The chapter 4.1 introduces analysis and development of advanced analytics technologies to optimize the process for recommending tourism services at our agency COMPANY-002. Based on previous research, where we design an overall analytic architecture and a method of automatic detection for potential customers this Section focuses about evaluating the offer documents sent to clients. The main aim is to check that recommendations are more precise and aligned with customers' wishes or needs.
To this end, they were analysed 50 randomly selected offer documents and their associated texts using natural language processing (PLN) techniques. Tools have been put into place. SOFTWARE-TH-002, SOFTWARE-TH-003 and SOFTWARE-TH-003 To conduct the recognition of designated entities (NER), enabling them to extract key terms and build a ratings dictionary. That dictionary was used to evaluate the relevance of offers as a result and depend on his or her willingness.
The chapter also includes a comparison of NER tools used, evaluating their accuracy and speed especially in Spanish text. That approach aims at not only improving the personalization of recommendations but also to identify most effective technological solutions for supporting travel agents in their decision-making.
4.1 Introduction
In previous research for our client COMPANY-002 we designed an overall analytic architecture to help the company improve its sales process
and we design or implement a method for the automatic detection of potential customers. In this research based on earlier knowledge, we focused with our offer documents sent to potential customers during the sales process so as To verify that recommendations best suit your client's wishes.
To study this recommendation process in detail, we select at random 50 offer documents and their associated text opportunities by the client. The analysis of text opportunities with PLN was initiated in order to remove the entities and set up a general dictionary that had an entity-related score, depending on their propensity. We then use that score dictionary as a basis for an in-depth assessment of the offer documents using three different methods: SOFTWARE-TH-003, SOFTWARE-TH-002 and one based on SOFTWARE-TH-003. We use all these different technologies to realize NER in the various texts. Name recognition (NER) is an important sub-task in the processing of natural language. Several NER systems have been developed in the last decade. They can be directed to different domains, using differing methodologies and working in a number of languages. To detect several types with an entity they will have been allowed at the same time as their entry or exit formats These conditions make it difficult for a technical user to select the appropriate NER tools with specific Jiang et al. ( 2016 ). We focused on making a Shelar et al comparison. ( 2020 ) with three different NER tools, using our client's information: COMPANY-002. Since our customers conduct Spanish text, we are trying to get the best solution for that Galicia-Haro and al. ( 2004 ).
4.2 Proposal
4.2.1 Analytical processes involved with a recommendation
Creation of the General Index Dictionary
To create a general score dictionary, we process all opportunities stored at CRM with NER functionality from SOFTWARE-TH-002. With that information about terms we create a table to store the ratio of term found with an opportunity result, purchase or desist.
We added that table using the count of terms and decision, purchase or desist. To quantify, we generate a simple score method by using the following formula:
if (purchase): count-id@-@ term * 2
Table 4.1: General score dictionary
Term Score Maldives 13 Madrid 8.5....... Tanzania - 1 Bangkok - 1
Table 4.2: Score of offers using an opportunity text
Offer-Id Destination Sbling 13485 Maldives 350 15839 Costa Rica 150....... 14592 Canary Islands 50 12394 Tanzania - 13.5
if (desist): 1 ) / 2
Where Count-id@-@ term represents the unique number of times that an id appears in a group table (decision, terms). With this quantification method for the three tools, we start to figure out how these document offers are.
We add again information with the sum of an earlier score for group (term), generating a general endpoint to one terms given any decision as shown in table: 4.1
Score of tender documents
We conduct the three PLN techniques about offering documents, stored under.docxformat. The idea is to get by using NER terms important for documents. The following terms have been tipped using the general dictionary as described in this Section, so finally from an opportunity text we get a set of recommendations (offer documents) sorted by score generated with his and her General Dictionary. 4.2
Using different NER tools
SOFTWARE-TH-002
SOFTWARE-TH-002 is an open source Python library for PLN, launched at 2015 and the creators of SOFTWARE-TH-002 describe his work as an industrial PLN. SOFTWARE-TH-002 can work with pretrained language models and word vectors for more 60 languages. The language models allow us to perform a variety of tasks
Table 4.3: Speed Benchmark SOFTWARE-TH-002 in November 2022
LIBRARY pipelines WPS CPU and GPS SOFTWARE-TH-002 at _ core] web 10,014 14,954 SOFTWARE-TH-002 at _ core carnatestrf 684 3,768 Stanza at _ ewt 878 2,180 Flair pos (-fast) & ner [- quick] 323 1,184 UDPipe english-ewt-you 2.5 1,101 N / A
PLN, including recognition of Vasiliev (NER) 2020 ).
Install SOFTWARE-TH-002 It's very easy using zip, later you have to download the model from SOFTWARE-TH-002 Several have been tested in this study, with the best at _ core] web's.sys but we are currently working on Spanish text as our goal was to detect most of them.
Also, we add a plugin called nedspac and an item of pipeline SOFTWARE-TH-002 to deny text concepts, based on algorithm SOFTWARE-TH-005 Chapman et al. 2001 ). With this pipeline we can manage the logic involved in different entities, detecting when an entity is denied and providing better recommendation.
In terms of velocity, SOFTWARE-TH-002 is the fastest PLN library, we can also have a look at its reference points [LINK-REDACTED] We compared the velocity of different PLN libraries, as measured by words per second (WPS): The higher and better they were evaluated at 10,000 Reddit's comments:
SOFTWARE-TH-003
At end of 2018, a group of lab scientists ORG-TH-008 under the leadership of Devlin, introduced a new linguistic model called SOFTWARE-TH-003. That model is intended for a first and profound training of bidirectional text representation to be later used in automatic learning models. The advantage of this model is its ease to use, which means adding only an output layer with the existing neural architecture for obtaining text models that outweigh all those currently available in several Koroteev natural texting problems ( 2021 ).
For use with SOFTWARE-TH-003 We work with the model: xlma-roberta -large- ner, this models is an adjusted XLM.Robata--large type for name entity recognition (NER) about Spanish part of dataset NLL-200 2. When evaluated in the test subset of this dataset, they get an F1 score from 89,17 being one of the best Spanish NER available
at the time of this experiment. The speed of SOFTWARE-TH-003 NER's slower than with SOFTWARE-TH-002 but appears to be more precise in recommendations when they are successful, than that of SOFTWARE-TH-002.
Fine Tuned SOFTWARE-TH-003 Based
Although profound learning models have reached the state of art at many NPL tasks, these model require training from scratch and demand great datasets as well with computer days to converge into a solution. Research at NLLP has focused mainly on transductive transfer. For inductiva transfer, fine tuning of embeddings with pretrained words - a simple transferring technique that only addresses the first layer and has had an important impact on practice. It is used in most last-generation models) Recent approaches that concede inlays from other tasks with entry into different layers continue to train the main task model at zero and treat preset incruvals as fixed parameters, thus limiting their usefulness. In light of the benefits from pre-training, we should be able to do something better than randomly initializing all other parameters for our models. Inductive transfer through fine tuning has however been unsuccessful for PLN, but as a first time the smooth titration of an (LM) language model may be best solution and requires million documents to achieve good performance with them. Howard and Rog ( 2018 )
We use SOFTWARE-TH-002 first to tuning our transformer. First, we did the following data labelling:
Funa TYPE de O Honey TYEP africa DESNINATION beach DEJN DIY
After some configuration steps, we trained the model SOFTWARE-TH-003. The countries...
Table 4.4: A case account to generate a confusing matrix
Real SOFTWARE-TH-002 SOFTWARE-TH-003 Fine Tuned Good Bad good bad 9 9 13 6 14 6 Predicted bad 15 7 11 10 10 10
The most complex of the three libraries are installation and production, but their results have been shown in this Section.
4.3 Experiments
4.3.1 Comparison of results using confusing matrix metrics
Since we are comparing three libraries with NER functionalities, they're completely different and have established that to make a comparison between the various techniques We needed an interplay matrix procedure. We categorize recommendations in terms of good or bad, so we define the binary matrix as Real Recommendations and Previews.
We generate a binary confusion matrix, the counted values are shown in The table 4.4 and the metrics obtained by each PLN technique are shown in table 4.5 these metrics show that best techniques are Fine-tuned, SOFTWARE-TH-003 and SOFTWARE-TH-002 in this order. Examining the metrics we can note that techniques of SOFTWARE-TH-003 and Fine Tuned are very similar, though the last one's better than his first. The technologies and their production SOFTWARE-TH-002 It's the worse, with a great difference from them but best to manage logic between entities.
Table 4.5: Metrics obtained for different PLN techniques
SOFTWARE-TH-002 SOFTWARE-TH-003 Accuracy 0.4000 0.5750 0.6000 Precision 0.5000 0.6842 0.7000 F1 Score 0.4286 0.6047 0.6364
4.3.2 Other PLN solutions: ORG-TH-007
We decided to test other PLN solutions for our purposes. We have been trying the services of ORG-TH-007 with model 3.5 for the pro-
NER cesus, with similar results. It's true that the entities recognised by ORG-TH-007 are more than with SOFTWARE-TH-002, SOFTWARE-TH-003 o Fine-Tuning but work with ORG-TH-007 creates new problems, its API cannot answer in Spanish and the logic of these entities isn't misconsidered or found as follows:
For the question:\ Answer in Spanish. Honey Moon, something like New York and NO to Maldives with a honeymoon for two people. 15 of August 2022 thank you.
The response from ORG-TH-007 es:\ "ю 124, {Maldives}.12). 15, 2022 {pos) 124;\ Also, for two people}
What a priori is an interesting functionality, could generate bad recommendation because the initial consultation said: Maldives No and ORG-TH-007 He's returning exactly the opposite, bad recommendation.
With our model SOFTWARE-TH-002 by using the negative plugin and used to deny text concepts with an algorithm SOFTWARE-TH-005 we can work better with logic between entities, especially by denying them. That type of capacity currently cannot be achieved with ORG-TH-007 and his model 3.5 and with SOFTWARE-TH-003 this is managed by association with the importance of a score term.
4.3.3 Future creation of a virtual travel agent through the multiagency system
In a previous investigation Moreno et al. ( 2022 ) we use MAS to distribute the system's workload. Here we also apply this with a bioinspired architecture design to organize the different processes involved and so that they can spread their skills or emulate them from future electronic travel agents.
As stated in the previous Section, multiagent systems are collections of independent intelligent entities that cooperate with them to solve complex problems jointly. Other previous research indicates why and how to apply them in the field of tourism X. Wang et al. ( 1997 ). Here we are focused on a specific, even future problem of this industry: how to emulate an agent for travel.
As we have noted, today travel agents could be seen as human recommends and they get an application. They process them with their best recommendation back to the list of countries that are currently undergoing a better decision by applying for information about how these measures will work out at national level or through local public authorities in South Africa's economic policies (see Section 3).The Commission has been following up on this issue: But from a technical modelling perspective, the recommendation process is complex and requires that many other processes - some analytics or more operations) participate in an integrated manner to achieve best success. We propose as a solution an innovative and bioinspired MAS based on the structure of ants.
The abstraction of an ant allows to organize the code as a framework, for instance where agent Reigns is similar with controller and soldiers like models. But as the ant works, it allows this type of organisation to be an intelligent way for instance by integrating more advanced features such and inter-agent negotiation. To improve agents' negotiating skills, an incremental balanced method could be used to reach agreement between actors but also with humans and agent. ( 2003 ).
4.4 Conclusions
The PLN is the best artificial intelligence technique that can be applied when data belong to text type. After trying three different open source technologies, nobody would use these solutions for problems where the solution needs to minimize false positives as we have shown in a table of counted cases. But after trying out other commercial solutions, such as ORG-TH-007, the area of open source solutions have still a lot to say in this field.
Open source technologies have been tested only against their NER functionalities, with more characteristics to be tried as PLN techniques but there's a great difference between basic solutions and SOFTWARE-TH-002 and basic solutions SOFTWARE-TH-003 and these last are best. The case of exposed use is a good example as to how PLN can help other companies (not necessarily related with tourism) achieve an upper level for text data.
The extension of the bioinspired MAS based on concrete structure to model an electronic travel agency is a new solution that shows not only how complete and versatile MOM architecture propose but also That divide-up design rule will help with complex system implementation.
Reported to our client: COMPANY-002 open source solution technologies provide a very good vision that could help travel agents quickly identify important entities while considering the text of client applications.