensions in a star schema, leading us toward conforming the BPO's into a data warehouse for analysis purposes.

Basic SAX Parsing.

The Python SAX and DOM libraries can be used to build a complete DOM from an XML document. For our purposes, we'll elide some features which aren't a common part of parsing spreadsheets. This xml.sax.ContentHandler does the minimal work required to create a DOM structure that contains our spreadsheet XML. We'll have to do more to make sense of this DOM structure.

import xml.sax, xml.dom.minidom

class DOMContentHandler( xml.sax.ContentHandler ):
    """A SAX ContentHandler which creates a simple DOM object."""
    def __init__( self ):
        """Initialize this handler."""
        self.dom= xml.dom.minidom.getDOMImplementation()
        self.document= self.dom.createDocument(None,None,None)
        self.context= [ self.document ]
    def startElement( self, name, attrs ):
        """Starts a new Element, pushing this onto the stack."""
        new= self.document.createElement( name )
        for k,v in attrs.items():
            new.setAttribute( k, v )
        self.context[-1].appendChild( new )
        self.context.append( new )
    def endElement( self, name ):
        """Ends an Element, popping the matching Element off the stack."""
        while self.context[-1].nodeName  != name:
            self.context.pop( -1 )
    def characters( self, content ):
        """Captures characters, creating a Text Node."""
        new= self.document.createTextNode( content )
        self.context[-1].appendChild( new )
    def getDocument( self ):
        """Returns the DOM object."""
        return self.document

To use this DOMContentHandler we have to create a SAX parser and initiate parsing on an input source. Note that this creates a generic DOM object, which we have to examine to locate the spreadsheet structure.

class XMLParser( FileParser ):
    """SAX2 parsing of entire DOM."""
    def __init__( self, aFile ):
        """Initialize parsing a given XML file which contains a SpreadSheet."""
        FileParser.__init__( self, aFile )
        handler= DOMContentHandler()
        myReader= xml.sax.parse( self.file, handler )
        self.xmlDoc= handler.getDocument()

The Spreadsheet Object Model (SSOM)

The low-level DOM structure is a large collection of Element and Text Nodes with Attribute values. It's much more pleasant to work with proper Worksheet, Row and Cell objects. We can condense the DOM into something more usable through a fairly simple algorithm that makes good use of Python's generators.

First, we'll need class definitions for the various spreadsheet entities we're going to deal with:
- Style. The spreadsheet style information is spread through a number of tags and attributes. Most of the tag values are single-occurence elements and the list-oriented getElementsByTagName() and getAttribute() aren't the most convenient API.
⁃ Workbook. The Workbook is a collection of Styles, Names, and "SupBook" references to external files. It is also a collection of Worksheets.
⁃ WorkbookLink. This is the Path information of a "SupBook" reference in a Workbook.
⁃ Worksheet. A Worksheet is essentially a container for Rows. It can also be looked at as a container for columns, but we won't often need to make use of this representation. Unlike the pure XML model, which is row-oriented, our spreadsheet object model can include additional indexing.
⁃ Row. A Row is a container for Cells.
⁃ Cell. A Cell has a number of attributes: data, an optional formula, and an optional style. We'll need these to parse the spreadsheet document. We can safely ignore any other attributes of a cell.
⁃ WorksheetReference. This is an reference embedded in a formula. These references are usually a subset of the "SubBook" references.

Most of these class definitions are relatively simple. They are containers with basic accessor methods to put and get specific components. A Workbook, for example, uses a number of dictionaries to keep Names, Styles and Worksheets. A Row is little more than a simple list of Cells.

However, since these are formal containers, the accessors create a very convenient API for accessing spreadsheet structure. Rather than a complex XPATH expression to locate a given cell of a given row of a given sheet, we can provide a pleasant method in the Workbook class to locate the Sheet, delegating the row and cell lookup. Within Worksheet we can locate the requested Row, and within Row we locate the requested Cell.

Generating SSOM Objects.

We can handle the creation of SSOM objects via Python generators. Here's a method that yields the top-level Worksheet instances. It also assures that each Worksheet is properly contained in the parent Workbook.

    def nextSheet( self ):
        """Generator which yields the next, empty L{Worksheet}."""
        wsList= self.xmlDoc.getElementsByTagName("Worksheet")
        for ws in wsList:
            self.currentWS= ws
            name= ws.getAttribute( "ss:Name" )
            self.sheet= Worksheet( name )
            self.document.addSheet( self.sheet )
            yield self.sheet

Our final application can use this method something like the following. This snippet looks for worksheets named "Assumptions", and examines only those pages of a Workbook.

    parse= ssDOM.XMLParser( file(aFile,'r') )
    doc= parse.getWorkbook()
    for ws in parse.nextSheet():
        if ws.name != "Assumptions":
        print ws

Generating the Rows (and Cells) is somewhat more complex because of the very rich information content in an individual Cell. However, the essential processing is pretty straightforward. We need to collect all the Cells within a row, along with any style or comments associated with the Cell. Then we bundle it into a SSDOM object that we can work with in our next level of parsing.

    def nextRow( self ):
        """Generator which yields a complete L{Row}."""
        for table in self.currentWS.getElementsByTagName("Table"):
            self.rowNumber= 0
            for row in table.getElementsByTagName("Row"):
                aRow= Row()
                for cell in row.getElementsByTagName("Cell"):
                    styleID= cell.getAttribute( "ss:StyleID" )
                    index= cell.getAttribute( "ss:Index" )
                    formula= cell.getAttribute( "ss:Formula" )
                    data = cell.getElementsByTagName("Data")+cell.getElementsByTagName("ss:Data")
					... some thrashing omitted ...
                    aRow.addCell( Cell(text,formula,style) )
                self.sheet.addRow( aRow )
                yield aRow
                self.rowNumber += 1

Business Process Objects.

Once we've got Rows, Cells and Worksheets, we can then do useful analysis of the resulting spreadsheet to locate the user's inputs. As one example, we'll peel the assumptions off the assumptions worksheet in the workbook.

In the following case, we're only looking at the Worksheet named assumptions. The worksheet contains a formula which repeats some identifying information on this worksheet. We'll need that to establish some business context for the following data.

The rows of the sheet have some instructions and examples, which we have to skip. Once we get to the "Summary P&L Assumptions" cell, everything below that will be user-entered Assumptions. The interesting part of the parsing is recognizing headings for areas that group the assumptions assumptions and the detailed assumptions within an area. The headings for an area have a style that involves a color plus a single underline. Unlike parsing text, where we simply compare strings, here we have to compare one attribute of a cell's style to see if the cell has special meaning.

    for ws in parse.nextSheet():
        if ws.name != "Assumptions":
        print ws
        rowIter= parse.nextRow()
        for row in rowIter:
            if row[0].formula == u"='Summary PNL LC'!RC":
                print "Title:", row[0].data
            if row[0].data == u"Summary P&L Assumptions":
                # Keep rows after the assumptions header.
        area= ""        
        for row in rowIter:
            if row[0].data.startswith( u"Examples " ):
            if len(row) == 0 or len(row[0].data) == 0:
            if len(row) == 1 and row[0].style.fontUnderline == u"Single":
                area= row[0].data
                txt= "; ".join( [ c.data for c in row if c.data ] )
                a= Assumption( area, txt, w_fk )
                print a

Since iterators maintain state, we can use the iterator to implement a very clean Skip The Headers design pattern. The first "for row in rowIter" loop will process rows until we find the last of the "overhead" rows. The second "for row in rowIter" loop will process the remaining rows; we skip blank rows and rows that contain examples.

Essential Business Entities.

Our Business Process Object may be the essential business entity or it may only be selected attributes of a more complex entity. In this case, the Assumption object that we parsed is just a puddle of text, and not very interesting. It is, however, a label on a more complete business model, which pervades the spreadsheet.

By accumulating the individual BPO's, we can accrete enough information to reconstruct the business model which is implemented as a spreadsheet. Extracting the parameters from this model is the heart of what our spreadsheet parser is doing.

Posted: Thursday - January 25, 2007 at 11:11 AM

Technorati Tags:
Technorati Cosmos: Technorati Cosmos
Technorati Watchlist: Technorati Watchlist
Add this entry to: