Files
logseq/pages/hls__TOSEM-2023-0032_Proof_hi_1680262636436_0.md
T
2025-06-02 17:15:13 +02:00

546 lines
20 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
file:: [TOSEM-2023-0032_Proof_hi_1680262636436_0.pdf](../assets/TOSEM-2023-0032_Proof_hi_1680262636436_0.pdf)
file-path:: ../assets/TOSEM-2023-0032_Proof_hi_1680262636436_0.pdf
- Measuring and Clustering Heterogeneous Chatbot Designs
ls-type:: annotation
hl-page:: 2
hl-color:: green
id:: 6426c651-96af-4297-8073-88830f4ece78
- suite of metrics for chatbot designs,
ls-type:: annotation
hl-page:: 2
hl-color:: green
id:: 6426c65b-52e5-4d0e-9bc9-83620aeb6a2c
- rouping chatbots along their conversation topics and design features
ls-type:: annotation
hl-page:: 2
hl-color:: green
id:: 6426c663-4983-4bb7-b7c2-3daf1ac4a64e
- 59 Dialogflow and Rasa chatbots
ls-type:: annotation
hl-page:: 2
hl-color:: green
id:: 6426c67d-4a4f-484e-960d-bc33397f0495
- improved
ls-type:: annotation
hl-page:: 2
hl-color:: green
id:: 6426c7b4-903e-4762-8f4b-faa9f53c7563
- We have refined our suite of metrics and extended it with a new readability metric, expanding the explanations and rationale of the metrics.
ls-type:: annotation
hl-page:: 3
hl-color:: green
id:: 64270a94-e163-4aa2-a155-66a778a97e27
- Francesco Basciani, Juri Di Rocco, Davide Di Ruscio, Ludovico Iovino, and Alfonso Pierantonio. 2016. Automated Clustering of Metamodel Repositories. In Proc. 28th Int. Conf. on Advanced Information Syst. Eng. (LNCS, Vol. 9694). Springer, 342358.
ls-type:: annotation
hl-page:: 38
hl-color:: green
id:: 64270ae3-06c3-4eb7-b8c5-4955f5045e74
- ESTHER GUERRA
ls-type:: annotation
hl-page:: 4
hl-color:: green
id:: 64270b12-3b66-4f3d-9c79-4b90a54a5a31
- Measuring and Clustering Heterogeneous Chatbot Designs
ls-type:: annotation
hl-page:: 4
hl-color:: green
id:: 64270b16-aa4c-44a4-be32-1aa114a999b7
- JUAN DE LARA
ls-type:: annotation
hl-page:: 4
hl-color:: green
id:: 64270b1f-e304-4d66-af8c-bdd9b073ee61
- there are hardly any mechanisms to compare and cluster chatbots developed with heterogeneous technologies
ls-type:: annotation
hl-page:: 4
hl-color:: yellow
id:: 64270c19-5357-4c40-a759-2ce6d8e898ee
hl-stamp:: 1680280620522
- 21 metrics for chatbot designs
ls-type:: annotation
hl-page:: 4
hl-color:: green
id:: 64270c5e-847d-46e8-b212-52d7e9c52f54
hl-stamp:: 1680280672632
- 259 Dialogflow and Rasa chatbots
ls-type:: annotation
hl-page:: 4
hl-color:: green
id:: 64270cdb-7b9c-4d56-95e1-581469801ba5
- early detection of quality issues
ls-type:: annotation
hl-page:: 4
hl-color:: yellow
id:: 64270d0d-3cb8-46db-ba98-f85e8ccdb270
hl-stamp:: 1680280847796
- chatbot comprehension, search and comparison
ls-type:: annotation
hl-page:: 4
hl-color:: green
id:: 64270d3e-3609-4143-acb2-2bf21a05aa9c
hl-stamp:: 1680280898517
- Chatbots are also used to assist workers in different domains, like software engineering
ls-type:: annotation
hl-page:: 5
hl-color:: green
id:: 64270e27-b4aa-4ec8-b175-babbfd85b037
- artner estimates that there are more than 2000 natural language technology providers, with a significant number of them offering facilities to create chatbots
ls-type:: annotation
hl-page:: 5
hl-color:: green
id:: 64270e35-49a7-407f-a40e-b02580704e6b
- tools for testing chatbots
ls-type:: annotation
hl-page:: 5
hl-color:: green
id:: 64270e85-54c0-4226-9008-ae0331127bab
- guide and control the quality of chatbots throughout their development and maintenance, becoming a complement to testing
ls-type:: annotation
hl-page:: 5
hl-color:: green
id:: 64270ec1-fa6a-4192-9b1e-b928d0e536a2
- there are hardly any metric proposals for chatbot designs
ls-type:: annotation
hl-page:: 5
hl-color:: purple
id:: 643071f5-9869-46bb-8b0f-aabd7c516732
- detect potential problems related to user experience (e.g., complex conversation flows, hard-to-read chatbot answers); as indicators of chatbot complexity; to compare properties of heterogeneous chatbots; to discover commonalities and cluster similar chatbots; and to understand how different implementation platforms can impact on the chatbot design
ls-type:: annotation
hl-page:: 5
hl-color:: purple
id:: 64307221-6d9a-47af-82e5-54a16fa70021
- This calls for mechanisms to measure the similarity or dissimilarity between chatbots, even if developed with different technologies, to understand their commonalities and enable searching for chatbots akin to a given one, e.g., as a first step towards the reuse of existing chatbots
ls-type:: annotation
hl-page:: 5
hl-color:: yellow
id:: 643072ad-0c81-45e5-b15e-949da1825196
- we propose a suite of 21 static metrics for chatbot designs, and two methods for clustering chatbots
ls-type:: annotation
hl-page:: 5
hl-color:: purple
id:: 6433ba23-b920-474f-ad4d-406cd9fa63da
- This is useful, e.g., to identify chatbots on the same topic for their reuse; to facilitate the construction of recommender systems for chatbots; to organise large chatbot repositories according to their domain; or to facilitate chatbot search.
ls-type:: annotation
hl-page:: 5
hl-color:: purple
id:: 6433ba51-985a-4318-a639-f0a7ab14412f
- metrics and the clustering methods over Dialogflow and Rasa chatbots
ls-type:: annotation
hl-page:: 6
hl-color:: green
id:: 6433bb84-0020-4fe7-99d8-cd95d5d5bc53
- 20 chatbot design metrics, proposed their technology-independent definition atop Conga, made their implementation available via an API, and evaluated their suitability on 12 chatbots.
ls-type:: annotation
hl-page:: 6
hl-color:: green
id:: 6433bc70-d72e-415b-977c-bc7ab83d864a
- principal component analysis.
ls-type:: annotation
hl-page:: 6
hl-color:: yellow
id:: 6433bc90-eb00-4b86-a2a7-282da8bdc294
- we are interested in task-oriented chatbots
ls-type:: annotation
hl-page:: 6
hl-color:: green
id:: 6433bd3b-1286-4416-8023-8d476c77c813
- [:span]
ls-type:: annotation
hl-page:: 7
hl-color:: green
id:: 6433bd5f-1c82-4f9a-94f3-a1f31e280fdc
hl-type:: area
hl-stamp:: 1681112412961
- ystemcentric
ls-type:: annotation
hl-page:: 7
hl-color:: green
id:: 6433be90-d066-4f2a-a9a0-14a21683fe72
- Content-centric
ls-type:: annotation
hl-page:: 7
hl-color:: green
id:: 6433be97-64ff-4dfe-ab62-25f0cb8cbd80
- Visualcentric
ls-type:: annotation
hl-page:: 7
hl-color:: green
id:: 6433be9f-6267-42bf-8aa4-4d0e7c7604a6
- conversation-centric
ls-type:: annotation
hl-page:: 7
hl-color:: green
id:: 6433bea2-ea0d-4b36-bcee-0c68120b001e
- conversation patterns [35] and conversation design principles
ls-type:: annotation
hl-page:: 8
hl-color:: green
id:: 6433bef4-67d3-4e53-b4d2-7a9af9d85ab5
- dynamic execution of the chatbot
ls-type:: annotation
hl-page:: 8
hl-color:: yellow
id:: 6433bf2a-89ee-467a-80a4-3d55bc1628ff
- A means to facilitate finding artefacts of interest (chatbots in our case) is to organise them into meaningful group
ls-type:: annotation
hl-page:: 8
hl-color:: green
id:: 6433c0e8-418f-4a75-9fc1-d0967df26e44
- reusing existing artefacts in a cluster of interest
ls-type:: annotation
hl-page:: 8
hl-color:: green
id:: 6433c0ef-d481-420e-a21a-1f3432ed1dfc
- K-means and hierarchical clustering require setting the number of clusters as hyperparameters, while DBSCAN has other parameters like the neighbourhood distance and the minimum number of points required to form a dense region
ls-type:: annotation
hl-page:: 8
hl-color:: green
id:: 6433c14d-36ef-4daf-a8f0-49d9e92b1a60
- In our proposal to chatbot clustering, we use similar encodings and techniques, but focus on chatbot designs instead of meta-models
ls-type:: annotation
hl-page:: 9
hl-color:: purple
id:: 6433c174-0930-4fac-a5f4-8b96565043c8
- clustering for chatbot designs has not been studied yet in the literature
ls-type:: annotation
hl-page:: 9
hl-color:: purple
id:: 6433c18f-fbe8-411f-85fc-ea07febb53ce
- Section 4.1 starts by introducing the chatbot design notation
ls-type:: annotation
hl-page:: 9
hl-color:: purple
id:: 6433c1ae-2826-4eee-b30d-ea8685bfb55c
- Section 4.2 details our proposed metric
ls-type:: annotation
hl-page:: 9
hl-color:: purple
id:: 6433c1b5-9ee7-4f3c-9e18-97b25c1c4943
- Conga
ls-type:: annotation
hl-page:: 9
hl-color:: purple
id:: 6433c1df-9e2f-451e-a94a-9ef997a9aebe
- [:span]
ls-type:: annotation
hl-page:: 10
hl-color:: green
id:: 6433c204-1b23-4b7d-9074-fd6f98b13553
hl-type:: area
hl-stamp:: 1681113602953
- Intents
ls-type:: annotation
hl-page:: 9
hl-color:: green
id:: 6433c22a-309c-49d9-b724-9fb588e852bf
- Entities
ls-type:: annotation
hl-page:: 9
hl-color:: green
id:: 6433c22d-c52c-46f7-81db-17b8f1176a99
- t Actions
ls-type:: annotation
hl-page:: 9
hl-color:: green
id:: 6433c233-5eb0-4e84-851d-023d0744ae3a
- Flow
ls-type:: annotation
hl-page:: 9
hl-color:: green
id:: 6433c235-d772-47d5-8225-7c15d957b235
- For example, a chatbot can declare a simple entity for drink sizes with literals small, medium and large in English, and additionally define synonyms regular for medium and big for large.
ls-type:: annotation
hl-page:: 9
hl-color:: green
id:: 6433c2c2-9143-4196-9872-e8190587bc6e
- [:span]
ls-type:: annotation
hl-page:: 11
hl-color:: green
id:: 6433c311-1f72-4b52-9c0b-b7bde69046dc
hl-type:: area
hl-stamp:: 1681113870406
- Global metrics. We start explaining global metrics. These measure the number of intents(INT), entities (ENT) and conversation flows (FLOW, PATH), and also include understanding and user experience metrics (CNF, SNT)
ls-type:: annotation
hl-page:: 11
hl-color:: green
id:: 6433c350-559c-4c7e-a70a-21105ef239a9
- PATH measures the number of paths a conversation may take, which is an indicator of conversation complexity
ls-type:: annotation
hl-page:: 11
hl-color:: green
id:: 6433c399-7b88-4bdf-98a6-fac791bb6df9
- The combined use of FLOW and PATH can help detecting deviations of some design principles
ls-type:: annotation
hl-page:: 11
hl-color:: green
id:: 6433c409-b375-49ff-bc86-3021d87424a9
- ntent metrics measure quality properties of each intent with respect to the expected user utterances and the bot output phrases
ls-type:: annotation
hl-page:: 12
hl-color:: green
id:: 6433c49c-3c5b-441a-a622-b2d1032cef34
- interaction complexity,
ls-type:: annotation
hl-page:: 12
hl-color:: green
id:: 6433c4c9-9554-44c9-8d26-e7185c6ee7d5
hl-stamp:: 1681114321426
- user cognitive load and speed up the completion of the intended tas
ls-type:: annotation
hl-page:: 13
hl-color:: green
id:: 6433c5e6-33cc-4b9f-bca3-b0d2de20c997
- chatbot vocabulary
ls-type:: annotation
hl-page:: 13
hl-color:: green
id:: 6433d39e-4d96-447e-af37-700174196584
- internal quality features (e.g., complex/simple conversations, large/succinct outputs). The latter groups chatbots by conversation top
ls-type:: annotation
hl-page:: 13
hl-color:: yellow
id:: 6433d3a9-c5be-479e-b8c0-a055bf88ffb2
- metric INT (i.e., number of intents) would create groups of chatbots with similar size complexit
ls-type:: annotation
hl-page:: 13
hl-color:: green
id:: 6433d3f9-5203-47b0-950d-a59c625eaebe
- [:span]
ls-type:: annotation
hl-page:: 14
hl-color:: green
id:: 6433d40e-0f51-4f8e-85aa-6822d3061732
hl-type:: area
hl-stamp:: 1681118220789
- principal component analysis (PCA) o
ls-type:: annotation
hl-page:: 14
hl-color:: green
id:: 6433d722-2ee5-40ec-a9d6-94714b27e880
- he latter is possible since PCA permits explaining how each metric contributes to each component, by providing the loading factor of the metric for the component. This effectively groups the metrics that contribute the most to each component.
ls-type:: annotation
hl-page:: 14
hl-color:: green
id:: 6433d739-6293-4cc8-a657-23c4145a4490
- PATH, INT, FLOW, WL, PPTP and FPATH.
ls-type:: annotation
hl-page:: 14
hl-color:: yellow
id:: 6433d749-9f82-4aaa-856d-3492003fd1b8
hl-stamp:: 1681119074022
- Vocabulary-based clustering
ls-type:: annotation
hl-page:: 15
hl-color:: green
id:: 64348546-40a4-4e18-9283-55d312c65b9f
- we provide another clustering method based on the chatbot conversation topics
ls-type:: annotation
hl-page:: 15
hl-color:: purple
id:: 64348554-62cf-489c-838f-d80ab3ccaf72
- training phrases of the intents, the chatbot output phrases, and the literals and synonyms in entities.
ls-type:: annotation
hl-page:: 15
hl-color:: green
id:: 64348598-0791-4c13-b203-55fe7b82ec0b
- Cluster projection into three dimensions,
ls-type:: annotation
hl-page:: 14
hl-color:: yellow
id:: 643567d9-76d1-4749-b66d-0eb46ea8c2b4
- In the graphic, each dot represents a chatbot, and the shape of the dot identifies the cluster where the chatbot belongs.
ls-type:: annotation
hl-page:: 14
hl-color:: yellow
id:: 6435689b-3e6c-4371-9a76-13a96f421f89
- Latent Semantic Analysis
ls-type:: annotation
hl-page:: 15
hl-color:: purple
id:: 64356a14-052b-48c8-b7e6-c804d35f246e
- set of chatbots
ls-type:: annotation
hl-page:: 15
hl-color:: yellow
id:: 64356a38-076f-4fa8-b085-ea63f91f0ca9
- relevance of a word to a chatbot,
ls-type:: annotation
hl-page:: 16
hl-color:: green
id:: 64356ab2-9bc1-4ee9-807d-2a6cb170f649
- TF-IDF weight of a word for a chatbot
ls-type:: annotation
hl-page:: 16
hl-color:: green
id:: 64356b1b-d19d-4f5b-acc7-8fa7e4995558
- Hence, the weights tend to filter out common terms.
ls-type:: annotation
hl-page:: 16
hl-color:: purple
id:: 64356b2c-d02c-46f9-b646-3d6f913092eb
- osine similarit
ls-type:: annotation
hl-page:: 16
hl-color:: green
id:: 64356b47-da7b-44f5-a50c-bc86f1d4ce01
- agglomerative clustering
ls-type:: annotation
hl-page:: 16
hl-color:: green
id:: 64356b61-884f-4f24-a58c-f986e290c0cf
- Clustered similarity matrix of chatbots.
ls-type:: annotation
hl-page:: 18
hl-color:: green
id:: 64356bec-b05f-4ec3-8a8e-b61e511d7d88
- Chatbot working schema
ls-type:: annotation
hl-page:: 7
hl-color:: yellow
id:: 64356f75-f0e5-4767-a91d-fa6ca0f3aa68
- across all chatbots in the repository.
ls-type:: annotation
hl-page:: 18
hl-color:: green
id:: 643570a6-f407-4bd3-aeb0-a45c198f270e
- clustering chatbots
ls-type:: annotation
hl-page:: 18
hl-color:: green
id:: 643570b9-1a4b-41f3-8d44-90156ad6c9ff
- his core has an extensible design, which makes it easy to add new types of metrics, clustering criteria and chatbot technologie
ls-type:: annotation
hl-page:: 19
hl-color:: green
id:: 643571d5-ad05-41a6-b3a5-82b342c89b9b
- This way, Asymob computes the metrics on Conga models, independently of any chatbot implementation platform.
ls-type:: annotation
hl-page:: 19
hl-color:: green
id:: 6435738e-3ebc-4e8c-b966-6562e955bf50
- Asymob supports some third-party technologies to simplify the implementation of new metrics
ls-type:: annotation
hl-page:: 19
hl-color:: green
id:: 643573a6-12d4-4dec-885e-9883de5c1c6c
- Dialogflow and Rasa) into Conga.
ls-type:: annotation
hl-page:: 20
hl-color:: green
id:: 643573c8-c6f3-4642-98dd-99927d75eb44
- Agent.json describes global chatbot features, like its name, definition languages, or connection data to external services (the webhook).
ls-type:: annotation
hl-page:: 20
hl-color:: green
id:: 6435746d-1727-46d2-99a9-591ec23c2e19
- Intents
ls-type:: annotation
hl-page:: 20
hl-color:: purple
id:: 64357482-5893-4b2a-a454-36f065ea7160
hl-stamp:: 1681224841238
- usefulness of our metrics
ls-type:: annotation
hl-page:: 23
hl-color:: green
id:: 64357500-0e59-4814-a243-eb231c28aa52
- efficacy of our vocabulary-based clustering
ls-type:: annotation
hl-page:: 23
hl-color:: green
id:: 64357505-9559-477a-844a-1424ed2318c3
- chatbots in the wild compare with respect to their size, conversation style, outputs, expected inputs and vocabulary?
ls-type:: annotation
hl-page:: 23
hl-color:: yellow
id:: 643575c8-5044-40fe-b3ad-830d3a8b896e
- quality issues in real chatbots
ls-type:: annotation
hl-page:: 23
hl-color:: yellow
id:: 6435769c-5592-45d1-883f-8919ffb5423e
- meaningful groups of semantically related chatbots?
ls-type:: annotation
hl-page:: 24
hl-color:: yellow
id:: 64357724-3fbd-4e25-a199-67b40ec7d292
- we compared different dimensions of the chatbots (size, conversation style, responses, expected user utterances and vocabulary) based on the chatbots metric values
ls-type:: annotation
hl-page:: 24
hl-color:: green
id:: 6435778a-fe77-422a-a711-10d5b2f8ff15
- We do not expect every discordant metric value to signal a real problem, but instead, we are interested in assessing the usefulness of metrics as an inexpensive chatbot quality assurance mechanism (e.g., compared to testing) that can be used early in the development process
ls-type:: annotation
hl-page:: 24
hl-color:: purple
id:: 643577ad-fb25-4a39-a817-ece0ecbc6e48
hl-stamp:: 1681225647635
- size
ls-type:: annotation
hl-page:: 25
hl-color:: green
id:: 643578f1-018b-461d-b6af-828dac934a0f
- conversation style
ls-type:: annotation
hl-page:: 25
hl-color:: green
id:: 643578f6-95b3-4109-8a69-26b255baa946
- outputs
ls-type:: annotation
hl-page:: 25
hl-color:: green
id:: 643578f8-f070-4e8d-aa8c-2b802ebdf903
- expected inputs, vocabulary (pertinent for RQ1) and implementation platform (RQ1.1)
ls-type:: annotation
hl-page:: 25
hl-color:: green
id:: 64357900-5693-45d4-9ba5-e6c62c6342c1
- Naturally, chatbots with more intents tend to have more flows, and therefore, more paths.
ls-type:: annotation
hl-page:: 26
hl-color:: purple
id:: 64357937-6d96-4dbd-9fe7-9e6c1be0cf5e
- 7.3.7 Answering RQ1 and RQ1.1. With respect to RQ1, as detailed in Sections 7.3.17.3.5, our metrics permit comparing chatbot size using INT; conversation style along Moore and Arars taxonomy [ 33] using FLOW, PATH, FPATH and CL; chatbot outputs using WPO, CPO, VPOP, OPRE and SNT; expected user inputs using WPTP, VPTP and PPTP; and chatbot vocabulary size using ENT, LPE, SPL and WL. Moreover, since metrics are defined over Conga, they are applicable to different technologies and enable the comparison of heterogeneous chatbots. Regarding RQ1.1, as described in Section 7.3.6, we found that, for our dataset of chatbots, all metrics but PATH and VPTP follow different distributions for Dialogflow and Rasa.
ls-type:: annotation
hl-page:: 29
hl-color:: yellow
id:: 643579b1-5765-46ee-a5fb-1f76f2975cd4
hl-stamp:: 1681226391481
- The extreme cases reveal complex chatbots and may signal redundant intents
ls-type:: annotation
hl-page:: 30
hl-color:: yellow
id:: 64357a61-0924-486b-8f07-5a592295d6ad
- low or high metric outliers to assess if they have problems.
ls-type:: annotation
hl-page:: 33
hl-color:: yellow
id:: 64357c1f-9bef-4344-8c43-b2f99e5bba28
hl-stamp:: 1681226801614
- 3 clusters
ls-type:: annotation
hl-page:: 34
hl-color:: green
id:: 64357cd5-91cb-41e8-9e03-88dcc182ef9c
- Answering RQ3. Overall, we can answer RQ3 positively: vocabulary-based clustering can create meaningful groups of semantically related chatbots (e.g., groups of chatbots about hotels, restaurants, or weather, among others). To have a quantitative assessment, we measured the fraction of correctly clustered chatbot pairs using balanced accuracy, and compared against three baselines obtaining a substantially higher balanced accuracy (0.664 vs 0.5)
ls-type:: annotation
hl-page:: 36
hl-color:: purple
id:: 64357d16-08e9-4663-b27e-8b62599e9b81
- quality assurance during chatbot development, and so, we want our metrics to detect problems related to incomplete designs.
ls-type:: annotation
hl-page:: 37
hl-color:: green
id:: 64357d38-5ab0-4fd4-b9a8-5065a3021f3f
- Related to RQ2, while metrics can detect problems in chatbots, problems need to be confirmed by inspection
ls-type:: annotation
hl-page:: 37
hl-color:: purple
id:: 64357d4a-19c0-4004-9039-51a5c1d60a17
- Chatbots are increasingly relevant nowadays, so techniques for assessing, comparing and clustering chatbots before their deployment are required.
ls-type:: annotation
hl-page:: 37
hl-color:: green
id:: 64357d7c-870f-456a-b317-9258fdf1479d