Data Ingestion of Guide Commericiale

Project Plan & Milestones

REQUIRED: Project plan and milestones (5%) (>300 words)

Timeline

Timeframe	Task	Completion
Week 4	Explore and define possible approaches to the problem Explore the existing Guida Commerciale documents	✓
Week 5	Establish a pipeline for the project Explore pricing plans for OpenAI GPT models and Geocoding platforms Set up the GitHub repository	✓
Week 6	Start to experiment with different OCR models and potential alternatives, such as PDF text extraction libraries Attempt to cluster pages based on different page types using an unsupervised learning approach	✓
Week 7	Autumn Break	✓
Week 8	Create a new approach/pipeline that relies on basic domain knowledge to obtain better results Set up the semantic annotation platform, INCEpTION, for the new approach	✓
Week 9	Complete midterm presentation Run OCR with Pytesseract Extract text using PDFPlumber Compare the two text extraction approaches, choose the most suitable approach, and pre-process the text output	✓
Week 10	Set up existing model of annotations completed on the 1864 Guida Commericiale Use INCEpTION to complete semantic annotation and named-entity recognition of 1853 Guida Commericiale Convert this into a CSV file for further processing	✓
Week 11	Complete processing and cleaning of the CSV file Occupation Broadcasting Logical Segmentation Parish/Church Reconciliation Complete mapping of parishes to districts (sestiere)	✓
Week 12	Plot the district and number on a map for all the entries in the 1853 Guida Commericiale Obtain accuracy metrics including accuracy of PDFPlumber and the entire pipeline Finalize potential approaches to analysis	✓
Week 13	Complete analysis based on the map entries plotted Derive conclusions based on the analysis and report findings
Week 14	Complete the written deliverables and the GitHub repository Prepare the final presentation discussing the results

Motivation & Deliverables

REQUIRED FOR PROJECT: Motivation and description of the deliverables (5%) (>300 words)

Methods

REQUIRED: Detailed description of the methods (5%) (>500 words)

During our work, we approached our problem in two ways: in a more generalizable way that could process any guide commercial, and a more streamlined way that would require more manual annotation to improve results. We started with our first approach, [and worked on this for the first couple months], but came to realize its margin for error was for too high, making us pivot toward our second approach. The points below highlight each step of the way for both these approaches.

Approach 1: General

Pipeline Overview

Pipeline Overview This process aims to extract, annotate, standardize, and map data from historical documents. By leveraging OCR, named entity recognition (NER), and geocoding, the pipeline converts unstructured text into geographically and semantically meaningful data.

Pipeline

1. Process pages by:
   1. Convert batch of pages into images
   2. Pre-process images for OCR using CV2 and Pillow
   3. Use Pytesseract for OCR to convert image to string
2. Perform Named Entity Recognition by:
   1. Prompt GPT-4o to identify names, professions, and addresses, and to turn this into entries in a table format.
   2. Standardize addresses that include abbreviations, shorthand, etc.
   3. Append “Venice, Italy” to the end of addresses to ensure a feasible location
3. Geocode addresses by:
   1. Use a geocoding API like LocationHQ to convert the address to coordinates
   2. Take top result of search, and append to the map

Divide Up Pages

Since we wanted this solution to work for any guide commercial, not just the ones provided to us, we intended to develop a way to parse through a document, identifying pages of interest that had usable data. While we got a working solution that was able to weed out pages with absolutely no information, we were never able to find a reliable solution for finding pages with the specific type of data we wanted. A possible idea we had was to train a model for identifying patterns within each page and grouping them off of that, but due to timeline restrictions we decided against it. Because of this, we never ended up finalizing this step for approach 1, and proceeded with the rest of our pipeline.

Perform OCR

To start, we used PDF Plumber for extracting pages from our document in order to turn them into images. With the goal of extracting text from these pages, we then used libraries such as CV2 and Pillow in order to pre-process each image with filtering and thresholding to improve clarity, and lastly performed OCR using Pytesseract, tweaking certain settings to accommodate the old scripture better.

Named Entity Recognition

After having extracted all text from a page, we can then prompt GPT-4o to identify names, professions, and addresses, and turn our text into a table format with an entry for each person. Furthermore, since these documents use abbreviations and shorthand for simplifying their writing process, we need to undo these by replacing them with their full meaning. Finally, to make geocoding in the future possible, we need to specify that each entry belongs to Venice by adding "Venice, Italy" to the end of each address.

Geocoding

For geocoding, we found LocationHQ to be a feasible solution for our cause, successfully mapping addresses to close by locations within Venice. In order to implement this into our project, we attempted to query each address in our table. While some queries narrowed down the request to a single address, some would result in multiple possible options. Our heuristic for these scenarios was to simply pick the top option, as it seemed to be the closest match each time. Finally, after gathering the coordinates for our addresses, we would map them onto a visual map that we ended up not proceeding with in the end.

Approach 2: Streamlined

Pipeline Overview

This approach outlines a systematic process for extracting, annotating, and analyzing data from historical Venetian guide commercials. The aim is to convert unstructured document data into a clean, structured dataset that supports geographic mapping and data analysis.

Pipeline

1. Manually inspect entire document for page ranges with digestible data
2. Perform text extraction using PDF Plumber
3. Semantic annotation with INCEpTION, separating our pages into entries with first names, last names, occupations, addresses, etc.
4. Clean and format data
   1. Logical segmentation
   2. Handling rows with multiple LOC_PAR and LOC_STR values
   3. Handling rows with multiple names
   4. Occupation broadcasting
   5. Parish and Church Cleaning/Reconciliation
5. Map parish to districts using dictionary
6. Plot on a map using parish and number from entry
7. Perform data analysis.

Find page ranges of interest

The main difference with this approach compared to approach 1 was that instead of relying on a program to group our pages, we manually went through our document and found a range of pages that were of interest to us. This allowed us to go into the next steps of our pipeline with confidence, knowing we don't need to deal with the uncertainty of whether our page had irrelevant data or not.

Perform Text Extraction

Furthermore, after testing both OCR and text extraction, we found that using PDF Plumber's function for extracting text from the PDF directly worked much more reliably. Because of this, we ended up using this for all of our pages, resulting in a similar output to the OCR with a much higher accuracy. This also freed us up from requiring to perform pre-processing for images, making the pipeline much simpler.

Semantic Annotation

For semantic analysis, we steered toward using INCEpTION in place of GPT-4o, largely due to previous projects already having trained a model for extracting names, addresses, and professions from similar text documents. We were able to leverage the existing model to speed up the semantic annotation process for the 1853 Guida Commerciale. The pre-trained model allowed us to annotate entries similarly formatted to the annotations in the model. Despite this, we did require some manual intervention when annotating occupations due to the variations in the format of the different Guida Commerciale documents. Another large benefit from this was that we no longer needed to rely on an external API for this, allowing us to bypass the limitations of financing our requests. As such, we proceeded with semantic annotation and entity recognition on the relevant pages. Using the text files produced from the text extraction process, we could tag keywords with tags related to an individual’s first name, last name, occupation, street location, parish name, and door number. After completing the semantic annotation, we saved the semantic annotations as a CONLL 2002 file to further clean, format, and process the data.

Data Clean Up

After performing the annotations and obtaining a CONLL 2002 file for each page in the 1853 Guida Commerciale on which we performed text extraction, we converted the CONLL 2002 files into a single CSV file using conll2002toCsv.py to further clean and format the data.

The resulting CSV file is loaded into a Pandas DataFrame. After this, duplicate rows in the DataFrame are dropped to ensure we have unique entries. After this, we set up a contextual order for each row, starting from the "FIELD" and "OCC" tags (representing occupations) and ending with the "NUM" tag (representing the door number) and reordered the DataFrame in this order. The specific order for entries is "["FIELD", "OCC", "LAST_N", "FIRST_N", "PER_TITLE", "PER_GRP", "PER_COMPL", "LOC_ISLA", "LOC_STR", "LOC_PAR", "NUM"]".

Next, we perform two rounds of logical segmentation, iterating over rows of the Pandas DataFrame and attempting to reconcile rows with incomplete or logically segmented data. This merges rows where address components such as "NUM," "LOC_STR," and "LOC_PAR" are split across multiple rows. Any empty or redundant rows are removed after merging. This process ensures that logically connected segments are consolidated into a single row.

Following this, we refine the segmentation process by addressing rows with multiple LOC_PAR and LOC_STR values. Since these are logical components of an address, we ensure these fragmented components are correctly associated with their respective entries.

Then, we handle cases with multiple individuals or names associated with a single entry. This ensures that all information related to individuals or groups of individuals is consolidated into a single row in the Pandas DataFrame.

The next step involves Occupation Broadcasting. In this step, we focus on broadcasting occupation information down to subsequent rows where it is missing, as long as the previous entry has a valid occupation entry. This ensures that every relevant entry has an occupation associated with it. This is an important step because the 1853 Guida Commerciale lists the occupation once and then lists all the individuals who work in that specific occupation.

Another important step is cleaning the LOC_PAR (parish) values to ensure we can accurately map the entries to their respective districts based on the parish-to-district dictionary. Some common discrepancies and errors exist due to the text extraction process, such as spelling mistakes and inaccurate text extraction. In addition to this, parishes and churches have different spellings that must be accounted for when a parish is extracted from PDFPlumber. Common errors also include inconsistent abbreviations such as "s .", "s.", "S .", "ss .", etc., were removed from the parish value. In addition to this, we ensured all parishes were in lowercase and stripped of any trailing spaces. After this, we added another column to represent the district and applied the existing parish-to-district dictionary to ensure all the parishes with corresponding values were mapped.

Plot on a Map

After mapping the parish and church names into districts, we plot the entries on the map by georeferencing with the door numbers. We begin by loading a GeoJSON file with single points for all the valid door numbers and district entries with their corresponding coordinates on a map of Venice. After this, we convert this GeoJSON file into a GeoDataFrame for data manipulation. In addition, we load an updated version of the entries that have been appropriately cleaned and formatted and consist of all the district mappings. We rename the columns in the DataFrame consisting of the 1853 Guida Commerciale entries to match the columns in the GeoJSON file with points for the door numbers to ensure we can merge the two DataFrames. After this, using an inner join, we merge the door GeoDataFrame with the 1853 Guida Commerciale entries DataFrame on the house number and district. We then filter out rows without an occupation, district, or house number to ensure we only plot relevant points to the 1853 Guida Commerciale. We update the GeoJSON to ensure each point contains all the other relevant information, such as name, occupation, and address. After this, we plot this GeoJSON file using Folium and utilize other software to view it for the data analysis component.

Quality Assessment & Limitations

REQUIRED: Quality assessment and discussion of limitations (5%) (>300 words)

Quality Assessment

To gain insight into the accuracy of our overall pipeline as well as each step of it, below we have listed each step and the results from manual testing.

Text Extraction

For our text extraction, we directly pulled the text from our PDF document using PDFPlumber's built-in functions. As these documents were previously parsed and OCRd by Google, the accuracy of the OCR is much higher compared to doing it ourselves. Even then, we still noticed errors, and below are our findings.

Page #	Total Words	Incorrect Words	Accuracy %
163	312	20	93.59
183	256	2	99.22
220	282	16	94.33
240	278	19	93.17
260	310	10	96.77
Total	1438	67	95.34

Table 1 Accuracy measures for 5 randomly sampled pages.

Looking at the performance of our text extraction on five randomly chosen pages, we manually cross-checked every word on these pages against the original document, and counted the number of errors we found. As for the heuristic of what constitutes as an error, the criteria goes as follows:

• If a word does not show up as it is in the original document, it counts as one error.
• If 4 words are clumped together in the PDFPlumber when they should have spaces in between, this counts as 4 errors.
• If a word has an incorrect letter, it counts as 1 error
• If any marks are generated that were not intentionally in the old document, that counts as 1 error.

Conclusion

Our final total accuracy comes out to 95.34%.

Most errors were due to certain words being hand written too closely together, and then the OCR done by Google analyzed them to be together when they should be two distinct words. Similarly, certain commas were written slightly too far from the word, leading to the OCR separating the comma from the word it should be in front of. Only two of these errors were due to a letter being misinterpreted, giving us good confidence that the OCR was reliable.

The figures below show examples of the errors we saw during our measuring process. Fig TODO and TODO show the two times that the OCR misinterpreted a character and gave an output that was different from the actual document. In Fig TODO, the "g" above the "Z" extended in a way where it was read as "Ž", and in Fig TODO, the "7" was written so low that the OCR read it as a dash ("-"). As for Fig TODO, even though it might not look too close to the human eye, the OCR interpreted these two words to be too close to each other, resulting in "s.Martino" instead of the correct "s. Martino".

Fig TODO: Z read as "Ž"

Fig TODO: 7 read as "-"

Fig TODO: Words being too close together

Semantic Annotation

For our semantic annotation, we decided to use INCEpTION for semantically annotating each of our pages, using a trained model provided to us by the instructors of the course. This model had been used on a similar Guide Commerciale file, and served as a great starting point for us to build off of.

Page #	Total Entries	Incorrect Entries	Accuracy %
163	39	2	94.87
183	31	2	93.55
220	37	1	97.30
240	30	0	100.00
260	38	0	100.00
Total	175	5	95.34

Table 1 Accuracy measures for 5 randomly sampled pages.

Looking at the performance of this model on five randomly chosen manually cleaned up pages from our text extraction process, we manually checked each entry on these pages, and counted the number of errors we found.

Conclusion

Our final total accuracy comes out to 97.14%

Mapping accuracy

In the final stages of the pipeline, two types of errors can occur:

1. The business fails to be mapped to a coordinate.
2. The business successfully maps to a coordinate, but the coordinate is incorrect.

Error 1: Failure to Map

First, it is important to understand the total number of entries that need mapping. To do so, we review the process of data cleanup and how that impacts the number of entries. Initially, the CSV file had 4357 entries, which was reduced to 4346 entries after dropping duplicate entries. After the first and second rounds of logical segmentation, we have 3811 and 3769 entries, respectively. After handling multiple LOC_STR and LOC_PAR values in one entry, we have 3696 entries. After handling multiple people in one entry, we have 3642 entries. After completing occupation broadcasting, we have 3604 entries. Finally, after performing other miscellaneous cleaning and parish/church reconciliation, we still have 3604 entries. Of the 3604 entries, 3602 entries have an occupation value. Some other interesting findings that showcase the limitations of the semantic annotation and PDFPlumber text extraction are that only 3486 entries have a last name, and 3395 entries have a first name. In addition to this, 3365 entries have a first and last name. Another interesting finding is that 3412 entries have a parish value.

During the initial phase of mapping parishes to district values, we realized that 645 entries with a parish value did not map to a district. During this process, we noticed some errors with our data cleaning as our cleaning step did not account for “m formosa” (136 entries), “m gloriosa” (94 entries), etc. They refer to “maria formosa” and “maria gloriosa dei frari” respectively. After manually inspecting the entries that did not map, we changed our cleaning and formatting steps to improve the % of parishes that are mapped to a district. After making fixes, we increased the % of entries with a parish that maps to a dictionary from 81.09% (2767 entries mapped) to 95.78% (3268 entries mapped).

The specific corrections included:

• 136 cases attributed to "m formosa" — corrected to "maria formosa" and mapped to Castello.

• 94 cases for "m gloriosa" — corrected to "maria gloriosa dei frari" and mapped to San Polo.

• 71 cases for "maria del gigliolio" — corrected to "maria del giglio" and mapped to San Marco.

• 63 cases for "m del carmine" — possibly related to "maria dei carmini", mapping uncertain.

• 42 cases for "m del rosario" — corrected to "madonna del rosario" and mapped to Dorsoduro.

• 31 cases for "pietro di" — corrected to "pietro di castello" or "pietro" and mapped to Castello.

• 21 cases for "gio battista in bragora" and 7 cases for "gio battista" — corrected to "giovanni battista in bragora" and mapped to Castello.

Following these manual corrections, only 4.22% (n = 144) of the data remained unmappable. These remaining errors were due to major spelling errors that occurred once. For time's sake, we avoided fixing 144 specific mistakes and wanted to showcase some errors due to text extraction and semantic annotation.

Some other interesting statistics are that we had 3093 entries with a house number and district and 3227 entries with a house number and parish. At the end of the pipeline, 3469 points were plotted. At first, this may seem alarming. However, after inspecting the existing single-door GeoJSON, we realized there are multiple entries with the same house number, district, first name, last name, occupation, parish, etc., as house numbers can be divided further into segments such as A, B, C, etc. In our model, we only merge the overall house number and district. As such, we plot a larger number of points. After further analyzing these points, we discovered that there are 501 unique duplicate entries. When summed up, this leads to 1304 total duplicate entries, as each unique duplicate entry can have multiple duplicates. Therefore, the extra number of entries that are plotted is equal to 803. Therefore, the total number of points/entries plotted is 2666. For all entries with a house number and district, 86.19% (accuracy) of them were plotted on the map. For all entries with a house number and parish, 82.61% (accuracy) of them were plotted. This discrepancy exists because not all parishes were mapped to the district. Of all the entries, including those without a parish, district, occupation, or house number value, 73.9% are plotted.

Due to PDF text extraction or semantic annotation errors, some entries could not be matched because they didn’t have a house number or a parish. The parish is needed to obtain the district. Other entries with a parish were not mapped because they don’t map to an existing district due to spelling errors, PDF text extraction errors, semantic annotation errors, or if the existing dictionary does not have the mapping. Some entries could not be mapped because they have an invalid number (string).

There are 3139 entries with a valid house number formatted correctly. This means there are 465 entries without a valid formatted house number. 295 of the entries had no house number value. 175 entries had invalid formatted house numbers due to semantic annotation or PDF text extraction errors. 30 of these errors with the entries were due to the text extraction, causing numbers to be recognized as letters and causing numbers to be concatenated with other characters. The remaining 145 errors were due to the semantic annotations being incorrect. There were 2940 entries with a district and a correctly formatted house number. Therefore, we achieved an accuracy of 90.68% for entries with a district and a correctly formatted house number. The remaining entries weren’t matched because there wasn’t a corresponding entry in the single-point GeoJSON file with the same district and house number.

Error 2: Incorrect Mapping

To evaluate if the data was plotted correctly, the 1808 guide commercial was used to validate the locations plotted in 1853. Firstly, a buffer zone was established around the known street locations in 1808. Business locations falling outside this zone were considered incorrectly plotted. Roughly 96% of the data was contained within a 5-metre buffer of the known street locations in 1808. Secondly, the expected location name was compared to the street name in the 1853 dataset. This validation involved checking if each location was within a 5-metre buffer and had the same street name, and using Levenshtein distance to account for small inaccuracies. With a Levenshtein distance of 3 and a buffer of 5 metres, only 10% of the data was completely correctly mapped. Closer examination revealed that some locations in 1808 were the same as those recorded differently in 1853, such as "Barbari" and "San Barbaria". These were manually checked and conservatively edited.

With the same distance and buffer settings, a total of 45% of the data was accurately plotted. The distribution of accuracies for varying Levenshtein distances and buffer lengths is shown in figure [PLEAEKAJSDFLKADSKLJFAL;KSDF;LKASDJF].

Limitations

OCR Results

In Approach 1, we used the pytesseract library for our OCR, but came to the realization that this was unfeasible due to the amount of mistakes it made. This led us to switching to PDFPlumber's text extraction utility, which let us extract the text straight from the PDF. The document we used were processed by Google, meaning their OCR will most likely be higher than most people's. This approach with text extraction will mean simply taking photos of a document will end up not working, and that the text needs to be part of the PDF.

Time and Money

In our Approach 1, we intended on using GPT-4o for natural entity recognition, as well as a geolocating API for mapping our addresses to the correct location in Venice. What we came to realize was that when we have thousands of entries to process, each requiring possibly multiple requests to each endpoint, this would not be possible without a financial burden or sacrifice of time. This became one of the leading causes we pivoted toward Approach 2, which would cut these dependencies out of our project.

Manual Inspection

In order for our Approach 2 to work, at two steps in our pipeline we needed to perform some level of manual work. Since our project was done using a smaller subset of a large document, this work was manageable with three people, but it's important to keep in mind for larger projects. The first was after our text extraction, where entries that had been divided onto more than 1 line had to be manually edited to fit onto a single line. The second was during our semantic annotation with INCEpTION, where there was a possibility for the program to flag occupations as names, resulting in us needing to correct all these cases by hand.

Language Barrier

Since none of us speak Italian, it is impossible for us to reliably verify that occupations, names, and other data from our document were correctly annotated. For our annotation work, we proceeded with the assumption that each entry followed the same pattern of listing the name first, then occupation, and finally the address.

Github Repository

Data Ingestion of Guide Commericiale